From cc9bfaeafad972cfe40e6cb804f60adba0c17be1 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Tue, 7 Oct 2014 17:12:32 -0400 Subject: Import synutils --- utils/synutils/Makefile.am | 16 + utils/synutils/alignment.h | 208 ++++++++ utils/synutils/argument_reorder_model.cc | 311 +++++++++++ utils/synutils/argument_reorder_model.h | 222 ++++++++ utils/synutils/constituent_reorder_model.cc | 796 ++++++++++++++++++++++++++++ utils/synutils/maxent-3.0/Makefile.am | 10 + utils/synutils/maxent-3.0/lbfgs.cpp | 110 ++++ utils/synutils/maxent-3.0/lbfgs.h | 21 + utils/synutils/maxent-3.0/mathvec.h | 97 ++++ utils/synutils/maxent-3.0/maxent.cpp | 698 ++++++++++++++++++++++++ utils/synutils/maxent-3.0/maxent.h | 395 ++++++++++++++ utils/synutils/maxent-3.0/owlqn.cpp | 137 +++++ utils/synutils/maxent-3.0/sgd.cpp | 180 +++++++ utils/synutils/srl_sentence.h | 215 ++++++++ utils/synutils/tree.h | 731 +++++++++++++++++++++++++ utils/synutils/tsuruoka_maxent.h | 160 ++++++ utils/synutils/utility.h | 129 +++++ 17 files changed, 4436 insertions(+) create mode 100644 utils/synutils/Makefile.am create mode 100644 utils/synutils/alignment.h create mode 100644 utils/synutils/argument_reorder_model.cc create mode 100644 utils/synutils/argument_reorder_model.h create mode 100644 utils/synutils/constituent_reorder_model.cc create mode 100644 utils/synutils/maxent-3.0/Makefile.am create mode 100644 utils/synutils/maxent-3.0/lbfgs.cpp create mode 100644 utils/synutils/maxent-3.0/lbfgs.h create mode 100644 utils/synutils/maxent-3.0/mathvec.h create mode 100644 utils/synutils/maxent-3.0/maxent.cpp create mode 100644 utils/synutils/maxent-3.0/maxent.h create mode 100644 utils/synutils/maxent-3.0/owlqn.cpp create mode 100644 utils/synutils/maxent-3.0/sgd.cpp create mode 100644 utils/synutils/srl_sentence.h create mode 100644 utils/synutils/tree.h create mode 100644 utils/synutils/tsuruoka_maxent.h create mode 100644 utils/synutils/utility.h (limited to 'utils') diff --git a/utils/synutils/Makefile.am b/utils/synutils/Makefile.am new file mode 100644 index 00000000..f87f1f8a --- /dev/null +++ b/utils/synutils/Makefile.am @@ -0,0 +1,16 @@ +bin_PROGRAMS = const_reorder_model_trainer + +const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc +const_reorder_model_trainer_LDADD = libfeature.a maxent-3.0/libtsuruoka_maxent.a -lgfortran + +noinst_LIBRARIES = libfeature.a + +libfeature_a_SOURCES = \ + alignment.h \ + srl_sentence.h \ + tree.h \ + utility.h \ + argument_reorder_model.h \ + tsuruoka_maxent.h + +AM_CPPFLAGS = -W -Wall -I$(top_srcdir) -I$(top_srcdir)/utils -I$(top_srcdir)/synutils/maxent-3.0 diff --git a/utils/synutils/alignment.h b/utils/synutils/alignment.h new file mode 100644 index 00000000..f8fe1993 --- /dev/null +++ b/utils/synutils/alignment.h @@ -0,0 +1,208 @@ +/* + * alignment.h + * + * Created on: May 23, 2013 + * Author: lijunhui + */ + +#ifndef ALIGNMENT_H_ +#define ALIGNMENT_H_ + +#include +#include +#include +#include + +#include "stringlib.h" + +using namespace std; + +/* + * Note: + * m_vec_s_align.size() may not be equal to the length of source side sentence + * due to the last words may not be aligned + * + */ +struct SAlignment { + typedef vector SingleAlign; + SAlignment(const char* pszAlign) { + fnInitializeAlignment(pszAlign); + } + ~SAlignment() { + + } + + bool fnIsAligned(int i, bool s) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + if ((*palign)[i].size() == 0) + return false; + return true; + } + + /* + * return true if [b, e] is aligned phrases on source side (if s==true) or on the target side (if s==false); + * return false, otherwise. + */ + bool fnIsAlignedPhrase(int b, int e, bool s, int *pob, int *poe) const { + int ob, oe; //[b, e] on the other side + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + + if (ob == -1) { + if (pob != NULL) (*pob) = -1; + if (poe != NULL) (*poe) = -1; + return false; //no aligned word among [b, e] + } + if (pob != NULL) (*pob) = ob; + if (poe != NULL) (*poe) = oe; + + int bb, be; //[b, e] back given [ob, oe] on the other side + if (s == true) + fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); + else + fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); + + if (bb < b || be > e) + return false; + return true; + } + + bool fnIsAlignedTightPhrase(int b, int e, bool s, int *pob, int *poe) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + + if ((*palign).size() <= e || (*palign)[b].size() == 0 || (*palign)[e].size() == 0) + return false; + + return fnIsAlignedPhrase(b, e, s, pob, poe); + } + + void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + } + + /* + * look the translation of source[b, e] is continuous or not + * 1) return "Unaligned": if the source[b, e] is translated silently; + * 2) return "Con't": if none of target words in target[.., ..] is exclusively aligned to any word outside source[b, e] + * 3) return "Discon't": otherwise; + */ + string fnIsContinuous(int b, int e) const { + int ob, oe; + fnGetLeftRightMost(b, e, true, ob, oe); + if (ob == -1) return "Unaligned"; + + for (int i = ob; i <= oe; i++) { + if (!fnIsAligned(i, false)) continue; + const SingleAlign& a = m_vec_t_align[i]; + int j; + for (j = 0; j < a.size(); j++) + if (a[j] >= b && a[j] <= e) break; + if (j == a.size()) return "Discon't"; + } + return "Con't"; + } + + const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { + if (s == true) { + if (i >= m_vec_s_align.size()) + return NULL; + return &(m_vec_s_align[i]); + } else { + if (i >= m_vec_t_align.size()) + return NULL; + return &(m_vec_t_align[i]); + } + } +private: + void fnGetLeftRightMost(int b, int e, const vector& align, int& ob, int& oe) const { + ob = oe = -1; + for (int i = b; i <= e && i < align.size(); i++) { + if (align[i].size() > 0) { + if (align[i][0] < ob || ob == -1) + ob = align[i][0]; + if (oe < align[i][align[i].size() - 1]) + oe = align[i][align[i].size() - 1]; + } + } + } + void fnInitializeAlignment(const char* pszAlign) { + m_vec_s_align.clear(); + m_vec_t_align.clear(); + + vector terms = SplitOnWhitespace(string(pszAlign)); + int si, ti; + for (size_t i = 0; i < terms.size(); i++) { + sscanf(terms[i].c_str(), "%d-%d", &si, &ti); + + while (m_vec_s_align.size() <= si) { + SingleAlign sa; + m_vec_s_align.push_back(sa); + } + while (m_vec_t_align.size() <= ti) { + SingleAlign sa; + m_vec_t_align.push_back(sa); + } + + m_vec_s_align[si].push_back(ti); + m_vec_t_align[ti].push_back(si); + } + + //sort + for (size_t i = 0; i < m_vec_s_align.size(); i++) { + std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); + } + for (size_t i = 0; i < m_vec_t_align.size(); i++) { + std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); + } + } + +private: + vector m_vec_s_align; //source side words' alignment + vector m_vec_t_align; //target side words' alignment +}; + +struct SAlignmentReader { + SAlignmentReader(const char *pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SAlignmentReader() { + if (m_fpIn != NULL) + fclose(m_fpIn); + } + SAlignment* fnReadNextAlignment() { + if (feof(m_fpIn) == true) + return NULL; + char *pszLine = new char[100001]; + pszLine[0] = '\0'; + fgets(pszLine, 10001, m_fpIn); + int iLen = strlen(pszLine); + if (iLen == 0) + return NULL; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + SAlignment *pAlign = new SAlignment(pszLine); + delete [] pszLine; + return pAlign; + } +private: + FILE *m_fpIn; +}; + + +#endif /* ALIGNMENT_H_ */ diff --git a/utils/synutils/argument_reorder_model.cc b/utils/synutils/argument_reorder_model.cc new file mode 100644 index 00000000..e30ee971 --- /dev/null +++ b/utils/synutils/argument_reorder_model.cc @@ -0,0 +1,311 @@ +/* + * argument_reorder_model.cc + * + * Created on: Dec 15, 2013 + * Author: lijunhui + */ + +#include +#include + +#include "argument_reorder_model.h" +#include "utility.h" +#include "tsuruoka_maxent.h" + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { + SFReader *pFReader = new STxtFileReader(pszFName); + char *pszLine = new char[ 100001 ]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) + continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE *fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) + continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + + delete [] pszLine; +} + +struct SArgumentReorderTrainer{ + SArgumentReorderTrainer(const char* pszSRLFname, //source-side srl tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszTopPredicateFname, //target file name + const char* pszInstanceFname, //training instance file name + const char* pszModelFname, //classifier model file name + int iCutoff) { + fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, pszTargetFname, pszTopPredicateFname, pszInstanceFname); + + string strInstanceFname, strModelFname; + strInstanceFname = string(pszInstanceFname) + string(".left"); + strModelFname = string(pszModelFname) + string(".left"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + strInstanceFname = string(pszInstanceFname) + string(".right"); + strModelFname = string(pszModelFname) + string(".right"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + } + + ~SArgumentReorderTrainer() { + + } + +private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { + char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete [] pszNewInstanceFName; + } + + void fnGenerateInstanceFiles(const char* pszSRLFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszTopPredicateFname, //top predicate file name (we only consider predicates with 100+ occurrences + const char* pszInstanceFname //training instance file name + ) { + SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); + SSrlSentenceReader *pSRLReader = new SSrlSentenceReader(pszSRLFname); + STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); + + Map *pMapPredicate; + if (pszTopPredicateFname != NULL) + pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); + else + pMapPredicate = NULL; + + char *pszLine = new char[50001]; + + FILE *fpLeftOut, *fpRightOut; + sprintf(pszLine, "%s.left", pszInstanceFname); + fpLeftOut = fopen(pszLine, "w"); + sprintf(pszLine, "%s.right", pszInstanceFname); + fpRightOut = fopen(pszLine, "w"); + + //read sentence by sentence + SAlignment *pAlign; + SSrlSentence *pSRL; + SParsedTree *pTree; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pSRL = pSRLReader->fnReadNextSrlSentence(); + assert(pSRL != NULL); + pTree = pSRL->m_pTree; + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + //vecTPOSTerms.size() == 0, given the case when an english sentence fails parsing + + if (pTree != NULL) { + for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { + SPredicate *pPred = pSRL->m_vecPred[i]; + if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition]->m_ptParent->m_pszTerm, "VA") == 0) continue; + string strPred = string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); + if (pMapPredicate != NULL) { + Map::iterator iter_map = pMapPredicate->find(strPred); + if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) + continue; + } + + SPredicateItem *pPredItem = new SPredicateItem(pTree, pPred); + + vector vecStrBlock; + for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { + SSRLItem *pItem1 = pPredItem->vec_items_[j]; + vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome(pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, pAlign)); + } + + vector vecStrLeftReorderType; + vector vecStrRightReorderType; + SArgumentReorderModel::fnGetReorderType(pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); + for (int j = 1; j < pPredItem->vec_items_.size(); j++) { + string strLeftOutcome, strRightOutcome; + strLeftOutcome = vecStrLeftReorderType[j - 1]; + strRightOutcome = vecStrRightReorderType[j - 1]; + ostringstream ostr; + SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, vecStrBlock[j - 1], vecStrBlock[j], ostr); + + //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + //fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), strOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); + fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), strRightOutcome.c_str()); + } + } + } + delete pSRL; + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) + fprintf(stderr, "#%d\n", iSentNum); + } + delete [] pszLine; + + fclose(fpLeftOut); + fclose(fpRightOut); + + + delete pAlignReader; + delete pSRLReader; + delete pTxtSReader; + delete pTxtTReader; + } + + + + Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { + if (pszTopPredicateFname == NULL) + return NULL; + + Map* pMapPredicate = new Map(); + STxtFileReader *pReader = new STxtFileReader(pszTopPredicateFname); + char *pszLine = new char[50001]; + int iNumCount = 0; + while (pReader->fnReadNextLine(pszLine, NULL)) { + if (pszLine[0] == '#') + continue; + char *p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + int iCount = atoi(p); + if (iCount < 100) + break; + (*pMapPredicate)[string(pszLine)] = iNumCount++; + } + delete pReader; + delete [] pszLine; + return pMapPredicate; + } +}; + + + +namespace po = boost::program_options; + +inline void print_options(std::ostream &out,po::options_description const& opts) { + typedef std::vector< boost::shared_ptr > Ds; + Ds const& ds=opts.options(); + out << '"'; + for (unsigned i=0;ilong_name(); + } + out << '"\n'; +} +inline string str(char const* name,po::variables_map const& conf) { + return conf[name].as(); +} + +//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file /scratch0/mt_exp/gale-align/gale-align.nw.al --source_file /scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file /scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file /scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix /scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 +//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file /scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file /scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file /scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file /scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix /scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options() + ("srl_file",po::value(),"srl file path (input)") + ("align_file",po::value(),"Alignment file path (input)") + ("source_file",po::value(),"Source text file path (input)") + ("target_file",po::value(),"Target text file path (input)") + ("instance_file",po::value(),"Instance file path (output)") + ("model_prefix",po::value(),"Model file path prefix (output): three files will be generated") + ("feature_cutoff",po::value()->default_value(100),"Feature cutoff threshold") + ("help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("srl_file") + || !vm.count("align_file") + || !vm.count("source_file") + || !vm.count("target_file") + || !vm.count("instance_file") + || !vm.count("model_prefix") + ) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + SArgumentReorderTrainer *pTrainer = new SArgumentReorderTrainer(str("srl_file", vm).c_str(), + str("align_file", vm).c_str(), + str("source_file", vm).c_str(), + str("target_file", vm).c_str(), + NULL, + str("instance_file", vm).c_str(), + str("model_prefix", vm).c_str(), + vm["feature_cutoff"].as()); + delete pTrainer; + + return 1; +} diff --git a/utils/synutils/argument_reorder_model.h b/utils/synutils/argument_reorder_model.h new file mode 100644 index 00000000..b553b34b --- /dev/null +++ b/utils/synutils/argument_reorder_model.h @@ -0,0 +1,222 @@ +/* + * argument_reorder_model.h + * + * Created on: Dec 15, 2013 + * Author: lijunhui + */ + +#ifndef ARGUMENT_REORDER_MODEL_H_ +#define ARGUMENT_REORDER_MODEL_H_ + +#include "alignment.h" +#include "tree.h" +#include "srl_sentence.h" + + +//an argument item or a predicate item (the verb itself) +struct SSRLItem{ + SSRLItem(const STreeItem *tree_item, string role): + tree_item_(tree_item), + role_(role) { + + } + ~SSRLItem() { + + } + const STreeItem *tree_item_; + const string role_; +}; + +struct SPredicateItem{ + SPredicateItem(const SParsedTree *tree, const SPredicate *pred): + pred_(pred) { + vec_items_.reserve(pred->m_vecArgt.size() + 1); + for (int i = 0; i < pred->m_vecArgt.size(); i++) { + vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, string(pred->m_vecArgt[i]->m_pszRole))); + } + vec_items_.push_back(new SSRLItem(tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); + sort(vec_items_.begin(), vec_items_.end(), SortFunction); + + begin_ = vec_items_[0]->tree_item_->m_iBegin; + end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; + } + + ~SPredicateItem() { + vec_items_.clear(); + } + + static bool SortFunction (SSRLItem *i, SSRLItem* j) { return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); } + + vector vec_items_; + int begin_; + int end_; + const SPredicate *pred_; +}; + +struct SArgumentReorderModel { +public: + static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { + return pAlign->fnIsContinuous(iBegin, iEnd); + } + static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, vector& vecStrLeftReorder, vector& vecStrRightReorder) { + vector vecLeft, vecRight; + for (int i = 0; i < pPredItem->vec_items_.size(); i++) { + const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vecStrLeftReorder.clear(); + vecStrRightReorder.clear(); + for (int i = 1; i < vecLeftPosition.size(); i++) { + string strOutcome; + fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); + vecStrLeftReorder.push_back(strOutcome); + fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); + vecStrRightReorder.push_back(strOutcome); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + * + * f1: (left_role, right_role, predicate_term) + * f2: (left_role, right_role, predicate_term, other_right_role) + * f3: (left_role, right_role, predicate_term, other_left_role) + * f4: (left_role, right_role, left_head_pos) + * f5: (left_role, right_role, left_head_word) + * f6: (left_role, right_role, left_syntactic_label) + * f7: (left_role, right_role, right_head_pos) + * f8: (left_role, right_role, right_head_word) + * f8: (left_role, right_role, right_syntactic_label) + * f8: (left_role, right_role, left_chunk_status) + * f9: (left_role, right_role, right_chunk_status) + * f10: (left_role, right_role, left_chunk_status) + * f11: (left_role, right_role, right_chunk_status) + * f12: (left_label, parent_label) + * f13: (right_label, parent_label) + */ + static void fnGenerateFeature(const SParsedTree *pTree, const SPredicate *pPred, const SPredicateItem *pPredItem, int iPos, const string& strBlock1, const string& strBlock2, ostringstream& ostr) { + SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; + SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; + const STreeItem *pCon1 = pSRLItem1->tree_item_; + const STreeItem *pCon2 = pSRLItem2->tree_item_; + + string left_role = pSRLItem1->role_; + string right_role = pSRLItem2->role_; + + string predicate_term = pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) + vec_other_right_sibling.push_back(string(pPredItem->vec_items_[i]->role_)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_right_sibling.push_back(string(pPredItem->vec_items_[i]->role_)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + + //generate features + //f1 + ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; + ostr << "f1=" << left_role << "_" << right_role; + + //f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) { + ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term << "_" << vec_other_right_sibling[i]; + ostr << " f2=" << left_role << "_" << right_role << "_" << vec_other_right_sibling[i]; + } + //f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) { + ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term << "_" << vec_other_left_sibling[i]; + ostr << " f3=" << left_role << "_" << right_role << "_" << vec_other_left_sibling[i]; + } + //f4 + ostr << " f4=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + //f5 + ostr << " f5=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; + //f6 + ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + //f7 + ostr << " f7=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + //f8 + ostr << " f8=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; + //f9 + ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + //f10 + ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; + //f11 + ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; + //f12 + ostr << " f12=" << left_role << "_" << predicate_term; + ostr << " f12=" << left_role; + //f13 + ostr << " f13=" << right_role << "_" << predicate_term; + ostr << " f13=" << right_role; + } + +private: + static void fnGetOutcome(int i1, int i2, string& strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) strOutcome = string("DM"); + else strOutcome = string("M"); + } else { + if (i1 > i2 + 1) strOutcome = string("DS"); + else strOutcome = string("S"); + } + } + + static void fnGetRelativePosition(const vector& vecLeft, vector& vecPosition) { + vecPosition.clear(); + + vector vec; + for (int i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i-1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (int i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (int j = 0; j < vecLeft.size(); j++) { + if ( j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } +}; + + +#endif /* ARGUMENT_REORDER_MODEL_H_ */ diff --git a/utils/synutils/constituent_reorder_model.cc b/utils/synutils/constituent_reorder_model.cc new file mode 100644 index 00000000..485c9667 --- /dev/null +++ b/utils/synutils/constituent_reorder_model.cc @@ -0,0 +1,796 @@ +/* + * constituent_reorder_model.cc + * + * Created on: Jul 10, 2013 + * Author: junhuili + */ + + +#include + +#include "alignment.h" +#include "tree.h" +#include "utility.h" +#include "tsuruoka_maxent.h" + +#include + +using namespace std; + + +typedef std::tr1::unordered_map Map; +typedef std::tr1::unordered_map::iterator Iterator; + +namespace po = boost::program_options; + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { + SFReader *pFReader = new STxtFileReader(pszFName); + char *pszLine = new char[ 100001 ]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) + continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE *fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) + continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + + delete [] pszLine; +} + +struct SConstReorderTrainer{ + SConstReorderTrainer(const char* pszSynFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszInstanceFname, //training instance file name + const char* pszModelPrefix, //classifier model file name prefix + int iClassifierType, //classifier type + int iCutoff, //feature count threshold + const char* pszOption //other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, pszTargetFname, pszInstanceFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + string strModelLeftFname = string(pszModelPrefix) + string(".left"); + string strModelRightFname = string(pszModelPrefix) + string(".right"); + + fprintf(stdout, "...Training the left ordering model\n"); + fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), iCutoff); + fprintf(stdout, "...Training the right ordering model\n"); + fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), iCutoff); + } + ~SConstReorderTrainer() { + + } + +private: + + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { + char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); + pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); + delete pZhangleMaxent;*/ + + Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete [] pszNewInstanceFName; + } + + inline bool fnIsVerbPOS(const char* pszTerm) { + if (strcmp(pszTerm, "VV") == 0 + || strcmp(pszTerm, "VA") == 0 + || strcmp(pszTerm, "VC") == 0 + || strcmp(pszTerm, "VE") == 0) + return true; + return false; + } + + inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, const SAlignment *pAlign, string& strOutcome) { + if (iL1 == -1 && iL2 == -1) + strOutcome = "BU"; //1. both are untranslated + else if (iL1 == -1) + strOutcome = "1U"; //2. XP1 is untranslated + else if (iL2 == -1) + strOutcome = "2U"; //3. XP2 is untranslated + else if (iL1 == iL2 && iR2 == iR2) + strOutcome = "SS"; //4. Have same scope + else if (iL1 <= iL2 && iR1 >= iR2) + strOutcome = "1C2"; //5. XP1's translation covers XP2's + else if (iL1 >= iL2 && iR1 <= iR2) + strOutcome = "2C1"; //6. XP2's translation covers XP1's + else if (iR1 < iL2) { + int i = iR1 + 1; + /*while (i < iL2) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL2) + strOutcome = "M"; //7. Monotone + else + strOutcome = "DM"; //8. Discontinuous monotone + } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) + strOutcome = "OM"; //9. Overlap monotone + else if (iR2 < iL1) { + int i = iR2 + 1; + /*while (i < iL1) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL1) + strOutcome = "S"; //10. Swap + else + strOutcome = "DS"; //11. Discontinuous swap + } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) + strOutcome = "OS"; //12. Overlap swap + else + assert(false); + } + + inline void fnGetOutcome(int i1, int i2, string& strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) strOutcome = string("DM"); + else strOutcome = string("M"); + } else { + if (i1 > i2 + 1) strOutcome = string("DS"); + else strOutcome = string("S"); + } + } + + inline void fnGetRelativePosition(const vector& vecLeft, vector& vecPosition) { + vecPosition.clear(); + + vector vec; + for (size_t i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i-1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (size_t i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (size_t j = 0; j < vecLeft.size(); j++) { + if ( j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + */ + void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pParent, int iPos, const vector& vecChunkStatus, const vector& vecPosition, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { + STreeItem *pCon1, *pCon2; + pCon1 = pParent->m_vecChildren[iPos - 1]; + pCon2 = pParent->m_vecChildren[iPos]; + + fnGetOutcome(vecPosition[iPos-1], vecPosition[iPos], strOutcome); + + string left_label = string(pCon1->m_pszTerm); + string right_label = string(pCon2->m_pszTerm); + string parent_label = string(pParent->m_pszTerm); + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) + vec_other_right_sibling.push_back(string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_left_sibling.push_back(string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + //generate features + //f1 + ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; + //f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) + ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label << "_" << vec_other_right_sibling[i]; + //f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) + ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label << "_" << vec_other_left_sibling[i]; + //f4 + ostr << " f4=" << left_label << "_" << right_label << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + //f5 + ostr << " f5=" << left_label << "_" << right_label << "_" << vecSTerms[pCon1->m_iHeadWord]; + //f6 + ostr << " f6=" << left_label << "_" << right_label << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + //f7 + ostr << " f7=" << left_label << "_" << right_label << "_" << vecSTerms[pCon2->m_iHeadWord]; + //f8 + ostr << " f8=" << left_label << "_" << right_label << "_" << vecChunkStatus[iPos - 1]; + //f9 + ostr << " f9=" << left_label << "_" << right_label << "_" << vecChunkStatus[iPos]; + //f10 + ostr << " f10=" << left_label << "_" << parent_label; + //f11 + ostr << " f11=" << right_label << "_" << parent_label; + } + + /* + * Source side (11 features): + * f1: the categories of XP1 and XP2 (f1_1, f1_2) + * f2: the head words of XP1 and XP2 (f2_1, f2_2) + * f3: the first and last word of XP1 (f3_f, f3_l) + * f4: the first and last word of XP2 (f4_f, f4_l) + * f5: is XP1 or XP2 the head node (f5_1, f5_2) + * f6: the category of the common parent + * Target side (6 features): + * f7: the first and the last word of XP1's translation (f7_f, f7_l) + * f8: the first and the last word of XP2's translation (f8_f, f8_l) + * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) + */ + void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pParent, const STreeItem *pCon1, const STreeItem *pCon2, const SAlignment *pAlign, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { + + int iLeft1, iRight1, iLeft2, iRight2; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); + pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, iRight2); + + fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); + + //generate features + //f1 + ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; + //f2 + ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" << vecSTerms[pCon2->m_iHeadWord]; + //f3 + ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] << " f3_l=" << vecSTerms[pCon1->m_iEnd]; + //f4 + ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] << " f4_l=" << vecSTerms[pCon2->m_iEnd]; + //f5 + if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) + ostr << " f5_1=1"; + else + ostr << " f5_1=0"; + if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) + ostr << " f5_2=1"; + else + ostr << " f5_2=0"; + //f6 + ostr << " f6=" << pParent->m_pszTerm; + + /*//f7 + if (iLeft1 != -1) { + ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << vecTTerms[iRight1]; + } + if (iLeft2 != -1) { + ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << vecTTerms[iRight2]; + } + + const vector* pvecTarget = pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); + string str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_1=" << str.substr(0, str.size()-1); + } + pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); + str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_2=" << str.substr(0, str.size()-1); + } */ + + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, vector& vecFocused){ + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem *pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + //if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + //do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + void fnGenerateInstanceFile(const char* pszSynFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszInstanceFname //training instance file name + ) { + SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader *pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + FILE *fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); + assert(fpLeftOut != NULL); + + FILE *fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); + assert(fpRightOut != NULL); + + //read sentence by sentence + SAlignment *pAlign; + SParsedTree *pTree; + char *pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size(); i++) { + + STreeItem *pParent = vecFocused[i]; + + vector vecLeft, vecRight; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + STreeItem *pCon1 = pParent->m_vecChildren[j]; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vector vecChunkStatus; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + string strOutcome = pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, pParent->m_vecChildren[j]->m_iEnd); + vecChunkStatus.push_back(strOutcome); + } + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + //children[j-1] vs. children[j] reordering + + string strLeftOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, j, vecChunkStatus, vecLeftPosition, vecSTerms, vecTTerms, strLeftOutcome, ostr); + + //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); + + string strRightOutcome; + fnGetOutcome(vecRightPosition[j-1], vecRightPosition[j], strRightOutcome); + fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str(), strRightOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) + fprintf(stderr, "#%d\n", iSentNum); + } + + + fclose(fpLeftOut); + fclose(fpRightOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete [] pszLine; + } + + void fnGenerateInstanceFile2(const char* pszSynFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszInstanceFname //training instance file name + ) { + SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader *pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE *fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + //read sentence by sentence + SAlignment *pAlign; + SParsedTree *pTree; + char *pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem *pParent = vecFocused[i]; + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + //children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j-1], pParent->m_vecChildren[j], pAlign, vecSTerms, vecTTerms, strOutcome, ostr); + + //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) + fprintf(stderr, "#%d\n", iSentNum); + } + + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete [] pszLine; + } +}; + +struct SConstContTrainer{ + SConstContTrainer(const char* pszFlattenedSynFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszInstanceFname, //training instance file name + const char* pszModelPrefix, //classifier model file name prefix + int iClassifierType, //classifier type + int iCutoff, //feature count threshold + const char* pszOption //other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, pszTargetFname, pszInstanceFname); + //fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, pszOption); + fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); + } + ~SConstContTrainer() { + + } + +private: + + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { + char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); + pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); + delete pZhangleMaxent;*/ + + Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete [] pszNewInstanceFName; + } + + + void fnGetFocusedParentNodes(const SParsedTree* pTree, vector& vecFocused){ + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem *pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + //if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + //do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + inline void fnGetOutcome(int iL1, int iR1, const SAlignment *pAlign, string& strOutcome) { + strOutcome = pAlign->fnIsContinuous(iL1, iR1); + } + + inline string fnGetLengthType(int iLen) { + if (iLen == 1) + return string("1"); + if (iLen == 2) + return string("2"); + if (iLen == 3) + return string("3"); + if (iLen < 6) + return string("4"); + if (iLen < 11) + return string("6"); + return string("11"); + } + + /* + * Source side (11 features): + * f1: the syntactic category + * f2: the syntactic category of its parent + * f3: the head word's pos + * f4: =1 if it's the head of its parent node + * or + * the head of its parent node + * f5: length type + */ + void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pCon1, const SAlignment *pAlign, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { + + fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); + + //generate features + //f1 + ostr << "f1=" << pCon1->m_pszTerm; + //f2 + ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; + //f3 + ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + //f4 + if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { + ostr << " f4=1"; + } else { + ostr << " f4=" << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild]->m_pszTerm; + } + //f5 + ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); + } + + void fnGenerateInstanceFile(const char* pszFlattenedSynFname, //source-side flattened parse tree file name + const char* pszAlignFname, //alignment filename + const char* pszSourceFname, //source file name + const char* pszTargetFname, //target file name + const char* pszInstanceFname //training instance file name + ) { + SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader *pParseReader = new SParseReader(pszFlattenedSynFname, true); + STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE *fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + //read sentence by sentence + SAlignment *pAlign; + SParsedTree *pTree; + char *pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTree != NULL); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem *pParent = vecFocused[i]; + + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + //children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, vecSTerms, vecTTerms, strOutcome, ostr); + + //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + + delete pAlign; + delete pTree; + iSentNum++; + + if (iSentNum % 100000 == 0) + fprintf(stderr, "#%d\n", iSentNum); + } + + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete [] pszLine; + } +}; + +inline void print_options(std::ostream &out,po::options_description const& opts) { + typedef std::vector< boost::shared_ptr > Ds; + Ds const& ds=opts.options(); + out << '"'; + for (unsigned i=0;ilong_name(); + } + out << '\n'; +} +inline string str(char const* name,po::variables_map const& conf) { + return conf[name].as(); +} + +//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file /scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file /scratch0/mt_exp/gq-ctb/data/train.cn --target_file /scratch0/mt_exp/gq-ctb/data/train.en --instance_file /scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix /scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 --classifier_type 1 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options() + ("parse_file",po::value(),"parse file path (input)") + ("align_file",po::value(),"Alignment file path (input)") + ("source_file",po::value(),"Source text file path (input)") + ("target_file",po::value(),"Target text file path (input)") + ("instance_file",po::value(),"Instance file path (output)") + ("model_prefix",po::value(),"Model file path prefix (output): three files will be generated") + ("classifier_type",po::value()->default_value(1),"Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for SVMLight") + ("feature_cutoff",po::value()->default_value(100),"Feature cutoff threshold") + ("svm_option",po::value(),"Parameters for SVMLight classifier") + ("help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("parse_file") + || !vm.count("align_file") + || !vm.count("source_file") + || !vm.count("target_file") + || !vm.count("instance_file") + || !vm.count("model_prefix") + ) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + const char *pOption; + if (vm.count("svm_option")) + pOption = str("svm_option", vm).c_str(); + else + pOption = NULL; + + SConstReorderTrainer *pTrainer = new SConstReorderTrainer(str("parse_file", vm).c_str(), + str("align_file", vm).c_str(), + str("source_file", vm).c_str(), + str("target_file", vm).c_str(), + str("instance_file", vm).c_str(), + str("model_prefix", vm).c_str(), + vm["classifier_type"].as(), + vm["feature_cutoff"].as(), + pOption); + delete pTrainer; + + return 1; + +} diff --git a/utils/synutils/maxent-3.0/Makefile.am b/utils/synutils/maxent-3.0/Makefile.am new file mode 100644 index 00000000..64bb038c --- /dev/null +++ b/utils/synutils/maxent-3.0/Makefile.am @@ -0,0 +1,10 @@ +noinst_LIBRARIES = libtsuruoka_maxent.a + +libtsuruoka_maxent_a_SOURCES = \ + lbfgs.cpp \ + maxent.cpp \ + owlqn.cpp \ + sgd.cpp + +AM_CPPFLAGS = -W -Wall + diff --git a/utils/synutils/maxent-3.0/lbfgs.cpp b/utils/synutils/maxent-3.0/lbfgs.cpp new file mode 100644 index 00000000..9eb04bef --- /dev/null +++ b/utils/synutils/maxent-3.0/lbfgs.cpp @@ -0,0 +1,110 @@ +#include +#include +#include +#include +#include "mathvec.h" +#include "lbfgs.h" +#include "maxent.h" + +using namespace std; + +const static int M = LBFGS_M; +const static double LINE_SEARCH_ALPHA = 0.1; +const static double LINE_SEARCH_BETA = 0.5; + +// stopping criteria +int LBFGS_MAX_ITER = 300; +const static double MIN_GRAD_NORM = 0.0001; + + +double +ME_Model::backtracking_line_search( + const Vec & x0, const Vec & grad0, const double f0, + const Vec & dx, Vec & x, Vec & grad1) +{ + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + f = FunctionGradient(x.STLVec(), grad1.STLVec()); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * t * dot_product(dx, grad0)); + + return f; +} + +// +// Jorge Nocedal, "Updating Quasi-Newton Matrices With Limited Storage", +// Mathematics of Computation, Vol. 35, No. 151, pp. 773-782, 1980. +// +Vec +approximate_Hg(const int iter, const Vec & grad, + const Vec s[], const Vec y[], const double z[]) +{ + int offset, bound; + if (iter <= M) { offset = 0; bound = iter; } + else { offset = iter - M; bound = M; } + + Vec q = grad; + double alpha[M], beta[M]; + for (int i = bound - 1; i >= 0; i--) { + const int j = (i + offset) % M; + alpha[i] = z[j] * dot_product(s[j], q); + q += -alpha[i] * y[j]; + } + if (iter > 0) { + const int j = (iter - 1) % M; + const double gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + // static double gamma; + // if (gamma == 0) gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + q *= gamma; + } + for (int i = 0; i <= bound - 1; i++) { + const int j = (i + offset) % M; + beta[i] = z[j] * dot_product(y[j], q); + q += s[j] * (alpha[i] - beta[i]); + } + + return q; +} + +vector +ME_Model::perform_LBFGS(const vector & x0) +{ + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < LBFGS_MAX_ITER; iter++) { + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(grad, grad)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, grad, s, y, z); + + Vec x1(dim), grad1(dim); + f = backtracking_line_search(x, grad, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + x = x1; + grad = grad1; + } + + return x.STLVec(); +} + diff --git a/utils/synutils/maxent-3.0/lbfgs.h b/utils/synutils/maxent-3.0/lbfgs.h new file mode 100644 index 00000000..113919a3 --- /dev/null +++ b/utils/synutils/maxent-3.0/lbfgs.h @@ -0,0 +1,21 @@ +#ifndef _LBFGS_H_ +#define _LBFGS_H_ + +//template +//std::vector +//perform_LBFGS(FuncGrad func_grad, const std::vector & x0); + +std::vector +perform_LBFGS(double (*func_grad)(const std::vector &, std::vector &), + const std::vector & x0); + + +std::vector +perform_OWLQN(double (*func_grad)(const std::vector &, std::vector &), + const std::vector & x0, + const double C); + +//const int LBFGS_M = 7; +const int LBFGS_M = 10; + +#endif diff --git a/utils/synutils/maxent-3.0/mathvec.h b/utils/synutils/maxent-3.0/mathvec.h new file mode 100644 index 00000000..4ec82797 --- /dev/null +++ b/utils/synutils/maxent-3.0/mathvec.h @@ -0,0 +1,97 @@ +#ifndef _MATH_VECTOR_H_ +#define _MATH_VECTOR_H_ + +#include +#include +#include + +class Vec +{ +private: + std::vector _v; +public: + Vec(const size_t n = 0, const double val = 0) { _v.resize(n, val); } + Vec(const std::vector & v) : _v(v) {} + const std::vector & STLVec() const { return _v; } + std::vector & STLVec() { return _v; } + size_t Size() const { return _v.size(); } + double & operator[](int i) { return _v[i]; } + const double & operator[](int i) const { return _v[i]; } + Vec & operator+=(const Vec & b) { + assert(b.Size() == _v.size()); + for (size_t i = 0; i < _v.size(); i++) { + _v[i] += b[i]; + } + return *this; + } + Vec & operator*=(const double c) { + for (size_t i = 0; i < _v.size(); i++) { + _v[i] *= c; + } + return *this; + } + void Project(const Vec & y) { + for (size_t i = 0; i < _v.size(); i++) { + // if (sign(_v[i]) != sign(y[i])) _v[i] = 0; + if (_v[i] * y[i] <=0) _v[i] = 0; + } + } +}; + +inline double dot_product(const Vec & a, const Vec & b) +{ + double sum = 0; + for (size_t i = 0; i < a.Size(); i++) { + sum += a[i] * b[i]; + } + return sum; +} + +inline std::ostream & operator<<(std::ostream & s, const Vec & a) +{ + s << "("; + for (size_t i = 0; i < a.Size(); i++) { + if (i != 0) s << ", "; + s << a[i]; + } + s << ")"; + return s; +} + +inline const Vec operator+(const Vec & a, const Vec & b) +{ + Vec v(a.Size()); + assert(a.Size() == b.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] + b[i]; + } + return v; +} + +inline const Vec operator-(const Vec & a, const Vec & b) +{ + Vec v(a.Size()); + assert(a.Size() == b.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] - b[i]; + } + return v; +} + +inline const Vec operator*(const Vec & a, const double c) +{ + Vec v(a.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] * c; + } + return v; +} + +inline const Vec operator*(const double c, const Vec & a) +{ + return a * c; +} + + + +#endif diff --git a/utils/synutils/maxent-3.0/maxent.cpp b/utils/synutils/maxent-3.0/maxent.cpp new file mode 100644 index 00000000..feb0efdc --- /dev/null +++ b/utils/synutils/maxent-3.0/maxent.cpp @@ -0,0 +1,698 @@ +/* + * $Id: maxent.cpp,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ + */ + +#include "maxent.h" +#include +#include +#include "lbfgs.h" + +using namespace std; + +double +ME_Model::FunctionGradient(const vector & x, vector & grad) +{ + assert((int)_fb.Size() == x.size()); + for (size_t i = 0; i < x.size(); i++) { + _vl[i] = x[i]; + } + + double score = update_model_expectation(); + + if (_l2reg == 0) { + for (size_t i = 0; i < x.size(); i++) { + grad[i] = -(_vee[i] - _vme[i]); + } + } else { + const double c = _l2reg * 2; + for (size_t i = 0; i < x.size(); i++) { + grad[i] = -(_vee[i] - _vme[i] - c * _vl[i]); + } + } + + return -score; +} + +int +ME_Model::perform_GIS(int C) +{ + cerr << "C = " << C << endl; + C = 1; + cerr << "performing AGIS" << endl; + vector pre_v; + double pre_logl = -999999; + for (int iter = 0; iter < 200; iter++) { + + double logl = update_model_expectation(); + fprintf(stderr, "iter = %2d C = %d f = %10.7f train_err = %7.5f", iter, C, logl, _train_error); + if (_heldout.size() > 0) { + double hlogl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", hlogl, _heldout_error); + } + cerr << endl; + + if (logl < pre_logl) { + C += 1; + _vl = pre_v; + iter--; + continue; + } + if (C > 1 && iter % 10 == 0) C--; + + pre_logl = logl; + pre_v = _vl; + for (int i = 0; i < _fb.Size(); i++) { + double coef = _vee[i] / _vme[i]; + _vl[i] += log(coef) / C; + } + } + cerr << endl; + + return 0; +} + +int +ME_Model::perform_QUASI_NEWTON() +{ + const int dim = _fb.Size(); + vector x0(dim); + + for (int i = 0; i < dim; i++) { x0[i] = _vl[i]; } + + vector x; + if (_l1reg > 0) { + cerr << "performing OWLQN" << endl; + x = perform_OWLQN(x0, _l1reg); + } else { + cerr << "performing LBFGS" << endl; + x = perform_LBFGS(x0); + } + + for (int i = 0; i < dim; i++) { _vl[i] = x[i]; } + + return 0; +} + +int +ME_Model::conditional_probability(const Sample & s, + std::vector & membp) const +{ + //int num_classes = membp.size(); + double sum = 0; + int max_label = -1; + // double maxp = 0; + + vector powv(_num_classes, 0.0); + for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ + for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + powv[_fb.Feature(*k).label()] += _vl[*k]; + } + } + for (vector >::const_iterator j = s.rvfeatures.begin(); j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { + powv[_fb.Feature(*k).label()] += _vl[*k] * j->second; + } + } + + std::vector::const_iterator pmax = max_element(powv.begin(), powv.end()); + double offset = max(0.0, *pmax - 700); // to avoid overflow + for (int label = 0; label < _num_classes; label++) { + double pow = powv[label] - offset; + double prod = exp(pow); + // cout << pow << " " << prod << ", "; + // if (_ref_modelp != NULL) prod *= _train_refpd[n][label]; + if (_ref_modelp != NULL) prod *= s.ref_pd[label]; + assert(prod != 0); + membp[label] = prod; + sum += prod; + } + for (int label = 0; label < _num_classes; label++) { + membp[label] /= sum; + if (membp[label] > membp[max_label]) max_label = label; + } + assert(max_label >= 0); + return max_label; +} + +int +ME_Model::make_feature_bag(const int cutoff) +{ + int max_num_features = 0; + + // count the occurrences of features +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + map_type count; + if (cutoff > 0) { + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++) { + for (std::vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++) { + count[ME_Feature(i->label, *j).body()]++; + } + for (std::vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + count[ME_Feature(i->label, j->first).body()]++; + } + } + } + + int n = 0; + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++, n++) { + max_num_features = max(max_num_features, (int)(i->positive_features.size())); + for (std::vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++) { + const ME_Feature feature(i->label, *j); + // if (cutoff > 0 && count[feature.body()] < cutoff) continue; + if (cutoff > 0 && count[feature.body()] <= cutoff) continue; + _fb.Put(feature); + // cout << i->label << "\t" << *j << "\t" << id << endl; + // feature2sample[id].push_back(n); + } + for (std::vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + const ME_Feature feature(i->label, j->first); + // if (cutoff > 0 && count[feature.body()] < cutoff) continue; + if (cutoff > 0 && count[feature.body()] <= cutoff) continue; + _fb.Put(feature); + } + } + count.clear(); + + // cerr << "num_classes = " << _num_classes << endl; + // cerr << "max_num_features = " << max_num_features << endl; + + init_feature2mef(); + + return max_num_features; +} + +double +ME_Model::heldout_likelihood() +{ + double logl = 0; + int ncorrect = 0; + for (std::vector::const_iterator i = _heldout.begin(); i != _heldout.end(); i++) { + vector membp(_num_classes); + int l = classify(*i, membp); + logl += log(membp[i->label]); + if (l == i->label) ncorrect++; + } + _heldout_error = 1 - (double)ncorrect / _heldout.size(); + + return logl /= _heldout.size(); +} + +double +ME_Model::update_model_expectation() +{ + double logl = 0; + int ncorrect = 0; + + _vme.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) _vme[i] = 0; + + int n = 0; + for (vector::const_iterator i = _vs.begin(); i != _vs.end(); i++, n++) { + vector membp(_num_classes); + int max_label = conditional_probability(*i, membp); + + logl += log(membp[i->label]); + // cout << membp[*i] << " " << logl << " "; + if (max_label == i->label) ncorrect++; + + // model_expectation + for (vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++){ + for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()]; + } + } + for (vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()] * j->second; + } + } + + } + + for (int i = 0; i < _fb.Size(); i++) { + _vme[i] /= _vs.size(); + } + + _train_error = 1 - (double)ncorrect / _vs.size(); + + logl /= _vs.size(); + + if (_l2reg > 0) { + const double c = _l2reg; + for (int i = 0; i < _fb.Size(); i++) { + logl -= _vl[i] * _vl[i] * c; + } + } + + //logl /= _vs.size(); + + // fprintf(stderr, "iter =%3d logl = %10.7f train_acc = %7.5f\n", iter, logl, (double)ncorrect/train.size()); + // fprintf(stderr, "logl = %10.7f train_acc = %7.5f\n", logl, (double)ncorrect/_train.size()); + + return logl; +} + +int +ME_Model::train(const vector & vms) +{ + _vs.clear(); + for (vector::const_iterator i = vms.begin(); i != vms.end(); i++) { + add_training_sample(*i); + } + + return train(); +} + +void +ME_Model::add_training_sample(const ME_Sample & mes) +{ + Sample s; + s.label = _label_bag.Put(mes.label); + if (s.label > ME_Feature::MAX_LABEL_TYPES) { + cerr << "error: too many types of labels." << endl; + exit(1); + } + for (vector::const_iterator j = mes.features.begin(); j != mes.features.end(); j++) { + s.positive_features.push_back(_featurename_bag.Put(*j)); + } + for (vector >::const_iterator j = mes.rvfeatures.begin(); j != mes.rvfeatures.end(); j++) { + s.rvfeatures.push_back(pair(_featurename_bag.Put(j->first), j->second)); + } + if (_ref_modelp != NULL) { + ME_Sample tmp = mes;; + s.ref_pd = _ref_modelp->classify(tmp); + } + // cout << s.label << "\t"; + // for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ + // cout << *j << " "; + // } + // cout << endl; + + _vs.push_back(s); +} + +int +ME_Model::train() +{ + if (_l1reg > 0 && _l2reg > 0) { + cerr << "error: L1 and L2 regularizers cannot be used simultaneously." << endl; + return 0; + } + if (_vs.size() == 0) { + cerr << "error: no training data." << endl; + return 0; + } + if (_nheldout >= (int)_vs.size()) { + cerr << "error: too much heldout data. no training data is available." << endl; + return 0; + } + // if (_nheldout > 0) random_shuffle(_vs.begin(), _vs.end()); + + int max_label = 0; + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++) { + max_label = max(max_label, i->label); + } + _num_classes = max_label + 1; + if (_num_classes != _label_bag.Size()) { + cerr << "warning: _num_class != _label_bag.Size()" << endl; + } + + if (_ref_modelp != NULL) { + cerr << "setting reference distribution..."; + for (int i = 0; i < _ref_modelp->num_classes(); i++) { + _label_bag.Put(_ref_modelp->get_class_label(i)); + } + _num_classes = _label_bag.Size(); + for (vector::iterator i = _vs.begin(); i != _vs.end(); i++) { + set_ref_dist(*i); + } + cerr << "done" << endl; + } + + for (int i = 0; i < _nheldout; i++) { + _heldout.push_back(_vs.back()); + _vs.pop_back(); + } + + sort(_vs.begin(), _vs.end()); + + int cutoff = 0; + if (cutoff > 0) cerr << "cutoff threshold = " << cutoff << endl; + if (_l1reg > 0) cerr << "L1 regularizer = " << _l1reg << endl; + if (_l2reg > 0) cerr << "L2 regularizer = " << _l2reg << endl; + + // normalize + _l1reg /= _vs.size(); + _l2reg /= _vs.size(); + + cerr << "preparing for estimation..."; + make_feature_bag(cutoff); + // _vs.clear(); + cerr << "done" << endl; + cerr << "number of samples = " << _vs.size() << endl; + cerr << "number of features = " << _fb.Size() << endl; + + cerr << "calculating empirical expectation..."; + _vee.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) { + _vee[i] = 0; + } + for (int n = 0; n < (int)_vs.size(); n++) { + const Sample * i = &_vs[n]; + for (vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++){ + for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += 1.0; + } + } + + for (vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += j->second; + } + } + + } + for (int i = 0; i < _fb.Size(); i++) { + _vee[i] /= _vs.size(); + } + cerr << "done" << endl; + + _vl.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) _vl[i] = 0.0; + + if (_optimization_method == SGD) { + perform_SGD(); + } else { + perform_QUASI_NEWTON(); + } + + int num_active = 0; + for (int i = 0; i < _fb.Size(); i++) { + if (_vl[i] != 0) num_active++; + } + cerr << "number of active features = " << num_active << endl; + + return 0; +} + +void +ME_Model::get_features(list< pair< pair, double> > & fl) +{ + fl.clear(); + // for (int i = 0; i < _fb.Size(); i++) { + // ME_Feature f = _fb.Feature(i); + // fl.push_back( make_pair(make_pair(_label_bag.Str(f.label()), _featurename_bag.Str(f.feature())), _vl[i])); + // } + for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); + i != _featurename_bag.end(); i++) { + for (int j = 0; j < _label_bag.Size(); j++) { + string label = _label_bag.Str(j); + string history = i->first; + int id = _fb.Id(ME_Feature(j, i->second)); + if (id < 0) continue; + fl.push_back( make_pair(make_pair(label, history), _vl[id]) ); + } + } +} + +void +ME_Model::clear() +{ + _vl.clear(); + _label_bag.Clear(); + _featurename_bag.Clear(); + _fb.Clear(); + _feature2mef.clear(); + _vee.clear(); + _vme.clear(); + _vs.clear(); + _heldout.clear(); +} + +bool +ME_Model::load_from_file(const string & filename) +{ + FILE * fp = fopen(filename.c_str(), "r"); + if (!fp) { + cerr << "error: cannot open " << filename << "!" << endl; + return false; + } + + _vl.clear(); + _label_bag.Clear(); + _featurename_bag.Clear(); + _fb.Clear(); + char buf[1024]; + while(fgets(buf, 1024, fp)) { + string line(buf); + string::size_type t1 = line.find_first_of('\t'); + string::size_type t2 = line.find_last_of('\t'); + string classname = line.substr(0, t1); + string featurename = line.substr(t1 + 1, t2 - (t1 + 1) ); + float lambda; + string w = line.substr(t2+1); + sscanf(w.c_str(), "%f", &lambda); + + int label = _label_bag.Put(classname); + int feature = _featurename_bag.Put(featurename); + _fb.Put(ME_Feature(label, feature)); + _vl.push_back(lambda); + } + + _num_classes = _label_bag.Size(); + + init_feature2mef(); + + fclose(fp); + + return true; +} + +void +ME_Model::init_feature2mef() +{ + _feature2mef.clear(); + for (int i = 0; i < _featurename_bag.Size(); i++) { + vector vi; + for (int k = 0; k < _num_classes; k++) { + int id = _fb.Id(ME_Feature(k, i)); + if (id >= 0) vi.push_back(id); + } + _feature2mef.push_back(vi); + } +} + +bool +ME_Model::load_from_array(const ME_Model_Data data[]) +{ + _vl.clear(); + for (int i = 0;; i++) { + if (string(data[i].label) == "///") break; + int label = _label_bag.Put(data[i].label); + int feature = _featurename_bag.Put(data[i].feature); + _fb.Put(ME_Feature(label, feature)); + _vl.push_back(data[i].weight); + } + _num_classes = _label_bag.Size(); + + init_feature2mef(); + + return true; +} + +bool +ME_Model::save_to_file(const string & filename, const double th) const +{ + FILE * fp = fopen(filename.c_str(), "w"); + if (!fp) { + cerr << "error: cannot open " << filename << "!" << endl; + return false; + } + + // for (int i = 0; i < _fb.Size(); i++) { + // if (_vl[i] == 0) continue; // ignore zero-weight features + // ME_Feature f = _fb.Feature(i); + // fprintf(fp, "%s\t%s\t%f\n", _label_bag.Str(f.label()).c_str(), _featurename_bag.Str(f.feature()).c_str(), _vl[i]); + // } + for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); + i != _featurename_bag.end(); i++) { + for (int j = 0; j < _label_bag.Size(); j++) { + string label = _label_bag.Str(j); + string history = i->first; + int id = _fb.Id(ME_Feature(j, i->second)); + if (id < 0) continue; + if (_vl[id] == 0) continue; // ignore zero-weight features + if (fabs(_vl[id]) < th) continue; // cut off low-weight features + fprintf(fp, "%s\t%s\t%f\n", label.c_str(), history.c_str(), _vl[id]); + } + } + + fclose(fp); + + return true; +} + +void +ME_Model::set_ref_dist(Sample & s) const +{ + vector v0 = s.ref_pd; + vector v(_num_classes); + for (unsigned int i = 0; i < v.size(); i++) { + v[i] = 0; + string label = get_class_label(i); + int id_ref = _ref_modelp->get_class_id(label); + if (id_ref != -1) { + v[i] = v0[id_ref]; + } + if (v[i] == 0) v[i] = 0.001; // to avoid -inf logl + } + s.ref_pd = v; +} + +int +ME_Model::classify(const Sample & nbs, vector & membp) const +{ + // vector membp(_num_classes); + assert(_num_classes == (int)membp.size()); + conditional_probability(nbs, membp); + int max_label = 0; + double max = 0.0; + for (int i = 0; i < (int)membp.size(); i++) { + // cout << membp[i] << " "; + if (membp[i] > max) { max_label = i; max = membp[i]; } + } + // cout << endl; + return max_label; +} + +vector +ME_Model::classify(ME_Sample & mes) const +{ + Sample s; + for (vector::const_iterator j = mes.features.begin(); j != mes.features.end(); j++) { + int id = _featurename_bag.Id(*j); + if (id >= 0) + s.positive_features.push_back(id); + } + for (vector >::const_iterator j = mes.rvfeatures.begin(); j != mes.rvfeatures.end(); j++) { + int id = _featurename_bag.Id(j->first); + if (id >= 0) { + s.rvfeatures.push_back(pair(id, j->second)); + } + } + if (_ref_modelp != NULL) { + s.ref_pd = _ref_modelp->classify(mes); + set_ref_dist(s); + } + + vector vp(_num_classes); + int label = classify(s, vp); + mes.label = get_class_label(label); + return vp; +} + + +/* + * $Log: maxent.cpp,v $ + * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida + * stepp tagger, by Okanohara and Tsuruoka + * + * Revision 1.28 2006/08/21 17:30:38 tsuruoka + * use MAX_LABEL_TYPES + * + * Revision 1.27 2006/07/25 13:19:53 tsuruoka + * sort _vs[] + * + * Revision 1.26 2006/07/18 11:13:15 tsuruoka + * modify comments + * + * Revision 1.25 2006/07/18 10:02:15 tsuruoka + * remove sample2feature[] + * speed up conditional_probability() + * + * Revision 1.24 2006/07/18 05:10:51 tsuruoka + * add ref_dist + * + * Revision 1.23 2005/12/24 07:05:32 tsuruoka + * modify conditional_probability() to avoid overflow + * + * Revision 1.22 2005/12/24 07:01:25 tsuruoka + * add cutoff for real-valued features + * + * Revision 1.21 2005/12/23 10:33:02 tsuruoka + * support real-valued features + * + * Revision 1.20 2005/12/23 09:15:29 tsuruoka + * modify _train to reduce memory consumption + * + * Revision 1.19 2005/10/28 13:10:14 tsuruoka + * fix for overflow (thanks to Ming Li) + * + * Revision 1.18 2005/10/28 13:03:07 tsuruoka + * add progress_bar + * + * Revision 1.17 2005/09/12 13:51:16 tsuruoka + * Sample: list -> vector + * + * Revision 1.16 2005/09/12 13:27:10 tsuruoka + * add add_training_sample() + * + * Revision 1.15 2005/04/27 11:22:27 tsuruoka + * bugfix + * ME_Sample: list -> vector + * + * Revision 1.14 2005/04/27 10:00:42 tsuruoka + * remove tmpfb + * + * Revision 1.13 2005/04/26 14:25:53 tsuruoka + * add MiniStringBag, USE_HASH_MAP + * + * Revision 1.12 2005/02/11 10:20:08 tsuruoka + * modify cutoff + * + * Revision 1.11 2004/10/04 05:50:25 tsuruoka + * add Clear() + * + * Revision 1.10 2004/08/26 16:52:26 tsuruoka + * fix load_from_file() + * + * Revision 1.9 2004/08/09 12:27:21 tsuruoka + * change messages + * + * Revision 1.8 2004/08/04 13:55:18 tsuruoka + * modify _sample2feature + * + * Revision 1.7 2004/07/28 13:42:58 tsuruoka + * add AGIS + * + * Revision 1.6 2004/07/28 05:54:13 tsuruoka + * get_class_name() -> get_class_label() + * ME_Feature: bugfix + * + * Revision 1.5 2004/07/27 16:58:47 tsuruoka + * modify the interface of classify() + * + * Revision 1.4 2004/07/26 17:23:46 tsuruoka + * _sample2feature: list -> vector + * + * Revision 1.3 2004/07/26 15:49:23 tsuruoka + * modify ME_Feature + * + * Revision 1.2 2004/07/26 13:52:18 tsuruoka + * modify cutoff + * + * Revision 1.1 2004/07/26 13:10:55 tsuruoka + * add files + * + * Revision 1.20 2004/07/22 08:34:45 tsuruoka + * modify _sample2feature[] + * + * Revision 1.19 2004/07/21 16:33:01 tsuruoka + * remove some comments + * + */ + diff --git a/utils/synutils/maxent-3.0/maxent.h b/utils/synutils/maxent-3.0/maxent.h new file mode 100644 index 00000000..a4391ead --- /dev/null +++ b/utils/synutils/maxent-3.0/maxent.h @@ -0,0 +1,395 @@ +/* + * $Id: maxent.h,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ + */ + +#ifndef __MAXENT_H_ +#define __MAXENT_H_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include "mathvec.h" + +#define USE_HASH_MAP // if you encounter errors with hash, try commenting out this line. (the program will be a bit slower, though) +#ifdef USE_HASH_MAP +#include +#endif + +// +// data format for each sample for training/testing +// +struct ME_Sample +{ +public: + ME_Sample() : label("") {}; + ME_Sample(const std::string & l) : label(l) {}; + void set_label(const std::string & l) { label = l; } + + // to add a binary feature + void add_feature(const std::string & f) { + features.push_back(f); + } + + // to add a real-valued feature + void add_feature(const std::string & s, const double d) { + rvfeatures.push_back(std::pair(s, d)); + } + +public: + std::string label; + std::vector features; + std::vector > rvfeatures; + + // obsolete + void add_feature(const std::pair & f) { + rvfeatures.push_back(f); // real-valued features + } +}; + + +// +// for those who want to use load_from_array() +// +typedef struct ME_Model_Data +{ + char * label; + char * feature; + double weight; +} ME_Model_Data; + + +class ME_Model +{ +public: + + void add_training_sample(const ME_Sample & s); + int train(); + std::vector classify(ME_Sample & s) const; + bool load_from_file(const std::string & filename); + bool save_to_file(const std::string & filename, const double th = 0) const; + int num_classes() const { return _num_classes; } + std::string get_class_label(int i) const { return _label_bag.Str(i); } + int get_class_id(const std::string & s) const { return _label_bag.Id(s); } + void get_features(std::list< std::pair< std::pair, double> > & fl); + void set_heldout(const int h, const int n = 0) { _nheldout = h; _early_stopping_n = n; }; + void use_l1_regularizer(const double v) { _l1reg = v; } + void use_l2_regularizer(const double v) { _l2reg = v; } + void use_SGD(int iter = 30, double eta0 = 1, double alpha = 0.85) { + _optimization_method = SGD; + SGD_ITER = iter; SGD_ETA0 = eta0; SGD_ALPHA = alpha; + } + bool load_from_array(const ME_Model_Data data[]); + void set_reference_model(const ME_Model & ref_model) { _ref_modelp = &ref_model; }; + void clear(); + + ME_Model() { + _l1reg = _l2reg = 0; + _nheldout = 0; + _early_stopping_n = 0; + _ref_modelp = NULL; + _optimization_method = LBFGS; + } + +public: + // obsolete. just for downward compatibility + int train(const std::vector & train); + +private: + + enum OPTIMIZATION_METHOD { LBFGS, OWLQN, SGD } _optimization_method; + // OWLQN and SGD are available only for L1-regularization + + int SGD_ITER; + double SGD_ETA0; + double SGD_ALPHA; + + double _l1reg, _l2reg; + + struct Sample { + int label; + std::vector positive_features; + std::vector > rvfeatures; + std::vector ref_pd; // reference probability distribution + bool operator<(const Sample & x) const { + for (unsigned int i = 0; i < positive_features.size(); i++) { + if (i >= x.positive_features.size()) return false; + int v0 = positive_features[i]; + int v1 = x.positive_features[i]; + if (v0 < v1) return true; + if (v0 > v1) return false; + } + return false; + } + }; + + struct ME_Feature + { + enum { MAX_LABEL_TYPES = 255 }; + + // ME_Feature(const int l, const int f) : _body((l << 24) + f) { + // assert(l >= 0 && l < 256); + // assert(f >= 0 && f <= 0xffffff); + // }; + // int label() const { return _body >> 24; } + // int feature() const { return _body & 0xffffff; } + ME_Feature(const int l, const int f) : _body((f << 8) + l) { + assert(l >= 0 && l <= MAX_LABEL_TYPES); + assert(f >= 0 && f <= 0xffffff); + }; + int label() const { return _body & 0xff; } + int feature() const { return _body >> 8; } + unsigned int body() const { return _body; } + private: + unsigned int _body; + }; + + struct ME_FeatureBag + { +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + map_type mef2id; + std::vector id2mef; + int Put(const ME_Feature & i) { + map_type::const_iterator j = mef2id.find(i.body()); + if (j == mef2id.end()) { + int id = id2mef.size(); + id2mef.push_back(i); + mef2id[i.body()] = id; + return id; + } + return j->second; + } + int Id(const ME_Feature & i) const { + map_type::const_iterator j = mef2id.find(i.body()); + if (j == mef2id.end()) { + return -1; + } + return j->second; + } + ME_Feature Feature(int id) const { + assert(id >= 0 && id < (int)id2mef.size()); + return id2mef[id]; + } + int Size() const { + return id2mef.size(); + } + void Clear() { + mef2id.clear(); + id2mef.clear(); + } + }; + + struct hashfun_str + { + size_t operator()(const std::string& s) const { + assert(sizeof(int) == 4 && sizeof(char) == 1); + const int* p = reinterpret_cast(s.c_str()); + size_t v = 0; + int n = s.size() / 4; + for (int i = 0; i < n; i++, p++) { + // v ^= *p; + v ^= *p << (4 * (i % 2)); // note) 0 <= char < 128 + } + int m = s.size() % 4; + for (int i = 0; i < m; i++) { + v ^= s[4 * n + i] << (i * 8); + } + return v; + } + }; + + struct MiniStringBag + { +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + int _size; + map_type str2id; + MiniStringBag() : _size(0) {} + int Put(const std::string & i) { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) { + int id = _size; + _size++; + str2id[i] = id; + return id; + } + return j->second; + } + int Id(const std::string & i) const { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) return -1; + return j->second; + } + int Size() const { return _size; } + void Clear() { str2id.clear(); _size = 0; } + map_type::const_iterator begin() const { return str2id.begin(); } + map_type::const_iterator end() const { return str2id.end(); } + }; + + struct StringBag : public MiniStringBag + { + std::vector id2str; + int Put(const std::string & i) { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) { + int id = id2str.size(); + id2str.push_back(i); + str2id[i] = id; + return id; + } + return j->second; + } + std::string Str(const int id) const { + assert(id >= 0 && id < (int)id2str.size()); + return id2str[id]; + } + int Size() const { return id2str.size(); } + void Clear() { + str2id.clear(); + id2str.clear(); + } + }; + + std::vector _vs; // vector of training_samples + StringBag _label_bag; + MiniStringBag _featurename_bag; + std::vector _vl; // vector of lambda + ME_FeatureBag _fb; + int _num_classes; + std::vector _vee; // empirical expectation + std::vector _vme; // empirical expectation + std::vector< std::vector< int > > _feature2mef; + std::vector< Sample > _heldout; + double _train_error; // current error rate on the training data + double _heldout_error; // current error rate on the heldout data + int _nheldout; + int _early_stopping_n; + std::vector _vhlogl; + const ME_Model * _ref_modelp; + + double heldout_likelihood(); + int conditional_probability(const Sample & nbs, std::vector & membp) const; + int make_feature_bag(const int cutoff); + int classify(const Sample & nbs, std::vector & membp) const; + double update_model_expectation(); + int perform_QUASI_NEWTON(); + int perform_SGD(); + int perform_GIS(int C); + std::vector perform_LBFGS(const std::vector & x0); + std::vector perform_OWLQN(const std::vector & x0, const double C); + double backtracking_line_search(const Vec & x0, const Vec & grad0, const double f0, const Vec & dx, Vec & x, Vec & grad1); + double regularized_func_grad(const double C, const Vec & x, Vec & grad); + double constrained_line_search(double C, const Vec & x0, const Vec & grad0, const double f0, const Vec & dx, Vec & x, Vec & grad1); + + + void set_ref_dist(Sample & s) const; + void init_feature2mef(); + + double FunctionGradient(const std::vector & x, std::vector & grad); + static double FunctionGradientWrapper(const std::vector & x, std::vector & grad); + +}; + + +#endif + + +/* + * $Log: maxent.h,v $ + * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida + * stepp tagger, by Okanohara and Tsuruoka + * + * Revision 1.24 2006/08/21 17:30:38 tsuruoka + * use MAX_LABEL_TYPES + * + * Revision 1.23 2006/07/25 13:19:53 tsuruoka + * sort _vs[] + * + * Revision 1.22 2006/07/18 11:13:15 tsuruoka + * modify comments + * + * Revision 1.21 2006/07/18 10:02:15 tsuruoka + * remove sample2feature[] + * speed up conditional_probability() + * + * Revision 1.20 2006/07/18 05:10:51 tsuruoka + * add ref_dist + * + * Revision 1.19 2005/12/23 10:33:02 tsuruoka + * support real-valued features + * + * Revision 1.18 2005/12/23 09:15:29 tsuruoka + * modify _train to reduce memory consumption + * + * Revision 1.17 2005/10/28 13:02:34 tsuruoka + * set_heldout(): add default value + * Feature() + * + * Revision 1.16 2005/09/12 13:51:16 tsuruoka + * Sample: list -> vector + * + * Revision 1.15 2005/09/12 13:27:10 tsuruoka + * add add_training_sample() + * + * Revision 1.14 2005/04/27 11:22:27 tsuruoka + * bugfix + * ME_Sample: list -> vector + * + * Revision 1.13 2005/04/27 10:20:19 tsuruoka + * MiniStringBag -> StringBag + * + * Revision 1.12 2005/04/27 10:00:42 tsuruoka + * remove tmpfb + * + * Revision 1.11 2005/04/26 14:25:53 tsuruoka + * add MiniStringBag, USE_HASH_MAP + * + * Revision 1.10 2004/10/04 05:50:25 tsuruoka + * add Clear() + * + * Revision 1.9 2004/08/09 12:27:21 tsuruoka + * change messages + * + * Revision 1.8 2004/08/04 13:55:19 tsuruoka + * modify _sample2feature + * + * Revision 1.7 2004/07/29 05:51:13 tsuruoka + * remove modeldata.h + * + * Revision 1.6 2004/07/28 13:42:58 tsuruoka + * add AGIS + * + * Revision 1.5 2004/07/28 05:54:14 tsuruoka + * get_class_name() -> get_class_label() + * ME_Feature: bugfix + * + * Revision 1.4 2004/07/27 16:58:47 tsuruoka + * modify the interface of classify() + * + * Revision 1.3 2004/07/26 17:23:46 tsuruoka + * _sample2feature: list -> vector + * + * Revision 1.2 2004/07/26 15:49:23 tsuruoka + * modify ME_Feature + * + * Revision 1.1 2004/07/26 13:10:55 tsuruoka + * add files + * + * Revision 1.18 2004/07/22 08:34:45 tsuruoka + * modify _sample2feature[] + * + * Revision 1.17 2004/07/21 16:33:01 tsuruoka + * remove some comments + * + */ diff --git a/utils/synutils/maxent-3.0/owlqn.cpp b/utils/synutils/maxent-3.0/owlqn.cpp new file mode 100644 index 00000000..7b2cea7d --- /dev/null +++ b/utils/synutils/maxent-3.0/owlqn.cpp @@ -0,0 +1,137 @@ +#include +#include +#include +#include +#include "mathvec.h" +#include "lbfgs.h" +#include "maxent.h" + +using namespace std; + +const static int M = LBFGS_M; +const static double LINE_SEARCH_ALPHA = 0.1; +const static double LINE_SEARCH_BETA = 0.5; + +// stopping criteria +int OWLQN_MAX_ITER = 300; +const static double MIN_GRAD_NORM = 0.0001; + + +Vec approximate_Hg(const int iter, const Vec & grad, + const Vec s[], const Vec y[], const double z[]); + + +inline int sign(double x) +{ + if (x > 0) return 1; + if (x < 0) return -1; + return 0; +}; + +static Vec +pseudo_gradient(const Vec & x, const Vec & grad0, const double C) +{ + Vec grad = grad0; + for (size_t i = 0; i < x.Size(); i++) { + if (x[i] != 0) { + grad[i] += C * sign(x[i]); + continue; + } + const double gm = grad0[i] - C; + if (gm > 0) { + grad[i] = gm; + continue; + } + const double gp = grad0[i] + C; + if (gp < 0) { + grad[i] = gp; + continue; + } + grad[i] = 0; + } + + return grad; +} + +double +ME_Model::regularized_func_grad(const double C, const Vec & x, Vec & grad) +{ + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + for (size_t i = 0; i < x.Size(); i++) { + f += C * fabs(x[i]); + } + + return f; +} + +double +ME_Model::constrained_line_search(double C, + const Vec & x0, const Vec & grad0, const double f0, + const Vec & dx, Vec & x, Vec & grad1) +{ + // compute the orthant to explore + Vec orthant = x0; + for (size_t i = 0; i < orthant.Size(); i++) { + if (orthant[i] == 0) orthant[i] = -grad0[i]; + } + + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + x.Project(orthant); + // for (size_t i = 0; i < x.Size(); i++) { + // if (x0[i] != 0 && sign(x[i]) != sign(x0[i])) x[i] = 0; + // } + + f = regularized_func_grad(C, x, grad1); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * dot_product(x - x0, grad0)); + + return f; +} + +vector +ME_Model::perform_OWLQN(const vector & x0, const double C) +{ + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = regularized_func_grad(C, x, grad); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < OWLQN_MAX_ITER; iter++) { + Vec pg = pseudo_gradient(x, grad, C); + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(pg, pg)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, pg, s, y, z); + if (dot_product(dx, pg) >= 0) + dx.Project(-1 * pg); + + Vec x1(dim), grad1(dim); + f = constrained_line_search(C, x, pg, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + + x = x1; + grad = grad1; + } + + return x.STLVec(); +} + diff --git a/utils/synutils/maxent-3.0/sgd.cpp b/utils/synutils/maxent-3.0/sgd.cpp new file mode 100644 index 00000000..6d28c23f --- /dev/null +++ b/utils/synutils/maxent-3.0/sgd.cpp @@ -0,0 +1,180 @@ +#include "maxent.h" +#include +#include + +using namespace std; + +//const double SGD_ETA0 = 1; +//const double SGD_ITER = 30; +//const double SGD_ALPHA = 0.85; + + +//#define FOLOS_NAIVE +//#define FOLOS_LAZY +#define SGD_CP + +inline void +apply_l1_penalty(const int i, const double u, + vector & _vl, vector & q) +{ + double & w = _vl[i]; + const double z = w; + double & qi = q[i]; + if (w > 0) { + w = max(0.0, w - (u + qi)); + } else if (w < 0) { + w = min(0.0, w + (u - qi)); + } + qi += w - z; +} + +static double +l1norm(const vector& v) +{ + double sum = 0; + for (size_t i = 0; i < v.size(); i++) sum += abs(v[i]); + return sum; +} + +inline void +update_folos_lazy(const int iter_sample, + const int k, vector & _vl, const vector & sum_eta, + vector & last_updated) +{ + const double penalty = sum_eta[iter_sample] - sum_eta[last_updated[k]]; + double & x = _vl[k]; + if (x > 0) x = max(0.0, x - penalty); + else x = min(0.0, x + penalty); + last_updated[k] = iter_sample; +} + +int +ME_Model::perform_SGD() +{ + if (_l2reg > 0) { + cerr << "error: L2 regularization is currently not supported in SGD mode." << endl; + exit(1); + } + + cerr << "performing SGD" << endl; + + const double l1param = _l1reg; + + const int d = _fb.Size(); + + vector ri(_vs.size()); + for (size_t i = 0; i < ri.size(); i++) ri[i] = i; + + vector grad(d); + int iter_sample = 0; + const double eta0 = SGD_ETA0; + + // cerr << "l1param = " << l1param << endl; + cerr << "eta0 = " << eta0 << " alpha = " << SGD_ALPHA << endl; + + double u = 0; + vector q(d, 0); + vector last_updated(d, 0); + vector sum_eta; + sum_eta.push_back(0); + + for (int iter = 0; iter < SGD_ITER; iter++) { + + random_shuffle(ri.begin(), ri.end()); + + double logl = 0; + int ncorrect = 0, ntotal = 0; + for (size_t i = 0; i < _vs.size(); i++, ntotal++, iter_sample++) { + const Sample & s = _vs[ri[i]]; + +#ifdef FOLOS_LAZY + for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ + for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); + } + } +#endif + + vector membp(_num_classes); + const int max_label = conditional_probability(s, membp); + + const double eta = eta0 * pow(SGD_ALPHA, (double)iter_sample / _vs.size()); // exponential decay + // const double eta = eta0 / (1.0 + (double)iter_sample / _vs.size()); + + // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << endl; + u += eta * l1param; + + sum_eta.push_back(sum_eta.back() + eta * l1param); + + logl += log(membp[s.label]); + if (max_label == s.label) ncorrect++; + + // binary features + for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ + for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee); + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + // real-valued features + for (vector >::const_iterator j = s.rvfeatures.begin(); j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee) * j->second; + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + +#ifdef FOLOS_NAIVE + for (size_t j = 0; j < d; j++) { + double & x = _vl[j]; + if (x > 0) x = max(0.0, x - eta * l1param); + else x = min(0.0, x + eta * l1param); + } +#endif + + } + logl /= _vs.size(); + // fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, (double)ncorrect / ntotal); + +#ifdef FOLOS_LAZY + if (l1param > 0) { + for (size_t j = 0; j < d; j++) + update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); + } +#endif + + double f = logl; + if (l1param > 0) { + const double l1 = l1norm(_vl); // this is not accurate when lazy update is used + // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " "; + f -= l1param * l1; + int nonzero = 0; + for (int j = 0; j < d; j++) if (_vl[j] != 0) nonzero++; + // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " << nonzero << endl; + } + // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, (double)ncorrect/ntotal); + // fprintf(stderr, "%4d obj = %f", iter+1, f); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, f, 1 - (double)ncorrect/ntotal); + + if (_nheldout > 0) { + double heldout_logl = heldout_likelihood(); + // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", heldout_logl, 1 - _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + } + fprintf(stderr, "\n"); + + + } + + return 0; +} diff --git a/utils/synutils/srl_sentence.h b/utils/synutils/srl_sentence.h new file mode 100644 index 00000000..6f00267a --- /dev/null +++ b/utils/synutils/srl_sentence.h @@ -0,0 +1,215 @@ +/* + * srl_sentence.h + * + * Created on: May 26, 2013 + * Author: junhuili + */ + +#ifndef SRL_SENTENCE_H_ +#define SRL_SENTENCE_H_ + +#include "tree.h" + +#include + +using namespace std; + +struct SArgument{ + SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { + m_pszRole = new char[strlen(pszRole) + 1]; + strcpy(m_pszRole, pszRole); + m_iBegin = iBegin; + m_iEnd = iEnd; + m_fProb = fProb; + m_pTreeItem = NULL; + } + ~SArgument() { + delete [] m_pszRole; + } + + void fnSetTreeItem(STreeItem *pTreeItem) { + m_pTreeItem = pTreeItem; + if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { + assert(m_pTreeItem->m_iBegin == m_iBegin); + assert(m_pTreeItem->m_iEnd == m_iEnd); + } + } + + char *m_pszRole; //argument rule, e.g., ARG0, ARGM-TMP + int m_iBegin; + int m_iEnd; //the span of the argument, [m_iBegin, m_iEnd] + float m_fProb; //the probability of this role, + STreeItem *m_pTreeItem; +}; + +struct SPredicate{ + SPredicate(const char* pszLemma, int iPosition) { + if (pszLemma != NULL) { + m_pszLemma = new char[strlen(pszLemma) + 1]; + strcpy(m_pszLemma, pszLemma); + } else + m_pszLemma = NULL; + m_iPosition = iPosition; + } + ~SPredicate() { + if (m_pszLemma != NULL) + delete [] m_pszLemma; + for (size_t i = 0; i < m_vecArgt.size(); i++) + delete m_vecArgt[i]; + } + int fnAppend(const char* pszRole, int iBegin, int iEnd) { + SArgument *pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); + return fnAppend(pArgt); + } + int fnAppend(SArgument *pArgt) { + m_vecArgt.push_back(pArgt); + int iPosition = m_vecArgt.size() - 1; + return iPosition; + } + + char *m_pszLemma; //lemma of the predicate, for Chinese, it's always as same as the predicate itself + int m_iPosition; //the position in sentence + vector m_vecArgt; //arguments associated to the predicate +}; + +struct SSrlSentence{ + SSrlSentence() { + m_pTree = NULL; + } + ~SSrlSentence() { + if (m_pTree != NULL) + delete m_pTree; + + for (size_t i = 0; i < m_vecPred.size(); i++) + delete m_vecPred[i]; + } + int fnAppend(const char* pszLemma, int iPosition) { + SPredicate *pPred = new SPredicate(pszLemma, iPosition); + return fnAppend(pPred); + } + int fnAppend(SPredicate* pPred) { + m_vecPred.push_back(pPred); + int iPosition = m_vecPred.size() - 1; + return iPosition; + } + int GetPredicateNum() { + return m_vecPred.size(); + } + + SParsedTree *m_pTree; + vector m_vecPred; +}; + +struct SSrlSentenceReader { + SSrlSentenceReader(const char* pszSrlFname) { + m_fpIn = fopen(pszSrlFname, "r"); + assert(m_fpIn != NULL); + } + ~SSrlSentenceReader() { + if (m_fpIn != NULL) + fclose(m_fpIn); + } + + inline void fnReplaceAll(std::string& str, const std::string& from, const std::string& to) { + size_t start_pos = 0; + while((start_pos = str.find(from, start_pos)) != std::string::npos) { + str.replace(start_pos, from.length(), to); + start_pos += to.length(); // In case 'to' contains 'from', like replacing 'x' with 'yx' + } + } + + //TODO: here only considers flat predicate-argument structure + // i.e., no overlap among them + SSrlSentence* fnReadNextSrlSentence() { + vector > vecContent; + if (fnReadNextContent(vecContent) == false) + return NULL; + + SSrlSentence *pSrlSentence = new SSrlSentence(); + int iSize = vecContent.size(); + //put together syntactic text + std::ostringstream ostr; + for (int i = 0; i < iSize; i++) { + string strSynSeg = vecContent[i][5]; //the 5th column is the syntactic segment + size_t iPosition = strSynSeg.find_first_of('*'); + assert(iPosition != string::npos); + ostringstream ostrTmp; + ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] << ")"; //the 2th column is POS-tag, and the 0th column is word + strSynSeg.replace(iPosition, 1, ostrTmp.str()); + fnReplaceAll(strSynSeg, "(", " ("); + ostr << strSynSeg; + } + string strSyn = ostr.str(); + pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); + pSrlSentence->m_pTree->fnSetHeadWord(); + pSrlSentence->m_pTree->fnSetSpanInfo(); + + //read predicate-argument structure + int iNumPred = vecContent[0].size() - 8; + for (int i = 0; i < iNumPred; i++) { + vector vecRole; + vector vecBegin; + vector vecEnd; + int iPred = -1; + for (int j = 0; j < iSize; j++) { + const char* p = vecContent[j][i + 8].c_str(); + const char* q; + if (p[0] == '(') { + //starting position of an argument(or predicate) + vecBegin.push_back(j); + q = strchr(p, '*'); + assert(q != NULL); + vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); + if (vecRole.back().compare("V") == 0) { + assert(iPred == -1); + iPred = vecRole.size() - 1; + } + } + if (p[strlen(p) - 1] == ')') { + //end position of an argument(or predicate) + vecEnd.push_back(j); + assert(vecBegin.size() == vecEnd.size()); + } + } + assert(iPred != -1); + SPredicate *pPred = new SPredicate(pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, vecBegin[iPred]); + pSrlSentence->fnAppend(pPred); + for (size_t j = 0; j < vecBegin.size(); j++) { + if (j == iPred) + continue; + pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); + pPred->m_vecArgt.back()->fnSetTreeItem(pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], false)); + } + } + return pSrlSentence; + } +private: + bool fnReadNextContent(vector >& vecContent) { + vecContent.clear(); + if (feof(m_fpIn) == true) + return false; + char *pszLine; + pszLine = new char[100001]; + pszLine[0] = '\0'; + int iLen; + while (!feof(m_fpIn)) { + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + if (iLen == 0) + break; //end of this sentence + + vector terms = SplitOnWhitespace(string(pszLine)); + assert(terms.size() > 7); + vecContent.push_back(terms); + } + delete [] pszLine; + return true; + } +private: + FILE *m_fpIn; +}; +#endif /* SRL_SENTENCE_H_ */ diff --git a/utils/synutils/tree.h b/utils/synutils/tree.h new file mode 100644 index 00000000..658a470d --- /dev/null +++ b/utils/synutils/tree.h @@ -0,0 +1,731 @@ +/* + * tree.h + * + * Created on: May 23, 2013 + * Author: lijunhui + */ + +#ifndef TREE_H_ +#define TREE_H_ + +#include +#include +#include + +using namespace std; + +struct STreeItem{ + STreeItem(const char* pszTerm) { + m_pszTerm = new char[ strlen( pszTerm ) + 1 ]; + strcpy( m_pszTerm, pszTerm ); + + m_ptParent = NULL; + m_iBegin = -1; + m_iEnd = -1; + m_iHeadChild = -1; + m_iHeadWord = -1; + m_iBrotherIndex = -1; + } + ~STreeItem( ) { + delete [] m_pszTerm; + for (size_t i = 0; i < m_vecChildren.size(); i++) + delete m_vecChildren[i]; + } + int fnAppend(STreeItem *ptChild) { + m_vecChildren.push_back(ptChild); + ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; + ptChild->m_ptParent = this; + return m_vecChildren.size() - 1; + } + int fnGetChildrenNum() { + return m_vecChildren.size(); + } + + bool fnIsPreTerminal( void ) { + int I; + if ( this == NULL || m_vecChildren.size() == 0 ) + return false; + + for ( I = 0; I < m_vecChildren.size(); I++ ) + if (m_vecChildren[I]->m_vecChildren.size() > 0 ) + return false; + + return true; + } +public: + char *m_pszTerm; + + vector m_vecChildren;//children items + STreeItem *m_ptParent;//the parent item + + int m_iBegin; + int m_iEnd; //the node span words[m_iBegin, m_iEnd] + int m_iHeadChild; //the index of its head child + int m_iHeadWord; //the index of its head word + int m_iBrotherIndex;//the index in his brothers +}; + + +struct SGetHeadWord{ + typedef vector CVectorStr; + SGetHeadWord() { + + } + ~SGetHeadWord() { + + } + int fnGetHeadWord( char *pszCFGLeft, CVectorStr vectRight ) { + //0 indicating from right to left while 1 indicating from left to right + char szaHeadLists[ 201 ] = "0"; + + /* //head rules for Egnlish + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN 0CD 0JJ JJS RB QP 0" ); + */ + + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0ADJP JJ 0AD NN CS 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "0ADVP AD 0" ); + else if( strcmp( pszCFGLeft, "CLP" ) == 0 ) + strcpy( szaHeadLists, "0CLP M 0" ); + else if( strcmp( pszCFGLeft, "CP" ) == 0 ) + strcpy( szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0" ); + else if( strcmp( pszCFGLeft, "DNP" ) == 0 ) + strcpy( szaHeadLists, "0DNP DEG 0DEC 0" ); + else if( strcmp( pszCFGLeft, "DVP" ) == 0 ) + strcpy( szaHeadLists, "0DVP DEV 0" ); + else if( strcmp( pszCFGLeft, "DP" ) == 0 ) + strcpy( szaHeadLists, "1DP DT 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "0VV NR NN 0" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0INTJ IJ 0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LST CD OD 1" ); + else if( strcmp( pszCFGLeft, "IP" ) == 0 ) + strcpy( szaHeadLists, "0IP VP 0VV 0" ); + //strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); + else if( strcmp( pszCFGLeft, "LCP" ) == 0 ) + strcpy( szaHeadLists, "0LCP LC 0" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NP NN NT NR QP 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1PP P 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "0 NP IP VP NT NR NN 0" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0QP CLP CD OD 0" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1" ); + else if( strcmp( pszCFGLeft, "VCD" ) == 0 ) + strcpy( szaHeadLists, "0VCD VV VA VC VE 0" ); + if( strcmp( pszCFGLeft, "VRD" ) == 0 ) + strcpy( szaHeadLists, "0VRD VV VA VC VE 0" ); + else if( strcmp( pszCFGLeft, "VSB" ) == 0 ) + strcpy( szaHeadLists, "0VSB VV VA VC VE 0" ); + else if( strcmp( pszCFGLeft, "VCP" ) == 0 ) + strcpy( szaHeadLists, "0VCP VV VA VC VE 0" ); + else if( strcmp( pszCFGLeft, "VNV" ) == 0 ) + strcpy( szaHeadLists, "0VNV VV VA VC VE 0" ); + else if( strcmp( pszCFGLeft, "VPT" ) == 0 ) + strcpy( szaHeadLists, "0VNV VV VA VC VE 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WHNP NP NN NT NR QP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1WHPP PP P 1" ); + + /* //head rules for GENIA corpus + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN 0CD 0JJ JJS RB QP 0" ); + */ + + return fnMyOwnHeadWordRule( szaHeadLists, vectRight ); + } + +private: + int fnMyOwnHeadWordRule( char *pszaHeadLists, CVectorStr vectRight ) { + char szHeadList[ 201 ], *p; + char szTerm[ 101 ]; + int J; + + p = pszaHeadLists; + + int iCountRight; + + iCountRight = vectRight.size( ); + + szHeadList[ 0 ] = '\0'; + while( 1 ){ + szTerm[ 0 ] = '\0'; + sscanf( p, "%s", szTerm ); + if( strlen( szHeadList ) == 0 ){ + if( strcmp( szTerm, "0" ) == 0 ){ + return iCountRight - 1; + } + if( strcmp( szTerm, "1" ) == 0 ){ + return 0; + } + + sprintf( szHeadList, "%c %s ", szTerm[ 0 ], szTerm + 1 ); + p = strstr( p, szTerm ); + p += strlen( szTerm ); + } + else{ + if( ( szTerm[ 0 ] == '0' ) + ||( szTerm[ 0 ] == '1' ) ){ + if( szHeadList[ 0 ] == '0' ){ + for( J = iCountRight - 1; J >= 0; J -- ){ + sprintf( szTerm, " %s ", vectRight.at( J ).c_str( ) ); + if( strstr( szHeadList, szTerm ) != NULL ) + return J; + } + } + else{ + for( J = 0; J < iCountRight; J ++ ){ + sprintf( szTerm, " %s ", vectRight.at( J ).c_str( ) ); + if( strstr( szHeadList, szTerm ) != NULL ) + return J; + } + } + + szHeadList[ 0 ] = '\0'; + } + else{ + strcat( szHeadList, szTerm ); + strcat( szHeadList, " " ); + + p = strstr( p, szTerm ); + p += strlen( szTerm ); + } + } + } + + return 0; + } + +}; + +struct SParsedTree{ + SParsedTree( ) { + m_ptRoot = NULL; + } + ~SParsedTree( ) { + if (m_ptRoot != NULL) + delete m_ptRoot; + } + static SParsedTree* fnConvertFromString(const char* pszStr) { + if (strcmp(pszStr, "(())") == 0) + return NULL; + SParsedTree* pTree = new SParsedTree(); + + vector vecSyn; + fnReadSyntactic(pszStr, vecSyn); + + int iLeft = 1, iRight = 1; //# left/right parenthesis + + STreeItem *pcurrent; + + pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); + + pcurrent = pTree->m_ptRoot; + + for (size_t i = 2; i < vecSyn.size() - 1; i++) { + if ( strcmp(vecSyn[i].c_str(), "(") == 0 ) + iLeft++; + else if (strcmp(vecSyn[i].c_str(), ")") == 0 ) { + iRight++; + if (pcurrent == NULL) { + //error + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + pcurrent = pcurrent->m_ptParent; + } else { + STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); + pcurrent->fnAppend( ptNewItem ); + pcurrent = ptNewItem; + + if (strcmp(vecSyn[i - 1].c_str(), "(" ) != 0 + && strcmp(vecSyn[i - 1].c_str(), ")" ) != 0 ) { + pTree->m_vecTerminals.push_back(ptNewItem); + pcurrent = pcurrent->m_ptParent; + } + } + } + + if ( iLeft != iRight ) { + //error + fprintf(stderr, "the left and right parentheses are not matched!"); + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + + return pTree; + } + + int fnGetNumWord() { + return m_vecTerminals.size(); + } + + void fnSetSpanInfo() { + int iNextNum = 0; + fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); + } + + void fnSetHeadWord() { + for (size_t i = 0; i < m_vecTerminals.size(); i++) + m_vecTerminals[i]->m_iHeadWord = i; + SGetHeadWord *pGetHeadWord = new SGetHeadWord(); + fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); + delete pGetHeadWord; + } + + STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { + STreeItem *pTreeItem = m_vecTerminals[iLeft]; + + while (pTreeItem->m_iEnd < iRight) { + pTreeItem = pTreeItem->m_ptParent; + if (pTreeItem == NULL) break; + } + if (pTreeItem == NULL) + return NULL; + if (pTreeItem->m_iEnd > iRight) + return NULL; + + assert(pTreeItem->m_iEnd == iRight); + if (bLowest) + return pTreeItem; + + while (pTreeItem->m_ptParent != NULL && pTreeItem->m_ptParent->fnGetChildrenNum() == 1) + pTreeItem = pTreeItem->m_ptParent; + + return pTreeItem; + } + +private: + void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int& iNextNum) { + int I; + int iNumChildren = ptItem->fnGetChildrenNum(); + for ( I = 0; I < iNumChildren; I++ ) + fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[ I ], iNextNum); + + if ( I == 0 ) + { + ptItem->m_iBegin = iNextNum; + ptItem->m_iEnd = iNextNum++; + } + else + { + ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; + ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; + } + } + + + void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, SGetHeadWord *pGetHeadWord) { + int I, iHeadchild; + + if ( ptItem->m_vecChildren.size() == 0 ) + return; + + for ( I = 0; I < ptItem->m_vecChildren.size(); I++ ) + fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); + + vector vecRight; + + + if ( ptItem->m_vecChildren.size() == 1 ) + iHeadchild = 0; + else + { + for ( I = 0; I < ptItem->m_vecChildren.size(); I++ ) + vecRight.push_back( string( ptItem->m_vecChildren[ I ]->m_pszTerm ) ); + + iHeadchild = pGetHeadWord->fnGetHeadWord( ptItem->m_pszTerm, vecRight ); + } + + ptItem->m_iHeadChild = iHeadchild; + ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + + + static void fnReadSyntactic(const char *pszSyn, vector& vec) { + char *p; + int I; + + int iLeftNum, iRightNum; + char *pszTmp, *pszTerm; + pszTmp = new char[strlen(pszSyn)]; + pszTerm = new char[strlen(pszSyn)]; + pszTmp[0] = pszTerm[0] = '\0'; + + vec.clear(); + + char *pszLine; + pszLine = new char[strlen(pszSyn) + 1]; + strcpy( pszLine, pszSyn ); + + char *pszLine2; + + while( 1 ) { + while(( strlen( pszLine ) > 0 ) + &&( pszLine[ strlen( pszLine ) - 1 ] > 0 ) + &&( pszLine[ strlen( pszLine ) - 1 ] <= ' ' ) ) + pszLine[ strlen( pszLine ) - 1 ] = '\0'; + + if( strlen( pszLine ) == 0 ) + break; + + //printf( "%s\n", pszLine ); + pszLine2 = pszLine; + while( pszLine2[ 0 ] <= ' ' ) + pszLine2 ++; + if( pszLine2[ 0 ] == '<' ) + continue; + + sscanf( pszLine2 + 1, "%s", pszTmp ); + + if ( pszLine2[ 0 ] == '(' ) { + iLeftNum = 0; + iRightNum = 0; + } + + p = pszLine2; + while ( 1 ) { + pszTerm[ 0 ] = '\0'; + sscanf( p, "%s", pszTerm ); + + if( strlen( pszTerm ) == 0 ) + break; + p = strstr( p, pszTerm ); + p += strlen( pszTerm ); + + if( ( pszTerm[ 0 ] == '(' ) + ||( pszTerm[ strlen( pszTerm ) - 1 ] == ')' ) ) { + if( pszTerm[ 0 ] == '(' ) { + vec.push_back(string("(")); + iLeftNum++; + + I = 1; + while ( pszTerm[ I ] == '(' && pszTerm[ I ]!= '\0' ) { + vec.push_back(string("(")); + iLeftNum++; + + I++; + } + + if( strlen( pszTerm ) > 1 ) + vec.push_back(string(pszTerm + I)); + } else { + char *pTmp; + pTmp = pszTerm + strlen( pszTerm ) - 1; + while( ( pTmp[ 0 ] == ')' ) && ( pTmp >= pszTerm ) ) + pTmp --; + pTmp[ 1 ] = '\0'; + + if( strlen( pszTerm ) > 0 ) + vec.push_back(string(pszTerm)); + pTmp += 2; + + for( I = 0; I <= (int)strlen( pTmp ); I ++ ) { + vec.push_back(string(")")); + iRightNum++; + } + } + } else { + char *q; + q = strchr( pszTerm, ')' ); + if ( q != NULL ) { + q[ 0 ] = '\0'; + if ( pszTerm[ 0 ] != '\0' ) + vec.push_back(string(pszTerm)); + vec.push_back(string(")")); + iRightNum++; + + q++; + while ( q[ 0 ] == ')' ) { + vec.push_back(string(")")); + q++; + iRightNum ++; + } + + while( q[ 0 ] == '(' ) { + vec.push_back(string("(")); + q++; + iLeftNum++; + } + + if ( q[ 0 ] != '\0' ) + vec.push_back(string(q)); + } + else + vec.push_back(string(pszTerm)); + } + } + + if (iLeftNum != iRightNum) { + fprintf(stderr, "%s\n", pszSyn); + assert(iLeftNum == iRightNum); + } + /*if ( iLeftNum != iRightNum ) { + printf( "ERROR: left( and right ) is not matched, %d ( and %d )\n", iLeftNum, iRightNum ); + return; + }*/ + + + if( vec.size() >= 2 + && strcmp( vec[1].c_str(), "(" ) == 0 ) { + //( (IP..) ) + std::vector::iterator it; + it = vec.begin(); + it++; + vec.insert(it, string("ROOT")); + } + + break; + } + + delete [] pszLine; + delete [] pszTmp; + delete [] pszTerm; + } + +public: + STreeItem *m_ptRoot; + vector m_vecTerminals; //the leaf nodes +}; + +struct SParseReader { + SParseReader(const char* pszParse_Fname, bool bFlattened = false) : + m_bFlattened(bFlattened){ + m_fpIn = fopen( pszParse_Fname, "r" ); + assert(m_fpIn != NULL); + } + ~SParseReader() { + if (m_fpIn != NULL) + fclose(m_fpIn); + } + + SParsedTree* fnReadNextParseTree( ) { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true ) { + if (iLen == 0) + continue; + + pTree = SParsedTree::fnConvertFromString(pszLine); + if (pTree == NULL) + break; + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete [] pszLine; + return pTree; + } + + SParsedTree* fnReadNextParseTreeWithProb(double *pProb) { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true ) { + if (iLen == 0) + continue; + + char *p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + if (pProb) + (*pProb) = atof(pszLine); + + pTree = SParsedTree::fnConvertFromString(p); + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete [] pszLine; + return pTree; + } +private: + /* + * since to the parse tree is a flattened tree, use the head mark to identify head info. + * the head node will be marked as "*XP*" + */ + void fnSetParseTreeHeadInfo(SParsedTree *pTree) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) + pTree->m_vecTerminals[i]->m_iHeadWord = i; + fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); + } + + void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { + if (pTreeItem->m_vecChildren.size() == 0) + return; + + for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++ ) + fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); + + vector vecRight; + + int iHeadchild; + + if (pTreeItem->fnIsPreTerminal()) { + iHeadchild = 0; + } else { + size_t i; + for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { + char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; + if (p[0] == '*' && p[strlen(p) - 1] == '*') { + iHeadchild = i; + p[strlen(p) - 1] = '\0'; + string str = p + 1; + strcpy(p, str.c_str()); //erase the "*..*" + break; + } + } + assert(i < pTreeItem->m_vecChildren.size()); + } + + pTreeItem->m_iHeadChild = iHeadchild; + pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + void fnPostProcessingFlattenedParse(SParsedTree *pTree) { + pTree->fnSetSpanInfo(); + fnSetParseTreeHeadInfo(pTree); + } + bool fnReadNextSentence(char *pszLine, int* piLength) { + if (feof(m_fpIn) == true) + return false; + + int iLen; + + pszLine[ 0 ] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + + if ( piLength != NULL ) + (*piLength) = iLen; + + return true; + } +private: + FILE *m_fpIn; + const bool m_bFlattened; +}; + +#endif /* TREE_H_ */ diff --git a/utils/synutils/tsuruoka_maxent.h b/utils/synutils/tsuruoka_maxent.h new file mode 100644 index 00000000..790876ab --- /dev/null +++ b/utils/synutils/tsuruoka_maxent.h @@ -0,0 +1,160 @@ +/* + * tsuruoka_maxent.h + * + */ + +#ifndef TSURUOKA_MAXENT_H_ +#define TSURUOKA_MAXENT_H_ + +#include "utility.h" +#include "stringlib.h" +#include "maxent-3.0/maxent.h" + +#include +#include +#include +#include +#include + +using namespace std; + + +typedef std::tr1::unordered_map Map; +typedef std::tr1::unordered_map::iterator Iterator; + + + +struct Tsuruoka_Maxent{ + Tsuruoka_Maxent(const char* pszModelFName) { + if (pszModelFName != NULL) { + m_pModel = new ME_Model(); + m_pModel->load_from_file(pszModelFName); + } else + m_pModel = NULL; + } + + ~Tsuruoka_Maxent() { + if (m_pModel != NULL) + delete m_pModel; + } + + void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, const char* pszModelFName, int iNumIteration) { + assert(strcmp(pszAlgorithm, "l1") == 0 + || strcmp(pszAlgorithm, "l2") == 0 + || strcmp(pszAlgorithm, "sgd") == 0 + || strcmp(pszAlgorithm, "SGD") == 0); + FILE *fpIn = fopen(pszInstanceFName, "r"); + + ME_Model *pModel = new ME_Model(); + + char *pszLine = new char[100001]; + int iNumInstances = 0; + int iLen; + while (!feof(fpIn)) { + pszLine[0] = '\0'; + fgets(pszLine, 20000, fpIn); + if (strlen(pszLine) == 0) { + continue; + } + + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + + + iNumInstances++; + + ME_Sample *pmes = new ME_Sample(); + + char *p = strrchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + vector vecContext; + SplitOnWhitespace(string(pszLine), &vecContext); + + pmes->label = string(p); + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + pModel->add_training_sample((*pmes)); + if (iNumInstances % 100000 == 0) + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + delete pmes; + } + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + fclose(fpIn); + + if (strcmp(pszAlgorithm, "l1") == 0) + pModel->use_l1_regularizer(1.0); + else if (strcmp(pszAlgorithm, "l2") == 0) + pModel->use_l2_regularizer(1.0); + else + pModel->use_SGD(); + + pModel->train(); + pModel->save_to_file(pszModelFName); + + delete pModel; + fprintf(stdout, "......Finished Training\n"); + fprintf(stdout, "......Model saved as %s\n", pszModelFName); + delete [] pszLine; + } + + double fnEval(const char* pszContext, const char* pszOutcome) const { + vector vecContext; + ME_Sample *pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + delete pmes; + int iLableID = m_pModel->get_class_id(pszOutcome); + return vecProb[iLableID]; + } + void fnEval(const char* pszContext, vector >& vecOutput) const { + vector vecContext; + ME_Sample *pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(make_pair(label, vecProb[i])); + } + delete pmes; + } + void fnEval(const char* pszContext, vector& vecOutput) const{ + vector vecContext; + ME_Sample *pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(vecProb[i]); + } + delete pmes; + } + int fnGetClassId(const string& strLabel) const { + return m_pModel->get_class_id(strLabel); + } +private: + ME_Model *m_pModel; +}; + + + +#endif /* TSURUOKA_MAXENT_H_ */ diff --git a/utils/synutils/utility.h b/utils/synutils/utility.h new file mode 100644 index 00000000..d3f52ee0 --- /dev/null +++ b/utils/synutils/utility.h @@ -0,0 +1,129 @@ +/* + * utility.h + * + * Created on: Jun 24, 2013 + * Author: lijunhui + */ + +#ifndef UTILITY_H_ +#define UTILITY_H_ + + +#include +#include +#include +#include +#include +#include + +using namespace std; + +typedef std::unordered_map MapString2Int; +typedef std::unordered_map MapString2Float; +typedef std::unordered_map::iterator MapString2FloatIterator; + +using namespace std; + +struct SFReader { + SFReader( ){ } + virtual ~SFReader( ){ } + + virtual bool fnReadNextLine( char *pszLine, int* piLength ) = 0; + virtual bool fnReadNextLine(string& strLine) = 0; +}; + +struct STxtFileReader: public SFReader { + STxtFileReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~STxtFileReader() { + if (m_fpIn != NULL) + fclose(m_fpIn); + } + + bool fnReadNextLine(char *pszLine, int* piLength) { + if (feof(m_fpIn) == true) + return false; + + int iLen; + + pszLine[ 0 ] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + if (iLen == 0) return false; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + + if ( piLength != NULL ) + (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char *pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) strLine = string(pszLine); + else strLine = string(""); + delete [] pszLine; + + return bOut; + } +private: + FILE *m_fpIn; +}; + +struct SGZFileReader: public SFReader { + SGZFileReader(const char* pszFname) { + m_fpIn = gzopen( pszFname, "r" ); + assert(m_fpIn != NULL); + } + ~SGZFileReader() { + if (m_fpIn != NULL) + gzclose(m_fpIn); + } + + bool fnReadNextLine(char *pszLine, int* piLength) { + if ( m_fpIn == NULL ) + exit( 0 ); + if ( gzeof( m_fpIn ) == true ) + return false; + + int iLen; + + pszLine[ 0 ] = '\0'; + + gzgets( m_fpIn, pszLine, 10001 ); + iLen = strlen( pszLine ); + while ( iLen > 0 && pszLine[ iLen - 1 ] > 0 && pszLine[ iLen -1 ] < 33 ) + { + pszLine[ iLen - 1 ] = '\0'; + iLen--; + } + + if ( piLength != NULL ) + (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char *pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) strLine = string(pszLine); + else strLine = string(""); + delete [] pszLine; + + return bOut; + } +private: + gzFile m_fpIn; +}; + + +#endif /* UTILITY_H_ */ + -- cgit v1.2.3 From 93c444e643e76fe18f852ce07635850e963eea16 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Tue, 7 Oct 2014 17:22:11 -0400 Subject: Apply clang-format --- utils/synutils/alignment.h | 348 +++--- utils/synutils/argument_reorder_model.cc | 572 +++++----- utils/synutils/argument_reorder_model.h | 417 +++---- utils/synutils/constituent_reorder_model.cc | 1562 ++++++++++++++------------- utils/synutils/maxent-3.0/lbfgs.cpp | 48 +- utils/synutils/maxent-3.0/lbfgs.h | 24 +- utils/synutils/maxent-3.0/mathvec.h | 48 +- utils/synutils/maxent-3.0/maxent.cpp | 291 ++--- utils/synutils/maxent-3.0/maxent.h | 187 ++-- utils/synutils/maxent-3.0/owlqn.cpp | 46 +- utils/synutils/maxent-3.0/sgd.cpp | 147 +-- utils/synutils/srl_sentence.h | 376 ++++--- utils/synutils/tree.h | 1377 ++++++++++++----------- utils/synutils/tsuruoka_maxent.h | 263 +++-- utils/synutils/utility.h | 188 ++-- 15 files changed, 2961 insertions(+), 2933 deletions(-) (limited to 'utils') diff --git a/utils/synutils/alignment.h b/utils/synutils/alignment.h index f8fe1993..c0648aab 100644 --- a/utils/synutils/alignment.h +++ b/utils/synutils/alignment.h @@ -19,190 +19,184 @@ using namespace std; /* * Note: - * m_vec_s_align.size() may not be equal to the length of source side sentence + * m_vec_s_align.size() may not be equal to the length of source side + *sentence * due to the last words may not be aligned * */ struct SAlignment { - typedef vector SingleAlign; - SAlignment(const char* pszAlign) { - fnInitializeAlignment(pszAlign); - } - ~SAlignment() { - - } - - bool fnIsAligned(int i, bool s) const { - const vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - if ((*palign)[i].size() == 0) - return false; - return true; - } - - /* - * return true if [b, e] is aligned phrases on source side (if s==true) or on the target side (if s==false); - * return false, otherwise. - */ - bool fnIsAlignedPhrase(int b, int e, bool s, int *pob, int *poe) const { - int ob, oe; //[b, e] on the other side - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - - if (ob == -1) { - if (pob != NULL) (*pob) = -1; - if (poe != NULL) (*poe) = -1; - return false; //no aligned word among [b, e] - } - if (pob != NULL) (*pob) = ob; - if (poe != NULL) (*poe) = oe; - - int bb, be; //[b, e] back given [ob, oe] on the other side - if (s == true) - fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); - else - fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); - - if (bb < b || be > e) - return false; - return true; - } - - bool fnIsAlignedTightPhrase(int b, int e, bool s, int *pob, int *poe) const { - const vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - - if ((*palign).size() <= e || (*palign)[b].size() == 0 || (*palign)[e].size() == 0) - return false; - - return fnIsAlignedPhrase(b, e, s, pob, poe); - } - - void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - } - - /* - * look the translation of source[b, e] is continuous or not - * 1) return "Unaligned": if the source[b, e] is translated silently; - * 2) return "Con't": if none of target words in target[.., ..] is exclusively aligned to any word outside source[b, e] - * 3) return "Discon't": otherwise; - */ - string fnIsContinuous(int b, int e) const { - int ob, oe; - fnGetLeftRightMost(b, e, true, ob, oe); - if (ob == -1) return "Unaligned"; - - for (int i = ob; i <= oe; i++) { - if (!fnIsAligned(i, false)) continue; - const SingleAlign& a = m_vec_t_align[i]; - int j; - for (j = 0; j < a.size(); j++) - if (a[j] >= b && a[j] <= e) break; - if (j == a.size()) return "Discon't"; - } - return "Con't"; - } - - const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { - if (s == true) { - if (i >= m_vec_s_align.size()) - return NULL; - return &(m_vec_s_align[i]); - } else { - if (i >= m_vec_t_align.size()) - return NULL; - return &(m_vec_t_align[i]); - } - } -private: - void fnGetLeftRightMost(int b, int e, const vector& align, int& ob, int& oe) const { - ob = oe = -1; - for (int i = b; i <= e && i < align.size(); i++) { - if (align[i].size() > 0) { - if (align[i][0] < ob || ob == -1) - ob = align[i][0]; - if (oe < align[i][align[i].size() - 1]) - oe = align[i][align[i].size() - 1]; - } - } - } - void fnInitializeAlignment(const char* pszAlign) { - m_vec_s_align.clear(); - m_vec_t_align.clear(); - - vector terms = SplitOnWhitespace(string(pszAlign)); - int si, ti; - for (size_t i = 0; i < terms.size(); i++) { - sscanf(terms[i].c_str(), "%d-%d", &si, &ti); - - while (m_vec_s_align.size() <= si) { - SingleAlign sa; - m_vec_s_align.push_back(sa); - } - while (m_vec_t_align.size() <= ti) { - SingleAlign sa; - m_vec_t_align.push_back(sa); - } - - m_vec_s_align[si].push_back(ti); - m_vec_t_align[ti].push_back(si); - } - - //sort - for (size_t i = 0; i < m_vec_s_align.size(); i++) { - std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); - } - for (size_t i = 0; i < m_vec_t_align.size(); i++) { - std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); - } - } - -private: - vector m_vec_s_align; //source side words' alignment - vector m_vec_t_align; //target side words' alignment + typedef vector SingleAlign; + SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } + ~SAlignment() {} + + bool fnIsAligned(int i, bool s) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + if ((*palign)[i].size() == 0) return false; + return true; + } + + /* + * return true if [b, e] is aligned phrases on source side (if s==true) or on + * the target side (if s==false); + * return false, otherwise. + */ + bool fnIsAlignedPhrase(int b, int e, bool s, int* pob, int* poe) const { + int ob, oe; //[b, e] on the other side + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + + if (ob == -1) { + if (pob != NULL) (*pob) = -1; + if (poe != NULL) (*poe) = -1; + return false; // no aligned word among [b, e] + } + if (pob != NULL) (*pob) = ob; + if (poe != NULL) (*poe) = oe; + + int bb, be; //[b, e] back given [ob, oe] on the other side + if (s == true) + fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); + else + fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); + + if (bb < b || be > e) return false; + return true; + } + + bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + + if ((*palign).size() <= e || (*palign)[b].size() == 0 || + (*palign)[e].size() == 0) + return false; + + return fnIsAlignedPhrase(b, e, s, pob, poe); + } + + void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + } + + /* + * look the translation of source[b, e] is continuous or not + * 1) return "Unaligned": if the source[b, e] is translated silently; + * 2) return "Con't": if none of target words in target[.., ..] is exclusively + * aligned to any word outside source[b, e] + * 3) return "Discon't": otherwise; + */ + string fnIsContinuous(int b, int e) const { + int ob, oe; + fnGetLeftRightMost(b, e, true, ob, oe); + if (ob == -1) return "Unaligned"; + + for (int i = ob; i <= oe; i++) { + if (!fnIsAligned(i, false)) continue; + const SingleAlign& a = m_vec_t_align[i]; + int j; + for (j = 0; j < a.size(); j++) + if (a[j] >= b && a[j] <= e) break; + if (j == a.size()) return "Discon't"; + } + return "Con't"; + } + + const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { + if (s == true) { + if (i >= m_vec_s_align.size()) return NULL; + return &(m_vec_s_align[i]); + } else { + if (i >= m_vec_t_align.size()) return NULL; + return &(m_vec_t_align[i]); + } + } + + private: + void fnGetLeftRightMost(int b, int e, const vector& align, + int& ob, int& oe) const { + ob = oe = -1; + for (int i = b; i <= e && i < align.size(); i++) { + if (align[i].size() > 0) { + if (align[i][0] < ob || ob == -1) ob = align[i][0]; + if (oe < align[i][align[i].size() - 1]) + oe = align[i][align[i].size() - 1]; + } + } + } + void fnInitializeAlignment(const char* pszAlign) { + m_vec_s_align.clear(); + m_vec_t_align.clear(); + + vector terms = SplitOnWhitespace(string(pszAlign)); + int si, ti; + for (size_t i = 0; i < terms.size(); i++) { + sscanf(terms[i].c_str(), "%d-%d", &si, &ti); + + while (m_vec_s_align.size() <= si) { + SingleAlign sa; + m_vec_s_align.push_back(sa); + } + while (m_vec_t_align.size() <= ti) { + SingleAlign sa; + m_vec_t_align.push_back(sa); + } + + m_vec_s_align[si].push_back(ti); + m_vec_t_align[ti].push_back(si); + } + + // sort + for (size_t i = 0; i < m_vec_s_align.size(); i++) { + std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); + } + for (size_t i = 0; i < m_vec_t_align.size(); i++) { + std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); + } + } + + private: + vector m_vec_s_align; // source side words' alignment + vector m_vec_t_align; // target side words' alignment }; struct SAlignmentReader { - SAlignmentReader(const char *pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~SAlignmentReader() { - if (m_fpIn != NULL) - fclose(m_fpIn); - } - SAlignment* fnReadNextAlignment() { - if (feof(m_fpIn) == true) - return NULL; - char *pszLine = new char[100001]; - pszLine[0] = '\0'; - fgets(pszLine, 10001, m_fpIn); - int iLen = strlen(pszLine); - if (iLen == 0) - return NULL; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - SAlignment *pAlign = new SAlignment(pszLine); - delete [] pszLine; - return pAlign; - } -private: - FILE *m_fpIn; + SAlignmentReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SAlignmentReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + SAlignment* fnReadNextAlignment() { + if (feof(m_fpIn) == true) return NULL; + char* pszLine = new char[100001]; + pszLine[0] = '\0'; + fgets(pszLine, 10001, m_fpIn); + int iLen = strlen(pszLine); + if (iLen == 0) return NULL; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + SAlignment* pAlign = new SAlignment(pszLine); + delete[] pszLine; + return pAlign; + } + + private: + FILE* m_fpIn; }; - #endif /* ALIGNMENT_H_ */ diff --git a/utils/synutils/argument_reorder_model.cc b/utils/synutils/argument_reorder_model.cc index e30ee971..2b6bcda1 100644 --- a/utils/synutils/argument_reorder_model.cc +++ b/utils/synutils/argument_reorder_model.cc @@ -12,300 +12,312 @@ #include "utility.h" #include "tsuruoka_maxent.h" -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { - SFReader *pFReader = new STxtFileReader(pszFName); - char *pszLine = new char[ 100001 ]; - int iLen; - Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) - continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; - } - } - } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE *fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) - continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; - } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); - } - } - fclose(fpOut); - delete pFReader; - - - delete [] pszLine; +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + SFReader* pFReader = new STxtFileReader(pszFName); + char* pszLine = new char[100001]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE* fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + delete[] pszLine; } -struct SArgumentReorderTrainer{ - SArgumentReorderTrainer(const char* pszSRLFname, //source-side srl tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszTopPredicateFname, //target file name - const char* pszInstanceFname, //training instance file name - const char* pszModelFname, //classifier model file name - int iCutoff) { - fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, pszTargetFname, pszTopPredicateFname, pszInstanceFname); - - string strInstanceFname, strModelFname; - strInstanceFname = string(pszInstanceFname) + string(".left"); - strModelFname = string(pszModelFname) + string(".left"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - strInstanceFname = string(pszInstanceFname) + string(".right"); - strModelFname = string(pszModelFname) + string(".right"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - } - - ~SArgumentReorderTrainer() { - - } - -private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { - char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete [] pszNewInstanceFName; - } - - void fnGenerateInstanceFiles(const char* pszSRLFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszTopPredicateFname, //top predicate file name (we only consider predicates with 100+ occurrences - const char* pszInstanceFname //training instance file name - ) { - SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); - SSrlSentenceReader *pSRLReader = new SSrlSentenceReader(pszSRLFname); - STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); - - Map *pMapPredicate; - if (pszTopPredicateFname != NULL) - pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); - else - pMapPredicate = NULL; - - char *pszLine = new char[50001]; - - FILE *fpLeftOut, *fpRightOut; - sprintf(pszLine, "%s.left", pszInstanceFname); - fpLeftOut = fopen(pszLine, "w"); - sprintf(pszLine, "%s.right", pszInstanceFname); - fpRightOut = fopen(pszLine, "w"); - - //read sentence by sentence - SAlignment *pAlign; - SSrlSentence *pSRL; - SParsedTree *pTree; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pSRL = pSRLReader->fnReadNextSrlSentence(); - assert(pSRL != NULL); - pTree = pSRL->m_pTree; - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - //vecTPOSTerms.size() == 0, given the case when an english sentence fails parsing - - if (pTree != NULL) { - for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { - SPredicate *pPred = pSRL->m_vecPred[i]; - if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition]->m_ptParent->m_pszTerm, "VA") == 0) continue; - string strPred = string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); - if (pMapPredicate != NULL) { - Map::iterator iter_map = pMapPredicate->find(strPred); - if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) - continue; - } - - SPredicateItem *pPredItem = new SPredicateItem(pTree, pPred); - - vector vecStrBlock; - for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { - SSRLItem *pItem1 = pPredItem->vec_items_[j]; - vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome(pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, pAlign)); - } - - vector vecStrLeftReorderType; - vector vecStrRightReorderType; - SArgumentReorderModel::fnGetReorderType(pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); - for (int j = 1; j < pPredItem->vec_items_.size(); j++) { - string strLeftOutcome, strRightOutcome; - strLeftOutcome = vecStrLeftReorderType[j - 1]; - strRightOutcome = vecStrRightReorderType[j - 1]; - ostringstream ostr; - SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, vecStrBlock[j - 1], vecStrBlock[j], ostr); - - //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - //fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), strOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); - fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), strRightOutcome.c_str()); - } - } - } - delete pSRL; - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) - fprintf(stderr, "#%d\n", iSentNum); - } - delete [] pszLine; - - fclose(fpLeftOut); - fclose(fpRightOut); - - - delete pAlignReader; - delete pSRLReader; - delete pTxtSReader; - delete pTxtTReader; - } - - - - Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { - if (pszTopPredicateFname == NULL) - return NULL; - - Map* pMapPredicate = new Map(); - STxtFileReader *pReader = new STxtFileReader(pszTopPredicateFname); - char *pszLine = new char[50001]; - int iNumCount = 0; - while (pReader->fnReadNextLine(pszLine, NULL)) { - if (pszLine[0] == '#') - continue; - char *p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - int iCount = atoi(p); - if (iCount < 100) - break; - (*pMapPredicate)[string(pszLine)] = iNumCount++; - } - delete pReader; - delete [] pszLine; - return pMapPredicate; - } -}; +struct SArgumentReorderTrainer { + SArgumentReorderTrainer( + const char* pszSRLFname, // source-side srl tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelFname, // classifier model file name + int iCutoff) { + fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszTopPredicateFname, + pszInstanceFname); + + string strInstanceFname, strModelFname; + strInstanceFname = string(pszInstanceFname) + string(".left"); + strModelFname = string(pszModelFname) + string(".left"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + strInstanceFname = string(pszInstanceFname) + string(".right"); + strModelFname = string(pszModelFname) + string(".right"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + } + ~SArgumentReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + void fnGenerateInstanceFiles( + const char* pszSRLFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // top predicate file name (we only + // consider predicates with 100+ + // occurrences + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + Map* pMapPredicate; + if (pszTopPredicateFname != NULL) + pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); + else + pMapPredicate = NULL; + + char* pszLine = new char[50001]; + + FILE* fpLeftOut, *fpRightOut; + sprintf(pszLine, "%s.left", pszInstanceFname); + fpLeftOut = fopen(pszLine, "w"); + sprintf(pszLine, "%s.right", pszInstanceFname); + fpRightOut = fopen(pszLine, "w"); + + // read sentence by sentence + SAlignment* pAlign; + SSrlSentence* pSRL; + SParsedTree* pTree; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pSRL = pSRLReader->fnReadNextSrlSentence(); + assert(pSRL != NULL); + pTree = pSRL->m_pTree; + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + // vecTPOSTerms.size() == 0, given the case when an english sentence fails + // parsing + + if (pTree != NULL) { + for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { + SPredicate* pPred = pSRL->m_vecPred[i]; + if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition] + ->m_ptParent->m_pszTerm, + "VA") == 0) + continue; + string strPred = + string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); + if (pMapPredicate != NULL) { + Map::iterator iter_map = pMapPredicate->find(strPred); + if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) + continue; + } + + SPredicateItem* pPredItem = new SPredicateItem(pTree, pPred); + + vector vecStrBlock; + for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { + SSRLItem* pItem1 = pPredItem->vec_items_[j]; + vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome( + pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, + pAlign)); + } + + vector vecStrLeftReorderType; + vector vecStrRightReorderType; + SArgumentReorderModel::fnGetReorderType( + pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); + for (int j = 1; j < pPredItem->vec_items_.size(); j++) { + string strLeftOutcome, strRightOutcome; + strLeftOutcome = vecStrLeftReorderType[j - 1]; + strRightOutcome = vecStrRightReorderType[j - 1]; + ostringstream ostr; + SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, + vecStrBlock[j - 1], + vecStrBlock[j], ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), + // strOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str()); + fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), + strRightOutcome.c_str()); + } + } + } + delete pSRL; + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + delete[] pszLine; + + fclose(fpLeftOut); + fclose(fpRightOut); + + delete pAlignReader; + delete pSRLReader; + delete pTxtSReader; + delete pTxtTReader; + } + + Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { + if (pszTopPredicateFname == NULL) return NULL; + + Map* pMapPredicate = new Map(); + STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); + char* pszLine = new char[50001]; + int iNumCount = 0; + while (pReader->fnReadNextLine(pszLine, NULL)) { + if (pszLine[0] == '#') continue; + char* p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + int iCount = atoi(p); + if (iCount < 100) break; + (*pMapPredicate)[string(pszLine)] = iNumCount++; + } + delete pReader; + delete[] pszLine; + return pMapPredicate; + } +}; namespace po = boost::program_options; -inline void print_options(std::ostream &out,po::options_description const& opts) { - typedef std::vector< boost::shared_ptr > Ds; - Ds const& ds=opts.options(); +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); out << '"'; - for (unsigned i=0;ilong_name(); + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); } out << '"\n'; } -inline string str(char const* name,po::variables_map const& conf) { +inline string str(char const* name, po::variables_map const& conf) { return conf[name].as(); } -//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file /scratch0/mt_exp/gale-align/gale-align.nw.al --source_file /scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file /scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file /scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix /scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 -//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file /scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file /scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file /scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file /scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix /scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 +//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file +///scratch0/mt_exp/gale-align/gale-align.nw.al --source_file +///scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file +///scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file +///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix +///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 +//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file +///scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file +///scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file +///scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file +///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix +///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 int main(int argc, char** argv) { - po::options_description opts("Configuration options"); - opts.add_options() - ("srl_file",po::value(),"srl file path (input)") - ("align_file",po::value(),"Alignment file path (input)") - ("source_file",po::value(),"Source text file path (input)") - ("target_file",po::value(),"Target text file path (input)") - ("instance_file",po::value(),"Instance file path (output)") - ("model_prefix",po::value(),"Model file path prefix (output): three files will be generated") - ("feature_cutoff",po::value()->default_value(100),"Feature cutoff threshold") - ("help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("srl_file") - || !vm.count("align_file") - || !vm.count("source_file") - || !vm.count("target_file") - || !vm.count("instance_file") - || !vm.count("model_prefix") - ) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - SArgumentReorderTrainer *pTrainer = new SArgumentReorderTrainer(str("srl_file", vm).c_str(), - str("align_file", vm).c_str(), - str("source_file", vm).c_str(), - str("target_file", vm).c_str(), - NULL, - str("instance_file", vm).c_str(), - str("model_prefix", vm).c_str(), - vm["feature_cutoff"].as()); - delete pTrainer; - - return 1; + po::options_description opts("Configuration options"); + opts.add_options()("srl_file", po::value(), "srl file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")("help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("srl_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + SArgumentReorderTrainer* pTrainer = new SArgumentReorderTrainer( + str("srl_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), NULL, + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["feature_cutoff"].as()); + delete pTrainer; + + return 1; } diff --git a/utils/synutils/argument_reorder_model.h b/utils/synutils/argument_reorder_model.h index b553b34b..062b8841 100644 --- a/utils/synutils/argument_reorder_model.h +++ b/utils/synutils/argument_reorder_model.h @@ -12,211 +12,228 @@ #include "tree.h" #include "srl_sentence.h" - -//an argument item or a predicate item (the verb itself) -struct SSRLItem{ - SSRLItem(const STreeItem *tree_item, string role): - tree_item_(tree_item), - role_(role) { - - } - ~SSRLItem() { - - } - const STreeItem *tree_item_; - const string role_; +// an argument item or a predicate item (the verb itself) +struct SSRLItem { + SSRLItem(const STreeItem *tree_item, string role) + : tree_item_(tree_item), role_(role) {} + ~SSRLItem() {} + const STreeItem *tree_item_; + const string role_; }; -struct SPredicateItem{ - SPredicateItem(const SParsedTree *tree, const SPredicate *pred): - pred_(pred) { - vec_items_.reserve(pred->m_vecArgt.size() + 1); - for (int i = 0; i < pred->m_vecArgt.size(); i++) { - vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, string(pred->m_vecArgt[i]->m_pszRole))); - } - vec_items_.push_back(new SSRLItem(tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); - sort(vec_items_.begin(), vec_items_.end(), SortFunction); - - begin_ = vec_items_[0]->tree_item_->m_iBegin; - end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; - } - - ~SPredicateItem() { - vec_items_.clear(); - } - - static bool SortFunction (SSRLItem *i, SSRLItem* j) { return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); } - - vector vec_items_; - int begin_; - int end_; - const SPredicate *pred_; +struct SPredicateItem { + SPredicateItem(const SParsedTree *tree, const SPredicate *pred) + : pred_(pred) { + vec_items_.reserve(pred->m_vecArgt.size() + 1); + for (int i = 0; i < pred->m_vecArgt.size(); i++) { + vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, + string(pred->m_vecArgt[i]->m_pszRole))); + } + vec_items_.push_back(new SSRLItem( + tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); + sort(vec_items_.begin(), vec_items_.end(), SortFunction); + + begin_ = vec_items_[0]->tree_item_->m_iBegin; + end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; + } + + ~SPredicateItem() { vec_items_.clear(); } + + static bool SortFunction(SSRLItem *i, SSRLItem *j) { + return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); + } + + vector vec_items_; + int begin_; + int end_; + const SPredicate *pred_; }; struct SArgumentReorderModel { -public: - static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { - return pAlign->fnIsContinuous(iBegin, iEnd); - } - static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, vector& vecStrLeftReorder, vector& vecStrRightReorder) { - vector vecLeft, vecRight; - for (int i = 0; i < pPredItem->vec_items_.size(); i++) { - const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vecStrLeftReorder.clear(); - vecStrRightReorder.clear(); - for (int i = 1; i < vecLeftPosition.size(); i++) { - string strOutcome; - fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); - vecStrLeftReorder.push_back(strOutcome); - fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); - vecStrRightReorder.push_back(strOutcome); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - * - * f1: (left_role, right_role, predicate_term) - * f2: (left_role, right_role, predicate_term, other_right_role) - * f3: (left_role, right_role, predicate_term, other_left_role) - * f4: (left_role, right_role, left_head_pos) - * f5: (left_role, right_role, left_head_word) - * f6: (left_role, right_role, left_syntactic_label) - * f7: (left_role, right_role, right_head_pos) - * f8: (left_role, right_role, right_head_word) - * f8: (left_role, right_role, right_syntactic_label) - * f8: (left_role, right_role, left_chunk_status) - * f9: (left_role, right_role, right_chunk_status) - * f10: (left_role, right_role, left_chunk_status) - * f11: (left_role, right_role, right_chunk_status) - * f12: (left_label, parent_label) - * f13: (right_label, parent_label) - */ - static void fnGenerateFeature(const SParsedTree *pTree, const SPredicate *pPred, const SPredicateItem *pPredItem, int iPos, const string& strBlock1, const string& strBlock2, ostringstream& ostr) { - SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; - SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; - const STreeItem *pCon1 = pSRLItem1->tree_item_; - const STreeItem *pCon2 = pSRLItem2->tree_item_; - - string left_role = pSRLItem1->role_; - string right_role = pSRLItem2->role_; - - string predicate_term = pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; - - vector vec_other_right_sibling; - for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) - vec_other_right_sibling.push_back(string(pPredItem->vec_items_[i]->role_)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); - - vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_right_sibling.push_back(string(pPredItem->vec_items_[i]->role_)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); - - - //generate features - //f1 - ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; - ostr << "f1=" << left_role << "_" << right_role; - - //f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) { - ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term << "_" << vec_other_right_sibling[i]; - ostr << " f2=" << left_role << "_" << right_role << "_" << vec_other_right_sibling[i]; - } - //f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) { - ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term << "_" << vec_other_left_sibling[i]; - ostr << " f3=" << left_role << "_" << right_role << "_" << vec_other_left_sibling[i]; - } - //f4 - ostr << " f4=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - //f5 - ostr << " f5=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; - //f6 - ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - //f7 - ostr << " f7=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - //f8 - ostr << " f8=" << left_role << "_" << right_role << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; - //f9 - ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - //f10 - ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; - //f11 - ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; - //f12 - ostr << " f12=" << left_role << "_" << predicate_term; - ostr << " f12=" << left_role; - //f13 - ostr << " f13=" << right_role << "_" << predicate_term; - ostr << " f13=" << right_role; - } - -private: - static void fnGetOutcome(int i1, int i2, string& strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) strOutcome = string("DM"); - else strOutcome = string("M"); - } else { - if (i1 > i2 + 1) strOutcome = string("DS"); - else strOutcome = string("S"); - } - } - - static void fnGetRelativePosition(const vector& vecLeft, vector& vecPosition) { - vecPosition.clear(); - - vector vec; - for (int i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i-1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (int i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (int j = 0; j < vecLeft.size(); j++) { - if ( j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } + public: + static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { + return pAlign->fnIsContinuous(iBegin, iEnd); + } + static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, + vector &vecStrLeftReorder, + vector &vecStrRightReorder) { + vector vecLeft, vecRight; + for (int i = 0; i < pPredItem->vec_items_.size(); i++) { + const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vecStrLeftReorder.clear(); + vecStrRightReorder.clear(); + for (int i = 1; i < vecLeftPosition.size(); i++) { + string strOutcome; + fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); + vecStrLeftReorder.push_back(strOutcome); + fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); + vecStrRightReorder.push_back(strOutcome); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + * + * f1: (left_role, right_role, predicate_term) + * f2: (left_role, right_role, predicate_term, other_right_role) + * f3: (left_role, right_role, predicate_term, other_left_role) + * f4: (left_role, right_role, left_head_pos) + * f5: (left_role, right_role, left_head_word) + * f6: (left_role, right_role, left_syntactic_label) + * f7: (left_role, right_role, right_head_pos) + * f8: (left_role, right_role, right_head_word) + * f8: (left_role, right_role, right_syntactic_label) + * f8: (left_role, right_role, left_chunk_status) + * f9: (left_role, right_role, right_chunk_status) + * f10: (left_role, right_role, left_chunk_status) + * f11: (left_role, right_role, right_chunk_status) + * f12: (left_label, parent_label) + * f13: (right_label, parent_label) + */ + static void fnGenerateFeature(const SParsedTree *pTree, + const SPredicate *pPred, + const SPredicateItem *pPredItem, int iPos, + const string &strBlock1, + const string &strBlock2, ostringstream &ostr) { + SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; + SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; + const STreeItem *pCon1 = pSRLItem1->tree_item_; + const STreeItem *pCon2 = pSRLItem2->tree_item_; + + string left_role = pSRLItem1->role_; + string right_role = pSRLItem2->role_; + + string predicate_term = + pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) + vec_other_right_sibling.push_back( + string(pPredItem->vec_items_[i]->role_)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_right_sibling.push_back( + string(pPredItem->vec_items_[i]->role_)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; + ostr << "f1=" << left_role << "_" << right_role; + + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) { + ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_right_sibling[i]; + ostr << " f2=" << left_role << "_" << right_role << "_" + << vec_other_right_sibling[i]; + } + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) { + ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_left_sibling[i]; + ostr << " f3=" << left_role << "_" << right_role << "_" + << vec_other_left_sibling[i]; + } + // f4 + ostr << " f4=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; + // f6 + ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f7 + ostr << " f7=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f8 + ostr << " f8=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; + // f9 + ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f10 + ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; + // f11 + ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; + // f12 + ostr << " f12=" << left_role << "_" << predicate_term; + ostr << " f12=" << left_role; + // f13 + ostr << " f13=" << right_role << "_" << predicate_term; + ostr << " f13=" << right_role; + } + + private: + static void fnGetOutcome(int i1, int i2, string &strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = string("DM"); + else + strOutcome = string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = string("DS"); + else + strOutcome = string("S"); + } + } + + static void fnGetRelativePosition(const vector &vecLeft, + vector &vecPosition) { + vecPosition.clear(); + + vector vec; + for (int i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (int i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (int j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } }; - #endif /* ARGUMENT_REORDER_MODEL_H_ */ diff --git a/utils/synutils/constituent_reorder_model.cc b/utils/synutils/constituent_reorder_model.cc index 485c9667..a4fb9627 100644 --- a/utils/synutils/constituent_reorder_model.cc +++ b/utils/synutils/constituent_reorder_model.cc @@ -5,7 +5,6 @@ * Author: junhuili */ - #include #include "alignment.h" @@ -17,780 +16,831 @@ using namespace std; - typedef std::tr1::unordered_map Map; typedef std::tr1::unordered_map::iterator Iterator; namespace po = boost::program_options; -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { - SFReader *pFReader = new STxtFileReader(pszFName); - char *pszLine = new char[ 100001 ]; - int iLen; - Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) - continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; - } - } - } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE *fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) - continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; - } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); - } - } - fclose(fpOut); - delete pFReader; - - - delete [] pszLine; +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + SFReader* pFReader = new STxtFileReader(pszFName); + char* pszLine = new char[100001]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE* fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + delete[] pszLine; } -struct SConstReorderTrainer{ - SConstReorderTrainer(const char* pszSynFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszInstanceFname, //training instance file name - const char* pszModelPrefix, //classifier model file name prefix - int iClassifierType, //classifier type - int iCutoff, //feature count threshold - const char* pszOption //other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, pszTargetFname, pszInstanceFname); - - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - string strModelLeftFname = string(pszModelPrefix) + string(".left"); - string strModelRightFname = string(pszModelPrefix) + string(".right"); - - fprintf(stdout, "...Training the left ordering model\n"); - fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), iCutoff); - fprintf(stdout, "...Training the right ordering model\n"); - fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), iCutoff); - } - ~SConstReorderTrainer() { - - } - -private: - - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { - char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); - pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); - delete pZhangleMaxent;*/ - - Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete [] pszNewInstanceFName; - } - - inline bool fnIsVerbPOS(const char* pszTerm) { - if (strcmp(pszTerm, "VV") == 0 - || strcmp(pszTerm, "VA") == 0 - || strcmp(pszTerm, "VC") == 0 - || strcmp(pszTerm, "VE") == 0) - return true; - return false; - } - - inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, const SAlignment *pAlign, string& strOutcome) { - if (iL1 == -1 && iL2 == -1) - strOutcome = "BU"; //1. both are untranslated - else if (iL1 == -1) - strOutcome = "1U"; //2. XP1 is untranslated - else if (iL2 == -1) - strOutcome = "2U"; //3. XP2 is untranslated - else if (iL1 == iL2 && iR2 == iR2) - strOutcome = "SS"; //4. Have same scope - else if (iL1 <= iL2 && iR1 >= iR2) - strOutcome = "1C2"; //5. XP1's translation covers XP2's - else if (iL1 >= iL2 && iR1 <= iR2) - strOutcome = "2C1"; //6. XP2's translation covers XP1's - else if (iR1 < iL2) { - int i = iR1 + 1; - /*while (i < iL2) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL2) - strOutcome = "M"; //7. Monotone - else - strOutcome = "DM"; //8. Discontinuous monotone - } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) - strOutcome = "OM"; //9. Overlap monotone - else if (iR2 < iL1) { - int i = iR2 + 1; - /*while (i < iL1) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL1) - strOutcome = "S"; //10. Swap - else - strOutcome = "DS"; //11. Discontinuous swap - } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) - strOutcome = "OS"; //12. Overlap swap - else - assert(false); - } - - inline void fnGetOutcome(int i1, int i2, string& strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) strOutcome = string("DM"); - else strOutcome = string("M"); - } else { - if (i1 > i2 + 1) strOutcome = string("DS"); - else strOutcome = string("S"); - } - } - - inline void fnGetRelativePosition(const vector& vecLeft, vector& vecPosition) { - vecPosition.clear(); - - vector vec; - for (size_t i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i-1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (size_t i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (size_t j = 0; j < vecLeft.size(); j++) { - if ( j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - */ - void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pParent, int iPos, const vector& vecChunkStatus, const vector& vecPosition, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { - STreeItem *pCon1, *pCon2; - pCon1 = pParent->m_vecChildren[iPos - 1]; - pCon2 = pParent->m_vecChildren[iPos]; - - fnGetOutcome(vecPosition[iPos-1], vecPosition[iPos], strOutcome); - - string left_label = string(pCon1->m_pszTerm); - string right_label = string(pCon2->m_pszTerm); - string parent_label = string(pParent->m_pszTerm); - - vector vec_other_right_sibling; - for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) - vec_other_right_sibling.push_back(string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); - vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_left_sibling.push_back(string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); - - //generate features - //f1 - ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; - //f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) - ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label << "_" << vec_other_right_sibling[i]; - //f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) - ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label << "_" << vec_other_left_sibling[i]; - //f4 - ostr << " f4=" << left_label << "_" << right_label << "_" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - //f5 - ostr << " f5=" << left_label << "_" << right_label << "_" << vecSTerms[pCon1->m_iHeadWord]; - //f6 - ostr << " f6=" << left_label << "_" << right_label << "_" << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - //f7 - ostr << " f7=" << left_label << "_" << right_label << "_" << vecSTerms[pCon2->m_iHeadWord]; - //f8 - ostr << " f8=" << left_label << "_" << right_label << "_" << vecChunkStatus[iPos - 1]; - //f9 - ostr << " f9=" << left_label << "_" << right_label << "_" << vecChunkStatus[iPos]; - //f10 - ostr << " f10=" << left_label << "_" << parent_label; - //f11 - ostr << " f11=" << right_label << "_" << parent_label; - } - - /* - * Source side (11 features): - * f1: the categories of XP1 and XP2 (f1_1, f1_2) - * f2: the head words of XP1 and XP2 (f2_1, f2_2) - * f3: the first and last word of XP1 (f3_f, f3_l) - * f4: the first and last word of XP2 (f4_f, f4_l) - * f5: is XP1 or XP2 the head node (f5_1, f5_2) - * f6: the category of the common parent - * Target side (6 features): - * f7: the first and the last word of XP1's translation (f7_f, f7_l) - * f8: the first and the last word of XP2's translation (f8_f, f8_l) - * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) - */ - void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pParent, const STreeItem *pCon1, const STreeItem *pCon2, const SAlignment *pAlign, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { - - int iLeft1, iRight1, iLeft2, iRight2; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); - pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, iRight2); - - fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); - - //generate features - //f1 - ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; - //f2 - ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" << vecSTerms[pCon2->m_iHeadWord]; - //f3 - ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] << " f3_l=" << vecSTerms[pCon1->m_iEnd]; - //f4 - ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] << " f4_l=" << vecSTerms[pCon2->m_iEnd]; - //f5 - if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) - ostr << " f5_1=1"; - else - ostr << " f5_1=0"; - if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) - ostr << " f5_2=1"; - else - ostr << " f5_2=0"; - //f6 - ostr << " f6=" << pParent->m_pszTerm; - - /*//f7 - if (iLeft1 != -1) { - ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << vecTTerms[iRight1]; - } - if (iLeft2 != -1) { - ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << vecTTerms[iRight2]; - } - - const vector* pvecTarget = pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); - string str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_1=" << str.substr(0, str.size()-1); - } - pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); - str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_2=" << str.substr(0, str.size()-1); - } */ - - } - - void fnGetFocusedParentNodes(const SParsedTree* pTree, vector& vecFocused){ - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem *pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - //if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - //do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - void fnGenerateInstanceFile(const char* pszSynFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszInstanceFname //training instance file name - ) { - SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader *pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); - - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - FILE *fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); - assert(fpLeftOut != NULL); - - FILE *fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); - assert(fpRightOut != NULL); - - //read sentence by sentence - SAlignment *pAlign; - SParsedTree *pTree; - char *pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; i < vecFocused.size(); i++) { - - STreeItem *pParent = vecFocused[i]; - - vector vecLeft, vecRight; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - STreeItem *pCon1 = pParent->m_vecChildren[j]; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vector vecChunkStatus; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - string strOutcome = pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, pParent->m_vecChildren[j]->m_iEnd); - vecChunkStatus.push_back(strOutcome); - } - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - //children[j-1] vs. children[j] reordering - - string strLeftOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, j, vecChunkStatus, vecLeftPosition, vecSTerms, vecTTerms, strLeftOutcome, ostr); - - //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str()); - - string strRightOutcome; - fnGetOutcome(vecRightPosition[j-1], vecRightPosition[j], strRightOutcome); - fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), strLeftOutcome.c_str(), strRightOutcome.c_str()); - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) - fprintf(stderr, "#%d\n", iSentNum); - } - - - fclose(fpLeftOut); - fclose(fpRightOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete [] pszLine; - } - - void fnGenerateInstanceFile2(const char* pszSynFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszInstanceFname //training instance file name - ) { - SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader *pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); - - FILE *fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); - - //read sentence by sentence - SAlignment *pAlign; - SParsedTree *pTree; - char *pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem *pParent = vecFocused[i]; - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - //children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j-1], pParent->m_vecChildren[j], pAlign, vecSTerms, vecTTerms, strOutcome, ostr); - - //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) - fprintf(stderr, "#%d\n", iSentNum); - } - - - fclose(fpOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete [] pszLine; - } +struct SConstReorderTrainer { + SConstReorderTrainer( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold + const char* pszOption // other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszInstanceFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + string strModelLeftFname = string(pszModelPrefix) + string(".left"); + string strModelRightFname = string(pszModelPrefix) + string(".right"); + + fprintf(stdout, "...Training the left ordering model\n"); + fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), + iCutoff); + fprintf(stdout, "...Training the right ordering model\n"); + fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), + iCutoff); + } + ~SConstReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); +pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); +delete pZhangleMaxent;*/ + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + inline bool fnIsVerbPOS(const char* pszTerm) { + if (strcmp(pszTerm, "VV") == 0 || strcmp(pszTerm, "VA") == 0 || + strcmp(pszTerm, "VC") == 0 || strcmp(pszTerm, "VE") == 0) + return true; + return false; + } + + inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, + const SAlignment* pAlign, string& strOutcome) { + if (iL1 == -1 && iL2 == -1) + strOutcome = "BU"; // 1. both are untranslated + else if (iL1 == -1) + strOutcome = "1U"; // 2. XP1 is untranslated + else if (iL2 == -1) + strOutcome = "2U"; // 3. XP2 is untranslated + else if (iL1 == iL2 && iR2 == iR2) + strOutcome = "SS"; // 4. Have same scope + else if (iL1 <= iL2 && iR1 >= iR2) + strOutcome = "1C2"; // 5. XP1's translation covers XP2's + else if (iL1 >= iL2 && iR1 <= iR2) + strOutcome = "2C1"; // 6. XP2's translation covers XP1's + else if (iR1 < iL2) { + int i = iR1 + 1; + /*while (i < iL2) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL2) + strOutcome = "M"; // 7. Monotone + else + strOutcome = "DM"; // 8. Discontinuous monotone + } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) + strOutcome = "OM"; // 9. Overlap monotone + else if (iR2 < iL1) { + int i = iR2 + 1; + /*while (i < iL1) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL1) + strOutcome = "S"; // 10. Swap + else + strOutcome = "DS"; // 11. Discontinuous swap + } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) + strOutcome = "OS"; // 12. Overlap swap + else + assert(false); + } + + inline void fnGetOutcome(int i1, int i2, string& strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = string("DM"); + else + strOutcome = string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = string("DS"); + else + strOutcome = string("S"); + } + } + + inline void fnGetRelativePosition(const vector& vecLeft, + vector& vecPosition) { + vecPosition.clear(); + + vector vec; + for (size_t i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (size_t i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (size_t j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + int iPos, const vector& vecChunkStatus, + const vector& vecPosition, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + STreeItem* pCon1, *pCon2; + pCon1 = pParent->m_vecChildren[iPos - 1]; + pCon2 = pParent->m_vecChildren[iPos]; + + fnGetOutcome(vecPosition[iPos - 1], vecPosition[iPos], strOutcome); + + string left_label = string(pCon1->m_pszTerm); + string right_label = string(pCon2->m_pszTerm); + string parent_label = string(pParent->m_pszTerm); + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) + vec_other_right_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_left_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) + ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_right_sibling[i]; + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) + ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_left_sibling[i]; + // f4 + ostr << " f4=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon1->m_iHeadWord]; + // f6 + ostr << " f6=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f7 + ostr << " f7=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon2->m_iHeadWord]; + // f8 + ostr << " f8=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos - 1]; + // f9 + ostr << " f9=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos]; + // f10 + ostr << " f10=" << left_label << "_" << parent_label; + // f11 + ostr << " f11=" << right_label << "_" << parent_label; + } + + /* + * Source side (11 features): + * f1: the categories of XP1 and XP2 (f1_1, f1_2) + * f2: the head words of XP1 and XP2 (f2_1, f2_2) + * f3: the first and last word of XP1 (f3_f, f3_l) + * f4: the first and last word of XP2 (f4_f, f4_l) + * f5: is XP1 or XP2 the head node (f5_1, f5_2) + * f6: the category of the common parent + * Target side (6 features): + * f7: the first and the last word of XP1's translation (f7_f, f7_l) + * f8: the first and the last word of XP2's translation (f8_f, f8_l) + * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + const STreeItem* pCon1, const STreeItem* pCon2, + const SAlignment* pAlign, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + + int iLeft1, iRight1, iLeft2, iRight2; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, + iRight2); + + fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); + + // generate features + // f1 + ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; + // f2 + ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" + << vecSTerms[pCon2->m_iHeadWord]; + // f3 + ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] + << " f3_l=" << vecSTerms[pCon1->m_iEnd]; + // f4 + ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] + << " f4_l=" << vecSTerms[pCon2->m_iEnd]; + // f5 + if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) + ostr << " f5_1=1"; + else + ostr << " f5_1=0"; + if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) + ostr << " f5_2=1"; + else + ostr << " f5_2=0"; + // f6 + ostr << " f6=" << pParent->m_pszTerm; + + /*//f7 + if (iLeft1 != -1) { + ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << + vecTTerms[iRight1]; + } + if (iLeft2 != -1) { + ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << + vecTTerms[iRight2]; + } + + const vector* pvecTarget = + pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); + string str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_1=" << str.substr(0, str.size()-1); + } + pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); + str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_2=" << str.substr(0, str.size()-1); + } */ + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, + vector& vecFocused) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - + // pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + // do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + void fnGenerateInstanceFile( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + FILE* fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); + assert(fpLeftOut != NULL); + + FILE* fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); + assert(fpRightOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size(); i++) { + + STreeItem* pParent = vecFocused[i]; + + vector vecLeft, vecRight; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + STreeItem* pCon1 = pParent->m_vecChildren[j]; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, + iLeft1, iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vector vecChunkStatus; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + string strOutcome = + pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, + pParent->m_vecChildren[j]->m_iEnd); + vecChunkStatus.push_back(strOutcome); + } + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strLeftOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, j, vecChunkStatus, + vecLeftPosition, vecSTerms, vecTTerms, + strLeftOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strLeftOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str()); + + string strRightOutcome; + fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], + strRightOutcome); + fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str(), strRightOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpLeftOut); + fclose(fpRightOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } + + void fnGenerateInstanceFile2( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE* fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; + i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem* pParent = vecFocused[i]; + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j - 1], + pParent->m_vecChildren[j], pAlign, vecSTerms, + vecTTerms, strOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } }; -struct SConstContTrainer{ - SConstContTrainer(const char* pszFlattenedSynFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszInstanceFname, //training instance file name - const char* pszModelPrefix, //classifier model file name prefix - int iClassifierType, //classifier type - int iCutoff, //feature count threshold - const char* pszOption //other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, pszTargetFname, pszInstanceFname); - //fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, pszOption); - fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); - } - ~SConstContTrainer() { - - } - -private: - - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, int iCutoff) { - char *pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); - pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); - delete pZhangleMaxent;*/ - - Tsuruoka_Maxent *pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete [] pszNewInstanceFName; - } - - - void fnGetFocusedParentNodes(const SParsedTree* pTree, vector& vecFocused){ - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem *pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - //if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - //do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - inline void fnGetOutcome(int iL1, int iR1, const SAlignment *pAlign, string& strOutcome) { - strOutcome = pAlign->fnIsContinuous(iL1, iR1); - } - - inline string fnGetLengthType(int iLen) { - if (iLen == 1) - return string("1"); - if (iLen == 2) - return string("2"); - if (iLen == 3) - return string("3"); - if (iLen < 6) - return string("4"); - if (iLen < 11) - return string("6"); - return string("11"); - } - - /* - * Source side (11 features): - * f1: the syntactic category - * f2: the syntactic category of its parent - * f3: the head word's pos - * f4: =1 if it's the head of its parent node - * or - * the head of its parent node - * f5: length type - */ - void fnGenerateInstance(const SParsedTree *pTree, const STreeItem *pCon1, const SAlignment *pAlign, const vector& vecSTerms, const vector& vecTTerms, string& strOutcome, ostringstream& ostr) { - - fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); - - //generate features - //f1 - ostr << "f1=" << pCon1->m_pszTerm; - //f2 - ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; - //f3 - ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - //f4 - if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { - ostr << " f4=1"; - } else { - ostr << " f4=" << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild]->m_pszTerm; - } - //f5 - ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); - } - - void fnGenerateInstanceFile(const char* pszFlattenedSynFname, //source-side flattened parse tree file name - const char* pszAlignFname, //alignment filename - const char* pszSourceFname, //source file name - const char* pszTargetFname, //target file name - const char* pszInstanceFname //training instance file name - ) { - SAlignmentReader *pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader *pParseReader = new SParseReader(pszFlattenedSynFname, true); - STxtFileReader *pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader *pTxtTReader = new STxtFileReader(pszTargetFname); - - FILE *fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); - - //read sentence by sentence - SAlignment *pAlign; - SParsedTree *pTree; - char *pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTree != NULL); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem *pParent = vecFocused[i]; - - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - //children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, vecSTerms, vecTTerms, strOutcome, ostr); - - //fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - } - } - - delete pAlign; - delete pTree; - iSentNum++; - - if (iSentNum % 100000 == 0) - fprintf(stderr, "#%d\n", iSentNum); - } - - - fclose(fpOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete [] pszLine; - } +struct SConstContTrainer { + SConstContTrainer( + const char* pszFlattenedSynFname, // source-side flattened parse tree + // file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold + const char* pszOption // other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszInstanceFname); + // fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, + // pszOption); + fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); + } + ~SConstContTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); + pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, + 2.0); + delete pZhangleMaxent;*/ + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, + vector& vecFocused) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - + // pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + // do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + inline void fnGetOutcome(int iL1, int iR1, const SAlignment* pAlign, + string& strOutcome) { + strOutcome = pAlign->fnIsContinuous(iL1, iR1); + } + + inline string fnGetLengthType(int iLen) { + if (iLen == 1) return string("1"); + if (iLen == 2) return string("2"); + if (iLen == 3) return string("3"); + if (iLen < 6) return string("4"); + if (iLen < 11) return string("6"); + return string("11"); + } + + /* + * Source side (11 features): + * f1: the syntactic category + * f2: the syntactic category of its parent + * f3: the head word's pos + * f4: =1 if it's the head of its parent node + * or + * the head of its parent node + * f5: length type + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pCon1, + const SAlignment* pAlign, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + + fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); + + // generate features + // f1 + ostr << "f1=" << pCon1->m_pszTerm; + // f2 + ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; + // f3 + ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord] + ->m_ptParent->m_pszTerm; + // f4 + if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { + ostr << " f4=1"; + } else { + ostr << " f4=" + << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild] + ->m_pszTerm; + } + // f5 + ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); + } + + void fnGenerateInstanceFile( + const char* pszFlattenedSynFname, // source-side flattened parse tree + // file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszFlattenedSynFname, true); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE* fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTree != NULL); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; + i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem* pParent = vecFocused[i]; + + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, + vecSTerms, vecTTerms, strOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + + delete pAlign; + delete pTree; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } }; -inline void print_options(std::ostream &out,po::options_description const& opts) { - typedef std::vector< boost::shared_ptr > Ds; - Ds const& ds=opts.options(); +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); out << '"'; - for (unsigned i=0;ilong_name(); + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); } out << '\n'; } -inline string str(char const* name,po::variables_map const& conf) { +inline string str(char const* name, po::variables_map const& conf) { return conf[name].as(); } -//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file /scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file /scratch0/mt_exp/gq-ctb/data/train.cn --target_file /scratch0/mt_exp/gq-ctb/data/train.en --instance_file /scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix /scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 --classifier_type 1 +//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file +///scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file +///scratch0/mt_exp/gq-ctb/data/train.cn --target_file +///scratch0/mt_exp/gq-ctb/data/train.en --instance_file +///scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix +///scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 +//--classifier_type 1 int main(int argc, char** argv) { - po::options_description opts("Configuration options"); - opts.add_options() - ("parse_file",po::value(),"parse file path (input)") - ("align_file",po::value(),"Alignment file path (input)") - ("source_file",po::value(),"Source text file path (input)") - ("target_file",po::value(),"Target text file path (input)") - ("instance_file",po::value(),"Instance file path (output)") - ("model_prefix",po::value(),"Model file path prefix (output): three files will be generated") - ("classifier_type",po::value()->default_value(1),"Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for SVMLight") - ("feature_cutoff",po::value()->default_value(100),"Feature cutoff threshold") - ("svm_option",po::value(),"Parameters for SVMLight classifier") - ("help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("parse_file") - || !vm.count("align_file") - || !vm.count("source_file") - || !vm.count("target_file") - || !vm.count("instance_file") - || !vm.count("model_prefix") - ) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - const char *pOption; - if (vm.count("svm_option")) - pOption = str("svm_option", vm).c_str(); - else - pOption = NULL; - - SConstReorderTrainer *pTrainer = new SConstReorderTrainer(str("parse_file", vm).c_str(), - str("align_file", vm).c_str(), - str("source_file", vm).c_str(), - str("target_file", vm).c_str(), - str("instance_file", vm).c_str(), - str("model_prefix", vm).c_str(), - vm["classifier_type"].as(), - vm["feature_cutoff"].as(), - pOption); - delete pTrainer; - - return 1; + po::options_description opts("Configuration options"); + opts.add_options()("parse_file", po::value(), + "parse file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "classifier_type", po::value()->default_value(1), + "Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for " + "SVMLight")("feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")( + "svm_option", po::value(), "Parameters for SVMLight classifier")( + "help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("parse_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + const char* pOption; + if (vm.count("svm_option")) + pOption = str("svm_option", vm).c_str(); + else + pOption = NULL; + + SConstReorderTrainer* pTrainer = new SConstReorderTrainer( + str("parse_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["classifier_type"].as(), vm["feature_cutoff"].as(), pOption); + delete pTrainer; + return 1; } diff --git a/utils/synutils/maxent-3.0/lbfgs.cpp b/utils/synutils/maxent-3.0/lbfgs.cpp index 9eb04bef..bd26f048 100644 --- a/utils/synutils/maxent-3.0/lbfgs.cpp +++ b/utils/synutils/maxent-3.0/lbfgs.cpp @@ -8,20 +8,17 @@ using namespace std; -const static int M = LBFGS_M; +const static int M = LBFGS_M; const static double LINE_SEARCH_ALPHA = 0.1; -const static double LINE_SEARCH_BETA = 0.5; +const static double LINE_SEARCH_BETA = 0.5; // stopping criteria -int LBFGS_MAX_ITER = 300; +int LBFGS_MAX_ITER = 300; const static double MIN_GRAD_NORM = 0.0001; - -double -ME_Model::backtracking_line_search( - const Vec & x0, const Vec & grad0, const double f0, - const Vec & dx, Vec & x, Vec & grad1) -{ +double ME_Model::backtracking_line_search(const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, + Vec& x, Vec& grad1) { double t = 1.0 / LINE_SEARCH_BETA; double f; @@ -39,20 +36,23 @@ ME_Model::backtracking_line_search( // Jorge Nocedal, "Updating Quasi-Newton Matrices With Limited Storage", // Mathematics of Computation, Vol. 35, No. 151, pp. 773-782, 1980. // -Vec -approximate_Hg(const int iter, const Vec & grad, - const Vec s[], const Vec y[], const double z[]) -{ +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]) { int offset, bound; - if (iter <= M) { offset = 0; bound = iter; } - else { offset = iter - M; bound = M; } + if (iter <= M) { + offset = 0; + bound = iter; + } else { + offset = iter - M; + bound = M; + } Vec q = grad; double alpha[M], beta[M]; for (int i = bound - 1; i >= 0; i--) { const int j = (i + offset) % M; - alpha[i] = z[j] * dot_product(s[j], q); - q += -alpha[i] * y[j]; + alpha[i] = z[j] * dot_product(s[j], q); + q += -alpha[i] * y[j]; } if (iter > 0) { const int j = (iter - 1) % M; @@ -63,16 +63,14 @@ approximate_Hg(const int iter, const Vec & grad, } for (int i = 0; i <= bound - 1; i++) { const int j = (i + offset) % M; - beta[i] = z[j] * dot_product(y[j], q); - q += s[j] * (alpha[i] - beta[i]); + beta[i] = z[j] * dot_product(y[j], q); + q += s[j] * (alpha[i] - beta[i]); } return q; } -vector -ME_Model::perform_LBFGS(const vector & x0) -{ +vector ME_Model::perform_LBFGS(const vector& x0) { const size_t dim = x0.size(); Vec x = x0; @@ -84,10 +82,11 @@ ME_Model::perform_LBFGS(const vector & x0) for (int iter = 0; iter < LBFGS_MAX_ITER; iter++) { - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, -f, _train_error); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); if (_nheldout > 0) { const double heldout_logl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); } fprintf(stderr, "\n"); @@ -107,4 +106,3 @@ ME_Model::perform_LBFGS(const vector & x0) return x.STLVec(); } - diff --git a/utils/synutils/maxent-3.0/lbfgs.h b/utils/synutils/maxent-3.0/lbfgs.h index 113919a3..ed5cd944 100644 --- a/utils/synutils/maxent-3.0/lbfgs.h +++ b/utils/synutils/maxent-3.0/lbfgs.h @@ -1,21 +1,19 @@ #ifndef _LBFGS_H_ #define _LBFGS_H_ -//template -//std::vector -//perform_LBFGS(FuncGrad func_grad, const std::vector & x0); +// template +// std::vector +// perform_LBFGS(FuncGrad func_grad, const std::vector & x0); -std::vector -perform_LBFGS(double (*func_grad)(const std::vector &, std::vector &), - const std::vector & x0); +std::vector perform_LBFGS( + double (*func_grad)(const std::vector &, std::vector &), + const std::vector &x0); +std::vector perform_OWLQN( + double (*func_grad)(const std::vector &, std::vector &), + const std::vector &x0, const double C); -std::vector -perform_OWLQN(double (*func_grad)(const std::vector &, std::vector &), - const std::vector & x0, - const double C); - -//const int LBFGS_M = 7; -const int LBFGS_M = 10; +// const int LBFGS_M = 7; +const int LBFGS_M = 10; #endif diff --git a/utils/synutils/maxent-3.0/mathvec.h b/utils/synutils/maxent-3.0/mathvec.h index 4ec82797..f8c60e5d 100644 --- a/utils/synutils/maxent-3.0/mathvec.h +++ b/utils/synutils/maxent-3.0/mathvec.h @@ -5,41 +5,40 @@ #include #include -class Vec -{ -private: +class Vec { + private: std::vector _v; -public: + + public: Vec(const size_t n = 0, const double val = 0) { _v.resize(n, val); } - Vec(const std::vector & v) : _v(v) {} - const std::vector & STLVec() const { return _v; } - std::vector & STLVec() { return _v; } + Vec(const std::vector& v) : _v(v) {} + const std::vector& STLVec() const { return _v; } + std::vector& STLVec() { return _v; } size_t Size() const { return _v.size(); } - double & operator[](int i) { return _v[i]; } - const double & operator[](int i) const { return _v[i]; } - Vec & operator+=(const Vec & b) { + double& operator[](int i) { return _v[i]; } + const double& operator[](int i) const { return _v[i]; } + Vec& operator+=(const Vec& b) { assert(b.Size() == _v.size()); for (size_t i = 0; i < _v.size(); i++) { _v[i] += b[i]; } return *this; } - Vec & operator*=(const double c) { + Vec& operator*=(const double c) { for (size_t i = 0; i < _v.size(); i++) { _v[i] *= c; } return *this; } - void Project(const Vec & y) { + void Project(const Vec& y) { for (size_t i = 0; i < _v.size(); i++) { // if (sign(_v[i]) != sign(y[i])) _v[i] = 0; - if (_v[i] * y[i] <=0) _v[i] = 0; + if (_v[i] * y[i] <= 0) _v[i] = 0; } } }; -inline double dot_product(const Vec & a, const Vec & b) -{ +inline double dot_product(const Vec& a, const Vec& b) { double sum = 0; for (size_t i = 0; i < a.Size(); i++) { sum += a[i] * b[i]; @@ -47,8 +46,7 @@ inline double dot_product(const Vec & a, const Vec & b) return sum; } -inline std::ostream & operator<<(std::ostream & s, const Vec & a) -{ +inline std::ostream& operator<<(std::ostream& s, const Vec& a) { s << "("; for (size_t i = 0; i < a.Size(); i++) { if (i != 0) s << ", "; @@ -58,8 +56,7 @@ inline std::ostream & operator<<(std::ostream & s, const Vec & a) return s; } -inline const Vec operator+(const Vec & a, const Vec & b) -{ +inline const Vec operator+(const Vec& a, const Vec& b) { Vec v(a.Size()); assert(a.Size() == b.Size()); for (size_t i = 0; i < a.Size(); i++) { @@ -68,8 +65,7 @@ inline const Vec operator+(const Vec & a, const Vec & b) return v; } -inline const Vec operator-(const Vec & a, const Vec & b) -{ +inline const Vec operator-(const Vec& a, const Vec& b) { Vec v(a.Size()); assert(a.Size() == b.Size()); for (size_t i = 0; i < a.Size(); i++) { @@ -78,8 +74,7 @@ inline const Vec operator-(const Vec & a, const Vec & b) return v; } -inline const Vec operator*(const Vec & a, const double c) -{ +inline const Vec operator*(const Vec& a, const double c) { Vec v(a.Size()); for (size_t i = 0; i < a.Size(); i++) { v[i] = a[i] * c; @@ -87,11 +82,6 @@ inline const Vec operator*(const Vec & a, const double c) return v; } -inline const Vec operator*(const double c, const Vec & a) -{ - return a * c; -} - - +inline const Vec operator*(const double c, const Vec& a) { return a * c; } #endif diff --git a/utils/synutils/maxent-3.0/maxent.cpp b/utils/synutils/maxent-3.0/maxent.cpp index feb0efdc..8d00ac1d 100644 --- a/utils/synutils/maxent-3.0/maxent.cpp +++ b/utils/synutils/maxent-3.0/maxent.cpp @@ -9,14 +9,13 @@ using namespace std; -double -ME_Model::FunctionGradient(const vector & x, vector & grad) -{ +double ME_Model::FunctionGradient(const vector& x, + vector& grad) { assert((int)_fb.Size() == x.size()); for (size_t i = 0; i < x.size(); i++) { _vl[i] = x[i]; } - + double score = update_model_expectation(); if (_l2reg == 0) { @@ -33,9 +32,7 @@ ME_Model::FunctionGradient(const vector & x, vector & grad) return -score; } -int -ME_Model::perform_GIS(int C) -{ +int ME_Model::perform_GIS(int C) { cerr << "C = " << C << endl; C = 1; cerr << "performing AGIS" << endl; @@ -43,11 +40,13 @@ ME_Model::perform_GIS(int C) double pre_logl = -999999; for (int iter = 0; iter < 200; iter++) { - double logl = update_model_expectation(); - fprintf(stderr, "iter = %2d C = %d f = %10.7f train_err = %7.5f", iter, C, logl, _train_error); + double logl = update_model_expectation(); + fprintf(stderr, "iter = %2d C = %d f = %10.7f train_err = %7.5f", iter, + C, logl, _train_error); if (_heldout.size() > 0) { double hlogl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", hlogl, _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", hlogl, + _heldout_error); } cerr << endl; @@ -71,13 +70,13 @@ ME_Model::perform_GIS(int C) return 0; } -int -ME_Model::perform_QUASI_NEWTON() -{ +int ME_Model::perform_QUASI_NEWTON() { const int dim = _fb.Size(); vector x0(dim); - for (int i = 0; i < dim; i++) { x0[i] = _vl[i]; } + for (int i = 0; i < dim; i++) { + x0[i] = _vl[i]; + } vector x; if (_l1reg > 0) { @@ -88,34 +87,39 @@ ME_Model::perform_QUASI_NEWTON() x = perform_LBFGS(x0); } - for (int i = 0; i < dim; i++) { _vl[i] = x[i]; } + for (int i = 0; i < dim; i++) { + _vl[i] = x[i]; + } return 0; } -int -ME_Model::conditional_probability(const Sample & s, - std::vector & membp) const -{ - //int num_classes = membp.size(); +int ME_Model::conditional_probability(const Sample& s, + std::vector& membp) const { + // int num_classes = membp.size(); double sum = 0; int max_label = -1; // double maxp = 0; vector powv(_num_classes, 0.0); - for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ - for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { powv[_fb.Feature(*k).label()] += _vl[*k]; } } - for (vector >::const_iterator j = s.rvfeatures.begin(); j != s.rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { + for (vector >::const_iterator j = s.rvfeatures.begin(); + j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { powv[_fb.Feature(*k).label()] += _vl[*k] * j->second; } } - std::vector::const_iterator pmax = max_element(powv.begin(), powv.end()); - double offset = max(0.0, *pmax - 700); // to avoid overflow + std::vector::const_iterator pmax = + max_element(powv.begin(), powv.end()); + double offset = max(0.0, *pmax - 700); // to avoid overflow for (int label = 0; label < _num_classes; label++) { double pow = powv[label] - offset; double prod = exp(pow); @@ -134,33 +138,38 @@ ME_Model::conditional_probability(const Sample & s, return max_label; } -int -ME_Model::make_feature_bag(const int cutoff) -{ +int ME_Model::make_feature_bag(const int cutoff) { int max_num_features = 0; - // count the occurrences of features +// count the occurrences of features #ifdef USE_HASH_MAP typedef __gnu_cxx::hash_map map_type; -#else +#else typedef std::map map_type; #endif map_type count; if (cutoff > 0) { - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++) { - for (std::vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++) { + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++) { + for (std::vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { count[ME_Feature(i->label, *j).body()]++; } - for (std::vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + for (std::vector >::const_iterator j = + i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { count[ME_Feature(i->label, j->first).body()]++; } } } - int n = 0; - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++, n++) { - max_num_features = max(max_num_features, (int)(i->positive_features.size())); - for (std::vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++) { + int n = 0; + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++, n++) { + max_num_features = + max(max_num_features, (int)(i->positive_features.size())); + for (std::vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { const ME_Feature feature(i->label, *j); // if (cutoff > 0 && count[feature.body()] < cutoff) continue; if (cutoff > 0 && count[feature.body()] <= cutoff) continue; @@ -168,7 +177,9 @@ ME_Model::make_feature_bag(const int cutoff) // cout << i->label << "\t" << *j << "\t" << id << endl; // feature2sample[id].push_back(n); } - for (std::vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { + for (std::vector >::const_iterator j = + i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { const ME_Feature feature(i->label, j->first); // if (cutoff > 0 && count[feature.body()] < cutoff) continue; if (cutoff > 0 && count[feature.body()] <= cutoff) continue; @@ -176,71 +187,72 @@ ME_Model::make_feature_bag(const int cutoff) } } count.clear(); - + // cerr << "num_classes = " << _num_classes << endl; // cerr << "max_num_features = " << max_num_features << endl; init_feature2mef(); - + return max_num_features; } -double -ME_Model::heldout_likelihood() -{ +double ME_Model::heldout_likelihood() { double logl = 0; int ncorrect = 0; - for (std::vector::const_iterator i = _heldout.begin(); i != _heldout.end(); i++) { + for (std::vector::const_iterator i = _heldout.begin(); + i != _heldout.end(); i++) { vector membp(_num_classes); int l = classify(*i, membp); logl += log(membp[i->label]); if (l == i->label) ncorrect++; } _heldout_error = 1 - (double)ncorrect / _heldout.size(); - + return logl /= _heldout.size(); } -double -ME_Model::update_model_expectation() -{ +double ME_Model::update_model_expectation() { double logl = 0; int ncorrect = 0; _vme.resize(_fb.Size()); for (int i = 0; i < _fb.Size(); i++) _vme[i] = 0; - + int n = 0; - for (vector::const_iterator i = _vs.begin(); i != _vs.end(); i++, n++) { + for (vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++, n++) { vector membp(_num_classes); int max_label = conditional_probability(*i, membp); - + logl += log(membp[i->label]); // cout << membp[*i] << " " << logl << " "; if (max_label == i->label) ncorrect++; // model_expectation - for (vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++){ - for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { - _vme[*k] += membp[_fb.Feature(*k).label()]; + for (vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()]; } } - for (vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { - _vme[*k] += membp[_fb.Feature(*k).label()] * j->second; + for (vector >::const_iterator j = i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()] * j->second; } } - } for (int i = 0; i < _fb.Size(); i++) { _vme[i] /= _vs.size(); } - + _train_error = 1 - (double)ncorrect / _vs.size(); logl /= _vs.size(); - + if (_l2reg > 0) { const double c = _l2reg; for (int i = 0; i < _fb.Size(); i++) { @@ -248,17 +260,17 @@ ME_Model::update_model_expectation() } } - //logl /= _vs.size(); - - // fprintf(stderr, "iter =%3d logl = %10.7f train_acc = %7.5f\n", iter, logl, (double)ncorrect/train.size()); - // fprintf(stderr, "logl = %10.7f train_acc = %7.5f\n", logl, (double)ncorrect/_train.size()); + // logl /= _vs.size(); + + // fprintf(stderr, "iter =%3d logl = %10.7f train_acc = %7.5f\n", iter, + // logl, (double)ncorrect/train.size()); + // fprintf(stderr, "logl = %10.7f train_acc = %7.5f\n", logl, + // (double)ncorrect/_train.size()); return logl; } -int -ME_Model::train(const vector & vms) -{ +int ME_Model::train(const vector& vms) { _vs.clear(); for (vector::const_iterator i = vms.begin(); i != vms.end(); i++) { add_training_sample(*i); @@ -267,39 +279,41 @@ ME_Model::train(const vector & vms) return train(); } -void -ME_Model::add_training_sample(const ME_Sample & mes) -{ +void ME_Model::add_training_sample(const ME_Sample& mes) { Sample s; s.label = _label_bag.Put(mes.label); if (s.label > ME_Feature::MAX_LABEL_TYPES) { cerr << "error: too many types of labels." << endl; exit(1); } - for (vector::const_iterator j = mes.features.begin(); j != mes.features.end(); j++) { + for (vector::const_iterator j = mes.features.begin(); + j != mes.features.end(); j++) { s.positive_features.push_back(_featurename_bag.Put(*j)); } - for (vector >::const_iterator j = mes.rvfeatures.begin(); j != mes.rvfeatures.end(); j++) { - s.rvfeatures.push_back(pair(_featurename_bag.Put(j->first), j->second)); + for (vector >::const_iterator j = mes.rvfeatures.begin(); + j != mes.rvfeatures.end(); j++) { + s.rvfeatures.push_back( + pair(_featurename_bag.Put(j->first), j->second)); } if (_ref_modelp != NULL) { - ME_Sample tmp = mes;; + ME_Sample tmp = mes; + ; s.ref_pd = _ref_modelp->classify(tmp); } // cout << s.label << "\t"; - // for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ + // for (vector::const_iterator j = s.positive_features.begin(); j != + // s.positive_features.end(); j++){ // cout << *j << " "; // } // cout << endl; - + _vs.push_back(s); } -int -ME_Model::train() -{ +int ME_Model::train() { if (_l1reg > 0 && _l2reg > 0) { - cerr << "error: L1 and L2 regularizers cannot be used simultaneously." << endl; + cerr << "error: L1 and L2 regularizers cannot be used simultaneously." + << endl; return 0; } if (_vs.size() == 0) { @@ -307,20 +321,22 @@ ME_Model::train() return 0; } if (_nheldout >= (int)_vs.size()) { - cerr << "error: too much heldout data. no training data is available." << endl; + cerr << "error: too much heldout data. no training data is available." + << endl; return 0; } // if (_nheldout > 0) random_shuffle(_vs.begin(), _vs.end()); int max_label = 0; - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); i++) { + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++) { max_label = max(max_label, i->label); } _num_classes = max_label + 1; if (_num_classes != _label_bag.Size()) { cerr << "warning: _num_class != _label_bag.Size()" << endl; } - + if (_ref_modelp != NULL) { cerr << "setting reference distribution..."; for (int i = 0; i < _ref_modelp->num_classes(); i++) { @@ -332,7 +348,7 @@ ME_Model::train() } cerr << "done" << endl; } - + for (int i = 0; i < _nheldout; i++) { _heldout.push_back(_vs.back()); _vs.pop_back(); @@ -362,25 +378,28 @@ ME_Model::train() _vee[i] = 0; } for (int n = 0; n < (int)_vs.size(); n++) { - const Sample * i = &_vs[n]; - for (vector::const_iterator j = i->positive_features.begin(); j != i->positive_features.end(); j++){ - for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { - if (_fb.Feature(*k).label() == i->label) _vee[*k] += 1.0; + const Sample* i = &_vs[n]; + for (vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += 1.0; } } - for (vector >::const_iterator j = i->rvfeatures.begin(); j != i->rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { - if (_fb.Feature(*k).label() == i->label) _vee[*k] += j->second; + for (vector >::const_iterator j = i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += j->second; } } - } for (int i = 0; i < _fb.Size(); i++) { _vee[i] /= _vs.size(); } cerr << "done" << endl; - + _vl.resize(_fb.Size()); for (int i = 0; i < _fb.Size(); i++) _vl[i] = 0.0; @@ -399,13 +418,12 @@ ME_Model::train() return 0; } -void -ME_Model::get_features(list< pair< pair, double> > & fl) -{ +void ME_Model::get_features(list, double> >& fl) { fl.clear(); // for (int i = 0; i < _fb.Size(); i++) { // ME_Feature f = _fb.Feature(i); - // fl.push_back( make_pair(make_pair(_label_bag.Str(f.label()), _featurename_bag.Str(f.feature())), _vl[i])); + // fl.push_back( make_pair(make_pair(_label_bag.Str(f.label()), + // _featurename_bag.Str(f.feature())), _vl[i])); // } for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); i != _featurename_bag.end(); i++) { @@ -414,14 +432,12 @@ ME_Model::get_features(list< pair< pair, double> > & fl) string history = i->first; int id = _fb.Id(ME_Feature(j, i->second)); if (id < 0) continue; - fl.push_back( make_pair(make_pair(label, history), _vl[id]) ); + fl.push_back(make_pair(make_pair(label, history), _vl[id])); } } } -void -ME_Model::clear() -{ +void ME_Model::clear() { _vl.clear(); _label_bag.Clear(); _featurename_bag.Clear(); @@ -433,10 +449,8 @@ ME_Model::clear() _heldout.clear(); } -bool -ME_Model::load_from_file(const string & filename) -{ - FILE * fp = fopen(filename.c_str(), "r"); +bool ME_Model::load_from_file(const string& filename) { + FILE* fp = fopen(filename.c_str(), "r"); if (!fp) { cerr << "error: cannot open " << filename << "!" << endl; return false; @@ -447,22 +461,22 @@ ME_Model::load_from_file(const string & filename) _featurename_bag.Clear(); _fb.Clear(); char buf[1024]; - while(fgets(buf, 1024, fp)) { + while (fgets(buf, 1024, fp)) { string line(buf); string::size_type t1 = line.find_first_of('\t'); string::size_type t2 = line.find_last_of('\t'); string classname = line.substr(0, t1); - string featurename = line.substr(t1 + 1, t2 - (t1 + 1) ); + string featurename = line.substr(t1 + 1, t2 - (t1 + 1)); float lambda; - string w = line.substr(t2+1); + string w = line.substr(t2 + 1); sscanf(w.c_str(), "%f", &lambda); - + int label = _label_bag.Put(classname); int feature = _featurename_bag.Put(featurename); _fb.Put(ME_Feature(label, feature)); _vl.push_back(lambda); } - + _num_classes = _label_bag.Size(); init_feature2mef(); @@ -472,9 +486,7 @@ ME_Model::load_from_file(const string & filename) return true; } -void -ME_Model::init_feature2mef() -{ +void ME_Model::init_feature2mef() { _feature2mef.clear(); for (int i = 0; i < _featurename_bag.Size(); i++) { vector vi; @@ -486,9 +498,7 @@ ME_Model::init_feature2mef() } } -bool -ME_Model::load_from_array(const ME_Model_Data data[]) -{ +bool ME_Model::load_from_array(const ME_Model_Data data[]) { _vl.clear(); for (int i = 0;; i++) { if (string(data[i].label) == "///") break; @@ -500,14 +510,12 @@ ME_Model::load_from_array(const ME_Model_Data data[]) _num_classes = _label_bag.Size(); init_feature2mef(); - + return true; } -bool -ME_Model::save_to_file(const string & filename, const double th) const -{ - FILE * fp = fopen(filename.c_str(), "w"); +bool ME_Model::save_to_file(const string& filename, const double th) const { + FILE* fp = fopen(filename.c_str(), "w"); if (!fp) { cerr << "error: cannot open " << filename << "!" << endl; return false; @@ -516,7 +524,8 @@ ME_Model::save_to_file(const string & filename, const double th) const // for (int i = 0; i < _fb.Size(); i++) { // if (_vl[i] == 0) continue; // ignore zero-weight features // ME_Feature f = _fb.Feature(i); - // fprintf(fp, "%s\t%s\t%f\n", _label_bag.Str(f.label()).c_str(), _featurename_bag.Str(f.feature()).c_str(), _vl[i]); + // fprintf(fp, "%s\t%s\t%f\n", _label_bag.Str(f.label()).c_str(), + // _featurename_bag.Str(f.feature()).c_str(), _vl[i]); // } for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); i != _featurename_bag.end(); i++) { @@ -525,8 +534,8 @@ ME_Model::save_to_file(const string & filename, const double th) const string history = i->first; int id = _fb.Id(ME_Feature(j, i->second)); if (id < 0) continue; - if (_vl[id] == 0) continue; // ignore zero-weight features - if (fabs(_vl[id]) < th) continue; // cut off low-weight features + if (_vl[id] == 0) continue; // ignore zero-weight features + if (fabs(_vl[id]) < th) continue; // cut off low-weight features fprintf(fp, "%s\t%s\t%f\n", label.c_str(), history.c_str(), _vl[id]); } } @@ -536,9 +545,7 @@ ME_Model::save_to_file(const string & filename, const double th) const return true; } -void -ME_Model::set_ref_dist(Sample & s) const -{ +void ME_Model::set_ref_dist(Sample& s) const { vector v0 = s.ref_pd; vector v(_num_classes); for (unsigned int i = 0; i < v.size(); i++) { @@ -548,14 +555,12 @@ ME_Model::set_ref_dist(Sample & s) const if (id_ref != -1) { v[i] = v0[id_ref]; } - if (v[i] == 0) v[i] = 0.001; // to avoid -inf logl + if (v[i] == 0) v[i] = 0.001; // to avoid -inf logl } s.ref_pd = v; } - -int -ME_Model::classify(const Sample & nbs, vector & membp) const -{ + +int ME_Model::classify(const Sample& nbs, vector& membp) const { // vector membp(_num_classes); assert(_num_classes == (int)membp.size()); conditional_probability(nbs, membp); @@ -563,22 +568,24 @@ ME_Model::classify(const Sample & nbs, vector & membp) const double max = 0.0; for (int i = 0; i < (int)membp.size(); i++) { // cout << membp[i] << " "; - if (membp[i] > max) { max_label = i; max = membp[i]; } + if (membp[i] > max) { + max_label = i; + max = membp[i]; + } } // cout << endl; return max_label; } -vector -ME_Model::classify(ME_Sample & mes) const -{ +vector ME_Model::classify(ME_Sample& mes) const { Sample s; - for (vector::const_iterator j = mes.features.begin(); j != mes.features.end(); j++) { + for (vector::const_iterator j = mes.features.begin(); + j != mes.features.end(); j++) { int id = _featurename_bag.Id(*j); - if (id >= 0) - s.positive_features.push_back(id); + if (id >= 0) s.positive_features.push_back(id); } - for (vector >::const_iterator j = mes.rvfeatures.begin(); j != mes.rvfeatures.end(); j++) { + for (vector >::const_iterator j = mes.rvfeatures.begin(); + j != mes.rvfeatures.end(); j++) { int id = _featurename_bag.Id(j->first); if (id >= 0) { s.rvfeatures.push_back(pair(id, j->second)); @@ -595,7 +602,6 @@ ME_Model::classify(ME_Sample & mes) const return vp; } - /* * $Log: maxent.cpp,v $ * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida @@ -695,4 +701,3 @@ ME_Model::classify(ME_Sample & mes) const * remove some comments * */ - diff --git a/utils/synutils/maxent-3.0/maxent.h b/utils/synutils/maxent-3.0/maxent.h index a4391ead..88a04e25 100644 --- a/utils/synutils/maxent-3.0/maxent.h +++ b/utils/synutils/maxent-3.0/maxent.h @@ -15,7 +15,8 @@ #include #include "mathvec.h" -#define USE_HASH_MAP // if you encounter errors with hash, try commenting out this line. (the program will be a bit slower, though) +#define USE_HASH_MAP // if you encounter errors with hash, try commenting out + // this line. (the program will be a bit slower, though) #ifdef USE_HASH_MAP #include #endif @@ -23,68 +24,68 @@ // // data format for each sample for training/testing // -struct ME_Sample -{ -public: +struct ME_Sample { + public: ME_Sample() : label("") {}; - ME_Sample(const std::string & l) : label(l) {}; - void set_label(const std::string & l) { label = l; } + ME_Sample(const std::string& l) : label(l) {}; + void set_label(const std::string& l) { label = l; } // to add a binary feature - void add_feature(const std::string & f) { - features.push_back(f); - } + void add_feature(const std::string& f) { features.push_back(f); } // to add a real-valued feature - void add_feature(const std::string & s, const double d) { - rvfeatures.push_back(std::pair(s, d)); + void add_feature(const std::string& s, const double d) { + rvfeatures.push_back(std::pair(s, d)); } -public: + public: std::string label; std::vector features; std::vector > rvfeatures; // obsolete - void add_feature(const std::pair & f) { - rvfeatures.push_back(f); // real-valued features + void add_feature(const std::pair& f) { + rvfeatures.push_back(f); // real-valued features } }; - // // for those who want to use load_from_array() // -typedef struct ME_Model_Data -{ - char * label; - char * feature; +typedef struct ME_Model_Data { + char* label; + char* feature; double weight; } ME_Model_Data; - -class ME_Model -{ -public: - - void add_training_sample(const ME_Sample & s); +class ME_Model { + public: + void add_training_sample(const ME_Sample& s); int train(); - std::vector classify(ME_Sample & s) const; - bool load_from_file(const std::string & filename); - bool save_to_file(const std::string & filename, const double th = 0) const; + std::vector classify(ME_Sample& s) const; + bool load_from_file(const std::string& filename); + bool save_to_file(const std::string& filename, const double th = 0) const; int num_classes() const { return _num_classes; } std::string get_class_label(int i) const { return _label_bag.Str(i); } - int get_class_id(const std::string & s) const { return _label_bag.Id(s); } - void get_features(std::list< std::pair< std::pair, double> > & fl); - void set_heldout(const int h, const int n = 0) { _nheldout = h; _early_stopping_n = n; }; + int get_class_id(const std::string& s) const { return _label_bag.Id(s); } + void get_features( + std::list, double> >& fl); + void set_heldout(const int h, const int n = 0) { + _nheldout = h; + _early_stopping_n = n; + }; void use_l1_regularizer(const double v) { _l1reg = v; } void use_l2_regularizer(const double v) { _l2reg = v; } void use_SGD(int iter = 30, double eta0 = 1, double alpha = 0.85) { _optimization_method = SGD; - SGD_ITER = iter; SGD_ETA0 = eta0; SGD_ALPHA = alpha; + SGD_ITER = iter; + SGD_ETA0 = eta0; + SGD_ALPHA = alpha; } bool load_from_array(const ME_Model_Data data[]); - void set_reference_model(const ME_Model & ref_model) { _ref_modelp = &ref_model; }; + void set_reference_model(const ME_Model& ref_model) { + _ref_modelp = &ref_model; + }; void clear(); ME_Model() { @@ -95,13 +96,16 @@ public: _optimization_method = LBFGS; } -public: + public: // obsolete. just for downward compatibility - int train(const std::vector & train); - -private: - - enum OPTIMIZATION_METHOD { LBFGS, OWLQN, SGD } _optimization_method; + int train(const std::vector& train); + + private: + enum OPTIMIZATION_METHOD { + LBFGS, + OWLQN, + SGD + } _optimization_method; // OWLQN and SGD are available only for L1-regularization int SGD_ITER; @@ -109,13 +113,13 @@ private: double SGD_ALPHA; double _l1reg, _l2reg; - + struct Sample { int label; std::vector positive_features; std::vector > rvfeatures; - std::vector ref_pd; // reference probability distribution - bool operator<(const Sample & x) const { + std::vector ref_pd; // reference probability distribution + bool operator<(const Sample& x) const { for (unsigned int i = 0; i < positive_features.size(); i++) { if (i >= x.positive_features.size()) return false; int v0 = positive_features[i]; @@ -127,10 +131,11 @@ private: } }; - struct ME_Feature - { - enum { MAX_LABEL_TYPES = 255 }; - + struct ME_Feature { + enum { + MAX_LABEL_TYPES = 255 + }; + // ME_Feature(const int l, const int f) : _body((l << 24) + f) { // assert(l >= 0 && l < 256); // assert(f >= 0 && f <= 0xffffff); @@ -144,20 +149,20 @@ private: int label() const { return _body & 0xff; } int feature() const { return _body >> 8; } unsigned int body() const { return _body; } - private: + + private: unsigned int _body; }; - struct ME_FeatureBag - { + struct ME_FeatureBag { #ifdef USE_HASH_MAP typedef __gnu_cxx::hash_map map_type; -#else +#else typedef std::map map_type; #endif map_type mef2id; std::vector id2mef; - int Put(const ME_Feature & i) { + int Put(const ME_Feature& i) { map_type::const_iterator j = mef2id.find(i.body()); if (j == mef2id.end()) { int id = id2mef.size(); @@ -167,7 +172,7 @@ private: } return j->second; } - int Id(const ME_Feature & i) const { + int Id(const ME_Feature& i) const { map_type::const_iterator j = mef2id.find(i.body()); if (j == mef2id.end()) { return -1; @@ -178,17 +183,14 @@ private: assert(id >= 0 && id < (int)id2mef.size()); return id2mef[id]; } - int Size() const { - return id2mef.size(); - } + int Size() const { return id2mef.size(); } void Clear() { mef2id.clear(); id2mef.clear(); } }; - struct hashfun_str - { + struct hashfun_str { size_t operator()(const std::string& s) const { assert(sizeof(int) == 4 && sizeof(char) == 1); const int* p = reinterpret_cast(s.c_str()); @@ -196,7 +198,7 @@ private: int n = s.size() / 4; for (int i = 0; i < n; i++, p++) { // v ^= *p; - v ^= *p << (4 * (i % 2)); // note) 0 <= char < 128 + v ^= *p << (4 * (i % 2)); // note) 0 <= char < 128 } int m = s.size() % 4; for (int i = 0; i < m; i++) { @@ -206,17 +208,16 @@ private: } }; - struct MiniStringBag - { + struct MiniStringBag { #ifdef USE_HASH_MAP typedef __gnu_cxx::hash_map map_type; -#else +#else typedef std::map map_type; #endif int _size; map_type str2id; MiniStringBag() : _size(0) {} - int Put(const std::string & i) { + int Put(const std::string& i) { map_type::const_iterator j = str2id.find(i); if (j == str2id.end()) { int id = _size; @@ -226,21 +227,23 @@ private: } return j->second; } - int Id(const std::string & i) const { + int Id(const std::string& i) const { map_type::const_iterator j = str2id.find(i); - if (j == str2id.end()) return -1; + if (j == str2id.end()) return -1; return j->second; } int Size() const { return _size; } - void Clear() { str2id.clear(); _size = 0; } + void Clear() { + str2id.clear(); + _size = 0; + } map_type::const_iterator begin() const { return str2id.begin(); } - map_type::const_iterator end() const { return str2id.end(); } + map_type::const_iterator end() const { return str2id.end(); } }; - struct StringBag : public MiniStringBag - { + struct StringBag : public MiniStringBag { std::vector id2str; - int Put(const std::string & i) { + int Put(const std::string& i) { map_type::const_iterator j = str2id.find(i); if (j == str2id.end()) { int id = id2str.size(); @@ -261,7 +264,7 @@ private: } }; - std::vector _vs; // vector of training_samples + std::vector _vs; // vector of training_samples StringBag _label_bag; MiniStringBag _featurename_bag; std::vector _vl; // vector of lambda @@ -269,42 +272,46 @@ private: int _num_classes; std::vector _vee; // empirical expectation std::vector _vme; // empirical expectation - std::vector< std::vector< int > > _feature2mef; - std::vector< Sample > _heldout; - double _train_error; // current error rate on the training data - double _heldout_error; // current error rate on the heldout data + std::vector > _feature2mef; + std::vector _heldout; + double _train_error; // current error rate on the training data + double _heldout_error; // current error rate on the heldout data int _nheldout; int _early_stopping_n; std::vector _vhlogl; - const ME_Model * _ref_modelp; + const ME_Model* _ref_modelp; double heldout_likelihood(); - int conditional_probability(const Sample & nbs, std::vector & membp) const; + int conditional_probability(const Sample& nbs, + std::vector& membp) const; int make_feature_bag(const int cutoff); - int classify(const Sample & nbs, std::vector & membp) const; + int classify(const Sample& nbs, std::vector& membp) const; double update_model_expectation(); int perform_QUASI_NEWTON(); int perform_SGD(); int perform_GIS(int C); - std::vector perform_LBFGS(const std::vector & x0); - std::vector perform_OWLQN(const std::vector & x0, const double C); - double backtracking_line_search(const Vec & x0, const Vec & grad0, const double f0, const Vec & dx, Vec & x, Vec & grad1); - double regularized_func_grad(const double C, const Vec & x, Vec & grad); - double constrained_line_search(double C, const Vec & x0, const Vec & grad0, const double f0, const Vec & dx, Vec & x, Vec & grad1); - - - void set_ref_dist(Sample & s) const; + std::vector perform_LBFGS(const std::vector& x0); + std::vector perform_OWLQN(const std::vector& x0, + const double C); + double backtracking_line_search(const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, Vec& x, + Vec& grad1); + double regularized_func_grad(const double C, const Vec& x, Vec& grad); + double constrained_line_search(double C, const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, Vec& x, + Vec& grad1); + + void set_ref_dist(Sample& s) const; void init_feature2mef(); - double FunctionGradient(const std::vector & x, std::vector & grad); - static double FunctionGradientWrapper(const std::vector & x, std::vector & grad); - + double FunctionGradient(const std::vector& x, + std::vector& grad); + static double FunctionGradientWrapper(const std::vector& x, + std::vector& grad); }; - #endif - /* * $Log: maxent.h,v $ * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida diff --git a/utils/synutils/maxent-3.0/owlqn.cpp b/utils/synutils/maxent-3.0/owlqn.cpp index 7b2cea7d..c3a0f0da 100644 --- a/utils/synutils/maxent-3.0/owlqn.cpp +++ b/utils/synutils/maxent-3.0/owlqn.cpp @@ -8,29 +8,24 @@ using namespace std; -const static int M = LBFGS_M; +const static int M = LBFGS_M; const static double LINE_SEARCH_ALPHA = 0.1; -const static double LINE_SEARCH_BETA = 0.5; +const static double LINE_SEARCH_BETA = 0.5; // stopping criteria -int OWLQN_MAX_ITER = 300; +int OWLQN_MAX_ITER = 300; const static double MIN_GRAD_NORM = 0.0001; +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]); -Vec approximate_Hg(const int iter, const Vec & grad, - const Vec s[], const Vec y[], const double z[]); - - -inline int sign(double x) -{ +inline int sign(double x) { if (x > 0) return 1; if (x < 0) return -1; return 0; }; -static Vec -pseudo_gradient(const Vec & x, const Vec & grad0, const double C) -{ +static Vec pseudo_gradient(const Vec& x, const Vec& grad0, const double C) { Vec grad = grad0; for (size_t i = 0; i < x.Size(); i++) { if (x[i] != 0) { @@ -53,9 +48,8 @@ pseudo_gradient(const Vec & x, const Vec & grad0, const double C) return grad; } -double -ME_Model::regularized_func_grad(const double C, const Vec & x, Vec & grad) -{ +double ME_Model::regularized_func_grad(const double C, const Vec& x, + Vec& grad) { double f = FunctionGradient(x.STLVec(), grad.STLVec()); for (size_t i = 0; i < x.Size(); i++) { f += C * fabs(x[i]); @@ -64,11 +58,9 @@ ME_Model::regularized_func_grad(const double C, const Vec & x, Vec & grad) return f; } -double -ME_Model::constrained_line_search(double C, - const Vec & x0, const Vec & grad0, const double f0, - const Vec & dx, Vec & x, Vec & grad1) -{ +double ME_Model::constrained_line_search(double C, const Vec& x0, + const Vec& grad0, const double f0, + const Vec& dx, Vec& x, Vec& grad1) { // compute the orthant to explore Vec orthant = x0; for (size_t i = 0; i < orthant.Size(); i++) { @@ -93,9 +85,8 @@ ME_Model::constrained_line_search(double C, return f; } -vector -ME_Model::perform_OWLQN(const vector & x0, const double C) -{ +vector ME_Model::perform_OWLQN(const vector& x0, + const double C) { const size_t dim = x0.size(); Vec x = x0; @@ -108,18 +99,18 @@ ME_Model::perform_OWLQN(const vector & x0, const double C) for (int iter = 0; iter < OWLQN_MAX_ITER; iter++) { Vec pg = pseudo_gradient(x, grad, C); - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, -f, _train_error); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); if (_nheldout > 0) { const double heldout_logl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); } fprintf(stderr, "\n"); if (sqrt(dot_product(pg, pg)) < MIN_GRAD_NORM) break; dx = -1 * approximate_Hg(iter, pg, s, y, z); - if (dot_product(dx, pg) >= 0) - dx.Project(-1 * pg); + if (dot_product(dx, pg) >= 0) dx.Project(-1 * pg); Vec x1(dim), grad1(dim); f = constrained_line_search(C, x, pg, f, dx, x1, grad1); @@ -134,4 +125,3 @@ ME_Model::perform_OWLQN(const vector & x0, const double C) return x.STLVec(); } - diff --git a/utils/synutils/maxent-3.0/sgd.cpp b/utils/synutils/maxent-3.0/sgd.cpp index 6d28c23f..8613edca 100644 --- a/utils/synutils/maxent-3.0/sgd.cpp +++ b/utils/synutils/maxent-3.0/sgd.cpp @@ -4,23 +4,20 @@ using namespace std; -//const double SGD_ETA0 = 1; -//const double SGD_ITER = 30; -//const double SGD_ALPHA = 0.85; - +// const double SGD_ETA0 = 1; +// const double SGD_ITER = 30; +// const double SGD_ALPHA = 0.85; //#define FOLOS_NAIVE //#define FOLOS_LAZY #define SGD_CP -inline void -apply_l1_penalty(const int i, const double u, - vector & _vl, vector & q) -{ - double & w = _vl[i]; +inline void apply_l1_penalty(const int i, const double u, vector& _vl, + vector& q) { + double& w = _vl[i]; const double z = w; - double & qi = q[i]; - if (w > 0) { + double& qi = q[i]; + if (w > 0) { w = max(0.0, w - (u + qi)); } else if (w < 0) { w = min(0.0, w + (u - qi)); @@ -28,31 +25,29 @@ apply_l1_penalty(const int i, const double u, qi += w - z; } -static double -l1norm(const vector& v) -{ +static double l1norm(const vector& v) { double sum = 0; for (size_t i = 0; i < v.size(); i++) sum += abs(v[i]); return sum; } -inline void -update_folos_lazy(const int iter_sample, - const int k, vector & _vl, const vector & sum_eta, - vector & last_updated) -{ +inline void update_folos_lazy(const int iter_sample, const int k, + vector& _vl, + const vector& sum_eta, + vector& last_updated) { const double penalty = sum_eta[iter_sample] - sum_eta[last_updated[k]]; - double & x = _vl[k]; - if (x > 0) x = max(0.0, x - penalty); - else x = min(0.0, x + penalty); + double& x = _vl[k]; + if (x > 0) + x = max(0.0, x - penalty); + else + x = min(0.0, x + penalty); last_updated[k] = iter_sample; } -int -ME_Model::perform_SGD() -{ +int ME_Model::perform_SGD() { if (_l2reg > 0) { - cerr << "error: L2 regularization is currently not supported in SGD mode." << endl; + cerr << "error: L2 regularization is currently not supported in SGD mode." + << endl; exit(1); } @@ -85,95 +80,113 @@ ME_Model::perform_SGD() double logl = 0; int ncorrect = 0, ntotal = 0; for (size_t i = 0; i < _vs.size(); i++, ntotal++, iter_sample++) { - const Sample & s = _vs[ri[i]]; + const Sample& s = _vs[ri[i]]; #ifdef FOLOS_LAZY - for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ - for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { - update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); - } + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); + } } #endif vector membp(_num_classes); const int max_label = conditional_probability(s, membp); - const double eta = eta0 * pow(SGD_ALPHA, (double)iter_sample / _vs.size()); // exponential decay - // const double eta = eta0 / (1.0 + (double)iter_sample / _vs.size()); + const double eta = + eta0 * pow(SGD_ALPHA, + (double)iter_sample / _vs.size()); // exponential decay + // const double eta = eta0 / (1.0 + (double)iter_sample / + // _vs.size()); - // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << endl; + // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << + // endl; u += eta * l1param; sum_eta.push_back(sum_eta.back() + eta * l1param); - + logl += log(membp[s.label]); if (max_label == s.label) ncorrect++; // binary features - for (vector::const_iterator j = s.positive_features.begin(); j != s.positive_features.end(); j++){ - for (vector::const_iterator k = _feature2mef[*j].begin(); k != _feature2mef[*j].end(); k++) { - const double me = membp[_fb.Feature(*k).label()]; - const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); - const double grad = (me - ee); - _vl[*k] -= eta * grad; + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee); + _vl[*k] -= eta * grad; #ifdef SGD_CP - apply_l1_penalty(*k, u, _vl, q); + apply_l1_penalty(*k, u, _vl, q); #endif - } + } } // real-valued features - for (vector >::const_iterator j = s.rvfeatures.begin(); j != s.rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); k != _feature2mef[j->first].end(); k++) { - const double me = membp[_fb.Feature(*k).label()]; - const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); - const double grad = (me - ee) * j->second; - _vl[*k] -= eta * grad; + for (vector >::const_iterator j = s.rvfeatures.begin(); + j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee) * j->second; + _vl[*k] -= eta * grad; #ifdef SGD_CP - apply_l1_penalty(*k, u, _vl, q); + apply_l1_penalty(*k, u, _vl, q); #endif - } + } } #ifdef FOLOS_NAIVE for (size_t j = 0; j < d; j++) { - double & x = _vl[j]; - if (x > 0) x = max(0.0, x - eta * l1param); - else x = min(0.0, x + eta * l1param); + double& x = _vl[j]; + if (x > 0) + x = max(0.0, x - eta * l1param); + else + x = min(0.0, x + eta * l1param); } #endif - } logl /= _vs.size(); - // fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, (double)ncorrect / ntotal); +// fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, +// (double)ncorrect / ntotal); #ifdef FOLOS_LAZY if (l1param > 0) { for (size_t j = 0; j < d; j++) - update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); + update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); } #endif double f = logl; if (l1param > 0) { - const double l1 = l1norm(_vl); // this is not accurate when lazy update is used - // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " "; + const double l1 = + l1norm(_vl); // this is not accurate when lazy update is used + // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " + // "; f -= l1param * l1; int nonzero = 0; - for (int j = 0; j < d; j++) if (_vl[j] != 0) nonzero++; - // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " << nonzero << endl; + for (int j = 0; j < d; j++) + if (_vl[j] != 0) nonzero++; + // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " + // << nonzero << endl; } - // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, (double)ncorrect/ntotal); + // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, + // (double)ncorrect/ntotal); // fprintf(stderr, "%4d obj = %f", iter+1, f); - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter+1, f, 1 - (double)ncorrect/ntotal); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, f, + 1 - (double)ncorrect / ntotal); if (_nheldout > 0) { double heldout_logl = heldout_likelihood(); - // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", heldout_logl, 1 - _heldout_error); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, _heldout_error); + // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", + // heldout_logl, 1 - _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); } fprintf(stderr, "\n"); - - } return 0; diff --git a/utils/synutils/srl_sentence.h b/utils/synutils/srl_sentence.h index 6f00267a..c08cc065 100644 --- a/utils/synutils/srl_sentence.h +++ b/utils/synutils/srl_sentence.h @@ -14,202 +14,198 @@ using namespace std; -struct SArgument{ - SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { - m_pszRole = new char[strlen(pszRole) + 1]; - strcpy(m_pszRole, pszRole); - m_iBegin = iBegin; - m_iEnd = iEnd; - m_fProb = fProb; - m_pTreeItem = NULL; - } - ~SArgument() { - delete [] m_pszRole; - } - - void fnSetTreeItem(STreeItem *pTreeItem) { - m_pTreeItem = pTreeItem; - if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { - assert(m_pTreeItem->m_iBegin == m_iBegin); - assert(m_pTreeItem->m_iEnd == m_iEnd); - } - } - - char *m_pszRole; //argument rule, e.g., ARG0, ARGM-TMP - int m_iBegin; - int m_iEnd; //the span of the argument, [m_iBegin, m_iEnd] - float m_fProb; //the probability of this role, - STreeItem *m_pTreeItem; +struct SArgument { + SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { + m_pszRole = new char[strlen(pszRole) + 1]; + strcpy(m_pszRole, pszRole); + m_iBegin = iBegin; + m_iEnd = iEnd; + m_fProb = fProb; + m_pTreeItem = NULL; + } + ~SArgument() { delete[] m_pszRole; } + + void fnSetTreeItem(STreeItem* pTreeItem) { + m_pTreeItem = pTreeItem; + if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { + assert(m_pTreeItem->m_iBegin == m_iBegin); + assert(m_pTreeItem->m_iEnd == m_iEnd); + } + } + + char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP + int m_iBegin; + int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] + float m_fProb; // the probability of this role, + STreeItem* m_pTreeItem; }; -struct SPredicate{ - SPredicate(const char* pszLemma, int iPosition) { - if (pszLemma != NULL) { - m_pszLemma = new char[strlen(pszLemma) + 1]; - strcpy(m_pszLemma, pszLemma); - } else - m_pszLemma = NULL; - m_iPosition = iPosition; - } - ~SPredicate() { - if (m_pszLemma != NULL) - delete [] m_pszLemma; - for (size_t i = 0; i < m_vecArgt.size(); i++) - delete m_vecArgt[i]; - } - int fnAppend(const char* pszRole, int iBegin, int iEnd) { - SArgument *pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); - return fnAppend(pArgt); - } - int fnAppend(SArgument *pArgt) { - m_vecArgt.push_back(pArgt); - int iPosition = m_vecArgt.size() - 1; - return iPosition; - } - - char *m_pszLemma; //lemma of the predicate, for Chinese, it's always as same as the predicate itself - int m_iPosition; //the position in sentence - vector m_vecArgt; //arguments associated to the predicate +struct SPredicate { + SPredicate(const char* pszLemma, int iPosition) { + if (pszLemma != NULL) { + m_pszLemma = new char[strlen(pszLemma) + 1]; + strcpy(m_pszLemma, pszLemma); + } else + m_pszLemma = NULL; + m_iPosition = iPosition; + } + ~SPredicate() { + if (m_pszLemma != NULL) delete[] m_pszLemma; + for (size_t i = 0; i < m_vecArgt.size(); i++) delete m_vecArgt[i]; + } + int fnAppend(const char* pszRole, int iBegin, int iEnd) { + SArgument* pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); + return fnAppend(pArgt); + } + int fnAppend(SArgument* pArgt) { + m_vecArgt.push_back(pArgt); + int iPosition = m_vecArgt.size() - 1; + return iPosition; + } + + char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same + // as the predicate itself + int m_iPosition; // the position in sentence + vector m_vecArgt; // arguments associated to the predicate }; -struct SSrlSentence{ - SSrlSentence() { - m_pTree = NULL; - } - ~SSrlSentence() { - if (m_pTree != NULL) - delete m_pTree; - - for (size_t i = 0; i < m_vecPred.size(); i++) - delete m_vecPred[i]; - } - int fnAppend(const char* pszLemma, int iPosition) { - SPredicate *pPred = new SPredicate(pszLemma, iPosition); - return fnAppend(pPred); - } - int fnAppend(SPredicate* pPred) { - m_vecPred.push_back(pPred); - int iPosition = m_vecPred.size() - 1; - return iPosition; - } - int GetPredicateNum() { - return m_vecPred.size(); - } - - SParsedTree *m_pTree; - vector m_vecPred; +struct SSrlSentence { + SSrlSentence() { m_pTree = NULL; } + ~SSrlSentence() { + if (m_pTree != NULL) delete m_pTree; + + for (size_t i = 0; i < m_vecPred.size(); i++) delete m_vecPred[i]; + } + int fnAppend(const char* pszLemma, int iPosition) { + SPredicate* pPred = new SPredicate(pszLemma, iPosition); + return fnAppend(pPred); + } + int fnAppend(SPredicate* pPred) { + m_vecPred.push_back(pPred); + int iPosition = m_vecPred.size() - 1; + return iPosition; + } + int GetPredicateNum() { return m_vecPred.size(); } + + SParsedTree* m_pTree; + vector m_vecPred; }; struct SSrlSentenceReader { - SSrlSentenceReader(const char* pszSrlFname) { - m_fpIn = fopen(pszSrlFname, "r"); - assert(m_fpIn != NULL); - } - ~SSrlSentenceReader() { - if (m_fpIn != NULL) - fclose(m_fpIn); - } - - inline void fnReplaceAll(std::string& str, const std::string& from, const std::string& to) { - size_t start_pos = 0; - while((start_pos = str.find(from, start_pos)) != std::string::npos) { - str.replace(start_pos, from.length(), to); - start_pos += to.length(); // In case 'to' contains 'from', like replacing 'x' with 'yx' - } - } - - //TODO: here only considers flat predicate-argument structure - // i.e., no overlap among them - SSrlSentence* fnReadNextSrlSentence() { - vector > vecContent; - if (fnReadNextContent(vecContent) == false) - return NULL; - - SSrlSentence *pSrlSentence = new SSrlSentence(); - int iSize = vecContent.size(); - //put together syntactic text - std::ostringstream ostr; - for (int i = 0; i < iSize; i++) { - string strSynSeg = vecContent[i][5]; //the 5th column is the syntactic segment - size_t iPosition = strSynSeg.find_first_of('*'); - assert(iPosition != string::npos); - ostringstream ostrTmp; - ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] << ")"; //the 2th column is POS-tag, and the 0th column is word - strSynSeg.replace(iPosition, 1, ostrTmp.str()); - fnReplaceAll(strSynSeg, "(", " ("); - ostr << strSynSeg; - } - string strSyn = ostr.str(); - pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); - pSrlSentence->m_pTree->fnSetHeadWord(); - pSrlSentence->m_pTree->fnSetSpanInfo(); - - //read predicate-argument structure - int iNumPred = vecContent[0].size() - 8; - for (int i = 0; i < iNumPred; i++) { - vector vecRole; - vector vecBegin; - vector vecEnd; - int iPred = -1; - for (int j = 0; j < iSize; j++) { - const char* p = vecContent[j][i + 8].c_str(); - const char* q; - if (p[0] == '(') { - //starting position of an argument(or predicate) - vecBegin.push_back(j); - q = strchr(p, '*'); - assert(q != NULL); - vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); - if (vecRole.back().compare("V") == 0) { - assert(iPred == -1); - iPred = vecRole.size() - 1; - } - } - if (p[strlen(p) - 1] == ')') { - //end position of an argument(or predicate) - vecEnd.push_back(j); - assert(vecBegin.size() == vecEnd.size()); - } - } - assert(iPred != -1); - SPredicate *pPred = new SPredicate(pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, vecBegin[iPred]); - pSrlSentence->fnAppend(pPred); - for (size_t j = 0; j < vecBegin.size(); j++) { - if (j == iPred) - continue; - pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); - pPred->m_vecArgt.back()->fnSetTreeItem(pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], false)); - } - } - return pSrlSentence; - } -private: - bool fnReadNextContent(vector >& vecContent) { - vecContent.clear(); - if (feof(m_fpIn) == true) - return false; - char *pszLine; - pszLine = new char[100001]; - pszLine[0] = '\0'; - int iLen; - while (!feof(m_fpIn)) { - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - if (iLen == 0) - break; //end of this sentence - - vector terms = SplitOnWhitespace(string(pszLine)); - assert(terms.size() > 7); - vecContent.push_back(terms); - } - delete [] pszLine; - return true; - } -private: - FILE *m_fpIn; + SSrlSentenceReader(const char* pszSrlFname) { + m_fpIn = fopen(pszSrlFname, "r"); + assert(m_fpIn != NULL); + } + ~SSrlSentenceReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + inline void fnReplaceAll(std::string& str, const std::string& from, + const std::string& to) { + size_t start_pos = 0; + while ((start_pos = str.find(from, start_pos)) != std::string::npos) { + str.replace(start_pos, from.length(), to); + start_pos += to.length(); // In case 'to' contains 'from', like replacing + // 'x' with 'yx' + } + } + + // TODO: here only considers flat predicate-argument structure + // i.e., no overlap among them + SSrlSentence* fnReadNextSrlSentence() { + vector > vecContent; + if (fnReadNextContent(vecContent) == false) return NULL; + + SSrlSentence* pSrlSentence = new SSrlSentence(); + int iSize = vecContent.size(); + // put together syntactic text + std::ostringstream ostr; + for (int i = 0; i < iSize; i++) { + string strSynSeg = + vecContent[i][5]; // the 5th column is the syntactic segment + size_t iPosition = strSynSeg.find_first_of('*'); + assert(iPosition != string::npos); + ostringstream ostrTmp; + ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] + << ")"; // the 2th column is POS-tag, and the 0th column is word + strSynSeg.replace(iPosition, 1, ostrTmp.str()); + fnReplaceAll(strSynSeg, "(", " ("); + ostr << strSynSeg; + } + string strSyn = ostr.str(); + pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); + pSrlSentence->m_pTree->fnSetHeadWord(); + pSrlSentence->m_pTree->fnSetSpanInfo(); + + // read predicate-argument structure + int iNumPred = vecContent[0].size() - 8; + for (int i = 0; i < iNumPred; i++) { + vector vecRole; + vector vecBegin; + vector vecEnd; + int iPred = -1; + for (int j = 0; j < iSize; j++) { + const char* p = vecContent[j][i + 8].c_str(); + const char* q; + if (p[0] == '(') { + // starting position of an argument(or predicate) + vecBegin.push_back(j); + q = strchr(p, '*'); + assert(q != NULL); + vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); + if (vecRole.back().compare("V") == 0) { + assert(iPred == -1); + iPred = vecRole.size() - 1; + } + } + if (p[strlen(p) - 1] == ')') { + // end position of an argument(or predicate) + vecEnd.push_back(j); + assert(vecBegin.size() == vecEnd.size()); + } + } + assert(iPred != -1); + SPredicate* pPred = new SPredicate( + pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, + vecBegin[iPred]); + pSrlSentence->fnAppend(pPred); + for (size_t j = 0; j < vecBegin.size(); j++) { + if (j == iPred) continue; + pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); + pPred->m_vecArgt.back()->fnSetTreeItem( + pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], + false)); + } + } + return pSrlSentence; + } + + private: + bool fnReadNextContent(vector >& vecContent) { + vecContent.clear(); + if (feof(m_fpIn) == true) return false; + char* pszLine; + pszLine = new char[100001]; + pszLine[0] = '\0'; + int iLen; + while (!feof(m_fpIn)) { + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + if (iLen == 0) break; // end of this sentence + + vector terms = SplitOnWhitespace(string(pszLine)); + assert(terms.size() > 7); + vecContent.push_back(terms); + } + delete[] pszLine; + return true; + } + + private: + FILE* m_fpIn; }; #endif /* SRL_SENTENCE_H_ */ diff --git a/utils/synutils/tree.h b/utils/synutils/tree.h index 658a470d..32521a3a 100644 --- a/utils/synutils/tree.h +++ b/utils/synutils/tree.h @@ -14,718 +14,685 @@ using namespace std; -struct STreeItem{ - STreeItem(const char* pszTerm) { - m_pszTerm = new char[ strlen( pszTerm ) + 1 ]; - strcpy( m_pszTerm, pszTerm ); - - m_ptParent = NULL; - m_iBegin = -1; - m_iEnd = -1; - m_iHeadChild = -1; - m_iHeadWord = -1; - m_iBrotherIndex = -1; - } - ~STreeItem( ) { - delete [] m_pszTerm; - for (size_t i = 0; i < m_vecChildren.size(); i++) - delete m_vecChildren[i]; - } - int fnAppend(STreeItem *ptChild) { - m_vecChildren.push_back(ptChild); - ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; - ptChild->m_ptParent = this; - return m_vecChildren.size() - 1; - } - int fnGetChildrenNum() { - return m_vecChildren.size(); - } - - bool fnIsPreTerminal( void ) { - int I; - if ( this == NULL || m_vecChildren.size() == 0 ) - return false; - - for ( I = 0; I < m_vecChildren.size(); I++ ) - if (m_vecChildren[I]->m_vecChildren.size() > 0 ) - return false; - - return true; - } -public: - char *m_pszTerm; - - vector m_vecChildren;//children items - STreeItem *m_ptParent;//the parent item - - int m_iBegin; - int m_iEnd; //the node span words[m_iBegin, m_iEnd] - int m_iHeadChild; //the index of its head child - int m_iHeadWord; //the index of its head word - int m_iBrotherIndex;//the index in his brothers +struct STreeItem { + STreeItem(const char *pszTerm) { + m_pszTerm = new char[strlen(pszTerm) + 1]; + strcpy(m_pszTerm, pszTerm); + + m_ptParent = NULL; + m_iBegin = -1; + m_iEnd = -1; + m_iHeadChild = -1; + m_iHeadWord = -1; + m_iBrotherIndex = -1; + } + ~STreeItem() { + delete[] m_pszTerm; + for (size_t i = 0; i < m_vecChildren.size(); i++) delete m_vecChildren[i]; + } + int fnAppend(STreeItem *ptChild) { + m_vecChildren.push_back(ptChild); + ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; + ptChild->m_ptParent = this; + return m_vecChildren.size() - 1; + } + int fnGetChildrenNum() { return m_vecChildren.size(); } + + bool fnIsPreTerminal(void) { + int I; + if (this == NULL || m_vecChildren.size() == 0) return false; + + for (I = 0; I < m_vecChildren.size(); I++) + if (m_vecChildren[I]->m_vecChildren.size() > 0) return false; + + return true; + } + + public: + char *m_pszTerm; + + vector m_vecChildren; // children items + STreeItem *m_ptParent; // the parent item + + int m_iBegin; + int m_iEnd; // the node span words[m_iBegin, m_iEnd] + int m_iHeadChild; // the index of its head child + int m_iHeadWord; // the index of its head word + int m_iBrotherIndex; // the index in his brothers }; - -struct SGetHeadWord{ - typedef vector CVectorStr; - SGetHeadWord() { - - } - ~SGetHeadWord() { - - } - int fnGetHeadWord( char *pszCFGLeft, CVectorStr vectRight ) { - //0 indicating from right to left while 1 indicating from left to right - char szaHeadLists[ 201 ] = "0"; - - /* //head rules for Egnlish - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN 0CD 0JJ JJS RB QP 0" ); - */ - - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0ADJP JJ 0AD NN CS 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "0ADVP AD 0" ); - else if( strcmp( pszCFGLeft, "CLP" ) == 0 ) - strcpy( szaHeadLists, "0CLP M 0" ); - else if( strcmp( pszCFGLeft, "CP" ) == 0 ) - strcpy( szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0" ); - else if( strcmp( pszCFGLeft, "DNP" ) == 0 ) - strcpy( szaHeadLists, "0DNP DEG 0DEC 0" ); - else if( strcmp( pszCFGLeft, "DVP" ) == 0 ) - strcpy( szaHeadLists, "0DVP DEV 0" ); - else if( strcmp( pszCFGLeft, "DP" ) == 0 ) - strcpy( szaHeadLists, "1DP DT 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "0VV NR NN 0" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0INTJ IJ 0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LST CD OD 1" ); - else if( strcmp( pszCFGLeft, "IP" ) == 0 ) - strcpy( szaHeadLists, "0IP VP 0VV 0" ); - //strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); - else if( strcmp( pszCFGLeft, "LCP" ) == 0 ) - strcpy( szaHeadLists, "0LCP LC 0" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NP NN NT NR QP 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1PP P 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "0 NP IP VP NT NR NN 0" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0QP CLP CD OD 0" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1" ); - else if( strcmp( pszCFGLeft, "VCD" ) == 0 ) - strcpy( szaHeadLists, "0VCD VV VA VC VE 0" ); - if( strcmp( pszCFGLeft, "VRD" ) == 0 ) - strcpy( szaHeadLists, "0VRD VV VA VC VE 0" ); - else if( strcmp( pszCFGLeft, "VSB" ) == 0 ) - strcpy( szaHeadLists, "0VSB VV VA VC VE 0" ); - else if( strcmp( pszCFGLeft, "VCP" ) == 0 ) - strcpy( szaHeadLists, "0VCP VV VA VC VE 0" ); - else if( strcmp( pszCFGLeft, "VNV" ) == 0 ) - strcpy( szaHeadLists, "0VNV VV VA VC VE 0" ); - else if( strcmp( pszCFGLeft, "VPT" ) == 0 ) - strcpy( szaHeadLists, "0VNV VV VA VC VE 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WHNP NP NN NT NR QP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1WHPP PP P 1" ); - - /* //head rules for GENIA corpus - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN 0CD 0JJ JJS RB QP 0" ); - */ - - return fnMyOwnHeadWordRule( szaHeadLists, vectRight ); - } - -private: - int fnMyOwnHeadWordRule( char *pszaHeadLists, CVectorStr vectRight ) { - char szHeadList[ 201 ], *p; - char szTerm[ 101 ]; - int J; - - p = pszaHeadLists; - - int iCountRight; - - iCountRight = vectRight.size( ); - - szHeadList[ 0 ] = '\0'; - while( 1 ){ - szTerm[ 0 ] = '\0'; - sscanf( p, "%s", szTerm ); - if( strlen( szHeadList ) == 0 ){ - if( strcmp( szTerm, "0" ) == 0 ){ - return iCountRight - 1; - } - if( strcmp( szTerm, "1" ) == 0 ){ - return 0; - } - - sprintf( szHeadList, "%c %s ", szTerm[ 0 ], szTerm + 1 ); - p = strstr( p, szTerm ); - p += strlen( szTerm ); - } - else{ - if( ( szTerm[ 0 ] == '0' ) - ||( szTerm[ 0 ] == '1' ) ){ - if( szHeadList[ 0 ] == '0' ){ - for( J = iCountRight - 1; J >= 0; J -- ){ - sprintf( szTerm, " %s ", vectRight.at( J ).c_str( ) ); - if( strstr( szHeadList, szTerm ) != NULL ) - return J; - } - } - else{ - for( J = 0; J < iCountRight; J ++ ){ - sprintf( szTerm, " %s ", vectRight.at( J ).c_str( ) ); - if( strstr( szHeadList, szTerm ) != NULL ) - return J; - } - } - - szHeadList[ 0 ] = '\0'; - } - else{ - strcat( szHeadList, szTerm ); - strcat( szHeadList, " " ); - - p = strstr( p, szTerm ); - p += strlen( szTerm ); - } - } - } - - return 0; - } - +struct SGetHeadWord { + typedef vector CVectorStr; + SGetHeadWord() {} + ~SGetHeadWord() {} + int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { + // 0 indicating from right to left while 1 indicating from left to right + char szaHeadLists[201] = "0"; + + /* //head rules for Egnlish + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + if (strcmp(pszCFGLeft, "ADJP") == 0) + strcpy(szaHeadLists, "0ADJP JJ 0AD NN CS 0"); + else if (strcmp(pszCFGLeft, "ADVP") == 0) + strcpy(szaHeadLists, "0ADVP AD 0"); + else if (strcmp(pszCFGLeft, "CLP") == 0) + strcpy(szaHeadLists, "0CLP M 0"); + else if (strcmp(pszCFGLeft, "CP") == 0) + strcpy(szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0"); + else if (strcmp(pszCFGLeft, "DNP") == 0) + strcpy(szaHeadLists, "0DNP DEG 0DEC 0"); + else if (strcmp(pszCFGLeft, "DVP") == 0) + strcpy(szaHeadLists, "0DVP DEV 0"); + else if (strcmp(pszCFGLeft, "DP") == 0) + strcpy(szaHeadLists, "1DP DT 1"); + else if (strcmp(pszCFGLeft, "FRAG") == 0) + strcpy(szaHeadLists, "0VV NR NN 0"); + else if (strcmp(pszCFGLeft, "INTJ") == 0) + strcpy(szaHeadLists, "0INTJ IJ 0"); + else if (strcmp(pszCFGLeft, "LST") == 0) + strcpy(szaHeadLists, "1LST CD OD 1"); + else if (strcmp(pszCFGLeft, "IP") == 0) + strcpy(szaHeadLists, "0IP VP 0VV 0"); + // strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); + else if (strcmp(pszCFGLeft, "LCP") == 0) + strcpy(szaHeadLists, "0LCP LC 0"); + else if (strcmp(pszCFGLeft, "NP") == 0) + strcpy(szaHeadLists, "0NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "PP") == 0) + strcpy(szaHeadLists, "1PP P 1"); + else if (strcmp(pszCFGLeft, "PRN") == 0) + strcpy(szaHeadLists, "0 NP IP VP NT NR NN 0"); + else if (strcmp(pszCFGLeft, "QP") == 0) + strcpy(szaHeadLists, "0QP CLP CD OD 0"); + else if (strcmp(pszCFGLeft, "VP") == 0) + strcpy(szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1"); + else if (strcmp(pszCFGLeft, "VCD") == 0) + strcpy(szaHeadLists, "0VCD VV VA VC VE 0"); + if (strcmp(pszCFGLeft, "VRD") == 0) + strcpy(szaHeadLists, "0VRD VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VSB") == 0) + strcpy(szaHeadLists, "0VSB VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VCP") == 0) + strcpy(szaHeadLists, "0VCP VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VNV") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VPT") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "UCP") == 0) + strcpy(szaHeadLists, "0"); + else if (strcmp(pszCFGLeft, "WHNP") == 0) + strcpy(szaHeadLists, "0WHNP NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "WHPP") == 0) + strcpy(szaHeadLists, "1WHPP PP P 1"); + + /* //head rules for GENIA corpus + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + return fnMyOwnHeadWordRule(szaHeadLists, vectRight); + } + + private: + int fnMyOwnHeadWordRule(char *pszaHeadLists, CVectorStr vectRight) { + char szHeadList[201], *p; + char szTerm[101]; + int J; + + p = pszaHeadLists; + + int iCountRight; + + iCountRight = vectRight.size(); + + szHeadList[0] = '\0'; + while (1) { + szTerm[0] = '\0'; + sscanf(p, "%s", szTerm); + if (strlen(szHeadList) == 0) { + if (strcmp(szTerm, "0") == 0) { + return iCountRight - 1; + } + if (strcmp(szTerm, "1") == 0) { + return 0; + } + + sprintf(szHeadList, "%c %s ", szTerm[0], szTerm + 1); + p = strstr(p, szTerm); + p += strlen(szTerm); + } else { + if ((szTerm[0] == '0') || (szTerm[0] == '1')) { + if (szHeadList[0] == '0') { + for (J = iCountRight - 1; J >= 0; J--) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } else { + for (J = 0; J < iCountRight; J++) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } + + szHeadList[0] = '\0'; + } else { + strcat(szHeadList, szTerm); + strcat(szHeadList, " "); + + p = strstr(p, szTerm); + p += strlen(szTerm); + } + } + } + + return 0; + } }; -struct SParsedTree{ - SParsedTree( ) { - m_ptRoot = NULL; - } - ~SParsedTree( ) { - if (m_ptRoot != NULL) - delete m_ptRoot; - } - static SParsedTree* fnConvertFromString(const char* pszStr) { - if (strcmp(pszStr, "(())") == 0) - return NULL; - SParsedTree* pTree = new SParsedTree(); - - vector vecSyn; - fnReadSyntactic(pszStr, vecSyn); - - int iLeft = 1, iRight = 1; //# left/right parenthesis - - STreeItem *pcurrent; - - pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); - - pcurrent = pTree->m_ptRoot; - - for (size_t i = 2; i < vecSyn.size() - 1; i++) { - if ( strcmp(vecSyn[i].c_str(), "(") == 0 ) - iLeft++; - else if (strcmp(vecSyn[i].c_str(), ")") == 0 ) { - iRight++; - if (pcurrent == NULL) { - //error - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - pcurrent = pcurrent->m_ptParent; - } else { - STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); - pcurrent->fnAppend( ptNewItem ); - pcurrent = ptNewItem; - - if (strcmp(vecSyn[i - 1].c_str(), "(" ) != 0 - && strcmp(vecSyn[i - 1].c_str(), ")" ) != 0 ) { - pTree->m_vecTerminals.push_back(ptNewItem); - pcurrent = pcurrent->m_ptParent; - } - } - } - - if ( iLeft != iRight ) { - //error - fprintf(stderr, "the left and right parentheses are not matched!"); - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - - return pTree; - } - - int fnGetNumWord() { - return m_vecTerminals.size(); - } - - void fnSetSpanInfo() { - int iNextNum = 0; - fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); - } - - void fnSetHeadWord() { - for (size_t i = 0; i < m_vecTerminals.size(); i++) - m_vecTerminals[i]->m_iHeadWord = i; - SGetHeadWord *pGetHeadWord = new SGetHeadWord(); - fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); - delete pGetHeadWord; - } - - STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { - STreeItem *pTreeItem = m_vecTerminals[iLeft]; - - while (pTreeItem->m_iEnd < iRight) { - pTreeItem = pTreeItem->m_ptParent; - if (pTreeItem == NULL) break; - } - if (pTreeItem == NULL) - return NULL; - if (pTreeItem->m_iEnd > iRight) - return NULL; - - assert(pTreeItem->m_iEnd == iRight); - if (bLowest) - return pTreeItem; - - while (pTreeItem->m_ptParent != NULL && pTreeItem->m_ptParent->fnGetChildrenNum() == 1) - pTreeItem = pTreeItem->m_ptParent; - - return pTreeItem; - } - -private: - void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int& iNextNum) { - int I; - int iNumChildren = ptItem->fnGetChildrenNum(); - for ( I = 0; I < iNumChildren; I++ ) - fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[ I ], iNextNum); - - if ( I == 0 ) - { - ptItem->m_iBegin = iNextNum; - ptItem->m_iEnd = iNextNum++; - } - else - { - ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; - ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; - } - } - - - void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, SGetHeadWord *pGetHeadWord) { - int I, iHeadchild; - - if ( ptItem->m_vecChildren.size() == 0 ) - return; - - for ( I = 0; I < ptItem->m_vecChildren.size(); I++ ) - fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); - - vector vecRight; - - - if ( ptItem->m_vecChildren.size() == 1 ) - iHeadchild = 0; - else - { - for ( I = 0; I < ptItem->m_vecChildren.size(); I++ ) - vecRight.push_back( string( ptItem->m_vecChildren[ I ]->m_pszTerm ) ); - - iHeadchild = pGetHeadWord->fnGetHeadWord( ptItem->m_pszTerm, vecRight ); - } - - ptItem->m_iHeadChild = iHeadchild; - ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - - - static void fnReadSyntactic(const char *pszSyn, vector& vec) { - char *p; - int I; - - int iLeftNum, iRightNum; - char *pszTmp, *pszTerm; - pszTmp = new char[strlen(pszSyn)]; - pszTerm = new char[strlen(pszSyn)]; - pszTmp[0] = pszTerm[0] = '\0'; - - vec.clear(); - - char *pszLine; - pszLine = new char[strlen(pszSyn) + 1]; - strcpy( pszLine, pszSyn ); - - char *pszLine2; - - while( 1 ) { - while(( strlen( pszLine ) > 0 ) - &&( pszLine[ strlen( pszLine ) - 1 ] > 0 ) - &&( pszLine[ strlen( pszLine ) - 1 ] <= ' ' ) ) - pszLine[ strlen( pszLine ) - 1 ] = '\0'; - - if( strlen( pszLine ) == 0 ) - break; - - //printf( "%s\n", pszLine ); - pszLine2 = pszLine; - while( pszLine2[ 0 ] <= ' ' ) - pszLine2 ++; - if( pszLine2[ 0 ] == '<' ) - continue; - - sscanf( pszLine2 + 1, "%s", pszTmp ); - - if ( pszLine2[ 0 ] == '(' ) { - iLeftNum = 0; - iRightNum = 0; - } - - p = pszLine2; - while ( 1 ) { - pszTerm[ 0 ] = '\0'; - sscanf( p, "%s", pszTerm ); - - if( strlen( pszTerm ) == 0 ) - break; - p = strstr( p, pszTerm ); - p += strlen( pszTerm ); - - if( ( pszTerm[ 0 ] == '(' ) - ||( pszTerm[ strlen( pszTerm ) - 1 ] == ')' ) ) { - if( pszTerm[ 0 ] == '(' ) { - vec.push_back(string("(")); - iLeftNum++; - - I = 1; - while ( pszTerm[ I ] == '(' && pszTerm[ I ]!= '\0' ) { - vec.push_back(string("(")); - iLeftNum++; - - I++; - } - - if( strlen( pszTerm ) > 1 ) - vec.push_back(string(pszTerm + I)); - } else { - char *pTmp; - pTmp = pszTerm + strlen( pszTerm ) - 1; - while( ( pTmp[ 0 ] == ')' ) && ( pTmp >= pszTerm ) ) - pTmp --; - pTmp[ 1 ] = '\0'; - - if( strlen( pszTerm ) > 0 ) - vec.push_back(string(pszTerm)); - pTmp += 2; - - for( I = 0; I <= (int)strlen( pTmp ); I ++ ) { - vec.push_back(string(")")); - iRightNum++; - } - } - } else { - char *q; - q = strchr( pszTerm, ')' ); - if ( q != NULL ) { - q[ 0 ] = '\0'; - if ( pszTerm[ 0 ] != '\0' ) - vec.push_back(string(pszTerm)); - vec.push_back(string(")")); - iRightNum++; - - q++; - while ( q[ 0 ] == ')' ) { - vec.push_back(string(")")); - q++; - iRightNum ++; - } - - while( q[ 0 ] == '(' ) { - vec.push_back(string("(")); - q++; - iLeftNum++; - } - - if ( q[ 0 ] != '\0' ) - vec.push_back(string(q)); - } - else - vec.push_back(string(pszTerm)); - } - } - - if (iLeftNum != iRightNum) { - fprintf(stderr, "%s\n", pszSyn); - assert(iLeftNum == iRightNum); - } - /*if ( iLeftNum != iRightNum ) { - printf( "ERROR: left( and right ) is not matched, %d ( and %d )\n", iLeftNum, iRightNum ); - return; - }*/ - - - if( vec.size() >= 2 - && strcmp( vec[1].c_str(), "(" ) == 0 ) { - //( (IP..) ) - std::vector::iterator it; - it = vec.begin(); - it++; - vec.insert(it, string("ROOT")); - } - - break; - } - - delete [] pszLine; - delete [] pszTmp; - delete [] pszTerm; - } - -public: - STreeItem *m_ptRoot; - vector m_vecTerminals; //the leaf nodes +struct SParsedTree { + SParsedTree() { m_ptRoot = NULL; } + ~SParsedTree() { + if (m_ptRoot != NULL) delete m_ptRoot; + } + static SParsedTree *fnConvertFromString(const char *pszStr) { + if (strcmp(pszStr, "(())") == 0) return NULL; + SParsedTree *pTree = new SParsedTree(); + + vector vecSyn; + fnReadSyntactic(pszStr, vecSyn); + + int iLeft = 1, iRight = 1; //# left/right parenthesis + + STreeItem *pcurrent; + + pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); + + pcurrent = pTree->m_ptRoot; + + for (size_t i = 2; i < vecSyn.size() - 1; i++) { + if (strcmp(vecSyn[i].c_str(), "(") == 0) + iLeft++; + else if (strcmp(vecSyn[i].c_str(), ")") == 0) { + iRight++; + if (pcurrent == NULL) { + // error + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + pcurrent = pcurrent->m_ptParent; + } else { + STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); + pcurrent->fnAppend(ptNewItem); + pcurrent = ptNewItem; + + if (strcmp(vecSyn[i - 1].c_str(), "(") != 0 && + strcmp(vecSyn[i - 1].c_str(), ")") != 0) { + pTree->m_vecTerminals.push_back(ptNewItem); + pcurrent = pcurrent->m_ptParent; + } + } + } + + if (iLeft != iRight) { + // error + fprintf(stderr, "the left and right parentheses are not matched!"); + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + + return pTree; + } + + int fnGetNumWord() { return m_vecTerminals.size(); } + + void fnSetSpanInfo() { + int iNextNum = 0; + fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); + } + + void fnSetHeadWord() { + for (size_t i = 0; i < m_vecTerminals.size(); i++) + m_vecTerminals[i]->m_iHeadWord = i; + SGetHeadWord *pGetHeadWord = new SGetHeadWord(); + fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); + delete pGetHeadWord; + } + + STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { + STreeItem *pTreeItem = m_vecTerminals[iLeft]; + + while (pTreeItem->m_iEnd < iRight) { + pTreeItem = pTreeItem->m_ptParent; + if (pTreeItem == NULL) break; + } + if (pTreeItem == NULL) return NULL; + if (pTreeItem->m_iEnd > iRight) return NULL; + + assert(pTreeItem->m_iEnd == iRight); + if (bLowest) return pTreeItem; + + while (pTreeItem->m_ptParent != NULL && + pTreeItem->m_ptParent->fnGetChildrenNum() == 1) + pTreeItem = pTreeItem->m_ptParent; + + return pTreeItem; + } + + private: + void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int &iNextNum) { + int I; + int iNumChildren = ptItem->fnGetChildrenNum(); + for (I = 0; I < iNumChildren; I++) + fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[I], iNextNum); + + if (I == 0) { + ptItem->m_iBegin = iNextNum; + ptItem->m_iEnd = iNextNum++; + } else { + ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; + ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; + } + } + + void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, + SGetHeadWord *pGetHeadWord) { + int I, iHeadchild; + + if (ptItem->m_vecChildren.size() == 0) return; + + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); + + vector vecRight; + + if (ptItem->m_vecChildren.size() == 1) + iHeadchild = 0; + else { + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + vecRight.push_back(string(ptItem->m_vecChildren[I]->m_pszTerm)); + + iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); + } + + ptItem->m_iHeadChild = iHeadchild; + ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + + static void fnReadSyntactic(const char *pszSyn, vector &vec) { + char *p; + int I; + + int iLeftNum, iRightNum; + char *pszTmp, *pszTerm; + pszTmp = new char[strlen(pszSyn)]; + pszTerm = new char[strlen(pszSyn)]; + pszTmp[0] = pszTerm[0] = '\0'; + + vec.clear(); + + char *pszLine; + pszLine = new char[strlen(pszSyn) + 1]; + strcpy(pszLine, pszSyn); + + char *pszLine2; + + while (1) { + while ((strlen(pszLine) > 0) && (pszLine[strlen(pszLine) - 1] > 0) && + (pszLine[strlen(pszLine) - 1] <= ' ')) + pszLine[strlen(pszLine) - 1] = '\0'; + + if (strlen(pszLine) == 0) break; + + // printf( "%s\n", pszLine ); + pszLine2 = pszLine; + while (pszLine2[0] <= ' ') pszLine2++; + if (pszLine2[0] == '<') continue; + + sscanf(pszLine2 + 1, "%s", pszTmp); + + if (pszLine2[0] == '(') { + iLeftNum = 0; + iRightNum = 0; + } + + p = pszLine2; + while (1) { + pszTerm[0] = '\0'; + sscanf(p, "%s", pszTerm); + + if (strlen(pszTerm) == 0) break; + p = strstr(p, pszTerm); + p += strlen(pszTerm); + + if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { + if (pszTerm[0] == '(') { + vec.push_back(string("(")); + iLeftNum++; + + I = 1; + while (pszTerm[I] == '(' && pszTerm[I] != '\0') { + vec.push_back(string("(")); + iLeftNum++; + + I++; + } + + if (strlen(pszTerm) > 1) vec.push_back(string(pszTerm + I)); + } else { + char *pTmp; + pTmp = pszTerm + strlen(pszTerm) - 1; + while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; + pTmp[1] = '\0'; + + if (strlen(pszTerm) > 0) vec.push_back(string(pszTerm)); + pTmp += 2; + + for (I = 0; I <= (int)strlen(pTmp); I++) { + vec.push_back(string(")")); + iRightNum++; + } + } + } else { + char *q; + q = strchr(pszTerm, ')'); + if (q != NULL) { + q[0] = '\0'; + if (pszTerm[0] != '\0') vec.push_back(string(pszTerm)); + vec.push_back(string(")")); + iRightNum++; + + q++; + while (q[0] == ')') { + vec.push_back(string(")")); + q++; + iRightNum++; + } + + while (q[0] == '(') { + vec.push_back(string("(")); + q++; + iLeftNum++; + } + + if (q[0] != '\0') vec.push_back(string(q)); + } else + vec.push_back(string(pszTerm)); + } + } + + if (iLeftNum != iRightNum) { + fprintf(stderr, "%s\n", pszSyn); + assert(iLeftNum == iRightNum); + } + /*if ( iLeftNum != iRightNum ) { + printf( "ERROR: left( and right ) is not matched, %d ( and %d + )\n", iLeftNum, iRightNum ); + return; + }*/ + + if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { + //( (IP..) ) + std::vector::iterator it; + it = vec.begin(); + it++; + vec.insert(it, string("ROOT")); + } + + break; + } + + delete[] pszLine; + delete[] pszTmp; + delete[] pszTerm; + } + + public: + STreeItem *m_ptRoot; + vector m_vecTerminals; // the leaf nodes }; struct SParseReader { - SParseReader(const char* pszParse_Fname, bool bFlattened = false) : - m_bFlattened(bFlattened){ - m_fpIn = fopen( pszParse_Fname, "r" ); - assert(m_fpIn != NULL); - } - ~SParseReader() { - if (m_fpIn != NULL) - fclose(m_fpIn); - } - - SParsedTree* fnReadNextParseTree( ) { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true ) { - if (iLen == 0) - continue; - - pTree = SParsedTree::fnConvertFromString(pszLine); - if (pTree == NULL) - break; - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete [] pszLine; - return pTree; - } - - SParsedTree* fnReadNextParseTreeWithProb(double *pProb) { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true ) { - if (iLen == 0) - continue; - - char *p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - if (pProb) - (*pProb) = atof(pszLine); - - pTree = SParsedTree::fnConvertFromString(p); - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete [] pszLine; - return pTree; - } -private: - /* - * since to the parse tree is a flattened tree, use the head mark to identify head info. - * the head node will be marked as "*XP*" - */ - void fnSetParseTreeHeadInfo(SParsedTree *pTree) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) - pTree->m_vecTerminals[i]->m_iHeadWord = i; - fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); - } - - void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { - if (pTreeItem->m_vecChildren.size() == 0) - return; - - for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++ ) - fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); - - vector vecRight; - - int iHeadchild; - - if (pTreeItem->fnIsPreTerminal()) { - iHeadchild = 0; - } else { - size_t i; - for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { - char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; - if (p[0] == '*' && p[strlen(p) - 1] == '*') { - iHeadchild = i; - p[strlen(p) - 1] = '\0'; - string str = p + 1; - strcpy(p, str.c_str()); //erase the "*..*" - break; - } - } - assert(i < pTreeItem->m_vecChildren.size()); - } - - pTreeItem->m_iHeadChild = iHeadchild; - pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - void fnPostProcessingFlattenedParse(SParsedTree *pTree) { - pTree->fnSetSpanInfo(); - fnSetParseTreeHeadInfo(pTree); - } - bool fnReadNextSentence(char *pszLine, int* piLength) { - if (feof(m_fpIn) == true) - return false; - - int iLen; - - pszLine[ 0 ] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - - if ( piLength != NULL ) - (*piLength) = iLen; - - return true; - } -private: - FILE *m_fpIn; - const bool m_bFlattened; + SParseReader(const char *pszParse_Fname, bool bFlattened = false) + : m_bFlattened(bFlattened) { + m_fpIn = fopen(pszParse_Fname, "r"); + assert(m_fpIn != NULL); + } + ~SParseReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + SParsedTree *fnReadNextParseTree() { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + pTree = SParsedTree::fnConvertFromString(pszLine); + if (pTree == NULL) break; + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + SParsedTree *fnReadNextParseTreeWithProb(double *pProb) { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + char *p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + if (pProb) (*pProb) = atof(pszLine); + + pTree = SParsedTree::fnConvertFromString(p); + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + private: + /* + * since to the parse tree is a flattened tree, use the head mark to identify + * head info. + * the head node will be marked as "*XP*" + */ + void fnSetParseTreeHeadInfo(SParsedTree *pTree) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) + pTree->m_vecTerminals[i]->m_iHeadWord = i; + fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); + } + + void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { + if (pTreeItem->m_vecChildren.size() == 0) return; + + for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) + fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); + + vector vecRight; + + int iHeadchild; + + if (pTreeItem->fnIsPreTerminal()) { + iHeadchild = 0; + } else { + size_t i; + for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { + char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; + if (p[0] == '*' && p[strlen(p) - 1] == '*') { + iHeadchild = i; + p[strlen(p) - 1] = '\0'; + string str = p + 1; + strcpy(p, str.c_str()); // erase the "*..*" + break; + } + } + assert(i < pTreeItem->m_vecChildren.size()); + } + + pTreeItem->m_iHeadChild = iHeadchild; + pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + void fnPostProcessingFlattenedParse(SParsedTree *pTree) { + pTree->fnSetSpanInfo(); + fnSetParseTreeHeadInfo(pTree); + } + bool fnReadNextSentence(char *pszLine, int *piLength) { + if (feof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + private: + FILE *m_fpIn; + const bool m_bFlattened; }; #endif /* TREE_H_ */ diff --git a/utils/synutils/tsuruoka_maxent.h b/utils/synutils/tsuruoka_maxent.h index 790876ab..b5a87404 100644 --- a/utils/synutils/tsuruoka_maxent.h +++ b/utils/synutils/tsuruoka_maxent.h @@ -18,143 +18,138 @@ using namespace std; - typedef std::tr1::unordered_map Map; typedef std::tr1::unordered_map::iterator Iterator; - - -struct Tsuruoka_Maxent{ - Tsuruoka_Maxent(const char* pszModelFName) { - if (pszModelFName != NULL) { - m_pModel = new ME_Model(); - m_pModel->load_from_file(pszModelFName); - } else - m_pModel = NULL; - } - - ~Tsuruoka_Maxent() { - if (m_pModel != NULL) - delete m_pModel; - } - - void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, const char* pszModelFName, int iNumIteration) { - assert(strcmp(pszAlgorithm, "l1") == 0 - || strcmp(pszAlgorithm, "l2") == 0 - || strcmp(pszAlgorithm, "sgd") == 0 - || strcmp(pszAlgorithm, "SGD") == 0); - FILE *fpIn = fopen(pszInstanceFName, "r"); - - ME_Model *pModel = new ME_Model(); - - char *pszLine = new char[100001]; - int iNumInstances = 0; - int iLen; - while (!feof(fpIn)) { - pszLine[0] = '\0'; - fgets(pszLine, 20000, fpIn); - if (strlen(pszLine) == 0) { - continue; - } - - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - - - iNumInstances++; - - ME_Sample *pmes = new ME_Sample(); - - char *p = strrchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - vector vecContext; - SplitOnWhitespace(string(pszLine), &vecContext); - - pmes->label = string(p); - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - pModel->add_training_sample((*pmes)); - if (iNumInstances % 100000 == 0) - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - delete pmes; - } - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - fclose(fpIn); - - if (strcmp(pszAlgorithm, "l1") == 0) - pModel->use_l1_regularizer(1.0); - else if (strcmp(pszAlgorithm, "l2") == 0) - pModel->use_l2_regularizer(1.0); - else - pModel->use_SGD(); - - pModel->train(); - pModel->save_to_file(pszModelFName); - - delete pModel; - fprintf(stdout, "......Finished Training\n"); - fprintf(stdout, "......Model saved as %s\n", pszModelFName); - delete [] pszLine; - } - - double fnEval(const char* pszContext, const char* pszOutcome) const { - vector vecContext; - ME_Sample *pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - delete pmes; - int iLableID = m_pModel->get_class_id(pszOutcome); - return vecProb[iLableID]; - } - void fnEval(const char* pszContext, vector >& vecOutput) const { - vector vecContext; - ME_Sample *pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); - vecOutput.push_back(make_pair(label, vecProb[i])); - } - delete pmes; - } - void fnEval(const char* pszContext, vector& vecOutput) const{ - vector vecContext; - ME_Sample *pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); - vecOutput.push_back(vecProb[i]); - } - delete pmes; - } - int fnGetClassId(const string& strLabel) const { - return m_pModel->get_class_id(strLabel); - } -private: - ME_Model *m_pModel; +struct Tsuruoka_Maxent { + Tsuruoka_Maxent(const char* pszModelFName) { + if (pszModelFName != NULL) { + m_pModel = new ME_Model(); + m_pModel->load_from_file(pszModelFName); + } else + m_pModel = NULL; + } + + ~Tsuruoka_Maxent() { + if (m_pModel != NULL) delete m_pModel; + } + + void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, + const char* pszModelFName, int iNumIteration) { + assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || + strcmp(pszAlgorithm, "sgd") == 0 || + strcmp(pszAlgorithm, "SGD") == 0); + FILE* fpIn = fopen(pszInstanceFName, "r"); + + ME_Model* pModel = new ME_Model(); + + char* pszLine = new char[100001]; + int iNumInstances = 0; + int iLen; + while (!feof(fpIn)) { + pszLine[0] = '\0'; + fgets(pszLine, 20000, fpIn); + if (strlen(pszLine) == 0) { + continue; + } + + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + iNumInstances++; + + ME_Sample* pmes = new ME_Sample(); + + char* p = strrchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + vector vecContext; + SplitOnWhitespace(string(pszLine), &vecContext); + + pmes->label = string(p); + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + pModel->add_training_sample((*pmes)); + if (iNumInstances % 100000 == 0) + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + delete pmes; + } + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + fclose(fpIn); + + if (strcmp(pszAlgorithm, "l1") == 0) + pModel->use_l1_regularizer(1.0); + else if (strcmp(pszAlgorithm, "l2") == 0) + pModel->use_l2_regularizer(1.0); + else + pModel->use_SGD(); + + pModel->train(); + pModel->save_to_file(pszModelFName); + + delete pModel; + fprintf(stdout, "......Finished Training\n"); + fprintf(stdout, "......Model saved as %s\n", pszModelFName); + delete[] pszLine; + } + + double fnEval(const char* pszContext, const char* pszOutcome) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + delete pmes; + int iLableID = m_pModel->get_class_id(pszOutcome); + return vecProb[iLableID]; + } + void fnEval(const char* pszContext, + vector >& vecOutput) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(make_pair(label, vecProb[i])); + } + delete pmes; + } + void fnEval(const char* pszContext, vector& vecOutput) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(vecProb[i]); + } + delete pmes; + } + int fnGetClassId(const string& strLabel) const { + return m_pModel->get_class_id(strLabel); + } + + private: + ME_Model* m_pModel; }; - - #endif /* TSURUOKA_MAXENT_H_ */ diff --git a/utils/synutils/utility.h b/utils/synutils/utility.h index d3f52ee0..791b2286 100644 --- a/utils/synutils/utility.h +++ b/utils/synutils/utility.h @@ -8,7 +8,6 @@ #ifndef UTILITY_H_ #define UTILITY_H_ - #include #include #include @@ -20,110 +19,107 @@ using namespace std; typedef std::unordered_map MapString2Int; typedef std::unordered_map MapString2Float; -typedef std::unordered_map::iterator MapString2FloatIterator; +typedef std::unordered_map::iterator + MapString2FloatIterator; using namespace std; struct SFReader { - SFReader( ){ } - virtual ~SFReader( ){ } + SFReader() {} + virtual ~SFReader() {} - virtual bool fnReadNextLine( char *pszLine, int* piLength ) = 0; - virtual bool fnReadNextLine(string& strLine) = 0; + virtual bool fnReadNextLine(char* pszLine, int* piLength) = 0; + virtual bool fnReadNextLine(string& strLine) = 0; }; -struct STxtFileReader: public SFReader { - STxtFileReader(const char* pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~STxtFileReader() { - if (m_fpIn != NULL) - fclose(m_fpIn); - } - - bool fnReadNextLine(char *pszLine, int* piLength) { - if (feof(m_fpIn) == true) - return false; - - int iLen; - - pszLine[ 0 ] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - if (iLen == 0) return false; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen -1] < 33) { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - - if ( piLength != NULL ) - (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(string& strLine) { - char *pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) strLine = string(pszLine); - else strLine = string(""); - delete [] pszLine; - - return bOut; - } -private: - FILE *m_fpIn; +struct STxtFileReader : public SFReader { + STxtFileReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~STxtFileReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + bool fnReadNextLine(char* pszLine, int* piLength) { + if (feof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + if (iLen == 0) return false; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char* pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) + strLine = string(pszLine); + else + strLine = string(""); + delete[] pszLine; + + return bOut; + } + + private: + FILE* m_fpIn; }; -struct SGZFileReader: public SFReader { - SGZFileReader(const char* pszFname) { - m_fpIn = gzopen( pszFname, "r" ); - assert(m_fpIn != NULL); - } - ~SGZFileReader() { - if (m_fpIn != NULL) - gzclose(m_fpIn); - } - - bool fnReadNextLine(char *pszLine, int* piLength) { - if ( m_fpIn == NULL ) - exit( 0 ); - if ( gzeof( m_fpIn ) == true ) - return false; - - int iLen; - - pszLine[ 0 ] = '\0'; - - gzgets( m_fpIn, pszLine, 10001 ); - iLen = strlen( pszLine ); - while ( iLen > 0 && pszLine[ iLen - 1 ] > 0 && pszLine[ iLen -1 ] < 33 ) - { - pszLine[ iLen - 1 ] = '\0'; - iLen--; - } - - if ( piLength != NULL ) - (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(string& strLine) { - char *pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) strLine = string(pszLine); - else strLine = string(""); - delete [] pszLine; - - return bOut; - } -private: - gzFile m_fpIn; +struct SGZFileReader : public SFReader { + SGZFileReader(const char* pszFname) { + m_fpIn = gzopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SGZFileReader() { + if (m_fpIn != NULL) gzclose(m_fpIn); + } + + bool fnReadNextLine(char* pszLine, int* piLength) { + if (m_fpIn == NULL) exit(0); + if (gzeof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + gzgets(m_fpIn, pszLine, 10001); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char* pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) + strLine = string(pszLine); + else + strLine = string(""); + delete[] pszLine; + + return bOut; + } + + private: + gzFile m_fpIn; }; - #endif /* UTILITY_H_ */ - -- cgit v1.2.3 From ebeda1e75c77dd9044f1b9902770896e3009ae55 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Tue, 7 Oct 2014 17:41:19 -0400 Subject: Make synutils compile --- .gitignore | 1 + Makefile.am | 3 ++- configure.ac | 9 +++++---- utils/synutils/Makefile.am | 14 ++------------ utils/synutils/alignment.h | 2 +- utils/synutils/constituent_reorder_model.cc | 6 +++--- utils/synutils/srl_sentence.h | 5 ++++- utils/synutils/tree.h | 4 +++- utils/synutils/tsuruoka_maxent.h | 10 +++++----- utils/synutils/utility.h | 2 ++ 10 files changed, 28 insertions(+), 28 deletions(-) (limited to 'utils') diff --git a/.gitignore b/.gitignore index 72f4997c..f9e17574 100644 --- a/.gitignore +++ b/.gitignore @@ -192,6 +192,7 @@ utils/phmt utils/reconstruct_weights utils/small_vector_test utils/sv_test +utils/synutils/const_reorder_model_trainer utils/ts utils/weights_test training/crf/mpi_adagrad_optimize diff --git a/Makefile.am b/Makefile.am index 88327477..ce49799a 100644 --- a/Makefile.am +++ b/Makefile.am @@ -3,6 +3,8 @@ # cyclic dependencies between these directories! SUBDIRS = \ utils \ + utils/synutils/maxent-3.0 \ + utils/synutils \ klm/util/double-conversion \ klm/util \ klm/util/stream \ @@ -21,4 +23,3 @@ EXTRA_DIST = corpus tests python/cdec python/tests python/examples compound-spli AUTOMAKE_OPTIONS = foreign ACLOCAL_AMFLAGS = -I m4 AM_CPPFLAGS = -D_GLIBCXX_PARALLEL -march=native -mtune=native -O2 -pipe -fomit-frame-pointer -Wall - diff --git a/configure.ac b/configure.ac index c5575f14..e1ec47ee 100644 --- a/configure.ac +++ b/configure.ac @@ -5,9 +5,9 @@ AM_INIT_AUTOMAKE AC_CONFIG_HEADERS(config.h) AC_PROG_LIBTOOL AC_PROG_LEX -case $LEX in -:) AC_MSG_ERROR([No lex (Flex, lex, etc.) program found]);; -esac +case $LEX in +:) AC_MSG_ERROR([No lex (Flex, lex, etc.) program found]);; +esac OLD_CXXFLAGS=$CXXFLAGS AC_PROG_CC AC_PROG_CXX @@ -208,6 +208,8 @@ fi # core cdec stuff AC_CONFIG_FILES([Makefile]) AC_CONFIG_FILES([utils/Makefile]) +AC_CONFIG_FILES([utils/synutils/maxent-3.0/Makefile]) +AC_CONFIG_FILES([utils/synutils/Makefile]) AC_CONFIG_FILES([mteval/Makefile]) AC_CONFIG_FILES([mteval/meteor_jar.cc]) AC_CONFIG_FILES([decoder/Makefile]) @@ -240,4 +242,3 @@ AC_CONFIG_FILES([training/dtrain/Makefile]) AC_CONFIG_FILES([example_extff/Makefile]) AC_OUTPUT - diff --git a/utils/synutils/Makefile.am b/utils/synutils/Makefile.am index f87f1f8a..ac76be40 100644 --- a/utils/synutils/Makefile.am +++ b/utils/synutils/Makefile.am @@ -1,16 +1,6 @@ bin_PROGRAMS = const_reorder_model_trainer const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc -const_reorder_model_trainer_LDADD = libfeature.a maxent-3.0/libtsuruoka_maxent.a -lgfortran +const_reorder_model_trainer_LDADD = maxent-3.0/libtsuruoka_maxent.a -noinst_LIBRARIES = libfeature.a - -libfeature_a_SOURCES = \ - alignment.h \ - srl_sentence.h \ - tree.h \ - utility.h \ - argument_reorder_model.h \ - tsuruoka_maxent.h - -AM_CPPFLAGS = -W -Wall -I$(top_srcdir) -I$(top_srcdir)/utils -I$(top_srcdir)/synutils/maxent-3.0 +AM_CPPFLAGS = -W -Wall -I$(top_srcdir) diff --git a/utils/synutils/alignment.h b/utils/synutils/alignment.h index c0648aab..6abd12bc 100644 --- a/utils/synutils/alignment.h +++ b/utils/synutils/alignment.h @@ -13,7 +13,7 @@ #include #include -#include "stringlib.h" +#include "utils/stringlib.h" using namespace std; diff --git a/utils/synutils/constituent_reorder_model.cc b/utils/synutils/constituent_reorder_model.cc index a4fb9627..78d61b90 100644 --- a/utils/synutils/constituent_reorder_model.cc +++ b/utils/synutils/constituent_reorder_model.cc @@ -12,12 +12,12 @@ #include "utility.h" #include "tsuruoka_maxent.h" -#include +#include using namespace std; -typedef std::tr1::unordered_map Map; -typedef std::tr1::unordered_map::iterator Iterator; +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; namespace po = boost::program_options; diff --git a/utils/synutils/srl_sentence.h b/utils/synutils/srl_sentence.h index c08cc065..1afdec45 100644 --- a/utils/synutils/srl_sentence.h +++ b/utils/synutils/srl_sentence.h @@ -8,10 +8,13 @@ #ifndef SRL_SENTENCE_H_ #define SRL_SENTENCE_H_ -#include "tree.h" +#include #include +#include "tree.h" +#include "utils/stringlib.h" + using namespace std; struct SArgument { diff --git a/utils/synutils/tree.h b/utils/synutils/tree.h index 32521a3a..8070f828 100644 --- a/utils/synutils/tree.h +++ b/utils/synutils/tree.h @@ -8,10 +8,12 @@ #ifndef TREE_H_ #define TREE_H_ -#include #include #include +#include +#include + using namespace std; struct STreeItem { diff --git a/utils/synutils/tsuruoka_maxent.h b/utils/synutils/tsuruoka_maxent.h index b5a87404..89b06b97 100644 --- a/utils/synutils/tsuruoka_maxent.h +++ b/utils/synutils/tsuruoka_maxent.h @@ -7,19 +7,19 @@ #define TSURUOKA_MAXENT_H_ #include "utility.h" -#include "stringlib.h" -#include "maxent-3.0/maxent.h" +#include "utils/stringlib.h" +#include "utils/synutils/maxent-3.0/maxent.h" #include #include #include #include -#include +#include using namespace std; -typedef std::tr1::unordered_map Map; -typedef std::tr1::unordered_map::iterator Iterator; +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; struct Tsuruoka_Maxent { Tsuruoka_Maxent(const char* pszModelFName) { diff --git a/utils/synutils/utility.h b/utils/synutils/utility.h index 791b2286..ef7b78b7 100644 --- a/utils/synutils/utility.h +++ b/utils/synutils/utility.h @@ -13,6 +13,8 @@ #include #include #include + +#include #include using namespace std; -- cgit v1.2.3 From 9ba88cc8f776d85ef821a88c72413b14484e6457 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Tue, 7 Oct 2014 18:44:05 -0400 Subject: Move synutils under utils --- .gitignore | 2 +- Makefile.am | 2 - configure.ac | 2 - decoder/Makefile.am | 2 +- decoder/ff_const_reorder.cc | 8 +- utils/Makefile.am | 18 +- utils/alignment.h | 202 +++++++ utils/argument_reorder_model.cc | 323 +++++++++++ utils/argument_reorder_model.h | 239 ++++++++ utils/constituent_reorder_model.cc | 846 ++++++++++++++++++++++++++++ utils/lbfgs.cpp | 108 ++++ utils/lbfgs.h | 19 + utils/mathvec.h | 87 +++ utils/maxent.cpp | 703 +++++++++++++++++++++++ utils/maxent.h | 402 +++++++++++++ utils/owlqn.cpp | 127 +++++ utils/sgd.cpp | 193 +++++++ utils/srl_sentence.h | 214 +++++++ utils/synutils.h | 127 +++++ utils/synutils/Makefile.am | 6 - utils/synutils/alignment.h | 202 ------- utils/synutils/argument_reorder_model.cc | 323 ----------- utils/synutils/argument_reorder_model.h | 239 -------- utils/synutils/constituent_reorder_model.cc | 846 ---------------------------- utils/synutils/maxent-3.0/Makefile.am | 10 - utils/synutils/maxent-3.0/lbfgs.cpp | 108 ---- utils/synutils/maxent-3.0/lbfgs.h | 19 - utils/synutils/maxent-3.0/mathvec.h | 87 --- utils/synutils/maxent-3.0/maxent.cpp | 703 ----------------------- utils/synutils/maxent-3.0/maxent.h | 402 ------------- utils/synutils/maxent-3.0/owlqn.cpp | 127 ----- utils/synutils/maxent-3.0/sgd.cpp | 193 ------- utils/synutils/srl_sentence.h | 214 ------- utils/synutils/tree.h | 700 ----------------------- utils/synutils/tsuruoka_maxent.h | 155 ----- utils/synutils/utility.h | 127 ----- utils/tree.h | 700 +++++++++++++++++++++++ utils/tsuruoka_maxent.h | 155 +++++ 38 files changed, 4467 insertions(+), 4473 deletions(-) create mode 100644 utils/alignment.h create mode 100644 utils/argument_reorder_model.cc create mode 100644 utils/argument_reorder_model.h create mode 100644 utils/constituent_reorder_model.cc create mode 100644 utils/lbfgs.cpp create mode 100644 utils/lbfgs.h create mode 100644 utils/mathvec.h create mode 100644 utils/maxent.cpp create mode 100644 utils/maxent.h create mode 100644 utils/owlqn.cpp create mode 100644 utils/sgd.cpp create mode 100644 utils/srl_sentence.h create mode 100644 utils/synutils.h delete mode 100644 utils/synutils/Makefile.am delete mode 100644 utils/synutils/alignment.h delete mode 100644 utils/synutils/argument_reorder_model.cc delete mode 100644 utils/synutils/argument_reorder_model.h delete mode 100644 utils/synutils/constituent_reorder_model.cc delete mode 100644 utils/synutils/maxent-3.0/Makefile.am delete mode 100644 utils/synutils/maxent-3.0/lbfgs.cpp delete mode 100644 utils/synutils/maxent-3.0/lbfgs.h delete mode 100644 utils/synutils/maxent-3.0/mathvec.h delete mode 100644 utils/synutils/maxent-3.0/maxent.cpp delete mode 100644 utils/synutils/maxent-3.0/maxent.h delete mode 100644 utils/synutils/maxent-3.0/owlqn.cpp delete mode 100644 utils/synutils/maxent-3.0/sgd.cpp delete mode 100644 utils/synutils/srl_sentence.h delete mode 100644 utils/synutils/tree.h delete mode 100644 utils/synutils/tsuruoka_maxent.h delete mode 100644 utils/synutils/utility.h create mode 100644 utils/tree.h create mode 100644 utils/tsuruoka_maxent.h (limited to 'utils') diff --git a/.gitignore b/.gitignore index f9e17574..4efcdd8d 100644 --- a/.gitignore +++ b/.gitignore @@ -183,6 +183,7 @@ training/plftools training/test_ngram utils/atools utils/bin/ +utils/const_reorder_model_trainer utils/crp_test utils/dict_test utils/logval_test @@ -192,7 +193,6 @@ utils/phmt utils/reconstruct_weights utils/small_vector_test utils/sv_test -utils/synutils/const_reorder_model_trainer utils/ts utils/weights_test training/crf/mpi_adagrad_optimize diff --git a/Makefile.am b/Makefile.am index ce49799a..a2d2f332 100644 --- a/Makefile.am +++ b/Makefile.am @@ -3,8 +3,6 @@ # cyclic dependencies between these directories! SUBDIRS = \ utils \ - utils/synutils/maxent-3.0 \ - utils/synutils \ klm/util/double-conversion \ klm/util \ klm/util/stream \ diff --git a/configure.ac b/configure.ac index e1ec47ee..0b1f04f0 100644 --- a/configure.ac +++ b/configure.ac @@ -208,8 +208,6 @@ fi # core cdec stuff AC_CONFIG_FILES([Makefile]) AC_CONFIG_FILES([utils/Makefile]) -AC_CONFIG_FILES([utils/synutils/maxent-3.0/Makefile]) -AC_CONFIG_FILES([utils/synutils/Makefile]) AC_CONFIG_FILES([mteval/Makefile]) AC_CONFIG_FILES([mteval/meteor_jar.cc]) AC_CONFIG_FILES([decoder/Makefile]) diff --git a/decoder/Makefile.am b/decoder/Makefile.am index 6577cc12..cc52ccd4 100644 --- a/decoder/Makefile.am +++ b/decoder/Makefile.am @@ -21,7 +21,7 @@ trule_test_LDADD = $(BOOST_UNIT_TEST_FRAMEWORK_LDFLAGS) $(BOOST_UNIT_TEST_FRAMEW cdec_SOURCES = cdec.cc cdec_LDFLAGS= -rdynamic $(STATIC_FLAGS) -cdec_LDADD = libcdec.a ../mteval/libmteval.a ../utils/libutils.a ../utils/synutils/maxent-3.0/libtsuruoka_maxent.a ../klm/search/libksearch.a ../klm/lm/libklm.a ../klm/util/libklm_util.a ../klm/util/double-conversion/libklm_util_double.a +cdec_LDADD = libcdec.a ../mteval/libmteval.a ../utils/libutils.a ../klm/search/libksearch.a ../klm/lm/libklm.a ../klm/util/libklm_util.a ../klm/util/double-conversion/libklm_util_double.a AM_CPPFLAGS = -DTEST_DATA=\"$(top_srcdir)/decoder/test_data\" -DBOOST_TEST_DYN_LINK -W -Wno-sign-compare -I$(top_srcdir) -I$(top_srcdir)/mteval -I$(top_srcdir)/utils -I$(top_srcdir)/klm diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc index 77582b7c..97b76f7f 100644 --- a/decoder/ff_const_reorder.cc +++ b/decoder/ff_const_reorder.cc @@ -3,11 +3,11 @@ #include "stringlib.h" #include "hg.h" #include "sentence_metadata.h" -#include "synutils/tree.h" -#include "synutils/srl_sentence.h" -#include "synutils/tsuruoka_maxent.h" +#include "tree.h" +#include "srl_sentence.h" +#include "tsuruoka_maxent.h" #include "hash.h" -#include "synutils/argument_reorder_model.h" +#include "argument_reorder_model.h" #include #include diff --git a/utils/Makefile.am b/utils/Makefile.am index 727fa8a5..0bd21b2b 100644 --- a/utils/Makefile.am +++ b/utils/Makefile.am @@ -1,4 +1,4 @@ -bin_PROGRAMS = reconstruct_weights atools +bin_PROGRAMS = reconstruct_weights atools const_reorder_model_trainer noinst_PROGRAMS = \ ts \ @@ -20,7 +20,9 @@ noinst_LIBRARIES = libutils.a libutils_a_SOURCES = \ test_data \ alias_sampler.h \ + alignment.h \ alignment_io.h \ + argument_reorder_model.h \ array2d.h \ b64tools.h \ batched_append.h \ @@ -38,13 +40,20 @@ libutils_a_SOURCES = \ have_64_bits.h \ indices_after.h \ kernel_string_subseq.h \ + lbfgs.h \ + lbfgs.cpp \ logval.h \ m.h \ + mathvec.h \ + maxent.h \ + maxent.cpp \ murmur_hash3.h \ murmur_hash3.cc \ named_enum.h \ null_deleter.h \ null_traits.h \ + owlqn.cpp \ + sgd.cpp \ perfect_hash.h \ prob.h \ sampler.h \ @@ -52,12 +61,15 @@ libutils_a_SOURCES = \ show.h \ small_vector.h \ sparse_vector.h \ + srl_sentence.h \ star.h \ static_utoa.h \ stringlib.h \ string_piece.hh \ tdict.h \ timing_stats.h \ + tree.h \ + tsuruoka_maxent.h \ utoa.h \ value_array.h \ verbose.h \ @@ -79,6 +91,7 @@ libutils_a_SOURCES = \ filelib.cc \ stringlib.cc \ string_piece.cc \ + synutils.h \ sparse_vector.cc \ timing_stats.cc \ verbose.cc \ @@ -93,6 +106,8 @@ reconstruct_weights_LDADD = libutils.a atools_SOURCES = atools.cc atools_LDADD = libutils.a atools_LDFLAGS = $(STATIC_FLAGS) +const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc +const_reorder_model_trainer_LDADD = libutils.a phmt_SOURCES = phmt.cc phmt_LDADD = libutils.a $(BOOST_UNIT_TEST_FRAMEWORK_LDFLAGS) $(BOOST_UNIT_TEST_FRAMEWORK_LIBS) @@ -117,4 +132,3 @@ stringlib_test_LDADD = libutils.a $(BOOST_UNIT_TEST_FRAMEWORK_LDFLAGS) $(BOOST_U # do NOT NOT NOT add any other -I includes NO NO NO NO NO ###### AM_CPPFLAGS = -DBOOST_TEST_DYN_LINK -W -Wall -I. -I$(top_srcdir) -DTEST_DATA=\"$(top_srcdir)/utils/test_data\" ################################################################ - diff --git a/utils/alignment.h b/utils/alignment.h new file mode 100644 index 00000000..c0648aab --- /dev/null +++ b/utils/alignment.h @@ -0,0 +1,202 @@ +/* + * alignment.h + * + * Created on: May 23, 2013 + * Author: lijunhui + */ + +#ifndef ALIGNMENT_H_ +#define ALIGNMENT_H_ + +#include +#include +#include +#include + +#include "stringlib.h" + +using namespace std; + +/* + * Note: + * m_vec_s_align.size() may not be equal to the length of source side + *sentence + * due to the last words may not be aligned + * + */ +struct SAlignment { + typedef vector SingleAlign; + SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } + ~SAlignment() {} + + bool fnIsAligned(int i, bool s) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + if ((*palign)[i].size() == 0) return false; + return true; + } + + /* + * return true if [b, e] is aligned phrases on source side (if s==true) or on + * the target side (if s==false); + * return false, otherwise. + */ + bool fnIsAlignedPhrase(int b, int e, bool s, int* pob, int* poe) const { + int ob, oe; //[b, e] on the other side + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + + if (ob == -1) { + if (pob != NULL) (*pob) = -1; + if (poe != NULL) (*poe) = -1; + return false; // no aligned word among [b, e] + } + if (pob != NULL) (*pob) = ob; + if (poe != NULL) (*poe) = oe; + + int bb, be; //[b, e] back given [ob, oe] on the other side + if (s == true) + fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); + else + fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); + + if (bb < b || be > e) return false; + return true; + } + + bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { + const vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + + if ((*palign).size() <= e || (*palign)[b].size() == 0 || + (*palign)[e].size() == 0) + return false; + + return fnIsAlignedPhrase(b, e, s, pob, poe); + } + + void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + } + + /* + * look the translation of source[b, e] is continuous or not + * 1) return "Unaligned": if the source[b, e] is translated silently; + * 2) return "Con't": if none of target words in target[.., ..] is exclusively + * aligned to any word outside source[b, e] + * 3) return "Discon't": otherwise; + */ + string fnIsContinuous(int b, int e) const { + int ob, oe; + fnGetLeftRightMost(b, e, true, ob, oe); + if (ob == -1) return "Unaligned"; + + for (int i = ob; i <= oe; i++) { + if (!fnIsAligned(i, false)) continue; + const SingleAlign& a = m_vec_t_align[i]; + int j; + for (j = 0; j < a.size(); j++) + if (a[j] >= b && a[j] <= e) break; + if (j == a.size()) return "Discon't"; + } + return "Con't"; + } + + const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { + if (s == true) { + if (i >= m_vec_s_align.size()) return NULL; + return &(m_vec_s_align[i]); + } else { + if (i >= m_vec_t_align.size()) return NULL; + return &(m_vec_t_align[i]); + } + } + + private: + void fnGetLeftRightMost(int b, int e, const vector& align, + int& ob, int& oe) const { + ob = oe = -1; + for (int i = b; i <= e && i < align.size(); i++) { + if (align[i].size() > 0) { + if (align[i][0] < ob || ob == -1) ob = align[i][0]; + if (oe < align[i][align[i].size() - 1]) + oe = align[i][align[i].size() - 1]; + } + } + } + void fnInitializeAlignment(const char* pszAlign) { + m_vec_s_align.clear(); + m_vec_t_align.clear(); + + vector terms = SplitOnWhitespace(string(pszAlign)); + int si, ti; + for (size_t i = 0; i < terms.size(); i++) { + sscanf(terms[i].c_str(), "%d-%d", &si, &ti); + + while (m_vec_s_align.size() <= si) { + SingleAlign sa; + m_vec_s_align.push_back(sa); + } + while (m_vec_t_align.size() <= ti) { + SingleAlign sa; + m_vec_t_align.push_back(sa); + } + + m_vec_s_align[si].push_back(ti); + m_vec_t_align[ti].push_back(si); + } + + // sort + for (size_t i = 0; i < m_vec_s_align.size(); i++) { + std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); + } + for (size_t i = 0; i < m_vec_t_align.size(); i++) { + std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); + } + } + + private: + vector m_vec_s_align; // source side words' alignment + vector m_vec_t_align; // target side words' alignment +}; + +struct SAlignmentReader { + SAlignmentReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SAlignmentReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + SAlignment* fnReadNextAlignment() { + if (feof(m_fpIn) == true) return NULL; + char* pszLine = new char[100001]; + pszLine[0] = '\0'; + fgets(pszLine, 10001, m_fpIn); + int iLen = strlen(pszLine); + if (iLen == 0) return NULL; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + SAlignment* pAlign = new SAlignment(pszLine); + delete[] pszLine; + return pAlign; + } + + private: + FILE* m_fpIn; +}; + +#endif /* ALIGNMENT_H_ */ diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc new file mode 100644 index 00000000..58886251 --- /dev/null +++ b/utils/argument_reorder_model.cc @@ -0,0 +1,323 @@ +/* + * argument_reorder_model.cc + * + * Created on: Dec 15, 2013 + * Author: lijunhui + */ + +#include +#include + +#include "argument_reorder_model.h" +#include "synutils.h" +#include "tsuruoka_maxent.h" + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + SFReader* pFReader = new STxtFileReader(pszFName); + char* pszLine = new char[100001]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE* fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + delete[] pszLine; +} + +struct SArgumentReorderTrainer { + SArgumentReorderTrainer( + const char* pszSRLFname, // source-side srl tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelFname, // classifier model file name + int iCutoff) { + fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszTopPredicateFname, + pszInstanceFname); + + string strInstanceFname, strModelFname; + strInstanceFname = string(pszInstanceFname) + string(".left"); + strModelFname = string(pszModelFname) + string(".left"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + strInstanceFname = string(pszInstanceFname) + string(".right"); + strModelFname = string(pszModelFname) + string(".right"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + } + + ~SArgumentReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + void fnGenerateInstanceFiles( + const char* pszSRLFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // top predicate file name (we only + // consider predicates with 100+ + // occurrences + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + Map* pMapPredicate; + if (pszTopPredicateFname != NULL) + pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); + else + pMapPredicate = NULL; + + char* pszLine = new char[50001]; + + FILE* fpLeftOut, *fpRightOut; + sprintf(pszLine, "%s.left", pszInstanceFname); + fpLeftOut = fopen(pszLine, "w"); + sprintf(pszLine, "%s.right", pszInstanceFname); + fpRightOut = fopen(pszLine, "w"); + + // read sentence by sentence + SAlignment* pAlign; + SSrlSentence* pSRL; + SParsedTree* pTree; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pSRL = pSRLReader->fnReadNextSrlSentence(); + assert(pSRL != NULL); + pTree = pSRL->m_pTree; + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + // vecTPOSTerms.size() == 0, given the case when an english sentence fails + // parsing + + if (pTree != NULL) { + for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { + SPredicate* pPred = pSRL->m_vecPred[i]; + if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition] + ->m_ptParent->m_pszTerm, + "VA") == 0) + continue; + string strPred = + string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); + if (pMapPredicate != NULL) { + Map::iterator iter_map = pMapPredicate->find(strPred); + if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) + continue; + } + + SPredicateItem* pPredItem = new SPredicateItem(pTree, pPred); + + vector vecStrBlock; + for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { + SSRLItem* pItem1 = pPredItem->vec_items_[j]; + vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome( + pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, + pAlign)); + } + + vector vecStrLeftReorderType; + vector vecStrRightReorderType; + SArgumentReorderModel::fnGetReorderType( + pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); + for (int j = 1; j < pPredItem->vec_items_.size(); j++) { + string strLeftOutcome, strRightOutcome; + strLeftOutcome = vecStrLeftReorderType[j - 1]; + strRightOutcome = vecStrRightReorderType[j - 1]; + ostringstream ostr; + SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, + vecStrBlock[j - 1], + vecStrBlock[j], ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), + // strOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str()); + fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), + strRightOutcome.c_str()); + } + } + } + delete pSRL; + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + delete[] pszLine; + + fclose(fpLeftOut); + fclose(fpRightOut); + + delete pAlignReader; + delete pSRLReader; + delete pTxtSReader; + delete pTxtTReader; + } + + Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { + if (pszTopPredicateFname == NULL) return NULL; + + Map* pMapPredicate = new Map(); + STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); + char* pszLine = new char[50001]; + int iNumCount = 0; + while (pReader->fnReadNextLine(pszLine, NULL)) { + if (pszLine[0] == '#') continue; + char* p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + int iCount = atoi(p); + if (iCount < 100) break; + (*pMapPredicate)[string(pszLine)] = iNumCount++; + } + delete pReader; + delete[] pszLine; + return pMapPredicate; + } +}; + +namespace po = boost::program_options; + +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); + out << '"'; + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); + } + out << '"\n'; +} +inline string str(char const* name, po::variables_map const& conf) { + return conf[name].as(); +} + +//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file +///scratch0/mt_exp/gale-align/gale-align.nw.al --source_file +///scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file +///scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file +///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix +///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 +//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file +///scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file +///scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file +///scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file +///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix +///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options()("srl_file", po::value(), "srl file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")("help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("srl_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + SArgumentReorderTrainer* pTrainer = new SArgumentReorderTrainer( + str("srl_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), NULL, + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["feature_cutoff"].as()); + delete pTrainer; + + return 1; +} diff --git a/utils/argument_reorder_model.h b/utils/argument_reorder_model.h new file mode 100644 index 00000000..062b8841 --- /dev/null +++ b/utils/argument_reorder_model.h @@ -0,0 +1,239 @@ +/* + * argument_reorder_model.h + * + * Created on: Dec 15, 2013 + * Author: lijunhui + */ + +#ifndef ARGUMENT_REORDER_MODEL_H_ +#define ARGUMENT_REORDER_MODEL_H_ + +#include "alignment.h" +#include "tree.h" +#include "srl_sentence.h" + +// an argument item or a predicate item (the verb itself) +struct SSRLItem { + SSRLItem(const STreeItem *tree_item, string role) + : tree_item_(tree_item), role_(role) {} + ~SSRLItem() {} + const STreeItem *tree_item_; + const string role_; +}; + +struct SPredicateItem { + SPredicateItem(const SParsedTree *tree, const SPredicate *pred) + : pred_(pred) { + vec_items_.reserve(pred->m_vecArgt.size() + 1); + for (int i = 0; i < pred->m_vecArgt.size(); i++) { + vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, + string(pred->m_vecArgt[i]->m_pszRole))); + } + vec_items_.push_back(new SSRLItem( + tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); + sort(vec_items_.begin(), vec_items_.end(), SortFunction); + + begin_ = vec_items_[0]->tree_item_->m_iBegin; + end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; + } + + ~SPredicateItem() { vec_items_.clear(); } + + static bool SortFunction(SSRLItem *i, SSRLItem *j) { + return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); + } + + vector vec_items_; + int begin_; + int end_; + const SPredicate *pred_; +}; + +struct SArgumentReorderModel { + public: + static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { + return pAlign->fnIsContinuous(iBegin, iEnd); + } + static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, + vector &vecStrLeftReorder, + vector &vecStrRightReorder) { + vector vecLeft, vecRight; + for (int i = 0; i < pPredItem->vec_items_.size(); i++) { + const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vecStrLeftReorder.clear(); + vecStrRightReorder.clear(); + for (int i = 1; i < vecLeftPosition.size(); i++) { + string strOutcome; + fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); + vecStrLeftReorder.push_back(strOutcome); + fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); + vecStrRightReorder.push_back(strOutcome); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + * + * f1: (left_role, right_role, predicate_term) + * f2: (left_role, right_role, predicate_term, other_right_role) + * f3: (left_role, right_role, predicate_term, other_left_role) + * f4: (left_role, right_role, left_head_pos) + * f5: (left_role, right_role, left_head_word) + * f6: (left_role, right_role, left_syntactic_label) + * f7: (left_role, right_role, right_head_pos) + * f8: (left_role, right_role, right_head_word) + * f8: (left_role, right_role, right_syntactic_label) + * f8: (left_role, right_role, left_chunk_status) + * f9: (left_role, right_role, right_chunk_status) + * f10: (left_role, right_role, left_chunk_status) + * f11: (left_role, right_role, right_chunk_status) + * f12: (left_label, parent_label) + * f13: (right_label, parent_label) + */ + static void fnGenerateFeature(const SParsedTree *pTree, + const SPredicate *pPred, + const SPredicateItem *pPredItem, int iPos, + const string &strBlock1, + const string &strBlock2, ostringstream &ostr) { + SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; + SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; + const STreeItem *pCon1 = pSRLItem1->tree_item_; + const STreeItem *pCon2 = pSRLItem2->tree_item_; + + string left_role = pSRLItem1->role_; + string right_role = pSRLItem2->role_; + + string predicate_term = + pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) + vec_other_right_sibling.push_back( + string(pPredItem->vec_items_[i]->role_)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_right_sibling.push_back( + string(pPredItem->vec_items_[i]->role_)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; + ostr << "f1=" << left_role << "_" << right_role; + + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) { + ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_right_sibling[i]; + ostr << " f2=" << left_role << "_" << right_role << "_" + << vec_other_right_sibling[i]; + } + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) { + ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_left_sibling[i]; + ostr << " f3=" << left_role << "_" << right_role << "_" + << vec_other_left_sibling[i]; + } + // f4 + ostr << " f4=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; + // f6 + ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f7 + ostr << " f7=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f8 + ostr << " f8=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; + // f9 + ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f10 + ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; + // f11 + ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; + // f12 + ostr << " f12=" << left_role << "_" << predicate_term; + ostr << " f12=" << left_role; + // f13 + ostr << " f13=" << right_role << "_" << predicate_term; + ostr << " f13=" << right_role; + } + + private: + static void fnGetOutcome(int i1, int i2, string &strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = string("DM"); + else + strOutcome = string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = string("DS"); + else + strOutcome = string("S"); + } + } + + static void fnGetRelativePosition(const vector &vecLeft, + vector &vecPosition) { + vecPosition.clear(); + + vector vec; + for (int i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (int i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (int j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } +}; + +#endif /* ARGUMENT_REORDER_MODEL_H_ */ diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc new file mode 100644 index 00000000..042c751b --- /dev/null +++ b/utils/constituent_reorder_model.cc @@ -0,0 +1,846 @@ +/* + * constituent_reorder_model.cc + * + * Created on: Jul 10, 2013 + * Author: junhuili + */ + +#include + +#include "alignment.h" +#include "tree.h" +#include "synutils.h" +#include "tsuruoka_maxent.h" + +#include + +using namespace std; + +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; + +namespace po = boost::program_options; + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + SFReader* pFReader = new STxtFileReader(pszFName); + char* pszLine = new char[100001]; + int iLen; + Map hashPredicate; + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + if (iter == hashPredicate.end()) { + hashPredicate[vecTerms[i]] = 1; + + } else { + iter->second++; + } + } + } + delete pFReader; + + pFReader = new STxtFileReader(pszFName); + FILE* fpOut = fopen(pszNewFName, "w"); + while (pFReader->fnReadNextLine(pszLine, &iLen)) { + if (iLen == 0) continue; + + vector vecTerms; + SplitOnWhitespace(string(pszLine), &vecTerms); + ostringstream ostr; + for (size_t i = 0; i < vecTerms.size() - 1; i++) { + Iterator iter = hashPredicate.find(vecTerms[i]); + assert(iter != hashPredicate.end()); + if (iter->second >= iCutoff) { + ostr << vecTerms[i] << " "; + } + } + if (ostr.str().length() > 0) { + ostr << vecTerms[vecTerms.size() - 1]; + fprintf(fpOut, "%s\n", ostr.str().c_str()); + } + } + fclose(fpOut); + delete pFReader; + + delete[] pszLine; +} + +struct SConstReorderTrainer { + SConstReorderTrainer( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold + const char* pszOption // other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszInstanceFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + string strModelLeftFname = string(pszModelPrefix) + string(".left"); + string strModelRightFname = string(pszModelPrefix) + string(".right"); + + fprintf(stdout, "...Training the left ordering model\n"); + fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), + iCutoff); + fprintf(stdout, "...Training the right ordering model\n"); + fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), + iCutoff); + } + ~SConstReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); +pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); +delete pZhangleMaxent;*/ + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + inline bool fnIsVerbPOS(const char* pszTerm) { + if (strcmp(pszTerm, "VV") == 0 || strcmp(pszTerm, "VA") == 0 || + strcmp(pszTerm, "VC") == 0 || strcmp(pszTerm, "VE") == 0) + return true; + return false; + } + + inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, + const SAlignment* pAlign, string& strOutcome) { + if (iL1 == -1 && iL2 == -1) + strOutcome = "BU"; // 1. both are untranslated + else if (iL1 == -1) + strOutcome = "1U"; // 2. XP1 is untranslated + else if (iL2 == -1) + strOutcome = "2U"; // 3. XP2 is untranslated + else if (iL1 == iL2 && iR2 == iR2) + strOutcome = "SS"; // 4. Have same scope + else if (iL1 <= iL2 && iR1 >= iR2) + strOutcome = "1C2"; // 5. XP1's translation covers XP2's + else if (iL1 >= iL2 && iR1 <= iR2) + strOutcome = "2C1"; // 6. XP2's translation covers XP1's + else if (iR1 < iL2) { + int i = iR1 + 1; + /*while (i < iL2) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL2) + strOutcome = "M"; // 7. Monotone + else + strOutcome = "DM"; // 8. Discontinuous monotone + } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) + strOutcome = "OM"; // 9. Overlap monotone + else if (iR2 < iL1) { + int i = iR2 + 1; + /*while (i < iL1) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL1) + strOutcome = "S"; // 10. Swap + else + strOutcome = "DS"; // 11. Discontinuous swap + } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) + strOutcome = "OS"; // 12. Overlap swap + else + assert(false); + } + + inline void fnGetOutcome(int i1, int i2, string& strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = string("DM"); + else + strOutcome = string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = string("DS"); + else + strOutcome = string("S"); + } + } + + inline void fnGetRelativePosition(const vector& vecLeft, + vector& vecPosition) { + vecPosition.clear(); + + vector vec; + for (size_t i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (size_t i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (size_t j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + int iPos, const vector& vecChunkStatus, + const vector& vecPosition, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + STreeItem* pCon1, *pCon2; + pCon1 = pParent->m_vecChildren[iPos - 1]; + pCon2 = pParent->m_vecChildren[iPos]; + + fnGetOutcome(vecPosition[iPos - 1], vecPosition[iPos], strOutcome); + + string left_label = string(pCon1->m_pszTerm); + string right_label = string(pCon2->m_pszTerm); + string parent_label = string(pParent->m_pszTerm); + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) + vec_other_right_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_left_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) + ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_right_sibling[i]; + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) + ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_left_sibling[i]; + // f4 + ostr << " f4=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon1->m_iHeadWord]; + // f6 + ostr << " f6=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f7 + ostr << " f7=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon2->m_iHeadWord]; + // f8 + ostr << " f8=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos - 1]; + // f9 + ostr << " f9=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos]; + // f10 + ostr << " f10=" << left_label << "_" << parent_label; + // f11 + ostr << " f11=" << right_label << "_" << parent_label; + } + + /* + * Source side (11 features): + * f1: the categories of XP1 and XP2 (f1_1, f1_2) + * f2: the head words of XP1 and XP2 (f2_1, f2_2) + * f3: the first and last word of XP1 (f3_f, f3_l) + * f4: the first and last word of XP2 (f4_f, f4_l) + * f5: is XP1 or XP2 the head node (f5_1, f5_2) + * f6: the category of the common parent + * Target side (6 features): + * f7: the first and the last word of XP1's translation (f7_f, f7_l) + * f8: the first and the last word of XP2's translation (f8_f, f8_l) + * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + const STreeItem* pCon1, const STreeItem* pCon2, + const SAlignment* pAlign, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + + int iLeft1, iRight1, iLeft2, iRight2; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, + iRight2); + + fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); + + // generate features + // f1 + ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; + // f2 + ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" + << vecSTerms[pCon2->m_iHeadWord]; + // f3 + ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] + << " f3_l=" << vecSTerms[pCon1->m_iEnd]; + // f4 + ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] + << " f4_l=" << vecSTerms[pCon2->m_iEnd]; + // f5 + if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) + ostr << " f5_1=1"; + else + ostr << " f5_1=0"; + if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) + ostr << " f5_2=1"; + else + ostr << " f5_2=0"; + // f6 + ostr << " f6=" << pParent->m_pszTerm; + + /*//f7 + if (iLeft1 != -1) { + ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << + vecTTerms[iRight1]; + } + if (iLeft2 != -1) { + ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << + vecTTerms[iRight2]; + } + + const vector* pvecTarget = + pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); + string str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_1=" << str.substr(0, str.size()-1); + } + pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); + str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_2=" << str.substr(0, str.size()-1); + } */ + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, + vector& vecFocused) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - + // pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + // do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + void fnGenerateInstanceFile( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + FILE* fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); + assert(fpLeftOut != NULL); + + FILE* fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); + assert(fpRightOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size(); i++) { + + STreeItem* pParent = vecFocused[i]; + + vector vecLeft, vecRight; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + STreeItem* pCon1 = pParent->m_vecChildren[j]; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, + iLeft1, iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vector vecChunkStatus; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + string strOutcome = + pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, + pParent->m_vecChildren[j]->m_iEnd); + vecChunkStatus.push_back(strOutcome); + } + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strLeftOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, j, vecChunkStatus, + vecLeftPosition, vecSTerms, vecTTerms, + strLeftOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strLeftOutcome.c_str()); + fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str()); + + string strRightOutcome; + fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], + strRightOutcome); + fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), + strLeftOutcome.c_str(), strRightOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpLeftOut); + fclose(fpRightOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } + + void fnGenerateInstanceFile2( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE* fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; + i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem* pParent = vecFocused[i]; + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j - 1], + pParent->m_vecChildren[j], pAlign, vecSTerms, + vecTTerms, strOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } +}; + +struct SConstContTrainer { + SConstContTrainer( + const char* pszFlattenedSynFname, // source-side flattened parse tree + // file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold + const char* pszOption // other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszInstanceFname); + // fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, + // pszOption); + fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); + } + ~SConstContTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); + pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, + 2.0); + delete pZhangleMaxent;*/ + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, + vector& vecFocused) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - + // pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + // do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + inline void fnGetOutcome(int iL1, int iR1, const SAlignment* pAlign, + string& strOutcome) { + strOutcome = pAlign->fnIsContinuous(iL1, iR1); + } + + inline string fnGetLengthType(int iLen) { + if (iLen == 1) return string("1"); + if (iLen == 2) return string("2"); + if (iLen == 3) return string("3"); + if (iLen < 6) return string("4"); + if (iLen < 11) return string("6"); + return string("11"); + } + + /* + * Source side (11 features): + * f1: the syntactic category + * f2: the syntactic category of its parent + * f3: the head word's pos + * f4: =1 if it's the head of its parent node + * or + * the head of its parent node + * f5: length type + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pCon1, + const SAlignment* pAlign, + const vector& vecSTerms, + const vector& vecTTerms, string& strOutcome, + ostringstream& ostr) { + + fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); + + // generate features + // f1 + ostr << "f1=" << pCon1->m_pszTerm; + // f2 + ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; + // f3 + ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord] + ->m_ptParent->m_pszTerm; + // f4 + if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { + ostr << " f4=1"; + } else { + ostr << " f4=" + << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild] + ->m_pszTerm; + } + // f5 + ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); + } + + void fnGenerateInstanceFile( + const char* pszFlattenedSynFname, // source-side flattened parse tree + // file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszFlattenedSynFname, true); + STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); + STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + + FILE* fpOut = fopen(pszInstanceFname, "w"); + assert(fpOut != NULL); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + char* pszLine = new char[50001]; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(pTree != NULL); + assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + vector vecSTerms; + SplitOnWhitespace(string(pszLine), &vecSTerms); + assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + vector vecTTerms; + SplitOnWhitespace(string(pszLine), &vecTTerms); + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; + i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem* pParent = vecFocused[i]; + + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, + vecSTerms, vecTTerms, strOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + } + } + + delete pAlign; + delete pTree; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + fclose(fpOut); + delete pAlignReader; + delete pParseReader; + delete pTxtSReader; + delete pTxtTReader; + delete[] pszLine; + } +}; + +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); + out << '"'; + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); + } + out << '\n'; +} +inline string str(char const* name, po::variables_map const& conf) { + return conf[name].as(); +} + +//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file +///scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file +///scratch0/mt_exp/gq-ctb/data/train.cn --target_file +///scratch0/mt_exp/gq-ctb/data/train.en --instance_file +///scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix +///scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 +//--classifier_type 1 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options()("parse_file", po::value(), + "parse file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "classifier_type", po::value()->default_value(1), + "Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for " + "SVMLight")("feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")( + "svm_option", po::value(), "Parameters for SVMLight classifier")( + "help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("parse_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + const char* pOption; + if (vm.count("svm_option")) + pOption = str("svm_option", vm).c_str(); + else + pOption = NULL; + + SConstReorderTrainer* pTrainer = new SConstReorderTrainer( + str("parse_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["classifier_type"].as(), vm["feature_cutoff"].as(), pOption); + delete pTrainer; + + return 1; +} diff --git a/utils/lbfgs.cpp b/utils/lbfgs.cpp new file mode 100644 index 00000000..bd26f048 --- /dev/null +++ b/utils/lbfgs.cpp @@ -0,0 +1,108 @@ +#include +#include +#include +#include +#include "mathvec.h" +#include "lbfgs.h" +#include "maxent.h" + +using namespace std; + +const static int M = LBFGS_M; +const static double LINE_SEARCH_ALPHA = 0.1; +const static double LINE_SEARCH_BETA = 0.5; + +// stopping criteria +int LBFGS_MAX_ITER = 300; +const static double MIN_GRAD_NORM = 0.0001; + +double ME_Model::backtracking_line_search(const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, + Vec& x, Vec& grad1) { + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + f = FunctionGradient(x.STLVec(), grad1.STLVec()); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * t * dot_product(dx, grad0)); + + return f; +} + +// +// Jorge Nocedal, "Updating Quasi-Newton Matrices With Limited Storage", +// Mathematics of Computation, Vol. 35, No. 151, pp. 773-782, 1980. +// +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]) { + int offset, bound; + if (iter <= M) { + offset = 0; + bound = iter; + } else { + offset = iter - M; + bound = M; + } + + Vec q = grad; + double alpha[M], beta[M]; + for (int i = bound - 1; i >= 0; i--) { + const int j = (i + offset) % M; + alpha[i] = z[j] * dot_product(s[j], q); + q += -alpha[i] * y[j]; + } + if (iter > 0) { + const int j = (iter - 1) % M; + const double gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + // static double gamma; + // if (gamma == 0) gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); + q *= gamma; + } + for (int i = 0; i <= bound - 1; i++) { + const int j = (i + offset) % M; + beta[i] = z[j] * dot_product(y[j], q); + q += s[j] * (alpha[i] - beta[i]); + } + + return q; +} + +vector ME_Model::perform_LBFGS(const vector& x0) { + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < LBFGS_MAX_ITER; iter++) { + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(grad, grad)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, grad, s, y, z); + + Vec x1(dim), grad1(dim); + f = backtracking_line_search(x, grad, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + x = x1; + grad = grad1; + } + + return x.STLVec(); +} diff --git a/utils/lbfgs.h b/utils/lbfgs.h new file mode 100644 index 00000000..ed5cd944 --- /dev/null +++ b/utils/lbfgs.h @@ -0,0 +1,19 @@ +#ifndef _LBFGS_H_ +#define _LBFGS_H_ + +// template +// std::vector +// perform_LBFGS(FuncGrad func_grad, const std::vector & x0); + +std::vector perform_LBFGS( + double (*func_grad)(const std::vector &, std::vector &), + const std::vector &x0); + +std::vector perform_OWLQN( + double (*func_grad)(const std::vector &, std::vector &), + const std::vector &x0, const double C); + +// const int LBFGS_M = 7; +const int LBFGS_M = 10; + +#endif diff --git a/utils/mathvec.h b/utils/mathvec.h new file mode 100644 index 00000000..f8c60e5d --- /dev/null +++ b/utils/mathvec.h @@ -0,0 +1,87 @@ +#ifndef _MATH_VECTOR_H_ +#define _MATH_VECTOR_H_ + +#include +#include +#include + +class Vec { + private: + std::vector _v; + + public: + Vec(const size_t n = 0, const double val = 0) { _v.resize(n, val); } + Vec(const std::vector& v) : _v(v) {} + const std::vector& STLVec() const { return _v; } + std::vector& STLVec() { return _v; } + size_t Size() const { return _v.size(); } + double& operator[](int i) { return _v[i]; } + const double& operator[](int i) const { return _v[i]; } + Vec& operator+=(const Vec& b) { + assert(b.Size() == _v.size()); + for (size_t i = 0; i < _v.size(); i++) { + _v[i] += b[i]; + } + return *this; + } + Vec& operator*=(const double c) { + for (size_t i = 0; i < _v.size(); i++) { + _v[i] *= c; + } + return *this; + } + void Project(const Vec& y) { + for (size_t i = 0; i < _v.size(); i++) { + // if (sign(_v[i]) != sign(y[i])) _v[i] = 0; + if (_v[i] * y[i] <= 0) _v[i] = 0; + } + } +}; + +inline double dot_product(const Vec& a, const Vec& b) { + double sum = 0; + for (size_t i = 0; i < a.Size(); i++) { + sum += a[i] * b[i]; + } + return sum; +} + +inline std::ostream& operator<<(std::ostream& s, const Vec& a) { + s << "("; + for (size_t i = 0; i < a.Size(); i++) { + if (i != 0) s << ", "; + s << a[i]; + } + s << ")"; + return s; +} + +inline const Vec operator+(const Vec& a, const Vec& b) { + Vec v(a.Size()); + assert(a.Size() == b.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] + b[i]; + } + return v; +} + +inline const Vec operator-(const Vec& a, const Vec& b) { + Vec v(a.Size()); + assert(a.Size() == b.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] - b[i]; + } + return v; +} + +inline const Vec operator*(const Vec& a, const double c) { + Vec v(a.Size()); + for (size_t i = 0; i < a.Size(); i++) { + v[i] = a[i] * c; + } + return v; +} + +inline const Vec operator*(const double c, const Vec& a) { return a * c; } + +#endif diff --git a/utils/maxent.cpp b/utils/maxent.cpp new file mode 100644 index 00000000..8d00ac1d --- /dev/null +++ b/utils/maxent.cpp @@ -0,0 +1,703 @@ +/* + * $Id: maxent.cpp,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ + */ + +#include "maxent.h" +#include +#include +#include "lbfgs.h" + +using namespace std; + +double ME_Model::FunctionGradient(const vector& x, + vector& grad) { + assert((int)_fb.Size() == x.size()); + for (size_t i = 0; i < x.size(); i++) { + _vl[i] = x[i]; + } + + double score = update_model_expectation(); + + if (_l2reg == 0) { + for (size_t i = 0; i < x.size(); i++) { + grad[i] = -(_vee[i] - _vme[i]); + } + } else { + const double c = _l2reg * 2; + for (size_t i = 0; i < x.size(); i++) { + grad[i] = -(_vee[i] - _vme[i] - c * _vl[i]); + } + } + + return -score; +} + +int ME_Model::perform_GIS(int C) { + cerr << "C = " << C << endl; + C = 1; + cerr << "performing AGIS" << endl; + vector pre_v; + double pre_logl = -999999; + for (int iter = 0; iter < 200; iter++) { + + double logl = update_model_expectation(); + fprintf(stderr, "iter = %2d C = %d f = %10.7f train_err = %7.5f", iter, + C, logl, _train_error); + if (_heldout.size() > 0) { + double hlogl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", hlogl, + _heldout_error); + } + cerr << endl; + + if (logl < pre_logl) { + C += 1; + _vl = pre_v; + iter--; + continue; + } + if (C > 1 && iter % 10 == 0) C--; + + pre_logl = logl; + pre_v = _vl; + for (int i = 0; i < _fb.Size(); i++) { + double coef = _vee[i] / _vme[i]; + _vl[i] += log(coef) / C; + } + } + cerr << endl; + + return 0; +} + +int ME_Model::perform_QUASI_NEWTON() { + const int dim = _fb.Size(); + vector x0(dim); + + for (int i = 0; i < dim; i++) { + x0[i] = _vl[i]; + } + + vector x; + if (_l1reg > 0) { + cerr << "performing OWLQN" << endl; + x = perform_OWLQN(x0, _l1reg); + } else { + cerr << "performing LBFGS" << endl; + x = perform_LBFGS(x0); + } + + for (int i = 0; i < dim; i++) { + _vl[i] = x[i]; + } + + return 0; +} + +int ME_Model::conditional_probability(const Sample& s, + std::vector& membp) const { + // int num_classes = membp.size(); + double sum = 0; + int max_label = -1; + // double maxp = 0; + + vector powv(_num_classes, 0.0); + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + powv[_fb.Feature(*k).label()] += _vl[*k]; + } + } + for (vector >::const_iterator j = s.rvfeatures.begin(); + j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + powv[_fb.Feature(*k).label()] += _vl[*k] * j->second; + } + } + + std::vector::const_iterator pmax = + max_element(powv.begin(), powv.end()); + double offset = max(0.0, *pmax - 700); // to avoid overflow + for (int label = 0; label < _num_classes; label++) { + double pow = powv[label] - offset; + double prod = exp(pow); + // cout << pow << " " << prod << ", "; + // if (_ref_modelp != NULL) prod *= _train_refpd[n][label]; + if (_ref_modelp != NULL) prod *= s.ref_pd[label]; + assert(prod != 0); + membp[label] = prod; + sum += prod; + } + for (int label = 0; label < _num_classes; label++) { + membp[label] /= sum; + if (membp[label] > membp[max_label]) max_label = label; + } + assert(max_label >= 0); + return max_label; +} + +int ME_Model::make_feature_bag(const int cutoff) { + int max_num_features = 0; + +// count the occurrences of features +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + map_type count; + if (cutoff > 0) { + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++) { + for (std::vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + count[ME_Feature(i->label, *j).body()]++; + } + for (std::vector >::const_iterator j = + i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + count[ME_Feature(i->label, j->first).body()]++; + } + } + } + + int n = 0; + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++, n++) { + max_num_features = + max(max_num_features, (int)(i->positive_features.size())); + for (std::vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + const ME_Feature feature(i->label, *j); + // if (cutoff > 0 && count[feature.body()] < cutoff) continue; + if (cutoff > 0 && count[feature.body()] <= cutoff) continue; + _fb.Put(feature); + // cout << i->label << "\t" << *j << "\t" << id << endl; + // feature2sample[id].push_back(n); + } + for (std::vector >::const_iterator j = + i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + const ME_Feature feature(i->label, j->first); + // if (cutoff > 0 && count[feature.body()] < cutoff) continue; + if (cutoff > 0 && count[feature.body()] <= cutoff) continue; + _fb.Put(feature); + } + } + count.clear(); + + // cerr << "num_classes = " << _num_classes << endl; + // cerr << "max_num_features = " << max_num_features << endl; + + init_feature2mef(); + + return max_num_features; +} + +double ME_Model::heldout_likelihood() { + double logl = 0; + int ncorrect = 0; + for (std::vector::const_iterator i = _heldout.begin(); + i != _heldout.end(); i++) { + vector membp(_num_classes); + int l = classify(*i, membp); + logl += log(membp[i->label]); + if (l == i->label) ncorrect++; + } + _heldout_error = 1 - (double)ncorrect / _heldout.size(); + + return logl /= _heldout.size(); +} + +double ME_Model::update_model_expectation() { + double logl = 0; + int ncorrect = 0; + + _vme.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) _vme[i] = 0; + + int n = 0; + for (vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++, n++) { + vector membp(_num_classes); + int max_label = conditional_probability(*i, membp); + + logl += log(membp[i->label]); + // cout << membp[*i] << " " << logl << " "; + if (max_label == i->label) ncorrect++; + + // model_expectation + for (vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()]; + } + } + for (vector >::const_iterator j = i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + _vme[*k] += membp[_fb.Feature(*k).label()] * j->second; + } + } + } + + for (int i = 0; i < _fb.Size(); i++) { + _vme[i] /= _vs.size(); + } + + _train_error = 1 - (double)ncorrect / _vs.size(); + + logl /= _vs.size(); + + if (_l2reg > 0) { + const double c = _l2reg; + for (int i = 0; i < _fb.Size(); i++) { + logl -= _vl[i] * _vl[i] * c; + } + } + + // logl /= _vs.size(); + + // fprintf(stderr, "iter =%3d logl = %10.7f train_acc = %7.5f\n", iter, + // logl, (double)ncorrect/train.size()); + // fprintf(stderr, "logl = %10.7f train_acc = %7.5f\n", logl, + // (double)ncorrect/_train.size()); + + return logl; +} + +int ME_Model::train(const vector& vms) { + _vs.clear(); + for (vector::const_iterator i = vms.begin(); i != vms.end(); i++) { + add_training_sample(*i); + } + + return train(); +} + +void ME_Model::add_training_sample(const ME_Sample& mes) { + Sample s; + s.label = _label_bag.Put(mes.label); + if (s.label > ME_Feature::MAX_LABEL_TYPES) { + cerr << "error: too many types of labels." << endl; + exit(1); + } + for (vector::const_iterator j = mes.features.begin(); + j != mes.features.end(); j++) { + s.positive_features.push_back(_featurename_bag.Put(*j)); + } + for (vector >::const_iterator j = mes.rvfeatures.begin(); + j != mes.rvfeatures.end(); j++) { + s.rvfeatures.push_back( + pair(_featurename_bag.Put(j->first), j->second)); + } + if (_ref_modelp != NULL) { + ME_Sample tmp = mes; + ; + s.ref_pd = _ref_modelp->classify(tmp); + } + // cout << s.label << "\t"; + // for (vector::const_iterator j = s.positive_features.begin(); j != + // s.positive_features.end(); j++){ + // cout << *j << " "; + // } + // cout << endl; + + _vs.push_back(s); +} + +int ME_Model::train() { + if (_l1reg > 0 && _l2reg > 0) { + cerr << "error: L1 and L2 regularizers cannot be used simultaneously." + << endl; + return 0; + } + if (_vs.size() == 0) { + cerr << "error: no training data." << endl; + return 0; + } + if (_nheldout >= (int)_vs.size()) { + cerr << "error: too much heldout data. no training data is available." + << endl; + return 0; + } + // if (_nheldout > 0) random_shuffle(_vs.begin(), _vs.end()); + + int max_label = 0; + for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); + i++) { + max_label = max(max_label, i->label); + } + _num_classes = max_label + 1; + if (_num_classes != _label_bag.Size()) { + cerr << "warning: _num_class != _label_bag.Size()" << endl; + } + + if (_ref_modelp != NULL) { + cerr << "setting reference distribution..."; + for (int i = 0; i < _ref_modelp->num_classes(); i++) { + _label_bag.Put(_ref_modelp->get_class_label(i)); + } + _num_classes = _label_bag.Size(); + for (vector::iterator i = _vs.begin(); i != _vs.end(); i++) { + set_ref_dist(*i); + } + cerr << "done" << endl; + } + + for (int i = 0; i < _nheldout; i++) { + _heldout.push_back(_vs.back()); + _vs.pop_back(); + } + + sort(_vs.begin(), _vs.end()); + + int cutoff = 0; + if (cutoff > 0) cerr << "cutoff threshold = " << cutoff << endl; + if (_l1reg > 0) cerr << "L1 regularizer = " << _l1reg << endl; + if (_l2reg > 0) cerr << "L2 regularizer = " << _l2reg << endl; + + // normalize + _l1reg /= _vs.size(); + _l2reg /= _vs.size(); + + cerr << "preparing for estimation..."; + make_feature_bag(cutoff); + // _vs.clear(); + cerr << "done" << endl; + cerr << "number of samples = " << _vs.size() << endl; + cerr << "number of features = " << _fb.Size() << endl; + + cerr << "calculating empirical expectation..."; + _vee.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) { + _vee[i] = 0; + } + for (int n = 0; n < (int)_vs.size(); n++) { + const Sample* i = &_vs[n]; + for (vector::const_iterator j = i->positive_features.begin(); + j != i->positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += 1.0; + } + } + + for (vector >::const_iterator j = i->rvfeatures.begin(); + j != i->rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + if (_fb.Feature(*k).label() == i->label) _vee[*k] += j->second; + } + } + } + for (int i = 0; i < _fb.Size(); i++) { + _vee[i] /= _vs.size(); + } + cerr << "done" << endl; + + _vl.resize(_fb.Size()); + for (int i = 0; i < _fb.Size(); i++) _vl[i] = 0.0; + + if (_optimization_method == SGD) { + perform_SGD(); + } else { + perform_QUASI_NEWTON(); + } + + int num_active = 0; + for (int i = 0; i < _fb.Size(); i++) { + if (_vl[i] != 0) num_active++; + } + cerr << "number of active features = " << num_active << endl; + + return 0; +} + +void ME_Model::get_features(list, double> >& fl) { + fl.clear(); + // for (int i = 0; i < _fb.Size(); i++) { + // ME_Feature f = _fb.Feature(i); + // fl.push_back( make_pair(make_pair(_label_bag.Str(f.label()), + // _featurename_bag.Str(f.feature())), _vl[i])); + // } + for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); + i != _featurename_bag.end(); i++) { + for (int j = 0; j < _label_bag.Size(); j++) { + string label = _label_bag.Str(j); + string history = i->first; + int id = _fb.Id(ME_Feature(j, i->second)); + if (id < 0) continue; + fl.push_back(make_pair(make_pair(label, history), _vl[id])); + } + } +} + +void ME_Model::clear() { + _vl.clear(); + _label_bag.Clear(); + _featurename_bag.Clear(); + _fb.Clear(); + _feature2mef.clear(); + _vee.clear(); + _vme.clear(); + _vs.clear(); + _heldout.clear(); +} + +bool ME_Model::load_from_file(const string& filename) { + FILE* fp = fopen(filename.c_str(), "r"); + if (!fp) { + cerr << "error: cannot open " << filename << "!" << endl; + return false; + } + + _vl.clear(); + _label_bag.Clear(); + _featurename_bag.Clear(); + _fb.Clear(); + char buf[1024]; + while (fgets(buf, 1024, fp)) { + string line(buf); + string::size_type t1 = line.find_first_of('\t'); + string::size_type t2 = line.find_last_of('\t'); + string classname = line.substr(0, t1); + string featurename = line.substr(t1 + 1, t2 - (t1 + 1)); + float lambda; + string w = line.substr(t2 + 1); + sscanf(w.c_str(), "%f", &lambda); + + int label = _label_bag.Put(classname); + int feature = _featurename_bag.Put(featurename); + _fb.Put(ME_Feature(label, feature)); + _vl.push_back(lambda); + } + + _num_classes = _label_bag.Size(); + + init_feature2mef(); + + fclose(fp); + + return true; +} + +void ME_Model::init_feature2mef() { + _feature2mef.clear(); + for (int i = 0; i < _featurename_bag.Size(); i++) { + vector vi; + for (int k = 0; k < _num_classes; k++) { + int id = _fb.Id(ME_Feature(k, i)); + if (id >= 0) vi.push_back(id); + } + _feature2mef.push_back(vi); + } +} + +bool ME_Model::load_from_array(const ME_Model_Data data[]) { + _vl.clear(); + for (int i = 0;; i++) { + if (string(data[i].label) == "///") break; + int label = _label_bag.Put(data[i].label); + int feature = _featurename_bag.Put(data[i].feature); + _fb.Put(ME_Feature(label, feature)); + _vl.push_back(data[i].weight); + } + _num_classes = _label_bag.Size(); + + init_feature2mef(); + + return true; +} + +bool ME_Model::save_to_file(const string& filename, const double th) const { + FILE* fp = fopen(filename.c_str(), "w"); + if (!fp) { + cerr << "error: cannot open " << filename << "!" << endl; + return false; + } + + // for (int i = 0; i < _fb.Size(); i++) { + // if (_vl[i] == 0) continue; // ignore zero-weight features + // ME_Feature f = _fb.Feature(i); + // fprintf(fp, "%s\t%s\t%f\n", _label_bag.Str(f.label()).c_str(), + // _featurename_bag.Str(f.feature()).c_str(), _vl[i]); + // } + for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); + i != _featurename_bag.end(); i++) { + for (int j = 0; j < _label_bag.Size(); j++) { + string label = _label_bag.Str(j); + string history = i->first; + int id = _fb.Id(ME_Feature(j, i->second)); + if (id < 0) continue; + if (_vl[id] == 0) continue; // ignore zero-weight features + if (fabs(_vl[id]) < th) continue; // cut off low-weight features + fprintf(fp, "%s\t%s\t%f\n", label.c_str(), history.c_str(), _vl[id]); + } + } + + fclose(fp); + + return true; +} + +void ME_Model::set_ref_dist(Sample& s) const { + vector v0 = s.ref_pd; + vector v(_num_classes); + for (unsigned int i = 0; i < v.size(); i++) { + v[i] = 0; + string label = get_class_label(i); + int id_ref = _ref_modelp->get_class_id(label); + if (id_ref != -1) { + v[i] = v0[id_ref]; + } + if (v[i] == 0) v[i] = 0.001; // to avoid -inf logl + } + s.ref_pd = v; +} + +int ME_Model::classify(const Sample& nbs, vector& membp) const { + // vector membp(_num_classes); + assert(_num_classes == (int)membp.size()); + conditional_probability(nbs, membp); + int max_label = 0; + double max = 0.0; + for (int i = 0; i < (int)membp.size(); i++) { + // cout << membp[i] << " "; + if (membp[i] > max) { + max_label = i; + max = membp[i]; + } + } + // cout << endl; + return max_label; +} + +vector ME_Model::classify(ME_Sample& mes) const { + Sample s; + for (vector::const_iterator j = mes.features.begin(); + j != mes.features.end(); j++) { + int id = _featurename_bag.Id(*j); + if (id >= 0) s.positive_features.push_back(id); + } + for (vector >::const_iterator j = mes.rvfeatures.begin(); + j != mes.rvfeatures.end(); j++) { + int id = _featurename_bag.Id(j->first); + if (id >= 0) { + s.rvfeatures.push_back(pair(id, j->second)); + } + } + if (_ref_modelp != NULL) { + s.ref_pd = _ref_modelp->classify(mes); + set_ref_dist(s); + } + + vector vp(_num_classes); + int label = classify(s, vp); + mes.label = get_class_label(label); + return vp; +} + +/* + * $Log: maxent.cpp,v $ + * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida + * stepp tagger, by Okanohara and Tsuruoka + * + * Revision 1.28 2006/08/21 17:30:38 tsuruoka + * use MAX_LABEL_TYPES + * + * Revision 1.27 2006/07/25 13:19:53 tsuruoka + * sort _vs[] + * + * Revision 1.26 2006/07/18 11:13:15 tsuruoka + * modify comments + * + * Revision 1.25 2006/07/18 10:02:15 tsuruoka + * remove sample2feature[] + * speed up conditional_probability() + * + * Revision 1.24 2006/07/18 05:10:51 tsuruoka + * add ref_dist + * + * Revision 1.23 2005/12/24 07:05:32 tsuruoka + * modify conditional_probability() to avoid overflow + * + * Revision 1.22 2005/12/24 07:01:25 tsuruoka + * add cutoff for real-valued features + * + * Revision 1.21 2005/12/23 10:33:02 tsuruoka + * support real-valued features + * + * Revision 1.20 2005/12/23 09:15:29 tsuruoka + * modify _train to reduce memory consumption + * + * Revision 1.19 2005/10/28 13:10:14 tsuruoka + * fix for overflow (thanks to Ming Li) + * + * Revision 1.18 2005/10/28 13:03:07 tsuruoka + * add progress_bar + * + * Revision 1.17 2005/09/12 13:51:16 tsuruoka + * Sample: list -> vector + * + * Revision 1.16 2005/09/12 13:27:10 tsuruoka + * add add_training_sample() + * + * Revision 1.15 2005/04/27 11:22:27 tsuruoka + * bugfix + * ME_Sample: list -> vector + * + * Revision 1.14 2005/04/27 10:00:42 tsuruoka + * remove tmpfb + * + * Revision 1.13 2005/04/26 14:25:53 tsuruoka + * add MiniStringBag, USE_HASH_MAP + * + * Revision 1.12 2005/02/11 10:20:08 tsuruoka + * modify cutoff + * + * Revision 1.11 2004/10/04 05:50:25 tsuruoka + * add Clear() + * + * Revision 1.10 2004/08/26 16:52:26 tsuruoka + * fix load_from_file() + * + * Revision 1.9 2004/08/09 12:27:21 tsuruoka + * change messages + * + * Revision 1.8 2004/08/04 13:55:18 tsuruoka + * modify _sample2feature + * + * Revision 1.7 2004/07/28 13:42:58 tsuruoka + * add AGIS + * + * Revision 1.6 2004/07/28 05:54:13 tsuruoka + * get_class_name() -> get_class_label() + * ME_Feature: bugfix + * + * Revision 1.5 2004/07/27 16:58:47 tsuruoka + * modify the interface of classify() + * + * Revision 1.4 2004/07/26 17:23:46 tsuruoka + * _sample2feature: list -> vector + * + * Revision 1.3 2004/07/26 15:49:23 tsuruoka + * modify ME_Feature + * + * Revision 1.2 2004/07/26 13:52:18 tsuruoka + * modify cutoff + * + * Revision 1.1 2004/07/26 13:10:55 tsuruoka + * add files + * + * Revision 1.20 2004/07/22 08:34:45 tsuruoka + * modify _sample2feature[] + * + * Revision 1.19 2004/07/21 16:33:01 tsuruoka + * remove some comments + * + */ diff --git a/utils/maxent.h b/utils/maxent.h new file mode 100644 index 00000000..88a04e25 --- /dev/null +++ b/utils/maxent.h @@ -0,0 +1,402 @@ +/* + * $Id: maxent.h,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ + */ + +#ifndef __MAXENT_H_ +#define __MAXENT_H_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include "mathvec.h" + +#define USE_HASH_MAP // if you encounter errors with hash, try commenting out + // this line. (the program will be a bit slower, though) +#ifdef USE_HASH_MAP +#include +#endif + +// +// data format for each sample for training/testing +// +struct ME_Sample { + public: + ME_Sample() : label("") {}; + ME_Sample(const std::string& l) : label(l) {}; + void set_label(const std::string& l) { label = l; } + + // to add a binary feature + void add_feature(const std::string& f) { features.push_back(f); } + + // to add a real-valued feature + void add_feature(const std::string& s, const double d) { + rvfeatures.push_back(std::pair(s, d)); + } + + public: + std::string label; + std::vector features; + std::vector > rvfeatures; + + // obsolete + void add_feature(const std::pair& f) { + rvfeatures.push_back(f); // real-valued features + } +}; + +// +// for those who want to use load_from_array() +// +typedef struct ME_Model_Data { + char* label; + char* feature; + double weight; +} ME_Model_Data; + +class ME_Model { + public: + void add_training_sample(const ME_Sample& s); + int train(); + std::vector classify(ME_Sample& s) const; + bool load_from_file(const std::string& filename); + bool save_to_file(const std::string& filename, const double th = 0) const; + int num_classes() const { return _num_classes; } + std::string get_class_label(int i) const { return _label_bag.Str(i); } + int get_class_id(const std::string& s) const { return _label_bag.Id(s); } + void get_features( + std::list, double> >& fl); + void set_heldout(const int h, const int n = 0) { + _nheldout = h; + _early_stopping_n = n; + }; + void use_l1_regularizer(const double v) { _l1reg = v; } + void use_l2_regularizer(const double v) { _l2reg = v; } + void use_SGD(int iter = 30, double eta0 = 1, double alpha = 0.85) { + _optimization_method = SGD; + SGD_ITER = iter; + SGD_ETA0 = eta0; + SGD_ALPHA = alpha; + } + bool load_from_array(const ME_Model_Data data[]); + void set_reference_model(const ME_Model& ref_model) { + _ref_modelp = &ref_model; + }; + void clear(); + + ME_Model() { + _l1reg = _l2reg = 0; + _nheldout = 0; + _early_stopping_n = 0; + _ref_modelp = NULL; + _optimization_method = LBFGS; + } + + public: + // obsolete. just for downward compatibility + int train(const std::vector& train); + + private: + enum OPTIMIZATION_METHOD { + LBFGS, + OWLQN, + SGD + } _optimization_method; + // OWLQN and SGD are available only for L1-regularization + + int SGD_ITER; + double SGD_ETA0; + double SGD_ALPHA; + + double _l1reg, _l2reg; + + struct Sample { + int label; + std::vector positive_features; + std::vector > rvfeatures; + std::vector ref_pd; // reference probability distribution + bool operator<(const Sample& x) const { + for (unsigned int i = 0; i < positive_features.size(); i++) { + if (i >= x.positive_features.size()) return false; + int v0 = positive_features[i]; + int v1 = x.positive_features[i]; + if (v0 < v1) return true; + if (v0 > v1) return false; + } + return false; + } + }; + + struct ME_Feature { + enum { + MAX_LABEL_TYPES = 255 + }; + + // ME_Feature(const int l, const int f) : _body((l << 24) + f) { + // assert(l >= 0 && l < 256); + // assert(f >= 0 && f <= 0xffffff); + // }; + // int label() const { return _body >> 24; } + // int feature() const { return _body & 0xffffff; } + ME_Feature(const int l, const int f) : _body((f << 8) + l) { + assert(l >= 0 && l <= MAX_LABEL_TYPES); + assert(f >= 0 && f <= 0xffffff); + }; + int label() const { return _body & 0xff; } + int feature() const { return _body >> 8; } + unsigned int body() const { return _body; } + + private: + unsigned int _body; + }; + + struct ME_FeatureBag { +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + map_type mef2id; + std::vector id2mef; + int Put(const ME_Feature& i) { + map_type::const_iterator j = mef2id.find(i.body()); + if (j == mef2id.end()) { + int id = id2mef.size(); + id2mef.push_back(i); + mef2id[i.body()] = id; + return id; + } + return j->second; + } + int Id(const ME_Feature& i) const { + map_type::const_iterator j = mef2id.find(i.body()); + if (j == mef2id.end()) { + return -1; + } + return j->second; + } + ME_Feature Feature(int id) const { + assert(id >= 0 && id < (int)id2mef.size()); + return id2mef[id]; + } + int Size() const { return id2mef.size(); } + void Clear() { + mef2id.clear(); + id2mef.clear(); + } + }; + + struct hashfun_str { + size_t operator()(const std::string& s) const { + assert(sizeof(int) == 4 && sizeof(char) == 1); + const int* p = reinterpret_cast(s.c_str()); + size_t v = 0; + int n = s.size() / 4; + for (int i = 0; i < n; i++, p++) { + // v ^= *p; + v ^= *p << (4 * (i % 2)); // note) 0 <= char < 128 + } + int m = s.size() % 4; + for (int i = 0; i < m; i++) { + v ^= s[4 * n + i] << (i * 8); + } + return v; + } + }; + + struct MiniStringBag { +#ifdef USE_HASH_MAP + typedef __gnu_cxx::hash_map map_type; +#else + typedef std::map map_type; +#endif + int _size; + map_type str2id; + MiniStringBag() : _size(0) {} + int Put(const std::string& i) { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) { + int id = _size; + _size++; + str2id[i] = id; + return id; + } + return j->second; + } + int Id(const std::string& i) const { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) return -1; + return j->second; + } + int Size() const { return _size; } + void Clear() { + str2id.clear(); + _size = 0; + } + map_type::const_iterator begin() const { return str2id.begin(); } + map_type::const_iterator end() const { return str2id.end(); } + }; + + struct StringBag : public MiniStringBag { + std::vector id2str; + int Put(const std::string& i) { + map_type::const_iterator j = str2id.find(i); + if (j == str2id.end()) { + int id = id2str.size(); + id2str.push_back(i); + str2id[i] = id; + return id; + } + return j->second; + } + std::string Str(const int id) const { + assert(id >= 0 && id < (int)id2str.size()); + return id2str[id]; + } + int Size() const { return id2str.size(); } + void Clear() { + str2id.clear(); + id2str.clear(); + } + }; + + std::vector _vs; // vector of training_samples + StringBag _label_bag; + MiniStringBag _featurename_bag; + std::vector _vl; // vector of lambda + ME_FeatureBag _fb; + int _num_classes; + std::vector _vee; // empirical expectation + std::vector _vme; // empirical expectation + std::vector > _feature2mef; + std::vector _heldout; + double _train_error; // current error rate on the training data + double _heldout_error; // current error rate on the heldout data + int _nheldout; + int _early_stopping_n; + std::vector _vhlogl; + const ME_Model* _ref_modelp; + + double heldout_likelihood(); + int conditional_probability(const Sample& nbs, + std::vector& membp) const; + int make_feature_bag(const int cutoff); + int classify(const Sample& nbs, std::vector& membp) const; + double update_model_expectation(); + int perform_QUASI_NEWTON(); + int perform_SGD(); + int perform_GIS(int C); + std::vector perform_LBFGS(const std::vector& x0); + std::vector perform_OWLQN(const std::vector& x0, + const double C); + double backtracking_line_search(const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, Vec& x, + Vec& grad1); + double regularized_func_grad(const double C, const Vec& x, Vec& grad); + double constrained_line_search(double C, const Vec& x0, const Vec& grad0, + const double f0, const Vec& dx, Vec& x, + Vec& grad1); + + void set_ref_dist(Sample& s) const; + void init_feature2mef(); + + double FunctionGradient(const std::vector& x, + std::vector& grad); + static double FunctionGradientWrapper(const std::vector& x, + std::vector& grad); +}; + +#endif + +/* + * $Log: maxent.h,v $ + * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida + * stepp tagger, by Okanohara and Tsuruoka + * + * Revision 1.24 2006/08/21 17:30:38 tsuruoka + * use MAX_LABEL_TYPES + * + * Revision 1.23 2006/07/25 13:19:53 tsuruoka + * sort _vs[] + * + * Revision 1.22 2006/07/18 11:13:15 tsuruoka + * modify comments + * + * Revision 1.21 2006/07/18 10:02:15 tsuruoka + * remove sample2feature[] + * speed up conditional_probability() + * + * Revision 1.20 2006/07/18 05:10:51 tsuruoka + * add ref_dist + * + * Revision 1.19 2005/12/23 10:33:02 tsuruoka + * support real-valued features + * + * Revision 1.18 2005/12/23 09:15:29 tsuruoka + * modify _train to reduce memory consumption + * + * Revision 1.17 2005/10/28 13:02:34 tsuruoka + * set_heldout(): add default value + * Feature() + * + * Revision 1.16 2005/09/12 13:51:16 tsuruoka + * Sample: list -> vector + * + * Revision 1.15 2005/09/12 13:27:10 tsuruoka + * add add_training_sample() + * + * Revision 1.14 2005/04/27 11:22:27 tsuruoka + * bugfix + * ME_Sample: list -> vector + * + * Revision 1.13 2005/04/27 10:20:19 tsuruoka + * MiniStringBag -> StringBag + * + * Revision 1.12 2005/04/27 10:00:42 tsuruoka + * remove tmpfb + * + * Revision 1.11 2005/04/26 14:25:53 tsuruoka + * add MiniStringBag, USE_HASH_MAP + * + * Revision 1.10 2004/10/04 05:50:25 tsuruoka + * add Clear() + * + * Revision 1.9 2004/08/09 12:27:21 tsuruoka + * change messages + * + * Revision 1.8 2004/08/04 13:55:19 tsuruoka + * modify _sample2feature + * + * Revision 1.7 2004/07/29 05:51:13 tsuruoka + * remove modeldata.h + * + * Revision 1.6 2004/07/28 13:42:58 tsuruoka + * add AGIS + * + * Revision 1.5 2004/07/28 05:54:14 tsuruoka + * get_class_name() -> get_class_label() + * ME_Feature: bugfix + * + * Revision 1.4 2004/07/27 16:58:47 tsuruoka + * modify the interface of classify() + * + * Revision 1.3 2004/07/26 17:23:46 tsuruoka + * _sample2feature: list -> vector + * + * Revision 1.2 2004/07/26 15:49:23 tsuruoka + * modify ME_Feature + * + * Revision 1.1 2004/07/26 13:10:55 tsuruoka + * add files + * + * Revision 1.18 2004/07/22 08:34:45 tsuruoka + * modify _sample2feature[] + * + * Revision 1.17 2004/07/21 16:33:01 tsuruoka + * remove some comments + * + */ diff --git a/utils/owlqn.cpp b/utils/owlqn.cpp new file mode 100644 index 00000000..c3a0f0da --- /dev/null +++ b/utils/owlqn.cpp @@ -0,0 +1,127 @@ +#include +#include +#include +#include +#include "mathvec.h" +#include "lbfgs.h" +#include "maxent.h" + +using namespace std; + +const static int M = LBFGS_M; +const static double LINE_SEARCH_ALPHA = 0.1; +const static double LINE_SEARCH_BETA = 0.5; + +// stopping criteria +int OWLQN_MAX_ITER = 300; +const static double MIN_GRAD_NORM = 0.0001; + +Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], + const Vec y[], const double z[]); + +inline int sign(double x) { + if (x > 0) return 1; + if (x < 0) return -1; + return 0; +}; + +static Vec pseudo_gradient(const Vec& x, const Vec& grad0, const double C) { + Vec grad = grad0; + for (size_t i = 0; i < x.Size(); i++) { + if (x[i] != 0) { + grad[i] += C * sign(x[i]); + continue; + } + const double gm = grad0[i] - C; + if (gm > 0) { + grad[i] = gm; + continue; + } + const double gp = grad0[i] + C; + if (gp < 0) { + grad[i] = gp; + continue; + } + grad[i] = 0; + } + + return grad; +} + +double ME_Model::regularized_func_grad(const double C, const Vec& x, + Vec& grad) { + double f = FunctionGradient(x.STLVec(), grad.STLVec()); + for (size_t i = 0; i < x.Size(); i++) { + f += C * fabs(x[i]); + } + + return f; +} + +double ME_Model::constrained_line_search(double C, const Vec& x0, + const Vec& grad0, const double f0, + const Vec& dx, Vec& x, Vec& grad1) { + // compute the orthant to explore + Vec orthant = x0; + for (size_t i = 0; i < orthant.Size(); i++) { + if (orthant[i] == 0) orthant[i] = -grad0[i]; + } + + double t = 1.0 / LINE_SEARCH_BETA; + + double f; + do { + t *= LINE_SEARCH_BETA; + x = x0 + t * dx; + x.Project(orthant); + // for (size_t i = 0; i < x.Size(); i++) { + // if (x0[i] != 0 && sign(x[i]) != sign(x0[i])) x[i] = 0; + // } + + f = regularized_func_grad(C, x, grad1); + // cout << "*"; + } while (f > f0 + LINE_SEARCH_ALPHA * dot_product(x - x0, grad0)); + + return f; +} + +vector ME_Model::perform_OWLQN(const vector& x0, + const double C) { + const size_t dim = x0.size(); + Vec x = x0; + + Vec grad(dim), dx(dim); + double f = regularized_func_grad(C, x, grad); + + Vec s[M], y[M]; + double z[M]; // rho + + for (int iter = 0; iter < OWLQN_MAX_ITER; iter++) { + Vec pg = pseudo_gradient(x, grad, C); + + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); + if (_nheldout > 0) { + const double heldout_logl = heldout_likelihood(); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + + if (sqrt(dot_product(pg, pg)) < MIN_GRAD_NORM) break; + + dx = -1 * approximate_Hg(iter, pg, s, y, z); + if (dot_product(dx, pg) >= 0) dx.Project(-1 * pg); + + Vec x1(dim), grad1(dim); + f = constrained_line_search(C, x, pg, f, dx, x1, grad1); + + s[iter % M] = x1 - x; + y[iter % M] = grad1 - grad; + z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); + + x = x1; + grad = grad1; + } + + return x.STLVec(); +} diff --git a/utils/sgd.cpp b/utils/sgd.cpp new file mode 100644 index 00000000..8613edca --- /dev/null +++ b/utils/sgd.cpp @@ -0,0 +1,193 @@ +#include "maxent.h" +#include +#include + +using namespace std; + +// const double SGD_ETA0 = 1; +// const double SGD_ITER = 30; +// const double SGD_ALPHA = 0.85; + +//#define FOLOS_NAIVE +//#define FOLOS_LAZY +#define SGD_CP + +inline void apply_l1_penalty(const int i, const double u, vector& _vl, + vector& q) { + double& w = _vl[i]; + const double z = w; + double& qi = q[i]; + if (w > 0) { + w = max(0.0, w - (u + qi)); + } else if (w < 0) { + w = min(0.0, w + (u - qi)); + } + qi += w - z; +} + +static double l1norm(const vector& v) { + double sum = 0; + for (size_t i = 0; i < v.size(); i++) sum += abs(v[i]); + return sum; +} + +inline void update_folos_lazy(const int iter_sample, const int k, + vector& _vl, + const vector& sum_eta, + vector& last_updated) { + const double penalty = sum_eta[iter_sample] - sum_eta[last_updated[k]]; + double& x = _vl[k]; + if (x > 0) + x = max(0.0, x - penalty); + else + x = min(0.0, x + penalty); + last_updated[k] = iter_sample; +} + +int ME_Model::perform_SGD() { + if (_l2reg > 0) { + cerr << "error: L2 regularization is currently not supported in SGD mode." + << endl; + exit(1); + } + + cerr << "performing SGD" << endl; + + const double l1param = _l1reg; + + const int d = _fb.Size(); + + vector ri(_vs.size()); + for (size_t i = 0; i < ri.size(); i++) ri[i] = i; + + vector grad(d); + int iter_sample = 0; + const double eta0 = SGD_ETA0; + + // cerr << "l1param = " << l1param << endl; + cerr << "eta0 = " << eta0 << " alpha = " << SGD_ALPHA << endl; + + double u = 0; + vector q(d, 0); + vector last_updated(d, 0); + vector sum_eta; + sum_eta.push_back(0); + + for (int iter = 0; iter < SGD_ITER; iter++) { + + random_shuffle(ri.begin(), ri.end()); + + double logl = 0; + int ncorrect = 0, ntotal = 0; + for (size_t i = 0; i < _vs.size(); i++, ntotal++, iter_sample++) { + const Sample& s = _vs[ri[i]]; + +#ifdef FOLOS_LAZY + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); + } + } +#endif + + vector membp(_num_classes); + const int max_label = conditional_probability(s, membp); + + const double eta = + eta0 * pow(SGD_ALPHA, + (double)iter_sample / _vs.size()); // exponential decay + // const double eta = eta0 / (1.0 + (double)iter_sample / + // _vs.size()); + + // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << + // endl; + u += eta * l1param; + + sum_eta.push_back(sum_eta.back() + eta * l1param); + + logl += log(membp[s.label]); + if (max_label == s.label) ncorrect++; + + // binary features + for (vector::const_iterator j = s.positive_features.begin(); + j != s.positive_features.end(); j++) { + for (vector::const_iterator k = _feature2mef[*j].begin(); + k != _feature2mef[*j].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee); + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + // real-valued features + for (vector >::const_iterator j = s.rvfeatures.begin(); + j != s.rvfeatures.end(); j++) { + for (vector::const_iterator k = _feature2mef[j->first].begin(); + k != _feature2mef[j->first].end(); k++) { + const double me = membp[_fb.Feature(*k).label()]; + const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); + const double grad = (me - ee) * j->second; + _vl[*k] -= eta * grad; +#ifdef SGD_CP + apply_l1_penalty(*k, u, _vl, q); +#endif + } + } + +#ifdef FOLOS_NAIVE + for (size_t j = 0; j < d; j++) { + double& x = _vl[j]; + if (x > 0) + x = max(0.0, x - eta * l1param); + else + x = min(0.0, x + eta * l1param); + } +#endif + } + logl /= _vs.size(); +// fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, +// (double)ncorrect / ntotal); + +#ifdef FOLOS_LAZY + if (l1param > 0) { + for (size_t j = 0; j < d; j++) + update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); + } +#endif + + double f = logl; + if (l1param > 0) { + const double l1 = + l1norm(_vl); // this is not accurate when lazy update is used + // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " + // "; + f -= l1param * l1; + int nonzero = 0; + for (int j = 0; j < d; j++) + if (_vl[j] != 0) nonzero++; + // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " + // << nonzero << endl; + } + // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, + // (double)ncorrect/ntotal); + // fprintf(stderr, "%4d obj = %f", iter+1, f); + fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, f, + 1 - (double)ncorrect / ntotal); + + if (_nheldout > 0) { + double heldout_logl = heldout_likelihood(); + // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", + // heldout_logl, 1 - _heldout_error); + fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, + _heldout_error); + } + fprintf(stderr, "\n"); + } + + return 0; +} diff --git a/utils/srl_sentence.h b/utils/srl_sentence.h new file mode 100644 index 00000000..61532fb2 --- /dev/null +++ b/utils/srl_sentence.h @@ -0,0 +1,214 @@ +/* + * srl_sentence.h + * + * Created on: May 26, 2013 + * Author: junhuili + */ + +#ifndef SRL_SENTENCE_H_ +#define SRL_SENTENCE_H_ + + +#include +#include + +#include "tree.h" +#include "stringlib.h" + +using namespace std; + +struct SArgument { + SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { + m_pszRole = new char[strlen(pszRole) + 1]; + strcpy(m_pszRole, pszRole); + m_iBegin = iBegin; + m_iEnd = iEnd; + m_fProb = fProb; + m_pTreeItem = NULL; + } + ~SArgument() { delete[] m_pszRole; } + + void fnSetTreeItem(STreeItem* pTreeItem) { + m_pTreeItem = pTreeItem; + if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { + assert(m_pTreeItem->m_iBegin == m_iBegin); + assert(m_pTreeItem->m_iEnd == m_iEnd); + } + } + + char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP + int m_iBegin; + int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] + float m_fProb; // the probability of this role, + STreeItem* m_pTreeItem; +}; + +struct SPredicate { + SPredicate(const char* pszLemma, int iPosition) { + if (pszLemma != NULL) { + m_pszLemma = new char[strlen(pszLemma) + 1]; + strcpy(m_pszLemma, pszLemma); + } else + m_pszLemma = NULL; + m_iPosition = iPosition; + } + ~SPredicate() { + if (m_pszLemma != NULL) delete[] m_pszLemma; + for (size_t i = 0; i < m_vecArgt.size(); i++) delete m_vecArgt[i]; + } + int fnAppend(const char* pszRole, int iBegin, int iEnd) { + SArgument* pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); + return fnAppend(pArgt); + } + int fnAppend(SArgument* pArgt) { + m_vecArgt.push_back(pArgt); + int iPosition = m_vecArgt.size() - 1; + return iPosition; + } + + char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same + // as the predicate itself + int m_iPosition; // the position in sentence + vector m_vecArgt; // arguments associated to the predicate +}; + +struct SSrlSentence { + SSrlSentence() { m_pTree = NULL; } + ~SSrlSentence() { + if (m_pTree != NULL) delete m_pTree; + + for (size_t i = 0; i < m_vecPred.size(); i++) delete m_vecPred[i]; + } + int fnAppend(const char* pszLemma, int iPosition) { + SPredicate* pPred = new SPredicate(pszLemma, iPosition); + return fnAppend(pPred); + } + int fnAppend(SPredicate* pPred) { + m_vecPred.push_back(pPred); + int iPosition = m_vecPred.size() - 1; + return iPosition; + } + int GetPredicateNum() { return m_vecPred.size(); } + + SParsedTree* m_pTree; + vector m_vecPred; +}; + +struct SSrlSentenceReader { + SSrlSentenceReader(const char* pszSrlFname) { + m_fpIn = fopen(pszSrlFname, "r"); + assert(m_fpIn != NULL); + } + ~SSrlSentenceReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + inline void fnReplaceAll(std::string& str, const std::string& from, + const std::string& to) { + size_t start_pos = 0; + while ((start_pos = str.find(from, start_pos)) != std::string::npos) { + str.replace(start_pos, from.length(), to); + start_pos += to.length(); // In case 'to' contains 'from', like replacing + // 'x' with 'yx' + } + } + + // TODO: here only considers flat predicate-argument structure + // i.e., no overlap among them + SSrlSentence* fnReadNextSrlSentence() { + vector > vecContent; + if (fnReadNextContent(vecContent) == false) return NULL; + + SSrlSentence* pSrlSentence = new SSrlSentence(); + int iSize = vecContent.size(); + // put together syntactic text + std::ostringstream ostr; + for (int i = 0; i < iSize; i++) { + string strSynSeg = + vecContent[i][5]; // the 5th column is the syntactic segment + size_t iPosition = strSynSeg.find_first_of('*'); + assert(iPosition != string::npos); + ostringstream ostrTmp; + ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] + << ")"; // the 2th column is POS-tag, and the 0th column is word + strSynSeg.replace(iPosition, 1, ostrTmp.str()); + fnReplaceAll(strSynSeg, "(", " ("); + ostr << strSynSeg; + } + string strSyn = ostr.str(); + pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); + pSrlSentence->m_pTree->fnSetHeadWord(); + pSrlSentence->m_pTree->fnSetSpanInfo(); + + // read predicate-argument structure + int iNumPred = vecContent[0].size() - 8; + for (int i = 0; i < iNumPred; i++) { + vector vecRole; + vector vecBegin; + vector vecEnd; + int iPred = -1; + for (int j = 0; j < iSize; j++) { + const char* p = vecContent[j][i + 8].c_str(); + const char* q; + if (p[0] == '(') { + // starting position of an argument(or predicate) + vecBegin.push_back(j); + q = strchr(p, '*'); + assert(q != NULL); + vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); + if (vecRole.back().compare("V") == 0) { + assert(iPred == -1); + iPred = vecRole.size() - 1; + } + } + if (p[strlen(p) - 1] == ')') { + // end position of an argument(or predicate) + vecEnd.push_back(j); + assert(vecBegin.size() == vecEnd.size()); + } + } + assert(iPred != -1); + SPredicate* pPred = new SPredicate( + pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, + vecBegin[iPred]); + pSrlSentence->fnAppend(pPred); + for (size_t j = 0; j < vecBegin.size(); j++) { + if (j == iPred) continue; + pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); + pPred->m_vecArgt.back()->fnSetTreeItem( + pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], + false)); + } + } + return pSrlSentence; + } + + private: + bool fnReadNextContent(vector >& vecContent) { + vecContent.clear(); + if (feof(m_fpIn) == true) return false; + char* pszLine; + pszLine = new char[100001]; + pszLine[0] = '\0'; + int iLen; + while (!feof(m_fpIn)) { + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + if (iLen == 0) break; // end of this sentence + + vector terms = SplitOnWhitespace(string(pszLine)); + assert(terms.size() > 7); + vecContent.push_back(terms); + } + delete[] pszLine; + return true; + } + + private: + FILE* m_fpIn; +}; +#endif /* SRL_SENTENCE_H_ */ diff --git a/utils/synutils.h b/utils/synutils.h new file mode 100644 index 00000000..ef7b78b7 --- /dev/null +++ b/utils/synutils.h @@ -0,0 +1,127 @@ +/* + * utility.h + * + * Created on: Jun 24, 2013 + * Author: lijunhui + */ + +#ifndef UTILITY_H_ +#define UTILITY_H_ + +#include +#include +#include +#include +#include + +#include +#include + +using namespace std; + +typedef std::unordered_map MapString2Int; +typedef std::unordered_map MapString2Float; +typedef std::unordered_map::iterator + MapString2FloatIterator; + +using namespace std; + +struct SFReader { + SFReader() {} + virtual ~SFReader() {} + + virtual bool fnReadNextLine(char* pszLine, int* piLength) = 0; + virtual bool fnReadNextLine(string& strLine) = 0; +}; + +struct STxtFileReader : public SFReader { + STxtFileReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~STxtFileReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + bool fnReadNextLine(char* pszLine, int* piLength) { + if (feof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + if (iLen == 0) return false; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char* pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) + strLine = string(pszLine); + else + strLine = string(""); + delete[] pszLine; + + return bOut; + } + + private: + FILE* m_fpIn; +}; + +struct SGZFileReader : public SFReader { + SGZFileReader(const char* pszFname) { + m_fpIn = gzopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SGZFileReader() { + if (m_fpIn != NULL) gzclose(m_fpIn); + } + + bool fnReadNextLine(char* pszLine, int* piLength) { + if (m_fpIn == NULL) exit(0); + if (gzeof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + gzgets(m_fpIn, pszLine, 10001); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + bool fnReadNextLine(string& strLine) { + char* pszLine = new char[10001]; + bool bOut = fnReadNextLine(pszLine, NULL); + if (bOut) + strLine = string(pszLine); + else + strLine = string(""); + delete[] pszLine; + + return bOut; + } + + private: + gzFile m_fpIn; +}; + +#endif /* UTILITY_H_ */ diff --git a/utils/synutils/Makefile.am b/utils/synutils/Makefile.am deleted file mode 100644 index ac76be40..00000000 --- a/utils/synutils/Makefile.am +++ /dev/null @@ -1,6 +0,0 @@ -bin_PROGRAMS = const_reorder_model_trainer - -const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc -const_reorder_model_trainer_LDADD = maxent-3.0/libtsuruoka_maxent.a - -AM_CPPFLAGS = -W -Wall -I$(top_srcdir) diff --git a/utils/synutils/alignment.h b/utils/synutils/alignment.h deleted file mode 100644 index 6abd12bc..00000000 --- a/utils/synutils/alignment.h +++ /dev/null @@ -1,202 +0,0 @@ -/* - * alignment.h - * - * Created on: May 23, 2013 - * Author: lijunhui - */ - -#ifndef ALIGNMENT_H_ -#define ALIGNMENT_H_ - -#include -#include -#include -#include - -#include "utils/stringlib.h" - -using namespace std; - -/* - * Note: - * m_vec_s_align.size() may not be equal to the length of source side - *sentence - * due to the last words may not be aligned - * - */ -struct SAlignment { - typedef vector SingleAlign; - SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } - ~SAlignment() {} - - bool fnIsAligned(int i, bool s) const { - const vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - if ((*palign)[i].size() == 0) return false; - return true; - } - - /* - * return true if [b, e] is aligned phrases on source side (if s==true) or on - * the target side (if s==false); - * return false, otherwise. - */ - bool fnIsAlignedPhrase(int b, int e, bool s, int* pob, int* poe) const { - int ob, oe; //[b, e] on the other side - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - - if (ob == -1) { - if (pob != NULL) (*pob) = -1; - if (poe != NULL) (*poe) = -1; - return false; // no aligned word among [b, e] - } - if (pob != NULL) (*pob) = ob; - if (poe != NULL) (*poe) = oe; - - int bb, be; //[b, e] back given [ob, oe] on the other side - if (s == true) - fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); - else - fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); - - if (bb < b || be > e) return false; - return true; - } - - bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { - const vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - - if ((*palign).size() <= e || (*palign)[b].size() == 0 || - (*palign)[e].size() == 0) - return false; - - return fnIsAlignedPhrase(b, e, s, pob, poe); - } - - void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - } - - /* - * look the translation of source[b, e] is continuous or not - * 1) return "Unaligned": if the source[b, e] is translated silently; - * 2) return "Con't": if none of target words in target[.., ..] is exclusively - * aligned to any word outside source[b, e] - * 3) return "Discon't": otherwise; - */ - string fnIsContinuous(int b, int e) const { - int ob, oe; - fnGetLeftRightMost(b, e, true, ob, oe); - if (ob == -1) return "Unaligned"; - - for (int i = ob; i <= oe; i++) { - if (!fnIsAligned(i, false)) continue; - const SingleAlign& a = m_vec_t_align[i]; - int j; - for (j = 0; j < a.size(); j++) - if (a[j] >= b && a[j] <= e) break; - if (j == a.size()) return "Discon't"; - } - return "Con't"; - } - - const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { - if (s == true) { - if (i >= m_vec_s_align.size()) return NULL; - return &(m_vec_s_align[i]); - } else { - if (i >= m_vec_t_align.size()) return NULL; - return &(m_vec_t_align[i]); - } - } - - private: - void fnGetLeftRightMost(int b, int e, const vector& align, - int& ob, int& oe) const { - ob = oe = -1; - for (int i = b; i <= e && i < align.size(); i++) { - if (align[i].size() > 0) { - if (align[i][0] < ob || ob == -1) ob = align[i][0]; - if (oe < align[i][align[i].size() - 1]) - oe = align[i][align[i].size() - 1]; - } - } - } - void fnInitializeAlignment(const char* pszAlign) { - m_vec_s_align.clear(); - m_vec_t_align.clear(); - - vector terms = SplitOnWhitespace(string(pszAlign)); - int si, ti; - for (size_t i = 0; i < terms.size(); i++) { - sscanf(terms[i].c_str(), "%d-%d", &si, &ti); - - while (m_vec_s_align.size() <= si) { - SingleAlign sa; - m_vec_s_align.push_back(sa); - } - while (m_vec_t_align.size() <= ti) { - SingleAlign sa; - m_vec_t_align.push_back(sa); - } - - m_vec_s_align[si].push_back(ti); - m_vec_t_align[ti].push_back(si); - } - - // sort - for (size_t i = 0; i < m_vec_s_align.size(); i++) { - std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); - } - for (size_t i = 0; i < m_vec_t_align.size(); i++) { - std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); - } - } - - private: - vector m_vec_s_align; // source side words' alignment - vector m_vec_t_align; // target side words' alignment -}; - -struct SAlignmentReader { - SAlignmentReader(const char* pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~SAlignmentReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - SAlignment* fnReadNextAlignment() { - if (feof(m_fpIn) == true) return NULL; - char* pszLine = new char[100001]; - pszLine[0] = '\0'; - fgets(pszLine, 10001, m_fpIn); - int iLen = strlen(pszLine); - if (iLen == 0) return NULL; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - SAlignment* pAlign = new SAlignment(pszLine); - delete[] pszLine; - return pAlign; - } - - private: - FILE* m_fpIn; -}; - -#endif /* ALIGNMENT_H_ */ diff --git a/utils/synutils/argument_reorder_model.cc b/utils/synutils/argument_reorder_model.cc deleted file mode 100644 index 2b6bcda1..00000000 --- a/utils/synutils/argument_reorder_model.cc +++ /dev/null @@ -1,323 +0,0 @@ -/* - * argument_reorder_model.cc - * - * Created on: Dec 15, 2013 - * Author: lijunhui - */ - -#include -#include - -#include "argument_reorder_model.h" -#include "utility.h" -#include "tsuruoka_maxent.h" - -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, - const char* pszNewFName) { - SFReader* pFReader = new STxtFileReader(pszFName); - char* pszLine = new char[100001]; - int iLen; - Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; - } - } - } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE* fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; - } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); - } - } - fclose(fpOut); - delete pFReader; - - delete[] pszLine; -} - -struct SArgumentReorderTrainer { - SArgumentReorderTrainer( - const char* pszSRLFname, // source-side srl tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszTopPredicateFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelFname, // classifier model file name - int iCutoff) { - fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszTopPredicateFname, - pszInstanceFname); - - string strInstanceFname, strModelFname; - strInstanceFname = string(pszInstanceFname) + string(".left"); - strModelFname = string(pszModelFname) + string(".left"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - strInstanceFname = string(pszInstanceFname) + string(".right"); - strModelFname = string(pszModelFname) + string(".right"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - } - - ~SArgumentReorderTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - void fnGenerateInstanceFiles( - const char* pszSRLFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszTopPredicateFname, // top predicate file name (we only - // consider predicates with 100+ - // occurrences - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - - Map* pMapPredicate; - if (pszTopPredicateFname != NULL) - pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); - else - pMapPredicate = NULL; - - char* pszLine = new char[50001]; - - FILE* fpLeftOut, *fpRightOut; - sprintf(pszLine, "%s.left", pszInstanceFname); - fpLeftOut = fopen(pszLine, "w"); - sprintf(pszLine, "%s.right", pszInstanceFname); - fpRightOut = fopen(pszLine, "w"); - - // read sentence by sentence - SAlignment* pAlign; - SSrlSentence* pSRL; - SParsedTree* pTree; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pSRL = pSRLReader->fnReadNextSrlSentence(); - assert(pSRL != NULL); - pTree = pSRL->m_pTree; - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - // vecTPOSTerms.size() == 0, given the case when an english sentence fails - // parsing - - if (pTree != NULL) { - for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { - SPredicate* pPred = pSRL->m_vecPred[i]; - if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition] - ->m_ptParent->m_pszTerm, - "VA") == 0) - continue; - string strPred = - string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); - if (pMapPredicate != NULL) { - Map::iterator iter_map = pMapPredicate->find(strPred); - if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) - continue; - } - - SPredicateItem* pPredItem = new SPredicateItem(pTree, pPred); - - vector vecStrBlock; - for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { - SSRLItem* pItem1 = pPredItem->vec_items_[j]; - vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome( - pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, - pAlign)); - } - - vector vecStrLeftReorderType; - vector vecStrRightReorderType; - SArgumentReorderModel::fnGetReorderType( - pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); - for (int j = 1; j < pPredItem->vec_items_.size(); j++) { - string strLeftOutcome, strRightOutcome; - strLeftOutcome = vecStrLeftReorderType[j - 1]; - strRightOutcome = vecStrRightReorderType[j - 1]; - ostringstream ostr; - SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, - vecStrBlock[j - 1], - vecStrBlock[j], ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strOutcome.c_str()); - // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), - // strOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str()); - fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), - strRightOutcome.c_str()); - } - } - } - delete pSRL; - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - delete[] pszLine; - - fclose(fpLeftOut); - fclose(fpRightOut); - - delete pAlignReader; - delete pSRLReader; - delete pTxtSReader; - delete pTxtTReader; - } - - Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { - if (pszTopPredicateFname == NULL) return NULL; - - Map* pMapPredicate = new Map(); - STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); - char* pszLine = new char[50001]; - int iNumCount = 0; - while (pReader->fnReadNextLine(pszLine, NULL)) { - if (pszLine[0] == '#') continue; - char* p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - int iCount = atoi(p); - if (iCount < 100) break; - (*pMapPredicate)[string(pszLine)] = iNumCount++; - } - delete pReader; - delete[] pszLine; - return pMapPredicate; - } -}; - -namespace po = boost::program_options; - -inline void print_options(std::ostream& out, - po::options_description const& opts) { - typedef std::vector > Ds; - Ds const& ds = opts.options(); - out << '"'; - for (unsigned i = 0; i < ds.size(); ++i) { - if (i) out << ' '; - out << "--" << ds[i]->long_name(); - } - out << '"\n'; -} -inline string str(char const* name, po::variables_map const& conf) { - return conf[name].as(); -} - -//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file -///scratch0/mt_exp/gale-align/gale-align.nw.al --source_file -///scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file -///scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file -///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix -///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 -//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file -///scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file -///scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file -///scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file -///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix -///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 -int main(int argc, char** argv) { - - po::options_description opts("Configuration options"); - opts.add_options()("srl_file", po::value(), "srl file path (input)")( - "align_file", po::value(), "Alignment file path (input)")( - "source_file", po::value(), "Source text file path (input)")( - "target_file", po::value(), "Target text file path (input)")( - "instance_file", po::value(), "Instance file path (output)")( - "model_prefix", po::value(), - "Model file path prefix (output): three files will be generated")( - "feature_cutoff", po::value()->default_value(100), - "Feature cutoff threshold")("help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("srl_file") || !vm.count("align_file") || - !vm.count("source_file") || !vm.count("target_file") || - !vm.count("instance_file") || !vm.count("model_prefix")) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - SArgumentReorderTrainer* pTrainer = new SArgumentReorderTrainer( - str("srl_file", vm).c_str(), str("align_file", vm).c_str(), - str("source_file", vm).c_str(), str("target_file", vm).c_str(), NULL, - str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), - vm["feature_cutoff"].as()); - delete pTrainer; - - return 1; -} diff --git a/utils/synutils/argument_reorder_model.h b/utils/synutils/argument_reorder_model.h deleted file mode 100644 index 062b8841..00000000 --- a/utils/synutils/argument_reorder_model.h +++ /dev/null @@ -1,239 +0,0 @@ -/* - * argument_reorder_model.h - * - * Created on: Dec 15, 2013 - * Author: lijunhui - */ - -#ifndef ARGUMENT_REORDER_MODEL_H_ -#define ARGUMENT_REORDER_MODEL_H_ - -#include "alignment.h" -#include "tree.h" -#include "srl_sentence.h" - -// an argument item or a predicate item (the verb itself) -struct SSRLItem { - SSRLItem(const STreeItem *tree_item, string role) - : tree_item_(tree_item), role_(role) {} - ~SSRLItem() {} - const STreeItem *tree_item_; - const string role_; -}; - -struct SPredicateItem { - SPredicateItem(const SParsedTree *tree, const SPredicate *pred) - : pred_(pred) { - vec_items_.reserve(pred->m_vecArgt.size() + 1); - for (int i = 0; i < pred->m_vecArgt.size(); i++) { - vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, - string(pred->m_vecArgt[i]->m_pszRole))); - } - vec_items_.push_back(new SSRLItem( - tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); - sort(vec_items_.begin(), vec_items_.end(), SortFunction); - - begin_ = vec_items_[0]->tree_item_->m_iBegin; - end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; - } - - ~SPredicateItem() { vec_items_.clear(); } - - static bool SortFunction(SSRLItem *i, SSRLItem *j) { - return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); - } - - vector vec_items_; - int begin_; - int end_; - const SPredicate *pred_; -}; - -struct SArgumentReorderModel { - public: - static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { - return pAlign->fnIsContinuous(iBegin, iEnd); - } - static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, - vector &vecStrLeftReorder, - vector &vecStrRightReorder) { - vector vecLeft, vecRight; - for (int i = 0; i < pPredItem->vec_items_.size(); i++) { - const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, - iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vecStrLeftReorder.clear(); - vecStrRightReorder.clear(); - for (int i = 1; i < vecLeftPosition.size(); i++) { - string strOutcome; - fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); - vecStrLeftReorder.push_back(strOutcome); - fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); - vecStrRightReorder.push_back(strOutcome); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - * - * f1: (left_role, right_role, predicate_term) - * f2: (left_role, right_role, predicate_term, other_right_role) - * f3: (left_role, right_role, predicate_term, other_left_role) - * f4: (left_role, right_role, left_head_pos) - * f5: (left_role, right_role, left_head_word) - * f6: (left_role, right_role, left_syntactic_label) - * f7: (left_role, right_role, right_head_pos) - * f8: (left_role, right_role, right_head_word) - * f8: (left_role, right_role, right_syntactic_label) - * f8: (left_role, right_role, left_chunk_status) - * f9: (left_role, right_role, right_chunk_status) - * f10: (left_role, right_role, left_chunk_status) - * f11: (left_role, right_role, right_chunk_status) - * f12: (left_label, parent_label) - * f13: (right_label, parent_label) - */ - static void fnGenerateFeature(const SParsedTree *pTree, - const SPredicate *pPred, - const SPredicateItem *pPredItem, int iPos, - const string &strBlock1, - const string &strBlock2, ostringstream &ostr) { - SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; - SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; - const STreeItem *pCon1 = pSRLItem1->tree_item_; - const STreeItem *pCon2 = pSRLItem2->tree_item_; - - string left_role = pSRLItem1->role_; - string right_role = pSRLItem2->role_; - - string predicate_term = - pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; - - vector vec_other_right_sibling; - for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) - vec_other_right_sibling.push_back( - string(pPredItem->vec_items_[i]->role_)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); - - vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_right_sibling.push_back( - string(pPredItem->vec_items_[i]->role_)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); - - // generate features - // f1 - ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; - ostr << "f1=" << left_role << "_" << right_role; - - // f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) { - ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term - << "_" << vec_other_right_sibling[i]; - ostr << " f2=" << left_role << "_" << right_role << "_" - << vec_other_right_sibling[i]; - } - // f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) { - ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term - << "_" << vec_other_left_sibling[i]; - ostr << " f3=" << left_role << "_" << right_role << "_" - << vec_other_left_sibling[i]; - } - // f4 - ostr << " f4=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - // f5 - ostr << " f5=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; - // f6 - ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - // f7 - ostr << " f7=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - // f8 - ostr << " f8=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; - // f9 - ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - // f10 - ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; - // f11 - ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; - // f12 - ostr << " f12=" << left_role << "_" << predicate_term; - ostr << " f12=" << left_role; - // f13 - ostr << " f13=" << right_role << "_" << predicate_term; - ostr << " f13=" << right_role; - } - - private: - static void fnGetOutcome(int i1, int i2, string &strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) - strOutcome = string("DM"); - else - strOutcome = string("M"); - } else { - if (i1 > i2 + 1) - strOutcome = string("DS"); - else - strOutcome = string("S"); - } - } - - static void fnGetRelativePosition(const vector &vecLeft, - vector &vecPosition) { - vecPosition.clear(); - - vector vec; - for (int i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i - 1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (int i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (int j = 0; j < vecLeft.size(); j++) { - if (j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } -}; - -#endif /* ARGUMENT_REORDER_MODEL_H_ */ diff --git a/utils/synutils/constituent_reorder_model.cc b/utils/synutils/constituent_reorder_model.cc deleted file mode 100644 index 78d61b90..00000000 --- a/utils/synutils/constituent_reorder_model.cc +++ /dev/null @@ -1,846 +0,0 @@ -/* - * constituent_reorder_model.cc - * - * Created on: Jul 10, 2013 - * Author: junhuili - */ - -#include - -#include "alignment.h" -#include "tree.h" -#include "utility.h" -#include "tsuruoka_maxent.h" - -#include - -using namespace std; - -typedef std::unordered_map Map; -typedef std::unordered_map::iterator Iterator; - -namespace po = boost::program_options; - -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, - const char* pszNewFName) { - SFReader* pFReader = new STxtFileReader(pszFName); - char* pszLine = new char[100001]; - int iLen; - Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; - } - } - } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE* fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; - } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); - } - } - fclose(fpOut); - delete pFReader; - - delete[] pszLine; -} - -struct SConstReorderTrainer { - SConstReorderTrainer( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type - int iCutoff, // feature count threshold - const char* pszOption // other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszInstanceFname); - - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - string strModelLeftFname = string(pszModelPrefix) + string(".left"); - string strModelRightFname = string(pszModelPrefix) + string(".right"); - - fprintf(stdout, "...Training the left ordering model\n"); - fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), - iCutoff); - fprintf(stdout, "...Training the right ordering model\n"); - fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), - iCutoff); - } - ~SConstReorderTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); -pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); -delete pZhangleMaxent;*/ - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - inline bool fnIsVerbPOS(const char* pszTerm) { - if (strcmp(pszTerm, "VV") == 0 || strcmp(pszTerm, "VA") == 0 || - strcmp(pszTerm, "VC") == 0 || strcmp(pszTerm, "VE") == 0) - return true; - return false; - } - - inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, - const SAlignment* pAlign, string& strOutcome) { - if (iL1 == -1 && iL2 == -1) - strOutcome = "BU"; // 1. both are untranslated - else if (iL1 == -1) - strOutcome = "1U"; // 2. XP1 is untranslated - else if (iL2 == -1) - strOutcome = "2U"; // 3. XP2 is untranslated - else if (iL1 == iL2 && iR2 == iR2) - strOutcome = "SS"; // 4. Have same scope - else if (iL1 <= iL2 && iR1 >= iR2) - strOutcome = "1C2"; // 5. XP1's translation covers XP2's - else if (iL1 >= iL2 && iR1 <= iR2) - strOutcome = "2C1"; // 6. XP2's translation covers XP1's - else if (iR1 < iL2) { - int i = iR1 + 1; - /*while (i < iL2) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL2) - strOutcome = "M"; // 7. Monotone - else - strOutcome = "DM"; // 8. Discontinuous monotone - } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) - strOutcome = "OM"; // 9. Overlap monotone - else if (iR2 < iL1) { - int i = iR2 + 1; - /*while (i < iL1) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL1) - strOutcome = "S"; // 10. Swap - else - strOutcome = "DS"; // 11. Discontinuous swap - } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) - strOutcome = "OS"; // 12. Overlap swap - else - assert(false); - } - - inline void fnGetOutcome(int i1, int i2, string& strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) - strOutcome = string("DM"); - else - strOutcome = string("M"); - } else { - if (i1 > i2 + 1) - strOutcome = string("DS"); - else - strOutcome = string("S"); - } - } - - inline void fnGetRelativePosition(const vector& vecLeft, - vector& vecPosition) { - vecPosition.clear(); - - vector vec; - for (size_t i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i - 1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (size_t i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (size_t j = 0; j < vecLeft.size(); j++) { - if (j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, - int iPos, const vector& vecChunkStatus, - const vector& vecPosition, - const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, - ostringstream& ostr) { - STreeItem* pCon1, *pCon2; - pCon1 = pParent->m_vecChildren[iPos - 1]; - pCon2 = pParent->m_vecChildren[iPos]; - - fnGetOutcome(vecPosition[iPos - 1], vecPosition[iPos], strOutcome); - - string left_label = string(pCon1->m_pszTerm); - string right_label = string(pCon2->m_pszTerm); - string parent_label = string(pParent->m_pszTerm); - - vector vec_other_right_sibling; - for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) - vec_other_right_sibling.push_back( - string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); - vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_left_sibling.push_back( - string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); - - // generate features - // f1 - ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; - // f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) - ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label - << "_" << vec_other_right_sibling[i]; - // f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) - ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label - << "_" << vec_other_left_sibling[i]; - // f4 - ostr << " f4=" << left_label << "_" << right_label << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - // f5 - ostr << " f5=" << left_label << "_" << right_label << "_" - << vecSTerms[pCon1->m_iHeadWord]; - // f6 - ostr << " f6=" << left_label << "_" << right_label << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - // f7 - ostr << " f7=" << left_label << "_" << right_label << "_" - << vecSTerms[pCon2->m_iHeadWord]; - // f8 - ostr << " f8=" << left_label << "_" << right_label << "_" - << vecChunkStatus[iPos - 1]; - // f9 - ostr << " f9=" << left_label << "_" << right_label << "_" - << vecChunkStatus[iPos]; - // f10 - ostr << " f10=" << left_label << "_" << parent_label; - // f11 - ostr << " f11=" << right_label << "_" << parent_label; - } - - /* - * Source side (11 features): - * f1: the categories of XP1 and XP2 (f1_1, f1_2) - * f2: the head words of XP1 and XP2 (f2_1, f2_2) - * f3: the first and last word of XP1 (f3_f, f3_l) - * f4: the first and last word of XP2 (f4_f, f4_l) - * f5: is XP1 or XP2 the head node (f5_1, f5_2) - * f6: the category of the common parent - * Target side (6 features): - * f7: the first and the last word of XP1's translation (f7_f, f7_l) - * f8: the first and the last word of XP2's translation (f8_f, f8_l) - * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) - */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, - const STreeItem* pCon1, const STreeItem* pCon2, - const SAlignment* pAlign, - const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, - ostringstream& ostr) { - - int iLeft1, iRight1, iLeft2, iRight2; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, - iRight1); - pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, - iRight2); - - fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); - - // generate features - // f1 - ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; - // f2 - ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" - << vecSTerms[pCon2->m_iHeadWord]; - // f3 - ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] - << " f3_l=" << vecSTerms[pCon1->m_iEnd]; - // f4 - ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] - << " f4_l=" << vecSTerms[pCon2->m_iEnd]; - // f5 - if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) - ostr << " f5_1=1"; - else - ostr << " f5_1=0"; - if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) - ostr << " f5_2=1"; - else - ostr << " f5_2=0"; - // f6 - ostr << " f6=" << pParent->m_pszTerm; - - /*//f7 - if (iLeft1 != -1) { - ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << - vecTTerms[iRight1]; - } - if (iLeft2 != -1) { - ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << - vecTTerms[iRight2]; - } - - const vector* pvecTarget = - pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); - string str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_1=" << str.substr(0, str.size()-1); - } - pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); - str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_2=" << str.substr(0, str.size()-1); - } */ - } - - void fnGetFocusedParentNodes(const SParsedTree* pTree, - vector& vecFocused) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - - // pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - // do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - void fnGenerateInstanceFile( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - FILE* fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); - assert(fpLeftOut != NULL); - - FILE* fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); - assert(fpRightOut != NULL); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - char* pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; i < vecFocused.size(); i++) { - - STreeItem* pParent = vecFocused[i]; - - vector vecLeft, vecRight; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - STreeItem* pCon1 = pParent->m_vecChildren[j]; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, - iLeft1, iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vector vecChunkStatus; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - string strOutcome = - pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, - pParent->m_vecChildren[j]->m_iEnd); - vecChunkStatus.push_back(strOutcome); - } - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strLeftOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, j, vecChunkStatus, - vecLeftPosition, vecSTerms, vecTTerms, - strLeftOutcome, ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strLeftOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str()); - - string strRightOutcome; - fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], - strRightOutcome); - fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str(), strRightOutcome.c_str()); - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - fclose(fpLeftOut); - fclose(fpRightOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; - } - - void fnGenerateInstanceFile2( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - - FILE* fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - char* pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; - i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem* pParent = vecFocused[i]; - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j - 1], - pParent->m_vecChildren[j], pAlign, vecSTerms, - vecTTerms, strOutcome, ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - fclose(fpOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; - } -}; - -struct SConstContTrainer { - SConstContTrainer( - const char* pszFlattenedSynFname, // source-side flattened parse tree - // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type - int iCutoff, // feature count threshold - const char* pszOption // other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszInstanceFname); - // fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, - // pszOption); - fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); - } - ~SConstContTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); - pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, - 2.0); - delete pZhangleMaxent;*/ - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - void fnGetFocusedParentNodes(const SParsedTree* pTree, - vector& vecFocused) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - - // pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - // do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - inline void fnGetOutcome(int iL1, int iR1, const SAlignment* pAlign, - string& strOutcome) { - strOutcome = pAlign->fnIsContinuous(iL1, iR1); - } - - inline string fnGetLengthType(int iLen) { - if (iLen == 1) return string("1"); - if (iLen == 2) return string("2"); - if (iLen == 3) return string("3"); - if (iLen < 6) return string("4"); - if (iLen < 11) return string("6"); - return string("11"); - } - - /* - * Source side (11 features): - * f1: the syntactic category - * f2: the syntactic category of its parent - * f3: the head word's pos - * f4: =1 if it's the head of its parent node - * or - * the head of its parent node - * f5: length type - */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pCon1, - const SAlignment* pAlign, - const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, - ostringstream& ostr) { - - fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); - - // generate features - // f1 - ostr << "f1=" << pCon1->m_pszTerm; - // f2 - ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; - // f3 - ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord] - ->m_ptParent->m_pszTerm; - // f4 - if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { - ostr << " f4=1"; - } else { - ostr << " f4=" - << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild] - ->m_pszTerm; - } - // f5 - ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); - } - - void fnGenerateInstanceFile( - const char* pszFlattenedSynFname, // source-side flattened parse tree - // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszFlattenedSynFname, true); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - - FILE* fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - char* pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTree != NULL); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; - i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem* pParent = vecFocused[i]; - - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, - vecSTerms, vecTTerms, strOutcome, ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - } - } - - delete pAlign; - delete pTree; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - fclose(fpOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; - } -}; - -inline void print_options(std::ostream& out, - po::options_description const& opts) { - typedef std::vector > Ds; - Ds const& ds = opts.options(); - out << '"'; - for (unsigned i = 0; i < ds.size(); ++i) { - if (i) out << ' '; - out << "--" << ds[i]->long_name(); - } - out << '\n'; -} -inline string str(char const* name, po::variables_map const& conf) { - return conf[name].as(); -} - -//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file -///scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file -///scratch0/mt_exp/gq-ctb/data/train.cn --target_file -///scratch0/mt_exp/gq-ctb/data/train.en --instance_file -///scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix -///scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 -//--classifier_type 1 -int main(int argc, char** argv) { - - po::options_description opts("Configuration options"); - opts.add_options()("parse_file", po::value(), - "parse file path (input)")( - "align_file", po::value(), "Alignment file path (input)")( - "source_file", po::value(), "Source text file path (input)")( - "target_file", po::value(), "Target text file path (input)")( - "instance_file", po::value(), "Instance file path (output)")( - "model_prefix", po::value(), - "Model file path prefix (output): three files will be generated")( - "classifier_type", po::value()->default_value(1), - "Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for " - "SVMLight")("feature_cutoff", po::value()->default_value(100), - "Feature cutoff threshold")( - "svm_option", po::value(), "Parameters for SVMLight classifier")( - "help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("parse_file") || !vm.count("align_file") || - !vm.count("source_file") || !vm.count("target_file") || - !vm.count("instance_file") || !vm.count("model_prefix")) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - const char* pOption; - if (vm.count("svm_option")) - pOption = str("svm_option", vm).c_str(); - else - pOption = NULL; - - SConstReorderTrainer* pTrainer = new SConstReorderTrainer( - str("parse_file", vm).c_str(), str("align_file", vm).c_str(), - str("source_file", vm).c_str(), str("target_file", vm).c_str(), - str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), - vm["classifier_type"].as(), vm["feature_cutoff"].as(), pOption); - delete pTrainer; - - return 1; -} diff --git a/utils/synutils/maxent-3.0/Makefile.am b/utils/synutils/maxent-3.0/Makefile.am deleted file mode 100644 index 64bb038c..00000000 --- a/utils/synutils/maxent-3.0/Makefile.am +++ /dev/null @@ -1,10 +0,0 @@ -noinst_LIBRARIES = libtsuruoka_maxent.a - -libtsuruoka_maxent_a_SOURCES = \ - lbfgs.cpp \ - maxent.cpp \ - owlqn.cpp \ - sgd.cpp - -AM_CPPFLAGS = -W -Wall - diff --git a/utils/synutils/maxent-3.0/lbfgs.cpp b/utils/synutils/maxent-3.0/lbfgs.cpp deleted file mode 100644 index bd26f048..00000000 --- a/utils/synutils/maxent-3.0/lbfgs.cpp +++ /dev/null @@ -1,108 +0,0 @@ -#include -#include -#include -#include -#include "mathvec.h" -#include "lbfgs.h" -#include "maxent.h" - -using namespace std; - -const static int M = LBFGS_M; -const static double LINE_SEARCH_ALPHA = 0.1; -const static double LINE_SEARCH_BETA = 0.5; - -// stopping criteria -int LBFGS_MAX_ITER = 300; -const static double MIN_GRAD_NORM = 0.0001; - -double ME_Model::backtracking_line_search(const Vec& x0, const Vec& grad0, - const double f0, const Vec& dx, - Vec& x, Vec& grad1) { - double t = 1.0 / LINE_SEARCH_BETA; - - double f; - do { - t *= LINE_SEARCH_BETA; - x = x0 + t * dx; - f = FunctionGradient(x.STLVec(), grad1.STLVec()); - // cout << "*"; - } while (f > f0 + LINE_SEARCH_ALPHA * t * dot_product(dx, grad0)); - - return f; -} - -// -// Jorge Nocedal, "Updating Quasi-Newton Matrices With Limited Storage", -// Mathematics of Computation, Vol. 35, No. 151, pp. 773-782, 1980. -// -Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], - const Vec y[], const double z[]) { - int offset, bound; - if (iter <= M) { - offset = 0; - bound = iter; - } else { - offset = iter - M; - bound = M; - } - - Vec q = grad; - double alpha[M], beta[M]; - for (int i = bound - 1; i >= 0; i--) { - const int j = (i + offset) % M; - alpha[i] = z[j] * dot_product(s[j], q); - q += -alpha[i] * y[j]; - } - if (iter > 0) { - const int j = (iter - 1) % M; - const double gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); - // static double gamma; - // if (gamma == 0) gamma = ((1.0 / z[j]) / dot_product(y[j], y[j])); - q *= gamma; - } - for (int i = 0; i <= bound - 1; i++) { - const int j = (i + offset) % M; - beta[i] = z[j] * dot_product(y[j], q); - q += s[j] * (alpha[i] - beta[i]); - } - - return q; -} - -vector ME_Model::perform_LBFGS(const vector& x0) { - const size_t dim = x0.size(); - Vec x = x0; - - Vec grad(dim), dx(dim); - double f = FunctionGradient(x.STLVec(), grad.STLVec()); - - Vec s[M], y[M]; - double z[M]; // rho - - for (int iter = 0; iter < LBFGS_MAX_ITER; iter++) { - - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); - if (_nheldout > 0) { - const double heldout_logl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, - _heldout_error); - } - fprintf(stderr, "\n"); - - if (sqrt(dot_product(grad, grad)) < MIN_GRAD_NORM) break; - - dx = -1 * approximate_Hg(iter, grad, s, y, z); - - Vec x1(dim), grad1(dim); - f = backtracking_line_search(x, grad, f, dx, x1, grad1); - - s[iter % M] = x1 - x; - y[iter % M] = grad1 - grad; - z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); - x = x1; - grad = grad1; - } - - return x.STLVec(); -} diff --git a/utils/synutils/maxent-3.0/lbfgs.h b/utils/synutils/maxent-3.0/lbfgs.h deleted file mode 100644 index ed5cd944..00000000 --- a/utils/synutils/maxent-3.0/lbfgs.h +++ /dev/null @@ -1,19 +0,0 @@ -#ifndef _LBFGS_H_ -#define _LBFGS_H_ - -// template -// std::vector -// perform_LBFGS(FuncGrad func_grad, const std::vector & x0); - -std::vector perform_LBFGS( - double (*func_grad)(const std::vector &, std::vector &), - const std::vector &x0); - -std::vector perform_OWLQN( - double (*func_grad)(const std::vector &, std::vector &), - const std::vector &x0, const double C); - -// const int LBFGS_M = 7; -const int LBFGS_M = 10; - -#endif diff --git a/utils/synutils/maxent-3.0/mathvec.h b/utils/synutils/maxent-3.0/mathvec.h deleted file mode 100644 index f8c60e5d..00000000 --- a/utils/synutils/maxent-3.0/mathvec.h +++ /dev/null @@ -1,87 +0,0 @@ -#ifndef _MATH_VECTOR_H_ -#define _MATH_VECTOR_H_ - -#include -#include -#include - -class Vec { - private: - std::vector _v; - - public: - Vec(const size_t n = 0, const double val = 0) { _v.resize(n, val); } - Vec(const std::vector& v) : _v(v) {} - const std::vector& STLVec() const { return _v; } - std::vector& STLVec() { return _v; } - size_t Size() const { return _v.size(); } - double& operator[](int i) { return _v[i]; } - const double& operator[](int i) const { return _v[i]; } - Vec& operator+=(const Vec& b) { - assert(b.Size() == _v.size()); - for (size_t i = 0; i < _v.size(); i++) { - _v[i] += b[i]; - } - return *this; - } - Vec& operator*=(const double c) { - for (size_t i = 0; i < _v.size(); i++) { - _v[i] *= c; - } - return *this; - } - void Project(const Vec& y) { - for (size_t i = 0; i < _v.size(); i++) { - // if (sign(_v[i]) != sign(y[i])) _v[i] = 0; - if (_v[i] * y[i] <= 0) _v[i] = 0; - } - } -}; - -inline double dot_product(const Vec& a, const Vec& b) { - double sum = 0; - for (size_t i = 0; i < a.Size(); i++) { - sum += a[i] * b[i]; - } - return sum; -} - -inline std::ostream& operator<<(std::ostream& s, const Vec& a) { - s << "("; - for (size_t i = 0; i < a.Size(); i++) { - if (i != 0) s << ", "; - s << a[i]; - } - s << ")"; - return s; -} - -inline const Vec operator+(const Vec& a, const Vec& b) { - Vec v(a.Size()); - assert(a.Size() == b.Size()); - for (size_t i = 0; i < a.Size(); i++) { - v[i] = a[i] + b[i]; - } - return v; -} - -inline const Vec operator-(const Vec& a, const Vec& b) { - Vec v(a.Size()); - assert(a.Size() == b.Size()); - for (size_t i = 0; i < a.Size(); i++) { - v[i] = a[i] - b[i]; - } - return v; -} - -inline const Vec operator*(const Vec& a, const double c) { - Vec v(a.Size()); - for (size_t i = 0; i < a.Size(); i++) { - v[i] = a[i] * c; - } - return v; -} - -inline const Vec operator*(const double c, const Vec& a) { return a * c; } - -#endif diff --git a/utils/synutils/maxent-3.0/maxent.cpp b/utils/synutils/maxent-3.0/maxent.cpp deleted file mode 100644 index 8d00ac1d..00000000 --- a/utils/synutils/maxent-3.0/maxent.cpp +++ /dev/null @@ -1,703 +0,0 @@ -/* - * $Id: maxent.cpp,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ - */ - -#include "maxent.h" -#include -#include -#include "lbfgs.h" - -using namespace std; - -double ME_Model::FunctionGradient(const vector& x, - vector& grad) { - assert((int)_fb.Size() == x.size()); - for (size_t i = 0; i < x.size(); i++) { - _vl[i] = x[i]; - } - - double score = update_model_expectation(); - - if (_l2reg == 0) { - for (size_t i = 0; i < x.size(); i++) { - grad[i] = -(_vee[i] - _vme[i]); - } - } else { - const double c = _l2reg * 2; - for (size_t i = 0; i < x.size(); i++) { - grad[i] = -(_vee[i] - _vme[i] - c * _vl[i]); - } - } - - return -score; -} - -int ME_Model::perform_GIS(int C) { - cerr << "C = " << C << endl; - C = 1; - cerr << "performing AGIS" << endl; - vector pre_v; - double pre_logl = -999999; - for (int iter = 0; iter < 200; iter++) { - - double logl = update_model_expectation(); - fprintf(stderr, "iter = %2d C = %d f = %10.7f train_err = %7.5f", iter, - C, logl, _train_error); - if (_heldout.size() > 0) { - double hlogl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", hlogl, - _heldout_error); - } - cerr << endl; - - if (logl < pre_logl) { - C += 1; - _vl = pre_v; - iter--; - continue; - } - if (C > 1 && iter % 10 == 0) C--; - - pre_logl = logl; - pre_v = _vl; - for (int i = 0; i < _fb.Size(); i++) { - double coef = _vee[i] / _vme[i]; - _vl[i] += log(coef) / C; - } - } - cerr << endl; - - return 0; -} - -int ME_Model::perform_QUASI_NEWTON() { - const int dim = _fb.Size(); - vector x0(dim); - - for (int i = 0; i < dim; i++) { - x0[i] = _vl[i]; - } - - vector x; - if (_l1reg > 0) { - cerr << "performing OWLQN" << endl; - x = perform_OWLQN(x0, _l1reg); - } else { - cerr << "performing LBFGS" << endl; - x = perform_LBFGS(x0); - } - - for (int i = 0; i < dim; i++) { - _vl[i] = x[i]; - } - - return 0; -} - -int ME_Model::conditional_probability(const Sample& s, - std::vector& membp) const { - // int num_classes = membp.size(); - double sum = 0; - int max_label = -1; - // double maxp = 0; - - vector powv(_num_classes, 0.0); - for (vector::const_iterator j = s.positive_features.begin(); - j != s.positive_features.end(); j++) { - for (vector::const_iterator k = _feature2mef[*j].begin(); - k != _feature2mef[*j].end(); k++) { - powv[_fb.Feature(*k).label()] += _vl[*k]; - } - } - for (vector >::const_iterator j = s.rvfeatures.begin(); - j != s.rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); - k != _feature2mef[j->first].end(); k++) { - powv[_fb.Feature(*k).label()] += _vl[*k] * j->second; - } - } - - std::vector::const_iterator pmax = - max_element(powv.begin(), powv.end()); - double offset = max(0.0, *pmax - 700); // to avoid overflow - for (int label = 0; label < _num_classes; label++) { - double pow = powv[label] - offset; - double prod = exp(pow); - // cout << pow << " " << prod << ", "; - // if (_ref_modelp != NULL) prod *= _train_refpd[n][label]; - if (_ref_modelp != NULL) prod *= s.ref_pd[label]; - assert(prod != 0); - membp[label] = prod; - sum += prod; - } - for (int label = 0; label < _num_classes; label++) { - membp[label] /= sum; - if (membp[label] > membp[max_label]) max_label = label; - } - assert(max_label >= 0); - return max_label; -} - -int ME_Model::make_feature_bag(const int cutoff) { - int max_num_features = 0; - -// count the occurrences of features -#ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; -#else - typedef std::map map_type; -#endif - map_type count; - if (cutoff > 0) { - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); - i++) { - for (std::vector::const_iterator j = i->positive_features.begin(); - j != i->positive_features.end(); j++) { - count[ME_Feature(i->label, *j).body()]++; - } - for (std::vector >::const_iterator j = - i->rvfeatures.begin(); - j != i->rvfeatures.end(); j++) { - count[ME_Feature(i->label, j->first).body()]++; - } - } - } - - int n = 0; - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); - i++, n++) { - max_num_features = - max(max_num_features, (int)(i->positive_features.size())); - for (std::vector::const_iterator j = i->positive_features.begin(); - j != i->positive_features.end(); j++) { - const ME_Feature feature(i->label, *j); - // if (cutoff > 0 && count[feature.body()] < cutoff) continue; - if (cutoff > 0 && count[feature.body()] <= cutoff) continue; - _fb.Put(feature); - // cout << i->label << "\t" << *j << "\t" << id << endl; - // feature2sample[id].push_back(n); - } - for (std::vector >::const_iterator j = - i->rvfeatures.begin(); - j != i->rvfeatures.end(); j++) { - const ME_Feature feature(i->label, j->first); - // if (cutoff > 0 && count[feature.body()] < cutoff) continue; - if (cutoff > 0 && count[feature.body()] <= cutoff) continue; - _fb.Put(feature); - } - } - count.clear(); - - // cerr << "num_classes = " << _num_classes << endl; - // cerr << "max_num_features = " << max_num_features << endl; - - init_feature2mef(); - - return max_num_features; -} - -double ME_Model::heldout_likelihood() { - double logl = 0; - int ncorrect = 0; - for (std::vector::const_iterator i = _heldout.begin(); - i != _heldout.end(); i++) { - vector membp(_num_classes); - int l = classify(*i, membp); - logl += log(membp[i->label]); - if (l == i->label) ncorrect++; - } - _heldout_error = 1 - (double)ncorrect / _heldout.size(); - - return logl /= _heldout.size(); -} - -double ME_Model::update_model_expectation() { - double logl = 0; - int ncorrect = 0; - - _vme.resize(_fb.Size()); - for (int i = 0; i < _fb.Size(); i++) _vme[i] = 0; - - int n = 0; - for (vector::const_iterator i = _vs.begin(); i != _vs.end(); - i++, n++) { - vector membp(_num_classes); - int max_label = conditional_probability(*i, membp); - - logl += log(membp[i->label]); - // cout << membp[*i] << " " << logl << " "; - if (max_label == i->label) ncorrect++; - - // model_expectation - for (vector::const_iterator j = i->positive_features.begin(); - j != i->positive_features.end(); j++) { - for (vector::const_iterator k = _feature2mef[*j].begin(); - k != _feature2mef[*j].end(); k++) { - _vme[*k] += membp[_fb.Feature(*k).label()]; - } - } - for (vector >::const_iterator j = i->rvfeatures.begin(); - j != i->rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); - k != _feature2mef[j->first].end(); k++) { - _vme[*k] += membp[_fb.Feature(*k).label()] * j->second; - } - } - } - - for (int i = 0; i < _fb.Size(); i++) { - _vme[i] /= _vs.size(); - } - - _train_error = 1 - (double)ncorrect / _vs.size(); - - logl /= _vs.size(); - - if (_l2reg > 0) { - const double c = _l2reg; - for (int i = 0; i < _fb.Size(); i++) { - logl -= _vl[i] * _vl[i] * c; - } - } - - // logl /= _vs.size(); - - // fprintf(stderr, "iter =%3d logl = %10.7f train_acc = %7.5f\n", iter, - // logl, (double)ncorrect/train.size()); - // fprintf(stderr, "logl = %10.7f train_acc = %7.5f\n", logl, - // (double)ncorrect/_train.size()); - - return logl; -} - -int ME_Model::train(const vector& vms) { - _vs.clear(); - for (vector::const_iterator i = vms.begin(); i != vms.end(); i++) { - add_training_sample(*i); - } - - return train(); -} - -void ME_Model::add_training_sample(const ME_Sample& mes) { - Sample s; - s.label = _label_bag.Put(mes.label); - if (s.label > ME_Feature::MAX_LABEL_TYPES) { - cerr << "error: too many types of labels." << endl; - exit(1); - } - for (vector::const_iterator j = mes.features.begin(); - j != mes.features.end(); j++) { - s.positive_features.push_back(_featurename_bag.Put(*j)); - } - for (vector >::const_iterator j = mes.rvfeatures.begin(); - j != mes.rvfeatures.end(); j++) { - s.rvfeatures.push_back( - pair(_featurename_bag.Put(j->first), j->second)); - } - if (_ref_modelp != NULL) { - ME_Sample tmp = mes; - ; - s.ref_pd = _ref_modelp->classify(tmp); - } - // cout << s.label << "\t"; - // for (vector::const_iterator j = s.positive_features.begin(); j != - // s.positive_features.end(); j++){ - // cout << *j << " "; - // } - // cout << endl; - - _vs.push_back(s); -} - -int ME_Model::train() { - if (_l1reg > 0 && _l2reg > 0) { - cerr << "error: L1 and L2 regularizers cannot be used simultaneously." - << endl; - return 0; - } - if (_vs.size() == 0) { - cerr << "error: no training data." << endl; - return 0; - } - if (_nheldout >= (int)_vs.size()) { - cerr << "error: too much heldout data. no training data is available." - << endl; - return 0; - } - // if (_nheldout > 0) random_shuffle(_vs.begin(), _vs.end()); - - int max_label = 0; - for (std::vector::const_iterator i = _vs.begin(); i != _vs.end(); - i++) { - max_label = max(max_label, i->label); - } - _num_classes = max_label + 1; - if (_num_classes != _label_bag.Size()) { - cerr << "warning: _num_class != _label_bag.Size()" << endl; - } - - if (_ref_modelp != NULL) { - cerr << "setting reference distribution..."; - for (int i = 0; i < _ref_modelp->num_classes(); i++) { - _label_bag.Put(_ref_modelp->get_class_label(i)); - } - _num_classes = _label_bag.Size(); - for (vector::iterator i = _vs.begin(); i != _vs.end(); i++) { - set_ref_dist(*i); - } - cerr << "done" << endl; - } - - for (int i = 0; i < _nheldout; i++) { - _heldout.push_back(_vs.back()); - _vs.pop_back(); - } - - sort(_vs.begin(), _vs.end()); - - int cutoff = 0; - if (cutoff > 0) cerr << "cutoff threshold = " << cutoff << endl; - if (_l1reg > 0) cerr << "L1 regularizer = " << _l1reg << endl; - if (_l2reg > 0) cerr << "L2 regularizer = " << _l2reg << endl; - - // normalize - _l1reg /= _vs.size(); - _l2reg /= _vs.size(); - - cerr << "preparing for estimation..."; - make_feature_bag(cutoff); - // _vs.clear(); - cerr << "done" << endl; - cerr << "number of samples = " << _vs.size() << endl; - cerr << "number of features = " << _fb.Size() << endl; - - cerr << "calculating empirical expectation..."; - _vee.resize(_fb.Size()); - for (int i = 0; i < _fb.Size(); i++) { - _vee[i] = 0; - } - for (int n = 0; n < (int)_vs.size(); n++) { - const Sample* i = &_vs[n]; - for (vector::const_iterator j = i->positive_features.begin(); - j != i->positive_features.end(); j++) { - for (vector::const_iterator k = _feature2mef[*j].begin(); - k != _feature2mef[*j].end(); k++) { - if (_fb.Feature(*k).label() == i->label) _vee[*k] += 1.0; - } - } - - for (vector >::const_iterator j = i->rvfeatures.begin(); - j != i->rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); - k != _feature2mef[j->first].end(); k++) { - if (_fb.Feature(*k).label() == i->label) _vee[*k] += j->second; - } - } - } - for (int i = 0; i < _fb.Size(); i++) { - _vee[i] /= _vs.size(); - } - cerr << "done" << endl; - - _vl.resize(_fb.Size()); - for (int i = 0; i < _fb.Size(); i++) _vl[i] = 0.0; - - if (_optimization_method == SGD) { - perform_SGD(); - } else { - perform_QUASI_NEWTON(); - } - - int num_active = 0; - for (int i = 0; i < _fb.Size(); i++) { - if (_vl[i] != 0) num_active++; - } - cerr << "number of active features = " << num_active << endl; - - return 0; -} - -void ME_Model::get_features(list, double> >& fl) { - fl.clear(); - // for (int i = 0; i < _fb.Size(); i++) { - // ME_Feature f = _fb.Feature(i); - // fl.push_back( make_pair(make_pair(_label_bag.Str(f.label()), - // _featurename_bag.Str(f.feature())), _vl[i])); - // } - for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); - i != _featurename_bag.end(); i++) { - for (int j = 0; j < _label_bag.Size(); j++) { - string label = _label_bag.Str(j); - string history = i->first; - int id = _fb.Id(ME_Feature(j, i->second)); - if (id < 0) continue; - fl.push_back(make_pair(make_pair(label, history), _vl[id])); - } - } -} - -void ME_Model::clear() { - _vl.clear(); - _label_bag.Clear(); - _featurename_bag.Clear(); - _fb.Clear(); - _feature2mef.clear(); - _vee.clear(); - _vme.clear(); - _vs.clear(); - _heldout.clear(); -} - -bool ME_Model::load_from_file(const string& filename) { - FILE* fp = fopen(filename.c_str(), "r"); - if (!fp) { - cerr << "error: cannot open " << filename << "!" << endl; - return false; - } - - _vl.clear(); - _label_bag.Clear(); - _featurename_bag.Clear(); - _fb.Clear(); - char buf[1024]; - while (fgets(buf, 1024, fp)) { - string line(buf); - string::size_type t1 = line.find_first_of('\t'); - string::size_type t2 = line.find_last_of('\t'); - string classname = line.substr(0, t1); - string featurename = line.substr(t1 + 1, t2 - (t1 + 1)); - float lambda; - string w = line.substr(t2 + 1); - sscanf(w.c_str(), "%f", &lambda); - - int label = _label_bag.Put(classname); - int feature = _featurename_bag.Put(featurename); - _fb.Put(ME_Feature(label, feature)); - _vl.push_back(lambda); - } - - _num_classes = _label_bag.Size(); - - init_feature2mef(); - - fclose(fp); - - return true; -} - -void ME_Model::init_feature2mef() { - _feature2mef.clear(); - for (int i = 0; i < _featurename_bag.Size(); i++) { - vector vi; - for (int k = 0; k < _num_classes; k++) { - int id = _fb.Id(ME_Feature(k, i)); - if (id >= 0) vi.push_back(id); - } - _feature2mef.push_back(vi); - } -} - -bool ME_Model::load_from_array(const ME_Model_Data data[]) { - _vl.clear(); - for (int i = 0;; i++) { - if (string(data[i].label) == "///") break; - int label = _label_bag.Put(data[i].label); - int feature = _featurename_bag.Put(data[i].feature); - _fb.Put(ME_Feature(label, feature)); - _vl.push_back(data[i].weight); - } - _num_classes = _label_bag.Size(); - - init_feature2mef(); - - return true; -} - -bool ME_Model::save_to_file(const string& filename, const double th) const { - FILE* fp = fopen(filename.c_str(), "w"); - if (!fp) { - cerr << "error: cannot open " << filename << "!" << endl; - return false; - } - - // for (int i = 0; i < _fb.Size(); i++) { - // if (_vl[i] == 0) continue; // ignore zero-weight features - // ME_Feature f = _fb.Feature(i); - // fprintf(fp, "%s\t%s\t%f\n", _label_bag.Str(f.label()).c_str(), - // _featurename_bag.Str(f.feature()).c_str(), _vl[i]); - // } - for (MiniStringBag::map_type::const_iterator i = _featurename_bag.begin(); - i != _featurename_bag.end(); i++) { - for (int j = 0; j < _label_bag.Size(); j++) { - string label = _label_bag.Str(j); - string history = i->first; - int id = _fb.Id(ME_Feature(j, i->second)); - if (id < 0) continue; - if (_vl[id] == 0) continue; // ignore zero-weight features - if (fabs(_vl[id]) < th) continue; // cut off low-weight features - fprintf(fp, "%s\t%s\t%f\n", label.c_str(), history.c_str(), _vl[id]); - } - } - - fclose(fp); - - return true; -} - -void ME_Model::set_ref_dist(Sample& s) const { - vector v0 = s.ref_pd; - vector v(_num_classes); - for (unsigned int i = 0; i < v.size(); i++) { - v[i] = 0; - string label = get_class_label(i); - int id_ref = _ref_modelp->get_class_id(label); - if (id_ref != -1) { - v[i] = v0[id_ref]; - } - if (v[i] == 0) v[i] = 0.001; // to avoid -inf logl - } - s.ref_pd = v; -} - -int ME_Model::classify(const Sample& nbs, vector& membp) const { - // vector membp(_num_classes); - assert(_num_classes == (int)membp.size()); - conditional_probability(nbs, membp); - int max_label = 0; - double max = 0.0; - for (int i = 0; i < (int)membp.size(); i++) { - // cout << membp[i] << " "; - if (membp[i] > max) { - max_label = i; - max = membp[i]; - } - } - // cout << endl; - return max_label; -} - -vector ME_Model::classify(ME_Sample& mes) const { - Sample s; - for (vector::const_iterator j = mes.features.begin(); - j != mes.features.end(); j++) { - int id = _featurename_bag.Id(*j); - if (id >= 0) s.positive_features.push_back(id); - } - for (vector >::const_iterator j = mes.rvfeatures.begin(); - j != mes.rvfeatures.end(); j++) { - int id = _featurename_bag.Id(j->first); - if (id >= 0) { - s.rvfeatures.push_back(pair(id, j->second)); - } - } - if (_ref_modelp != NULL) { - s.ref_pd = _ref_modelp->classify(mes); - set_ref_dist(s); - } - - vector vp(_num_classes); - int label = classify(s, vp); - mes.label = get_class_label(label); - return vp; -} - -/* - * $Log: maxent.cpp,v $ - * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida - * stepp tagger, by Okanohara and Tsuruoka - * - * Revision 1.28 2006/08/21 17:30:38 tsuruoka - * use MAX_LABEL_TYPES - * - * Revision 1.27 2006/07/25 13:19:53 tsuruoka - * sort _vs[] - * - * Revision 1.26 2006/07/18 11:13:15 tsuruoka - * modify comments - * - * Revision 1.25 2006/07/18 10:02:15 tsuruoka - * remove sample2feature[] - * speed up conditional_probability() - * - * Revision 1.24 2006/07/18 05:10:51 tsuruoka - * add ref_dist - * - * Revision 1.23 2005/12/24 07:05:32 tsuruoka - * modify conditional_probability() to avoid overflow - * - * Revision 1.22 2005/12/24 07:01:25 tsuruoka - * add cutoff for real-valued features - * - * Revision 1.21 2005/12/23 10:33:02 tsuruoka - * support real-valued features - * - * Revision 1.20 2005/12/23 09:15:29 tsuruoka - * modify _train to reduce memory consumption - * - * Revision 1.19 2005/10/28 13:10:14 tsuruoka - * fix for overflow (thanks to Ming Li) - * - * Revision 1.18 2005/10/28 13:03:07 tsuruoka - * add progress_bar - * - * Revision 1.17 2005/09/12 13:51:16 tsuruoka - * Sample: list -> vector - * - * Revision 1.16 2005/09/12 13:27:10 tsuruoka - * add add_training_sample() - * - * Revision 1.15 2005/04/27 11:22:27 tsuruoka - * bugfix - * ME_Sample: list -> vector - * - * Revision 1.14 2005/04/27 10:00:42 tsuruoka - * remove tmpfb - * - * Revision 1.13 2005/04/26 14:25:53 tsuruoka - * add MiniStringBag, USE_HASH_MAP - * - * Revision 1.12 2005/02/11 10:20:08 tsuruoka - * modify cutoff - * - * Revision 1.11 2004/10/04 05:50:25 tsuruoka - * add Clear() - * - * Revision 1.10 2004/08/26 16:52:26 tsuruoka - * fix load_from_file() - * - * Revision 1.9 2004/08/09 12:27:21 tsuruoka - * change messages - * - * Revision 1.8 2004/08/04 13:55:18 tsuruoka - * modify _sample2feature - * - * Revision 1.7 2004/07/28 13:42:58 tsuruoka - * add AGIS - * - * Revision 1.6 2004/07/28 05:54:13 tsuruoka - * get_class_name() -> get_class_label() - * ME_Feature: bugfix - * - * Revision 1.5 2004/07/27 16:58:47 tsuruoka - * modify the interface of classify() - * - * Revision 1.4 2004/07/26 17:23:46 tsuruoka - * _sample2feature: list -> vector - * - * Revision 1.3 2004/07/26 15:49:23 tsuruoka - * modify ME_Feature - * - * Revision 1.2 2004/07/26 13:52:18 tsuruoka - * modify cutoff - * - * Revision 1.1 2004/07/26 13:10:55 tsuruoka - * add files - * - * Revision 1.20 2004/07/22 08:34:45 tsuruoka - * modify _sample2feature[] - * - * Revision 1.19 2004/07/21 16:33:01 tsuruoka - * remove some comments - * - */ diff --git a/utils/synutils/maxent-3.0/maxent.h b/utils/synutils/maxent-3.0/maxent.h deleted file mode 100644 index 88a04e25..00000000 --- a/utils/synutils/maxent-3.0/maxent.h +++ /dev/null @@ -1,402 +0,0 @@ -/* - * $Id: maxent.h,v 1.1.1.1 2007/05/15 08:30:35 kyoshida Exp $ - */ - -#ifndef __MAXENT_H_ -#define __MAXENT_H_ - -#include -#include -#include -#include -#include -#include -#include -#include -#include "mathvec.h" - -#define USE_HASH_MAP // if you encounter errors with hash, try commenting out - // this line. (the program will be a bit slower, though) -#ifdef USE_HASH_MAP -#include -#endif - -// -// data format for each sample for training/testing -// -struct ME_Sample { - public: - ME_Sample() : label("") {}; - ME_Sample(const std::string& l) : label(l) {}; - void set_label(const std::string& l) { label = l; } - - // to add a binary feature - void add_feature(const std::string& f) { features.push_back(f); } - - // to add a real-valued feature - void add_feature(const std::string& s, const double d) { - rvfeatures.push_back(std::pair(s, d)); - } - - public: - std::string label; - std::vector features; - std::vector > rvfeatures; - - // obsolete - void add_feature(const std::pair& f) { - rvfeatures.push_back(f); // real-valued features - } -}; - -// -// for those who want to use load_from_array() -// -typedef struct ME_Model_Data { - char* label; - char* feature; - double weight; -} ME_Model_Data; - -class ME_Model { - public: - void add_training_sample(const ME_Sample& s); - int train(); - std::vector classify(ME_Sample& s) const; - bool load_from_file(const std::string& filename); - bool save_to_file(const std::string& filename, const double th = 0) const; - int num_classes() const { return _num_classes; } - std::string get_class_label(int i) const { return _label_bag.Str(i); } - int get_class_id(const std::string& s) const { return _label_bag.Id(s); } - void get_features( - std::list, double> >& fl); - void set_heldout(const int h, const int n = 0) { - _nheldout = h; - _early_stopping_n = n; - }; - void use_l1_regularizer(const double v) { _l1reg = v; } - void use_l2_regularizer(const double v) { _l2reg = v; } - void use_SGD(int iter = 30, double eta0 = 1, double alpha = 0.85) { - _optimization_method = SGD; - SGD_ITER = iter; - SGD_ETA0 = eta0; - SGD_ALPHA = alpha; - } - bool load_from_array(const ME_Model_Data data[]); - void set_reference_model(const ME_Model& ref_model) { - _ref_modelp = &ref_model; - }; - void clear(); - - ME_Model() { - _l1reg = _l2reg = 0; - _nheldout = 0; - _early_stopping_n = 0; - _ref_modelp = NULL; - _optimization_method = LBFGS; - } - - public: - // obsolete. just for downward compatibility - int train(const std::vector& train); - - private: - enum OPTIMIZATION_METHOD { - LBFGS, - OWLQN, - SGD - } _optimization_method; - // OWLQN and SGD are available only for L1-regularization - - int SGD_ITER; - double SGD_ETA0; - double SGD_ALPHA; - - double _l1reg, _l2reg; - - struct Sample { - int label; - std::vector positive_features; - std::vector > rvfeatures; - std::vector ref_pd; // reference probability distribution - bool operator<(const Sample& x) const { - for (unsigned int i = 0; i < positive_features.size(); i++) { - if (i >= x.positive_features.size()) return false; - int v0 = positive_features[i]; - int v1 = x.positive_features[i]; - if (v0 < v1) return true; - if (v0 > v1) return false; - } - return false; - } - }; - - struct ME_Feature { - enum { - MAX_LABEL_TYPES = 255 - }; - - // ME_Feature(const int l, const int f) : _body((l << 24) + f) { - // assert(l >= 0 && l < 256); - // assert(f >= 0 && f <= 0xffffff); - // }; - // int label() const { return _body >> 24; } - // int feature() const { return _body & 0xffffff; } - ME_Feature(const int l, const int f) : _body((f << 8) + l) { - assert(l >= 0 && l <= MAX_LABEL_TYPES); - assert(f >= 0 && f <= 0xffffff); - }; - int label() const { return _body & 0xff; } - int feature() const { return _body >> 8; } - unsigned int body() const { return _body; } - - private: - unsigned int _body; - }; - - struct ME_FeatureBag { -#ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; -#else - typedef std::map map_type; -#endif - map_type mef2id; - std::vector id2mef; - int Put(const ME_Feature& i) { - map_type::const_iterator j = mef2id.find(i.body()); - if (j == mef2id.end()) { - int id = id2mef.size(); - id2mef.push_back(i); - mef2id[i.body()] = id; - return id; - } - return j->second; - } - int Id(const ME_Feature& i) const { - map_type::const_iterator j = mef2id.find(i.body()); - if (j == mef2id.end()) { - return -1; - } - return j->second; - } - ME_Feature Feature(int id) const { - assert(id >= 0 && id < (int)id2mef.size()); - return id2mef[id]; - } - int Size() const { return id2mef.size(); } - void Clear() { - mef2id.clear(); - id2mef.clear(); - } - }; - - struct hashfun_str { - size_t operator()(const std::string& s) const { - assert(sizeof(int) == 4 && sizeof(char) == 1); - const int* p = reinterpret_cast(s.c_str()); - size_t v = 0; - int n = s.size() / 4; - for (int i = 0; i < n; i++, p++) { - // v ^= *p; - v ^= *p << (4 * (i % 2)); // note) 0 <= char < 128 - } - int m = s.size() % 4; - for (int i = 0; i < m; i++) { - v ^= s[4 * n + i] << (i * 8); - } - return v; - } - }; - - struct MiniStringBag { -#ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; -#else - typedef std::map map_type; -#endif - int _size; - map_type str2id; - MiniStringBag() : _size(0) {} - int Put(const std::string& i) { - map_type::const_iterator j = str2id.find(i); - if (j == str2id.end()) { - int id = _size; - _size++; - str2id[i] = id; - return id; - } - return j->second; - } - int Id(const std::string& i) const { - map_type::const_iterator j = str2id.find(i); - if (j == str2id.end()) return -1; - return j->second; - } - int Size() const { return _size; } - void Clear() { - str2id.clear(); - _size = 0; - } - map_type::const_iterator begin() const { return str2id.begin(); } - map_type::const_iterator end() const { return str2id.end(); } - }; - - struct StringBag : public MiniStringBag { - std::vector id2str; - int Put(const std::string& i) { - map_type::const_iterator j = str2id.find(i); - if (j == str2id.end()) { - int id = id2str.size(); - id2str.push_back(i); - str2id[i] = id; - return id; - } - return j->second; - } - std::string Str(const int id) const { - assert(id >= 0 && id < (int)id2str.size()); - return id2str[id]; - } - int Size() const { return id2str.size(); } - void Clear() { - str2id.clear(); - id2str.clear(); - } - }; - - std::vector _vs; // vector of training_samples - StringBag _label_bag; - MiniStringBag _featurename_bag; - std::vector _vl; // vector of lambda - ME_FeatureBag _fb; - int _num_classes; - std::vector _vee; // empirical expectation - std::vector _vme; // empirical expectation - std::vector > _feature2mef; - std::vector _heldout; - double _train_error; // current error rate on the training data - double _heldout_error; // current error rate on the heldout data - int _nheldout; - int _early_stopping_n; - std::vector _vhlogl; - const ME_Model* _ref_modelp; - - double heldout_likelihood(); - int conditional_probability(const Sample& nbs, - std::vector& membp) const; - int make_feature_bag(const int cutoff); - int classify(const Sample& nbs, std::vector& membp) const; - double update_model_expectation(); - int perform_QUASI_NEWTON(); - int perform_SGD(); - int perform_GIS(int C); - std::vector perform_LBFGS(const std::vector& x0); - std::vector perform_OWLQN(const std::vector& x0, - const double C); - double backtracking_line_search(const Vec& x0, const Vec& grad0, - const double f0, const Vec& dx, Vec& x, - Vec& grad1); - double regularized_func_grad(const double C, const Vec& x, Vec& grad); - double constrained_line_search(double C, const Vec& x0, const Vec& grad0, - const double f0, const Vec& dx, Vec& x, - Vec& grad1); - - void set_ref_dist(Sample& s) const; - void init_feature2mef(); - - double FunctionGradient(const std::vector& x, - std::vector& grad); - static double FunctionGradientWrapper(const std::vector& x, - std::vector& grad); -}; - -#endif - -/* - * $Log: maxent.h,v $ - * Revision 1.1.1.1 2007/05/15 08:30:35 kyoshida - * stepp tagger, by Okanohara and Tsuruoka - * - * Revision 1.24 2006/08/21 17:30:38 tsuruoka - * use MAX_LABEL_TYPES - * - * Revision 1.23 2006/07/25 13:19:53 tsuruoka - * sort _vs[] - * - * Revision 1.22 2006/07/18 11:13:15 tsuruoka - * modify comments - * - * Revision 1.21 2006/07/18 10:02:15 tsuruoka - * remove sample2feature[] - * speed up conditional_probability() - * - * Revision 1.20 2006/07/18 05:10:51 tsuruoka - * add ref_dist - * - * Revision 1.19 2005/12/23 10:33:02 tsuruoka - * support real-valued features - * - * Revision 1.18 2005/12/23 09:15:29 tsuruoka - * modify _train to reduce memory consumption - * - * Revision 1.17 2005/10/28 13:02:34 tsuruoka - * set_heldout(): add default value - * Feature() - * - * Revision 1.16 2005/09/12 13:51:16 tsuruoka - * Sample: list -> vector - * - * Revision 1.15 2005/09/12 13:27:10 tsuruoka - * add add_training_sample() - * - * Revision 1.14 2005/04/27 11:22:27 tsuruoka - * bugfix - * ME_Sample: list -> vector - * - * Revision 1.13 2005/04/27 10:20:19 tsuruoka - * MiniStringBag -> StringBag - * - * Revision 1.12 2005/04/27 10:00:42 tsuruoka - * remove tmpfb - * - * Revision 1.11 2005/04/26 14:25:53 tsuruoka - * add MiniStringBag, USE_HASH_MAP - * - * Revision 1.10 2004/10/04 05:50:25 tsuruoka - * add Clear() - * - * Revision 1.9 2004/08/09 12:27:21 tsuruoka - * change messages - * - * Revision 1.8 2004/08/04 13:55:19 tsuruoka - * modify _sample2feature - * - * Revision 1.7 2004/07/29 05:51:13 tsuruoka - * remove modeldata.h - * - * Revision 1.6 2004/07/28 13:42:58 tsuruoka - * add AGIS - * - * Revision 1.5 2004/07/28 05:54:14 tsuruoka - * get_class_name() -> get_class_label() - * ME_Feature: bugfix - * - * Revision 1.4 2004/07/27 16:58:47 tsuruoka - * modify the interface of classify() - * - * Revision 1.3 2004/07/26 17:23:46 tsuruoka - * _sample2feature: list -> vector - * - * Revision 1.2 2004/07/26 15:49:23 tsuruoka - * modify ME_Feature - * - * Revision 1.1 2004/07/26 13:10:55 tsuruoka - * add files - * - * Revision 1.18 2004/07/22 08:34:45 tsuruoka - * modify _sample2feature[] - * - * Revision 1.17 2004/07/21 16:33:01 tsuruoka - * remove some comments - * - */ diff --git a/utils/synutils/maxent-3.0/owlqn.cpp b/utils/synutils/maxent-3.0/owlqn.cpp deleted file mode 100644 index c3a0f0da..00000000 --- a/utils/synutils/maxent-3.0/owlqn.cpp +++ /dev/null @@ -1,127 +0,0 @@ -#include -#include -#include -#include -#include "mathvec.h" -#include "lbfgs.h" -#include "maxent.h" - -using namespace std; - -const static int M = LBFGS_M; -const static double LINE_SEARCH_ALPHA = 0.1; -const static double LINE_SEARCH_BETA = 0.5; - -// stopping criteria -int OWLQN_MAX_ITER = 300; -const static double MIN_GRAD_NORM = 0.0001; - -Vec approximate_Hg(const int iter, const Vec& grad, const Vec s[], - const Vec y[], const double z[]); - -inline int sign(double x) { - if (x > 0) return 1; - if (x < 0) return -1; - return 0; -}; - -static Vec pseudo_gradient(const Vec& x, const Vec& grad0, const double C) { - Vec grad = grad0; - for (size_t i = 0; i < x.Size(); i++) { - if (x[i] != 0) { - grad[i] += C * sign(x[i]); - continue; - } - const double gm = grad0[i] - C; - if (gm > 0) { - grad[i] = gm; - continue; - } - const double gp = grad0[i] + C; - if (gp < 0) { - grad[i] = gp; - continue; - } - grad[i] = 0; - } - - return grad; -} - -double ME_Model::regularized_func_grad(const double C, const Vec& x, - Vec& grad) { - double f = FunctionGradient(x.STLVec(), grad.STLVec()); - for (size_t i = 0; i < x.Size(); i++) { - f += C * fabs(x[i]); - } - - return f; -} - -double ME_Model::constrained_line_search(double C, const Vec& x0, - const Vec& grad0, const double f0, - const Vec& dx, Vec& x, Vec& grad1) { - // compute the orthant to explore - Vec orthant = x0; - for (size_t i = 0; i < orthant.Size(); i++) { - if (orthant[i] == 0) orthant[i] = -grad0[i]; - } - - double t = 1.0 / LINE_SEARCH_BETA; - - double f; - do { - t *= LINE_SEARCH_BETA; - x = x0 + t * dx; - x.Project(orthant); - // for (size_t i = 0; i < x.Size(); i++) { - // if (x0[i] != 0 && sign(x[i]) != sign(x0[i])) x[i] = 0; - // } - - f = regularized_func_grad(C, x, grad1); - // cout << "*"; - } while (f > f0 + LINE_SEARCH_ALPHA * dot_product(x - x0, grad0)); - - return f; -} - -vector ME_Model::perform_OWLQN(const vector& x0, - const double C) { - const size_t dim = x0.size(); - Vec x = x0; - - Vec grad(dim), dx(dim); - double f = regularized_func_grad(C, x, grad); - - Vec s[M], y[M]; - double z[M]; // rho - - for (int iter = 0; iter < OWLQN_MAX_ITER; iter++) { - Vec pg = pseudo_gradient(x, grad, C); - - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, -f, _train_error); - if (_nheldout > 0) { - const double heldout_logl = heldout_likelihood(); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, - _heldout_error); - } - fprintf(stderr, "\n"); - - if (sqrt(dot_product(pg, pg)) < MIN_GRAD_NORM) break; - - dx = -1 * approximate_Hg(iter, pg, s, y, z); - if (dot_product(dx, pg) >= 0) dx.Project(-1 * pg); - - Vec x1(dim), grad1(dim); - f = constrained_line_search(C, x, pg, f, dx, x1, grad1); - - s[iter % M] = x1 - x; - y[iter % M] = grad1 - grad; - z[iter % M] = 1.0 / dot_product(y[iter % M], s[iter % M]); - - x = x1; - grad = grad1; - } - - return x.STLVec(); -} diff --git a/utils/synutils/maxent-3.0/sgd.cpp b/utils/synutils/maxent-3.0/sgd.cpp deleted file mode 100644 index 8613edca..00000000 --- a/utils/synutils/maxent-3.0/sgd.cpp +++ /dev/null @@ -1,193 +0,0 @@ -#include "maxent.h" -#include -#include - -using namespace std; - -// const double SGD_ETA0 = 1; -// const double SGD_ITER = 30; -// const double SGD_ALPHA = 0.85; - -//#define FOLOS_NAIVE -//#define FOLOS_LAZY -#define SGD_CP - -inline void apply_l1_penalty(const int i, const double u, vector& _vl, - vector& q) { - double& w = _vl[i]; - const double z = w; - double& qi = q[i]; - if (w > 0) { - w = max(0.0, w - (u + qi)); - } else if (w < 0) { - w = min(0.0, w + (u - qi)); - } - qi += w - z; -} - -static double l1norm(const vector& v) { - double sum = 0; - for (size_t i = 0; i < v.size(); i++) sum += abs(v[i]); - return sum; -} - -inline void update_folos_lazy(const int iter_sample, const int k, - vector& _vl, - const vector& sum_eta, - vector& last_updated) { - const double penalty = sum_eta[iter_sample] - sum_eta[last_updated[k]]; - double& x = _vl[k]; - if (x > 0) - x = max(0.0, x - penalty); - else - x = min(0.0, x + penalty); - last_updated[k] = iter_sample; -} - -int ME_Model::perform_SGD() { - if (_l2reg > 0) { - cerr << "error: L2 regularization is currently not supported in SGD mode." - << endl; - exit(1); - } - - cerr << "performing SGD" << endl; - - const double l1param = _l1reg; - - const int d = _fb.Size(); - - vector ri(_vs.size()); - for (size_t i = 0; i < ri.size(); i++) ri[i] = i; - - vector grad(d); - int iter_sample = 0; - const double eta0 = SGD_ETA0; - - // cerr << "l1param = " << l1param << endl; - cerr << "eta0 = " << eta0 << " alpha = " << SGD_ALPHA << endl; - - double u = 0; - vector q(d, 0); - vector last_updated(d, 0); - vector sum_eta; - sum_eta.push_back(0); - - for (int iter = 0; iter < SGD_ITER; iter++) { - - random_shuffle(ri.begin(), ri.end()); - - double logl = 0; - int ncorrect = 0, ntotal = 0; - for (size_t i = 0; i < _vs.size(); i++, ntotal++, iter_sample++) { - const Sample& s = _vs[ri[i]]; - -#ifdef FOLOS_LAZY - for (vector::const_iterator j = s.positive_features.begin(); - j != s.positive_features.end(); j++) { - for (vector::const_iterator k = _feature2mef[*j].begin(); - k != _feature2mef[*j].end(); k++) { - update_folos_lazy(iter_sample, *k, _vl, sum_eta, last_updated); - } - } -#endif - - vector membp(_num_classes); - const int max_label = conditional_probability(s, membp); - - const double eta = - eta0 * pow(SGD_ALPHA, - (double)iter_sample / _vs.size()); // exponential decay - // const double eta = eta0 / (1.0 + (double)iter_sample / - // _vs.size()); - - // if (iter_sample % _vs.size() == 0) cerr << "eta = " << eta << - // endl; - u += eta * l1param; - - sum_eta.push_back(sum_eta.back() + eta * l1param); - - logl += log(membp[s.label]); - if (max_label == s.label) ncorrect++; - - // binary features - for (vector::const_iterator j = s.positive_features.begin(); - j != s.positive_features.end(); j++) { - for (vector::const_iterator k = _feature2mef[*j].begin(); - k != _feature2mef[*j].end(); k++) { - const double me = membp[_fb.Feature(*k).label()]; - const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); - const double grad = (me - ee); - _vl[*k] -= eta * grad; -#ifdef SGD_CP - apply_l1_penalty(*k, u, _vl, q); -#endif - } - } - // real-valued features - for (vector >::const_iterator j = s.rvfeatures.begin(); - j != s.rvfeatures.end(); j++) { - for (vector::const_iterator k = _feature2mef[j->first].begin(); - k != _feature2mef[j->first].end(); k++) { - const double me = membp[_fb.Feature(*k).label()]; - const double ee = (_fb.Feature(*k).label() == s.label ? 1.0 : 0); - const double grad = (me - ee) * j->second; - _vl[*k] -= eta * grad; -#ifdef SGD_CP - apply_l1_penalty(*k, u, _vl, q); -#endif - } - } - -#ifdef FOLOS_NAIVE - for (size_t j = 0; j < d; j++) { - double& x = _vl[j]; - if (x > 0) - x = max(0.0, x - eta * l1param); - else - x = min(0.0, x + eta * l1param); - } -#endif - } - logl /= _vs.size(); -// fprintf(stderr, "%4d logl = %8.3f acc = %6.4f ", iter, logl, -// (double)ncorrect / ntotal); - -#ifdef FOLOS_LAZY - if (l1param > 0) { - for (size_t j = 0; j < d; j++) - update_folos_lazy(iter_sample, j, _vl, sum_eta, last_updated); - } -#endif - - double f = logl; - if (l1param > 0) { - const double l1 = - l1norm(_vl); // this is not accurate when lazy update is used - // cerr << "f0 = " << update_model_expectation() - l1param * l1 << " - // "; - f -= l1param * l1; - int nonzero = 0; - for (int j = 0; j < d; j++) - if (_vl[j] != 0) nonzero++; - // cerr << " f = " << f << " l1 = " << l1 << " nonzero_features = " - // << nonzero << endl; - } - // fprintf(stderr, "%4d obj = %7.3f acc = %6.4f", iter+1, f, - // (double)ncorrect/ntotal); - // fprintf(stderr, "%4d obj = %f", iter+1, f); - fprintf(stderr, "%3d obj(err) = %f (%6.4f)", iter + 1, f, - 1 - (double)ncorrect / ntotal); - - if (_nheldout > 0) { - double heldout_logl = heldout_likelihood(); - // fprintf(stderr, " heldout_logl = %f acc = %6.4f\n", - // heldout_logl, 1 - _heldout_error); - fprintf(stderr, " heldout_logl(err) = %f (%6.4f)", heldout_logl, - _heldout_error); - } - fprintf(stderr, "\n"); - } - - return 0; -} diff --git a/utils/synutils/srl_sentence.h b/utils/synutils/srl_sentence.h deleted file mode 100644 index 1afdec45..00000000 --- a/utils/synutils/srl_sentence.h +++ /dev/null @@ -1,214 +0,0 @@ -/* - * srl_sentence.h - * - * Created on: May 26, 2013 - * Author: junhuili - */ - -#ifndef SRL_SENTENCE_H_ -#define SRL_SENTENCE_H_ - - -#include -#include - -#include "tree.h" -#include "utils/stringlib.h" - -using namespace std; - -struct SArgument { - SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { - m_pszRole = new char[strlen(pszRole) + 1]; - strcpy(m_pszRole, pszRole); - m_iBegin = iBegin; - m_iEnd = iEnd; - m_fProb = fProb; - m_pTreeItem = NULL; - } - ~SArgument() { delete[] m_pszRole; } - - void fnSetTreeItem(STreeItem* pTreeItem) { - m_pTreeItem = pTreeItem; - if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { - assert(m_pTreeItem->m_iBegin == m_iBegin); - assert(m_pTreeItem->m_iEnd == m_iEnd); - } - } - - char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP - int m_iBegin; - int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] - float m_fProb; // the probability of this role, - STreeItem* m_pTreeItem; -}; - -struct SPredicate { - SPredicate(const char* pszLemma, int iPosition) { - if (pszLemma != NULL) { - m_pszLemma = new char[strlen(pszLemma) + 1]; - strcpy(m_pszLemma, pszLemma); - } else - m_pszLemma = NULL; - m_iPosition = iPosition; - } - ~SPredicate() { - if (m_pszLemma != NULL) delete[] m_pszLemma; - for (size_t i = 0; i < m_vecArgt.size(); i++) delete m_vecArgt[i]; - } - int fnAppend(const char* pszRole, int iBegin, int iEnd) { - SArgument* pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); - return fnAppend(pArgt); - } - int fnAppend(SArgument* pArgt) { - m_vecArgt.push_back(pArgt); - int iPosition = m_vecArgt.size() - 1; - return iPosition; - } - - char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same - // as the predicate itself - int m_iPosition; // the position in sentence - vector m_vecArgt; // arguments associated to the predicate -}; - -struct SSrlSentence { - SSrlSentence() { m_pTree = NULL; } - ~SSrlSentence() { - if (m_pTree != NULL) delete m_pTree; - - for (size_t i = 0; i < m_vecPred.size(); i++) delete m_vecPred[i]; - } - int fnAppend(const char* pszLemma, int iPosition) { - SPredicate* pPred = new SPredicate(pszLemma, iPosition); - return fnAppend(pPred); - } - int fnAppend(SPredicate* pPred) { - m_vecPred.push_back(pPred); - int iPosition = m_vecPred.size() - 1; - return iPosition; - } - int GetPredicateNum() { return m_vecPred.size(); } - - SParsedTree* m_pTree; - vector m_vecPred; -}; - -struct SSrlSentenceReader { - SSrlSentenceReader(const char* pszSrlFname) { - m_fpIn = fopen(pszSrlFname, "r"); - assert(m_fpIn != NULL); - } - ~SSrlSentenceReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - inline void fnReplaceAll(std::string& str, const std::string& from, - const std::string& to) { - size_t start_pos = 0; - while ((start_pos = str.find(from, start_pos)) != std::string::npos) { - str.replace(start_pos, from.length(), to); - start_pos += to.length(); // In case 'to' contains 'from', like replacing - // 'x' with 'yx' - } - } - - // TODO: here only considers flat predicate-argument structure - // i.e., no overlap among them - SSrlSentence* fnReadNextSrlSentence() { - vector > vecContent; - if (fnReadNextContent(vecContent) == false) return NULL; - - SSrlSentence* pSrlSentence = new SSrlSentence(); - int iSize = vecContent.size(); - // put together syntactic text - std::ostringstream ostr; - for (int i = 0; i < iSize; i++) { - string strSynSeg = - vecContent[i][5]; // the 5th column is the syntactic segment - size_t iPosition = strSynSeg.find_first_of('*'); - assert(iPosition != string::npos); - ostringstream ostrTmp; - ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] - << ")"; // the 2th column is POS-tag, and the 0th column is word - strSynSeg.replace(iPosition, 1, ostrTmp.str()); - fnReplaceAll(strSynSeg, "(", " ("); - ostr << strSynSeg; - } - string strSyn = ostr.str(); - pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); - pSrlSentence->m_pTree->fnSetHeadWord(); - pSrlSentence->m_pTree->fnSetSpanInfo(); - - // read predicate-argument structure - int iNumPred = vecContent[0].size() - 8; - for (int i = 0; i < iNumPred; i++) { - vector vecRole; - vector vecBegin; - vector vecEnd; - int iPred = -1; - for (int j = 0; j < iSize; j++) { - const char* p = vecContent[j][i + 8].c_str(); - const char* q; - if (p[0] == '(') { - // starting position of an argument(or predicate) - vecBegin.push_back(j); - q = strchr(p, '*'); - assert(q != NULL); - vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); - if (vecRole.back().compare("V") == 0) { - assert(iPred == -1); - iPred = vecRole.size() - 1; - } - } - if (p[strlen(p) - 1] == ')') { - // end position of an argument(or predicate) - vecEnd.push_back(j); - assert(vecBegin.size() == vecEnd.size()); - } - } - assert(iPred != -1); - SPredicate* pPred = new SPredicate( - pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, - vecBegin[iPred]); - pSrlSentence->fnAppend(pPred); - for (size_t j = 0; j < vecBegin.size(); j++) { - if (j == iPred) continue; - pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); - pPred->m_vecArgt.back()->fnSetTreeItem( - pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], - false)); - } - } - return pSrlSentence; - } - - private: - bool fnReadNextContent(vector >& vecContent) { - vecContent.clear(); - if (feof(m_fpIn) == true) return false; - char* pszLine; - pszLine = new char[100001]; - pszLine[0] = '\0'; - int iLen; - while (!feof(m_fpIn)) { - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - if (iLen == 0) break; // end of this sentence - - vector terms = SplitOnWhitespace(string(pszLine)); - assert(terms.size() > 7); - vecContent.push_back(terms); - } - delete[] pszLine; - return true; - } - - private: - FILE* m_fpIn; -}; -#endif /* SRL_SENTENCE_H_ */ diff --git a/utils/synutils/tree.h b/utils/synutils/tree.h deleted file mode 100644 index 8070f828..00000000 --- a/utils/synutils/tree.h +++ /dev/null @@ -1,700 +0,0 @@ -/* - * tree.h - * - * Created on: May 23, 2013 - * Author: lijunhui - */ - -#ifndef TREE_H_ -#define TREE_H_ - -#include -#include - -#include -#include - -using namespace std; - -struct STreeItem { - STreeItem(const char *pszTerm) { - m_pszTerm = new char[strlen(pszTerm) + 1]; - strcpy(m_pszTerm, pszTerm); - - m_ptParent = NULL; - m_iBegin = -1; - m_iEnd = -1; - m_iHeadChild = -1; - m_iHeadWord = -1; - m_iBrotherIndex = -1; - } - ~STreeItem() { - delete[] m_pszTerm; - for (size_t i = 0; i < m_vecChildren.size(); i++) delete m_vecChildren[i]; - } - int fnAppend(STreeItem *ptChild) { - m_vecChildren.push_back(ptChild); - ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; - ptChild->m_ptParent = this; - return m_vecChildren.size() - 1; - } - int fnGetChildrenNum() { return m_vecChildren.size(); } - - bool fnIsPreTerminal(void) { - int I; - if (this == NULL || m_vecChildren.size() == 0) return false; - - for (I = 0; I < m_vecChildren.size(); I++) - if (m_vecChildren[I]->m_vecChildren.size() > 0) return false; - - return true; - } - - public: - char *m_pszTerm; - - vector m_vecChildren; // children items - STreeItem *m_ptParent; // the parent item - - int m_iBegin; - int m_iEnd; // the node span words[m_iBegin, m_iEnd] - int m_iHeadChild; // the index of its head child - int m_iHeadWord; // the index of its head word - int m_iBrotherIndex; // the index in his brothers -}; - -struct SGetHeadWord { - typedef vector CVectorStr; - SGetHeadWord() {} - ~SGetHeadWord() {} - int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { - // 0 indicating from right to left while 1 indicating from left to right - char szaHeadLists[201] = "0"; - - /* //head rules for Egnlish - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP - 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN - 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP - 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR - 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ - 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP - 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP - 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN - 0CD 0JJ JJS RB QP 0" ); - */ - - if (strcmp(pszCFGLeft, "ADJP") == 0) - strcpy(szaHeadLists, "0ADJP JJ 0AD NN CS 0"); - else if (strcmp(pszCFGLeft, "ADVP") == 0) - strcpy(szaHeadLists, "0ADVP AD 0"); - else if (strcmp(pszCFGLeft, "CLP") == 0) - strcpy(szaHeadLists, "0CLP M 0"); - else if (strcmp(pszCFGLeft, "CP") == 0) - strcpy(szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0"); - else if (strcmp(pszCFGLeft, "DNP") == 0) - strcpy(szaHeadLists, "0DNP DEG 0DEC 0"); - else if (strcmp(pszCFGLeft, "DVP") == 0) - strcpy(szaHeadLists, "0DVP DEV 0"); - else if (strcmp(pszCFGLeft, "DP") == 0) - strcpy(szaHeadLists, "1DP DT 1"); - else if (strcmp(pszCFGLeft, "FRAG") == 0) - strcpy(szaHeadLists, "0VV NR NN 0"); - else if (strcmp(pszCFGLeft, "INTJ") == 0) - strcpy(szaHeadLists, "0INTJ IJ 0"); - else if (strcmp(pszCFGLeft, "LST") == 0) - strcpy(szaHeadLists, "1LST CD OD 1"); - else if (strcmp(pszCFGLeft, "IP") == 0) - strcpy(szaHeadLists, "0IP VP 0VV 0"); - // strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); - else if (strcmp(pszCFGLeft, "LCP") == 0) - strcpy(szaHeadLists, "0LCP LC 0"); - else if (strcmp(pszCFGLeft, "NP") == 0) - strcpy(szaHeadLists, "0NP NN NT NR QP 0"); - else if (strcmp(pszCFGLeft, "PP") == 0) - strcpy(szaHeadLists, "1PP P 1"); - else if (strcmp(pszCFGLeft, "PRN") == 0) - strcpy(szaHeadLists, "0 NP IP VP NT NR NN 0"); - else if (strcmp(pszCFGLeft, "QP") == 0) - strcpy(szaHeadLists, "0QP CLP CD OD 0"); - else if (strcmp(pszCFGLeft, "VP") == 0) - strcpy(szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1"); - else if (strcmp(pszCFGLeft, "VCD") == 0) - strcpy(szaHeadLists, "0VCD VV VA VC VE 0"); - if (strcmp(pszCFGLeft, "VRD") == 0) - strcpy(szaHeadLists, "0VRD VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VSB") == 0) - strcpy(szaHeadLists, "0VSB VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VCP") == 0) - strcpy(szaHeadLists, "0VCP VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VNV") == 0) - strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VPT") == 0) - strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "UCP") == 0) - strcpy(szaHeadLists, "0"); - else if (strcmp(pszCFGLeft, "WHNP") == 0) - strcpy(szaHeadLists, "0WHNP NP NN NT NR QP 0"); - else if (strcmp(pszCFGLeft, "WHPP") == 0) - strcpy(szaHeadLists, "1WHPP PP P 1"); - - /* //head rules for GENIA corpus - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP - 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN - 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP - 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR - 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ - 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP - 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP - 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN - 0CD 0JJ JJS RB QP 0" ); - */ - - return fnMyOwnHeadWordRule(szaHeadLists, vectRight); - } - - private: - int fnMyOwnHeadWordRule(char *pszaHeadLists, CVectorStr vectRight) { - char szHeadList[201], *p; - char szTerm[101]; - int J; - - p = pszaHeadLists; - - int iCountRight; - - iCountRight = vectRight.size(); - - szHeadList[0] = '\0'; - while (1) { - szTerm[0] = '\0'; - sscanf(p, "%s", szTerm); - if (strlen(szHeadList) == 0) { - if (strcmp(szTerm, "0") == 0) { - return iCountRight - 1; - } - if (strcmp(szTerm, "1") == 0) { - return 0; - } - - sprintf(szHeadList, "%c %s ", szTerm[0], szTerm + 1); - p = strstr(p, szTerm); - p += strlen(szTerm); - } else { - if ((szTerm[0] == '0') || (szTerm[0] == '1')) { - if (szHeadList[0] == '0') { - for (J = iCountRight - 1; J >= 0; J--) { - sprintf(szTerm, " %s ", vectRight.at(J).c_str()); - if (strstr(szHeadList, szTerm) != NULL) return J; - } - } else { - for (J = 0; J < iCountRight; J++) { - sprintf(szTerm, " %s ", vectRight.at(J).c_str()); - if (strstr(szHeadList, szTerm) != NULL) return J; - } - } - - szHeadList[0] = '\0'; - } else { - strcat(szHeadList, szTerm); - strcat(szHeadList, " "); - - p = strstr(p, szTerm); - p += strlen(szTerm); - } - } - } - - return 0; - } -}; - -struct SParsedTree { - SParsedTree() { m_ptRoot = NULL; } - ~SParsedTree() { - if (m_ptRoot != NULL) delete m_ptRoot; - } - static SParsedTree *fnConvertFromString(const char *pszStr) { - if (strcmp(pszStr, "(())") == 0) return NULL; - SParsedTree *pTree = new SParsedTree(); - - vector vecSyn; - fnReadSyntactic(pszStr, vecSyn); - - int iLeft = 1, iRight = 1; //# left/right parenthesis - - STreeItem *pcurrent; - - pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); - - pcurrent = pTree->m_ptRoot; - - for (size_t i = 2; i < vecSyn.size() - 1; i++) { - if (strcmp(vecSyn[i].c_str(), "(") == 0) - iLeft++; - else if (strcmp(vecSyn[i].c_str(), ")") == 0) { - iRight++; - if (pcurrent == NULL) { - // error - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - pcurrent = pcurrent->m_ptParent; - } else { - STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); - pcurrent->fnAppend(ptNewItem); - pcurrent = ptNewItem; - - if (strcmp(vecSyn[i - 1].c_str(), "(") != 0 && - strcmp(vecSyn[i - 1].c_str(), ")") != 0) { - pTree->m_vecTerminals.push_back(ptNewItem); - pcurrent = pcurrent->m_ptParent; - } - } - } - - if (iLeft != iRight) { - // error - fprintf(stderr, "the left and right parentheses are not matched!"); - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - - return pTree; - } - - int fnGetNumWord() { return m_vecTerminals.size(); } - - void fnSetSpanInfo() { - int iNextNum = 0; - fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); - } - - void fnSetHeadWord() { - for (size_t i = 0; i < m_vecTerminals.size(); i++) - m_vecTerminals[i]->m_iHeadWord = i; - SGetHeadWord *pGetHeadWord = new SGetHeadWord(); - fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); - delete pGetHeadWord; - } - - STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { - STreeItem *pTreeItem = m_vecTerminals[iLeft]; - - while (pTreeItem->m_iEnd < iRight) { - pTreeItem = pTreeItem->m_ptParent; - if (pTreeItem == NULL) break; - } - if (pTreeItem == NULL) return NULL; - if (pTreeItem->m_iEnd > iRight) return NULL; - - assert(pTreeItem->m_iEnd == iRight); - if (bLowest) return pTreeItem; - - while (pTreeItem->m_ptParent != NULL && - pTreeItem->m_ptParent->fnGetChildrenNum() == 1) - pTreeItem = pTreeItem->m_ptParent; - - return pTreeItem; - } - - private: - void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int &iNextNum) { - int I; - int iNumChildren = ptItem->fnGetChildrenNum(); - for (I = 0; I < iNumChildren; I++) - fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[I], iNextNum); - - if (I == 0) { - ptItem->m_iBegin = iNextNum; - ptItem->m_iEnd = iNextNum++; - } else { - ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; - ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; - } - } - - void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, - SGetHeadWord *pGetHeadWord) { - int I, iHeadchild; - - if (ptItem->m_vecChildren.size() == 0) return; - - for (I = 0; I < ptItem->m_vecChildren.size(); I++) - fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); - - vector vecRight; - - if (ptItem->m_vecChildren.size() == 1) - iHeadchild = 0; - else { - for (I = 0; I < ptItem->m_vecChildren.size(); I++) - vecRight.push_back(string(ptItem->m_vecChildren[I]->m_pszTerm)); - - iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); - } - - ptItem->m_iHeadChild = iHeadchild; - ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - - static void fnReadSyntactic(const char *pszSyn, vector &vec) { - char *p; - int I; - - int iLeftNum, iRightNum; - char *pszTmp, *pszTerm; - pszTmp = new char[strlen(pszSyn)]; - pszTerm = new char[strlen(pszSyn)]; - pszTmp[0] = pszTerm[0] = '\0'; - - vec.clear(); - - char *pszLine; - pszLine = new char[strlen(pszSyn) + 1]; - strcpy(pszLine, pszSyn); - - char *pszLine2; - - while (1) { - while ((strlen(pszLine) > 0) && (pszLine[strlen(pszLine) - 1] > 0) && - (pszLine[strlen(pszLine) - 1] <= ' ')) - pszLine[strlen(pszLine) - 1] = '\0'; - - if (strlen(pszLine) == 0) break; - - // printf( "%s\n", pszLine ); - pszLine2 = pszLine; - while (pszLine2[0] <= ' ') pszLine2++; - if (pszLine2[0] == '<') continue; - - sscanf(pszLine2 + 1, "%s", pszTmp); - - if (pszLine2[0] == '(') { - iLeftNum = 0; - iRightNum = 0; - } - - p = pszLine2; - while (1) { - pszTerm[0] = '\0'; - sscanf(p, "%s", pszTerm); - - if (strlen(pszTerm) == 0) break; - p = strstr(p, pszTerm); - p += strlen(pszTerm); - - if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { - if (pszTerm[0] == '(') { - vec.push_back(string("(")); - iLeftNum++; - - I = 1; - while (pszTerm[I] == '(' && pszTerm[I] != '\0') { - vec.push_back(string("(")); - iLeftNum++; - - I++; - } - - if (strlen(pszTerm) > 1) vec.push_back(string(pszTerm + I)); - } else { - char *pTmp; - pTmp = pszTerm + strlen(pszTerm) - 1; - while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; - pTmp[1] = '\0'; - - if (strlen(pszTerm) > 0) vec.push_back(string(pszTerm)); - pTmp += 2; - - for (I = 0; I <= (int)strlen(pTmp); I++) { - vec.push_back(string(")")); - iRightNum++; - } - } - } else { - char *q; - q = strchr(pszTerm, ')'); - if (q != NULL) { - q[0] = '\0'; - if (pszTerm[0] != '\0') vec.push_back(string(pszTerm)); - vec.push_back(string(")")); - iRightNum++; - - q++; - while (q[0] == ')') { - vec.push_back(string(")")); - q++; - iRightNum++; - } - - while (q[0] == '(') { - vec.push_back(string("(")); - q++; - iLeftNum++; - } - - if (q[0] != '\0') vec.push_back(string(q)); - } else - vec.push_back(string(pszTerm)); - } - } - - if (iLeftNum != iRightNum) { - fprintf(stderr, "%s\n", pszSyn); - assert(iLeftNum == iRightNum); - } - /*if ( iLeftNum != iRightNum ) { - printf( "ERROR: left( and right ) is not matched, %d ( and %d - )\n", iLeftNum, iRightNum ); - return; - }*/ - - if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { - //( (IP..) ) - std::vector::iterator it; - it = vec.begin(); - it++; - vec.insert(it, string("ROOT")); - } - - break; - } - - delete[] pszLine; - delete[] pszTmp; - delete[] pszTerm; - } - - public: - STreeItem *m_ptRoot; - vector m_vecTerminals; // the leaf nodes -}; - -struct SParseReader { - SParseReader(const char *pszParse_Fname, bool bFlattened = false) - : m_bFlattened(bFlattened) { - m_fpIn = fopen(pszParse_Fname, "r"); - assert(m_fpIn != NULL); - } - ~SParseReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - SParsedTree *fnReadNextParseTree() { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true) { - if (iLen == 0) continue; - - pTree = SParsedTree::fnConvertFromString(pszLine); - if (pTree == NULL) break; - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete[] pszLine; - return pTree; - } - - SParsedTree *fnReadNextParseTreeWithProb(double *pProb) { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true) { - if (iLen == 0) continue; - - char *p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - if (pProb) (*pProb) = atof(pszLine); - - pTree = SParsedTree::fnConvertFromString(p); - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete[] pszLine; - return pTree; - } - - private: - /* - * since to the parse tree is a flattened tree, use the head mark to identify - * head info. - * the head node will be marked as "*XP*" - */ - void fnSetParseTreeHeadInfo(SParsedTree *pTree) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) - pTree->m_vecTerminals[i]->m_iHeadWord = i; - fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); - } - - void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { - if (pTreeItem->m_vecChildren.size() == 0) return; - - for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) - fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); - - vector vecRight; - - int iHeadchild; - - if (pTreeItem->fnIsPreTerminal()) { - iHeadchild = 0; - } else { - size_t i; - for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { - char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; - if (p[0] == '*' && p[strlen(p) - 1] == '*') { - iHeadchild = i; - p[strlen(p) - 1] = '\0'; - string str = p + 1; - strcpy(p, str.c_str()); // erase the "*..*" - break; - } - } - assert(i < pTreeItem->m_vecChildren.size()); - } - - pTreeItem->m_iHeadChild = iHeadchild; - pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - void fnPostProcessingFlattenedParse(SParsedTree *pTree) { - pTree->fnSetSpanInfo(); - fnSetParseTreeHeadInfo(pTree); - } - bool fnReadNextSentence(char *pszLine, int *piLength) { - if (feof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - private: - FILE *m_fpIn; - const bool m_bFlattened; -}; - -#endif /* TREE_H_ */ diff --git a/utils/synutils/tsuruoka_maxent.h b/utils/synutils/tsuruoka_maxent.h deleted file mode 100644 index 89b06b97..00000000 --- a/utils/synutils/tsuruoka_maxent.h +++ /dev/null @@ -1,155 +0,0 @@ -/* - * tsuruoka_maxent.h - * - */ - -#ifndef TSURUOKA_MAXENT_H_ -#define TSURUOKA_MAXENT_H_ - -#include "utility.h" -#include "utils/stringlib.h" -#include "utils/synutils/maxent-3.0/maxent.h" - -#include -#include -#include -#include -#include - -using namespace std; - -typedef std::unordered_map Map; -typedef std::unordered_map::iterator Iterator; - -struct Tsuruoka_Maxent { - Tsuruoka_Maxent(const char* pszModelFName) { - if (pszModelFName != NULL) { - m_pModel = new ME_Model(); - m_pModel->load_from_file(pszModelFName); - } else - m_pModel = NULL; - } - - ~Tsuruoka_Maxent() { - if (m_pModel != NULL) delete m_pModel; - } - - void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, - const char* pszModelFName, int iNumIteration) { - assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || - strcmp(pszAlgorithm, "sgd") == 0 || - strcmp(pszAlgorithm, "SGD") == 0); - FILE* fpIn = fopen(pszInstanceFName, "r"); - - ME_Model* pModel = new ME_Model(); - - char* pszLine = new char[100001]; - int iNumInstances = 0; - int iLen; - while (!feof(fpIn)) { - pszLine[0] = '\0'; - fgets(pszLine, 20000, fpIn); - if (strlen(pszLine) == 0) { - continue; - } - - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - iNumInstances++; - - ME_Sample* pmes = new ME_Sample(); - - char* p = strrchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - vector vecContext; - SplitOnWhitespace(string(pszLine), &vecContext); - - pmes->label = string(p); - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - pModel->add_training_sample((*pmes)); - if (iNumInstances % 100000 == 0) - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - delete pmes; - } - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - fclose(fpIn); - - if (strcmp(pszAlgorithm, "l1") == 0) - pModel->use_l1_regularizer(1.0); - else if (strcmp(pszAlgorithm, "l2") == 0) - pModel->use_l2_regularizer(1.0); - else - pModel->use_SGD(); - - pModel->train(); - pModel->save_to_file(pszModelFName); - - delete pModel; - fprintf(stdout, "......Finished Training\n"); - fprintf(stdout, "......Model saved as %s\n", pszModelFName); - delete[] pszLine; - } - - double fnEval(const char* pszContext, const char* pszOutcome) const { - vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - delete pmes; - int iLableID = m_pModel->get_class_id(pszOutcome); - return vecProb[iLableID]; - } - void fnEval(const char* pszContext, - vector >& vecOutput) const { - vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); - vecOutput.push_back(make_pair(label, vecProb[i])); - } - delete pmes; - } - void fnEval(const char* pszContext, vector& vecOutput) const { - vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); - vecOutput.push_back(vecProb[i]); - } - delete pmes; - } - int fnGetClassId(const string& strLabel) const { - return m_pModel->get_class_id(strLabel); - } - - private: - ME_Model* m_pModel; -}; - -#endif /* TSURUOKA_MAXENT_H_ */ diff --git a/utils/synutils/utility.h b/utils/synutils/utility.h deleted file mode 100644 index ef7b78b7..00000000 --- a/utils/synutils/utility.h +++ /dev/null @@ -1,127 +0,0 @@ -/* - * utility.h - * - * Created on: Jun 24, 2013 - * Author: lijunhui - */ - -#ifndef UTILITY_H_ -#define UTILITY_H_ - -#include -#include -#include -#include -#include - -#include -#include - -using namespace std; - -typedef std::unordered_map MapString2Int; -typedef std::unordered_map MapString2Float; -typedef std::unordered_map::iterator - MapString2FloatIterator; - -using namespace std; - -struct SFReader { - SFReader() {} - virtual ~SFReader() {} - - virtual bool fnReadNextLine(char* pszLine, int* piLength) = 0; - virtual bool fnReadNextLine(string& strLine) = 0; -}; - -struct STxtFileReader : public SFReader { - STxtFileReader(const char* pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~STxtFileReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - bool fnReadNextLine(char* pszLine, int* piLength) { - if (feof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - if (iLen == 0) return false; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(string& strLine) { - char* pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) - strLine = string(pszLine); - else - strLine = string(""); - delete[] pszLine; - - return bOut; - } - - private: - FILE* m_fpIn; -}; - -struct SGZFileReader : public SFReader { - SGZFileReader(const char* pszFname) { - m_fpIn = gzopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~SGZFileReader() { - if (m_fpIn != NULL) gzclose(m_fpIn); - } - - bool fnReadNextLine(char* pszLine, int* piLength) { - if (m_fpIn == NULL) exit(0); - if (gzeof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - gzgets(m_fpIn, pszLine, 10001); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(string& strLine) { - char* pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) - strLine = string(pszLine); - else - strLine = string(""); - delete[] pszLine; - - return bOut; - } - - private: - gzFile m_fpIn; -}; - -#endif /* UTILITY_H_ */ diff --git a/utils/tree.h b/utils/tree.h new file mode 100644 index 00000000..8070f828 --- /dev/null +++ b/utils/tree.h @@ -0,0 +1,700 @@ +/* + * tree.h + * + * Created on: May 23, 2013 + * Author: lijunhui + */ + +#ifndef TREE_H_ +#define TREE_H_ + +#include +#include + +#include +#include + +using namespace std; + +struct STreeItem { + STreeItem(const char *pszTerm) { + m_pszTerm = new char[strlen(pszTerm) + 1]; + strcpy(m_pszTerm, pszTerm); + + m_ptParent = NULL; + m_iBegin = -1; + m_iEnd = -1; + m_iHeadChild = -1; + m_iHeadWord = -1; + m_iBrotherIndex = -1; + } + ~STreeItem() { + delete[] m_pszTerm; + for (size_t i = 0; i < m_vecChildren.size(); i++) delete m_vecChildren[i]; + } + int fnAppend(STreeItem *ptChild) { + m_vecChildren.push_back(ptChild); + ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; + ptChild->m_ptParent = this; + return m_vecChildren.size() - 1; + } + int fnGetChildrenNum() { return m_vecChildren.size(); } + + bool fnIsPreTerminal(void) { + int I; + if (this == NULL || m_vecChildren.size() == 0) return false; + + for (I = 0; I < m_vecChildren.size(); I++) + if (m_vecChildren[I]->m_vecChildren.size() > 0) return false; + + return true; + } + + public: + char *m_pszTerm; + + vector m_vecChildren; // children items + STreeItem *m_ptParent; // the parent item + + int m_iBegin; + int m_iEnd; // the node span words[m_iBegin, m_iEnd] + int m_iHeadChild; // the index of its head child + int m_iHeadWord; // the index of its head word + int m_iBrotherIndex; // the index in his brothers +}; + +struct SGetHeadWord { + typedef vector CVectorStr; + SGetHeadWord() {} + ~SGetHeadWord() {} + int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { + // 0 indicating from right to left while 1 indicating from left to right + char szaHeadLists[201] = "0"; + + /* //head rules for Egnlish + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + if (strcmp(pszCFGLeft, "ADJP") == 0) + strcpy(szaHeadLists, "0ADJP JJ 0AD NN CS 0"); + else if (strcmp(pszCFGLeft, "ADVP") == 0) + strcpy(szaHeadLists, "0ADVP AD 0"); + else if (strcmp(pszCFGLeft, "CLP") == 0) + strcpy(szaHeadLists, "0CLP M 0"); + else if (strcmp(pszCFGLeft, "CP") == 0) + strcpy(szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0"); + else if (strcmp(pszCFGLeft, "DNP") == 0) + strcpy(szaHeadLists, "0DNP DEG 0DEC 0"); + else if (strcmp(pszCFGLeft, "DVP") == 0) + strcpy(szaHeadLists, "0DVP DEV 0"); + else if (strcmp(pszCFGLeft, "DP") == 0) + strcpy(szaHeadLists, "1DP DT 1"); + else if (strcmp(pszCFGLeft, "FRAG") == 0) + strcpy(szaHeadLists, "0VV NR NN 0"); + else if (strcmp(pszCFGLeft, "INTJ") == 0) + strcpy(szaHeadLists, "0INTJ IJ 0"); + else if (strcmp(pszCFGLeft, "LST") == 0) + strcpy(szaHeadLists, "1LST CD OD 1"); + else if (strcmp(pszCFGLeft, "IP") == 0) + strcpy(szaHeadLists, "0IP VP 0VV 0"); + // strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); + else if (strcmp(pszCFGLeft, "LCP") == 0) + strcpy(szaHeadLists, "0LCP LC 0"); + else if (strcmp(pszCFGLeft, "NP") == 0) + strcpy(szaHeadLists, "0NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "PP") == 0) + strcpy(szaHeadLists, "1PP P 1"); + else if (strcmp(pszCFGLeft, "PRN") == 0) + strcpy(szaHeadLists, "0 NP IP VP NT NR NN 0"); + else if (strcmp(pszCFGLeft, "QP") == 0) + strcpy(szaHeadLists, "0QP CLP CD OD 0"); + else if (strcmp(pszCFGLeft, "VP") == 0) + strcpy(szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1"); + else if (strcmp(pszCFGLeft, "VCD") == 0) + strcpy(szaHeadLists, "0VCD VV VA VC VE 0"); + if (strcmp(pszCFGLeft, "VRD") == 0) + strcpy(szaHeadLists, "0VRD VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VSB") == 0) + strcpy(szaHeadLists, "0VSB VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VCP") == 0) + strcpy(szaHeadLists, "0VCP VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VNV") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VPT") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "UCP") == 0) + strcpy(szaHeadLists, "0"); + else if (strcmp(pszCFGLeft, "WHNP") == 0) + strcpy(szaHeadLists, "0WHNP NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "WHPP") == 0) + strcpy(szaHeadLists, "1WHPP PP P 1"); + + /* //head rules for GENIA corpus + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + return fnMyOwnHeadWordRule(szaHeadLists, vectRight); + } + + private: + int fnMyOwnHeadWordRule(char *pszaHeadLists, CVectorStr vectRight) { + char szHeadList[201], *p; + char szTerm[101]; + int J; + + p = pszaHeadLists; + + int iCountRight; + + iCountRight = vectRight.size(); + + szHeadList[0] = '\0'; + while (1) { + szTerm[0] = '\0'; + sscanf(p, "%s", szTerm); + if (strlen(szHeadList) == 0) { + if (strcmp(szTerm, "0") == 0) { + return iCountRight - 1; + } + if (strcmp(szTerm, "1") == 0) { + return 0; + } + + sprintf(szHeadList, "%c %s ", szTerm[0], szTerm + 1); + p = strstr(p, szTerm); + p += strlen(szTerm); + } else { + if ((szTerm[0] == '0') || (szTerm[0] == '1')) { + if (szHeadList[0] == '0') { + for (J = iCountRight - 1; J >= 0; J--) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } else { + for (J = 0; J < iCountRight; J++) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } + + szHeadList[0] = '\0'; + } else { + strcat(szHeadList, szTerm); + strcat(szHeadList, " "); + + p = strstr(p, szTerm); + p += strlen(szTerm); + } + } + } + + return 0; + } +}; + +struct SParsedTree { + SParsedTree() { m_ptRoot = NULL; } + ~SParsedTree() { + if (m_ptRoot != NULL) delete m_ptRoot; + } + static SParsedTree *fnConvertFromString(const char *pszStr) { + if (strcmp(pszStr, "(())") == 0) return NULL; + SParsedTree *pTree = new SParsedTree(); + + vector vecSyn; + fnReadSyntactic(pszStr, vecSyn); + + int iLeft = 1, iRight = 1; //# left/right parenthesis + + STreeItem *pcurrent; + + pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); + + pcurrent = pTree->m_ptRoot; + + for (size_t i = 2; i < vecSyn.size() - 1; i++) { + if (strcmp(vecSyn[i].c_str(), "(") == 0) + iLeft++; + else if (strcmp(vecSyn[i].c_str(), ")") == 0) { + iRight++; + if (pcurrent == NULL) { + // error + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + pcurrent = pcurrent->m_ptParent; + } else { + STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); + pcurrent->fnAppend(ptNewItem); + pcurrent = ptNewItem; + + if (strcmp(vecSyn[i - 1].c_str(), "(") != 0 && + strcmp(vecSyn[i - 1].c_str(), ")") != 0) { + pTree->m_vecTerminals.push_back(ptNewItem); + pcurrent = pcurrent->m_ptParent; + } + } + } + + if (iLeft != iRight) { + // error + fprintf(stderr, "the left and right parentheses are not matched!"); + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + + return pTree; + } + + int fnGetNumWord() { return m_vecTerminals.size(); } + + void fnSetSpanInfo() { + int iNextNum = 0; + fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); + } + + void fnSetHeadWord() { + for (size_t i = 0; i < m_vecTerminals.size(); i++) + m_vecTerminals[i]->m_iHeadWord = i; + SGetHeadWord *pGetHeadWord = new SGetHeadWord(); + fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); + delete pGetHeadWord; + } + + STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { + STreeItem *pTreeItem = m_vecTerminals[iLeft]; + + while (pTreeItem->m_iEnd < iRight) { + pTreeItem = pTreeItem->m_ptParent; + if (pTreeItem == NULL) break; + } + if (pTreeItem == NULL) return NULL; + if (pTreeItem->m_iEnd > iRight) return NULL; + + assert(pTreeItem->m_iEnd == iRight); + if (bLowest) return pTreeItem; + + while (pTreeItem->m_ptParent != NULL && + pTreeItem->m_ptParent->fnGetChildrenNum() == 1) + pTreeItem = pTreeItem->m_ptParent; + + return pTreeItem; + } + + private: + void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int &iNextNum) { + int I; + int iNumChildren = ptItem->fnGetChildrenNum(); + for (I = 0; I < iNumChildren; I++) + fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[I], iNextNum); + + if (I == 0) { + ptItem->m_iBegin = iNextNum; + ptItem->m_iEnd = iNextNum++; + } else { + ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; + ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; + } + } + + void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, + SGetHeadWord *pGetHeadWord) { + int I, iHeadchild; + + if (ptItem->m_vecChildren.size() == 0) return; + + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); + + vector vecRight; + + if (ptItem->m_vecChildren.size() == 1) + iHeadchild = 0; + else { + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + vecRight.push_back(string(ptItem->m_vecChildren[I]->m_pszTerm)); + + iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); + } + + ptItem->m_iHeadChild = iHeadchild; + ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + + static void fnReadSyntactic(const char *pszSyn, vector &vec) { + char *p; + int I; + + int iLeftNum, iRightNum; + char *pszTmp, *pszTerm; + pszTmp = new char[strlen(pszSyn)]; + pszTerm = new char[strlen(pszSyn)]; + pszTmp[0] = pszTerm[0] = '\0'; + + vec.clear(); + + char *pszLine; + pszLine = new char[strlen(pszSyn) + 1]; + strcpy(pszLine, pszSyn); + + char *pszLine2; + + while (1) { + while ((strlen(pszLine) > 0) && (pszLine[strlen(pszLine) - 1] > 0) && + (pszLine[strlen(pszLine) - 1] <= ' ')) + pszLine[strlen(pszLine) - 1] = '\0'; + + if (strlen(pszLine) == 0) break; + + // printf( "%s\n", pszLine ); + pszLine2 = pszLine; + while (pszLine2[0] <= ' ') pszLine2++; + if (pszLine2[0] == '<') continue; + + sscanf(pszLine2 + 1, "%s", pszTmp); + + if (pszLine2[0] == '(') { + iLeftNum = 0; + iRightNum = 0; + } + + p = pszLine2; + while (1) { + pszTerm[0] = '\0'; + sscanf(p, "%s", pszTerm); + + if (strlen(pszTerm) == 0) break; + p = strstr(p, pszTerm); + p += strlen(pszTerm); + + if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { + if (pszTerm[0] == '(') { + vec.push_back(string("(")); + iLeftNum++; + + I = 1; + while (pszTerm[I] == '(' && pszTerm[I] != '\0') { + vec.push_back(string("(")); + iLeftNum++; + + I++; + } + + if (strlen(pszTerm) > 1) vec.push_back(string(pszTerm + I)); + } else { + char *pTmp; + pTmp = pszTerm + strlen(pszTerm) - 1; + while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; + pTmp[1] = '\0'; + + if (strlen(pszTerm) > 0) vec.push_back(string(pszTerm)); + pTmp += 2; + + for (I = 0; I <= (int)strlen(pTmp); I++) { + vec.push_back(string(")")); + iRightNum++; + } + } + } else { + char *q; + q = strchr(pszTerm, ')'); + if (q != NULL) { + q[0] = '\0'; + if (pszTerm[0] != '\0') vec.push_back(string(pszTerm)); + vec.push_back(string(")")); + iRightNum++; + + q++; + while (q[0] == ')') { + vec.push_back(string(")")); + q++; + iRightNum++; + } + + while (q[0] == '(') { + vec.push_back(string("(")); + q++; + iLeftNum++; + } + + if (q[0] != '\0') vec.push_back(string(q)); + } else + vec.push_back(string(pszTerm)); + } + } + + if (iLeftNum != iRightNum) { + fprintf(stderr, "%s\n", pszSyn); + assert(iLeftNum == iRightNum); + } + /*if ( iLeftNum != iRightNum ) { + printf( "ERROR: left( and right ) is not matched, %d ( and %d + )\n", iLeftNum, iRightNum ); + return; + }*/ + + if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { + //( (IP..) ) + std::vector::iterator it; + it = vec.begin(); + it++; + vec.insert(it, string("ROOT")); + } + + break; + } + + delete[] pszLine; + delete[] pszTmp; + delete[] pszTerm; + } + + public: + STreeItem *m_ptRoot; + vector m_vecTerminals; // the leaf nodes +}; + +struct SParseReader { + SParseReader(const char *pszParse_Fname, bool bFlattened = false) + : m_bFlattened(bFlattened) { + m_fpIn = fopen(pszParse_Fname, "r"); + assert(m_fpIn != NULL); + } + ~SParseReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + SParsedTree *fnReadNextParseTree() { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + pTree = SParsedTree::fnConvertFromString(pszLine); + if (pTree == NULL) break; + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + SParsedTree *fnReadNextParseTreeWithProb(double *pProb) { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + char *p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + if (pProb) (*pProb) = atof(pszLine); + + pTree = SParsedTree::fnConvertFromString(p); + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + private: + /* + * since to the parse tree is a flattened tree, use the head mark to identify + * head info. + * the head node will be marked as "*XP*" + */ + void fnSetParseTreeHeadInfo(SParsedTree *pTree) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) + pTree->m_vecTerminals[i]->m_iHeadWord = i; + fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); + } + + void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { + if (pTreeItem->m_vecChildren.size() == 0) return; + + for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) + fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); + + vector vecRight; + + int iHeadchild; + + if (pTreeItem->fnIsPreTerminal()) { + iHeadchild = 0; + } else { + size_t i; + for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { + char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; + if (p[0] == '*' && p[strlen(p) - 1] == '*') { + iHeadchild = i; + p[strlen(p) - 1] = '\0'; + string str = p + 1; + strcpy(p, str.c_str()); // erase the "*..*" + break; + } + } + assert(i < pTreeItem->m_vecChildren.size()); + } + + pTreeItem->m_iHeadChild = iHeadchild; + pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + void fnPostProcessingFlattenedParse(SParsedTree *pTree) { + pTree->fnSetSpanInfo(); + fnSetParseTreeHeadInfo(pTree); + } + bool fnReadNextSentence(char *pszLine, int *piLength) { + if (feof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + private: + FILE *m_fpIn; + const bool m_bFlattened; +}; + +#endif /* TREE_H_ */ diff --git a/utils/tsuruoka_maxent.h b/utils/tsuruoka_maxent.h new file mode 100644 index 00000000..e6bef232 --- /dev/null +++ b/utils/tsuruoka_maxent.h @@ -0,0 +1,155 @@ +/* + * tsuruoka_maxent.h + * + */ + +#ifndef TSURUOKA_MAXENT_H_ +#define TSURUOKA_MAXENT_H_ + +#include "synutils.h" +#include "stringlib.h" +#include "maxent.h" + +#include +#include +#include +#include +#include + +using namespace std; + +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; + +struct Tsuruoka_Maxent { + Tsuruoka_Maxent(const char* pszModelFName) { + if (pszModelFName != NULL) { + m_pModel = new ME_Model(); + m_pModel->load_from_file(pszModelFName); + } else + m_pModel = NULL; + } + + ~Tsuruoka_Maxent() { + if (m_pModel != NULL) delete m_pModel; + } + + void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, + const char* pszModelFName, int iNumIteration) { + assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || + strcmp(pszAlgorithm, "sgd") == 0 || + strcmp(pszAlgorithm, "SGD") == 0); + FILE* fpIn = fopen(pszInstanceFName, "r"); + + ME_Model* pModel = new ME_Model(); + + char* pszLine = new char[100001]; + int iNumInstances = 0; + int iLen; + while (!feof(fpIn)) { + pszLine[0] = '\0'; + fgets(pszLine, 20000, fpIn); + if (strlen(pszLine) == 0) { + continue; + } + + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + iNumInstances++; + + ME_Sample* pmes = new ME_Sample(); + + char* p = strrchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + vector vecContext; + SplitOnWhitespace(string(pszLine), &vecContext); + + pmes->label = string(p); + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + pModel->add_training_sample((*pmes)); + if (iNumInstances % 100000 == 0) + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + delete pmes; + } + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + fclose(fpIn); + + if (strcmp(pszAlgorithm, "l1") == 0) + pModel->use_l1_regularizer(1.0); + else if (strcmp(pszAlgorithm, "l2") == 0) + pModel->use_l2_regularizer(1.0); + else + pModel->use_SGD(); + + pModel->train(); + pModel->save_to_file(pszModelFName); + + delete pModel; + fprintf(stdout, "......Finished Training\n"); + fprintf(stdout, "......Model saved as %s\n", pszModelFName); + delete[] pszLine; + } + + double fnEval(const char* pszContext, const char* pszOutcome) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + delete pmes; + int iLableID = m_pModel->get_class_id(pszOutcome); + return vecProb[iLableID]; + } + void fnEval(const char* pszContext, + vector >& vecOutput) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(make_pair(label, vecProb[i])); + } + delete pmes; + } + void fnEval(const char* pszContext, vector& vecOutput) const { + vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + string label = m_pModel->get_class_label(i); + vecOutput.push_back(vecProb[i]); + } + delete pmes; + } + int fnGetClassId(const string& strLabel) const { + return m_pModel->get_class_id(strLabel); + } + + private: + ME_Model* m_pModel; +}; + +#endif /* TSURUOKA_MAXENT_H_ */ -- cgit v1.2.3 From 778df1879a0f38e4a8684e14e325f0a805ee8dcf Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Wed, 8 Oct 2014 07:32:42 -0400 Subject: Fix an out-of-bound access bug in maxent --- utils/maxent.cpp | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) (limited to 'utils') diff --git a/utils/maxent.cpp b/utils/maxent.cpp index 8d00ac1d..9115f6f2 100644 --- a/utils/maxent.cpp +++ b/utils/maxent.cpp @@ -98,7 +98,7 @@ int ME_Model::conditional_probability(const Sample& s, std::vector& membp) const { // int num_classes = membp.size(); double sum = 0; - int max_label = -1; + int max_label = 0; // double maxp = 0; vector powv(_num_classes, 0.0); @@ -134,7 +134,6 @@ int ME_Model::conditional_probability(const Sample& s, membp[label] /= sum; if (membp[label] > membp[max_label]) max_label = label; } - assert(max_label >= 0); return max_label; } -- cgit v1.2.3 From 5405acffd9c86ec80be4306807eda8e6ad284b89 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 10:33:23 -0400 Subject: Replace deprecated hash_map with unordred_map --- utils/maxent.cpp | 2 +- utils/maxent.h | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) (limited to 'utils') diff --git a/utils/maxent.cpp b/utils/maxent.cpp index 9115f6f2..0f49ee9d 100644 --- a/utils/maxent.cpp +++ b/utils/maxent.cpp @@ -142,7 +142,7 @@ int ME_Model::make_feature_bag(const int cutoff) { // count the occurrences of features #ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; + typedef std::unordered_map map_type; #else typedef std::map map_type; #endif diff --git a/utils/maxent.h b/utils/maxent.h index 88a04e25..b1efd88e 100644 --- a/utils/maxent.h +++ b/utils/maxent.h @@ -18,7 +18,7 @@ #define USE_HASH_MAP // if you encounter errors with hash, try commenting out // this line. (the program will be a bit slower, though) #ifdef USE_HASH_MAP -#include +#include #endif // @@ -156,7 +156,7 @@ class ME_Model { struct ME_FeatureBag { #ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; + typedef std::unordered_map map_type; #else typedef std::map map_type; #endif @@ -210,7 +210,7 @@ class ME_Model { struct MiniStringBag { #ifdef USE_HASH_MAP - typedef __gnu_cxx::hash_map map_type; + typedef std::unordered_map map_type; #else typedef std::map map_type; #endif -- cgit v1.2.3 From 8458a1851c91ae3b1effaa0f389bbb84c4fa795a Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 16:55:48 -0400 Subject: Remove using namespace std --- utils/alignment.h | 18 +++++------ utils/argument_reorder_model.h | 73 +++++++++++++++++++++++------------------- utils/srl_sentence.h | 26 +++++++-------- utils/synutils.h | 18 ++++------- utils/tree.h | 47 +++++++++++++-------------- utils/tsuruoka_maxent.h | 47 +++++++++++++-------------- 6 files changed, 113 insertions(+), 116 deletions(-) (limited to 'utils') diff --git a/utils/alignment.h b/utils/alignment.h index c0648aab..456577ca 100644 --- a/utils/alignment.h +++ b/utils/alignment.h @@ -15,8 +15,6 @@ #include "stringlib.h" -using namespace std; - /* * Note: * m_vec_s_align.size() may not be equal to the length of source side @@ -25,12 +23,12 @@ using namespace std; * */ struct SAlignment { - typedef vector SingleAlign; + typedef std::vector SingleAlign; SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } ~SAlignment() {} bool fnIsAligned(int i, bool s) const { - const vector* palign; + const std::vector* palign; if (s == true) palign = &m_vec_s_align; else @@ -70,7 +68,7 @@ struct SAlignment { } bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { - const vector* palign; + const std::vector* palign; if (s == true) palign = &m_vec_s_align; else @@ -97,7 +95,7 @@ struct SAlignment { * aligned to any word outside source[b, e] * 3) return "Discon't": otherwise; */ - string fnIsContinuous(int b, int e) const { + std::string fnIsContinuous(int b, int e) const { int ob, oe; fnGetLeftRightMost(b, e, true, ob, oe); if (ob == -1) return "Unaligned"; @@ -124,7 +122,7 @@ struct SAlignment { } private: - void fnGetLeftRightMost(int b, int e, const vector& align, + void fnGetLeftRightMost(int b, int e, const std::vector& align, int& ob, int& oe) const { ob = oe = -1; for (int i = b; i <= e && i < align.size(); i++) { @@ -139,7 +137,7 @@ struct SAlignment { m_vec_s_align.clear(); m_vec_t_align.clear(); - vector terms = SplitOnWhitespace(string(pszAlign)); + std::vector terms = SplitOnWhitespace(std::string(pszAlign)); int si, ti; for (size_t i = 0; i < terms.size(); i++) { sscanf(terms[i].c_str(), "%d-%d", &si, &ti); @@ -167,8 +165,8 @@ struct SAlignment { } private: - vector m_vec_s_align; // source side words' alignment - vector m_vec_t_align; // target side words' alignment + std::vector m_vec_s_align; // source side words' alignment + std::vector m_vec_t_align; // target side words' alignment }; struct SAlignmentReader { diff --git a/utils/argument_reorder_model.h b/utils/argument_reorder_model.h index 062b8841..077fa5ba 100644 --- a/utils/argument_reorder_model.h +++ b/utils/argument_reorder_model.h @@ -8,17 +8,20 @@ #ifndef ARGUMENT_REORDER_MODEL_H_ #define ARGUMENT_REORDER_MODEL_H_ +#include +#include + #include "alignment.h" #include "tree.h" #include "srl_sentence.h" // an argument item or a predicate item (the verb itself) struct SSRLItem { - SSRLItem(const STreeItem *tree_item, string role) + SSRLItem(const STreeItem *tree_item, std::string role) : tree_item_(tree_item), role_(role) {} ~SSRLItem() {} const STreeItem *tree_item_; - const string role_; + const std::string role_; }; struct SPredicateItem { @@ -26,11 +29,13 @@ struct SPredicateItem { : pred_(pred) { vec_items_.reserve(pred->m_vecArgt.size() + 1); for (int i = 0; i < pred->m_vecArgt.size(); i++) { - vec_items_.push_back(new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, - string(pred->m_vecArgt[i]->m_pszRole))); + vec_items_.push_back( + new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, + std::string(pred->m_vecArgt[i]->m_pszRole))); } - vec_items_.push_back(new SSRLItem( - tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, string("Pred"))); + vec_items_.push_back( + new SSRLItem(tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, + std::string("Pred"))); sort(vec_items_.begin(), vec_items_.end(), SortFunction); begin_ = vec_items_[0]->tree_item_->m_iBegin; @@ -43,7 +48,7 @@ struct SPredicateItem { return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); } - vector vec_items_; + std::vector vec_items_; int begin_; int end_; const SPredicate *pred_; @@ -51,13 +56,14 @@ struct SPredicateItem { struct SArgumentReorderModel { public: - static string fnGetBlockOutcome(int iBegin, int iEnd, SAlignment *pAlign) { + static std::string fnGetBlockOutcome(int iBegin, int iEnd, + SAlignment *pAlign) { return pAlign->fnIsContinuous(iBegin, iEnd); } static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, - vector &vecStrLeftReorder, - vector &vecStrRightReorder) { - vector vecLeft, vecRight; + std::vector &vecStrLeftReorder, + std::vector &vecStrRightReorder) { + std::vector vecLeft, vecRight; for (int i = 0; i < pPredItem->vec_items_.size(); i++) { const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; int iLeft1, iRight1; @@ -66,15 +72,15 @@ struct SArgumentReorderModel { vecLeft.push_back(iLeft1); vecRight.push_back(iRight1); } - vector vecLeftPosition; + std::vector vecLeftPosition; fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; + std::vector vecRightPosition; fnGetRelativePosition(vecRight, vecRightPosition); vecStrLeftReorder.clear(); vecStrRightReorder.clear(); for (int i = 1; i < vecLeftPosition.size(); i++) { - string strOutcome; + std::string strOutcome; fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); vecStrLeftReorder.push_back(strOutcome); fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); @@ -115,32 +121,33 @@ struct SArgumentReorderModel { static void fnGenerateFeature(const SParsedTree *pTree, const SPredicate *pPred, const SPredicateItem *pPredItem, int iPos, - const string &strBlock1, - const string &strBlock2, ostringstream &ostr) { + const std::string &strBlock1, + const std::string &strBlock2, + std::ostringstream &ostr) { SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; const STreeItem *pCon1 = pSRLItem1->tree_item_; const STreeItem *pCon2 = pSRLItem2->tree_item_; - string left_role = pSRLItem1->role_; - string right_role = pSRLItem2->role_; + std::string left_role = pSRLItem1->role_; + std::string right_role = pSRLItem2->role_; - string predicate_term = + std::string predicate_term = pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; - vector vec_other_right_sibling; + std::vector vec_other_right_sibling; for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) vec_other_right_sibling.push_back( - string(pPredItem->vec_items_[i]->role_)); + std::string(pPredItem->vec_items_[i]->role_)); if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); + vec_other_right_sibling.push_back(std::string("NULL")); - vector vec_other_left_sibling; + std::vector vec_other_left_sibling; for (int i = 0; i < iPos - 1; i++) vec_other_right_sibling.push_back( - string(pPredItem->vec_items_[i]->role_)); + std::string(pPredItem->vec_items_[i]->role_)); if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); + vec_other_left_sibling.push_back(std::string("NULL")); // generate features // f1 @@ -190,26 +197,26 @@ struct SArgumentReorderModel { } private: - static void fnGetOutcome(int i1, int i2, string &strOutcome) { + static void fnGetOutcome(int i1, int i2, std::string &strOutcome) { assert(i1 != i2); if (i1 < i2) { if (i2 > i1 + 1) - strOutcome = string("DM"); + strOutcome = std::string("DM"); else - strOutcome = string("M"); + strOutcome = std::string("M"); } else { if (i1 > i2 + 1) - strOutcome = string("DS"); + strOutcome = std::string("DS"); else - strOutcome = string("S"); + strOutcome = std::string("S"); } } - static void fnGetRelativePosition(const vector &vecLeft, - vector &vecPosition) { + static void fnGetRelativePosition(const std::vector &vecLeft, + std::vector &vecPosition) { vecPosition.clear(); - vector vec; + std::vector vec; for (int i = 0; i < vecLeft.size(); i++) { if (vecLeft[i] == -1) { if (i == 0) diff --git a/utils/srl_sentence.h b/utils/srl_sentence.h index 61532fb2..10f0414b 100644 --- a/utils/srl_sentence.h +++ b/utils/srl_sentence.h @@ -15,8 +15,6 @@ #include "tree.h" #include "stringlib.h" -using namespace std; - struct SArgument { SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { m_pszRole = new char[strlen(pszRole) + 1]; @@ -69,7 +67,7 @@ struct SPredicate { char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same // as the predicate itself int m_iPosition; // the position in sentence - vector m_vecArgt; // arguments associated to the predicate + std::vector m_vecArgt; // arguments associated to the predicate }; struct SSrlSentence { @@ -91,7 +89,7 @@ struct SSrlSentence { int GetPredicateNum() { return m_vecPred.size(); } SParsedTree* m_pTree; - vector m_vecPred; + std::vector m_vecPred; }; struct SSrlSentenceReader { @@ -116,7 +114,7 @@ struct SSrlSentenceReader { // TODO: here only considers flat predicate-argument structure // i.e., no overlap among them SSrlSentence* fnReadNextSrlSentence() { - vector > vecContent; + std::vector > vecContent; if (fnReadNextContent(vecContent) == false) return NULL; SSrlSentence* pSrlSentence = new SSrlSentence(); @@ -124,18 +122,18 @@ struct SSrlSentenceReader { // put together syntactic text std::ostringstream ostr; for (int i = 0; i < iSize; i++) { - string strSynSeg = + std::string strSynSeg = vecContent[i][5]; // the 5th column is the syntactic segment size_t iPosition = strSynSeg.find_first_of('*'); - assert(iPosition != string::npos); - ostringstream ostrTmp; + assert(iPosition != std::string::npos); + std::ostringstream ostrTmp; ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] << ")"; // the 2th column is POS-tag, and the 0th column is word strSynSeg.replace(iPosition, 1, ostrTmp.str()); fnReplaceAll(strSynSeg, "(", " ("); ostr << strSynSeg; } - string strSyn = ostr.str(); + std::string strSyn = ostr.str(); pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); pSrlSentence->m_pTree->fnSetHeadWord(); pSrlSentence->m_pTree->fnSetSpanInfo(); @@ -143,9 +141,9 @@ struct SSrlSentenceReader { // read predicate-argument structure int iNumPred = vecContent[0].size() - 8; for (int i = 0; i < iNumPred; i++) { - vector vecRole; - vector vecBegin; - vector vecEnd; + std::vector vecRole; + std::vector vecBegin; + std::vector vecEnd; int iPred = -1; for (int j = 0; j < iSize; j++) { const char* p = vecContent[j][i + 8].c_str(); @@ -184,7 +182,7 @@ struct SSrlSentenceReader { } private: - bool fnReadNextContent(vector >& vecContent) { + bool fnReadNextContent(std::vector >& vecContent) { vecContent.clear(); if (feof(m_fpIn) == true) return false; char* pszLine; @@ -200,7 +198,7 @@ struct SSrlSentenceReader { } if (iLen == 0) break; // end of this sentence - vector terms = SplitOnWhitespace(string(pszLine)); + std::vector terms = SplitOnWhitespace(std::string(pszLine)); assert(terms.size() > 7); vecContent.push_back(terms); } diff --git a/utils/synutils.h b/utils/synutils.h index ef7b78b7..f611553e 100644 --- a/utils/synutils.h +++ b/utils/synutils.h @@ -17,21 +17,17 @@ #include #include -using namespace std; - typedef std::unordered_map MapString2Int; typedef std::unordered_map MapString2Float; typedef std::unordered_map::iterator MapString2FloatIterator; -using namespace std; - struct SFReader { SFReader() {} virtual ~SFReader() {} virtual bool fnReadNextLine(char* pszLine, int* piLength) = 0; - virtual bool fnReadNextLine(string& strLine) = 0; + virtual bool fnReadNextLine(std::string& strLine) = 0; }; struct STxtFileReader : public SFReader { @@ -63,13 +59,13 @@ struct STxtFileReader : public SFReader { return true; } - bool fnReadNextLine(string& strLine) { + bool fnReadNextLine(std::string& strLine) { char* pszLine = new char[10001]; bool bOut = fnReadNextLine(pszLine, NULL); if (bOut) - strLine = string(pszLine); + strLine = std::string(pszLine); else - strLine = string(""); + strLine = std::string(""); delete[] pszLine; return bOut; @@ -108,13 +104,13 @@ struct SGZFileReader : public SFReader { return true; } - bool fnReadNextLine(string& strLine) { + bool fnReadNextLine(std::string& strLine) { char* pszLine = new char[10001]; bool bOut = fnReadNextLine(pszLine, NULL); if (bOut) - strLine = string(pszLine); + strLine = std::string(pszLine); else - strLine = string(""); + strLine = std::string(""); delete[] pszLine; return bOut; diff --git a/utils/tree.h b/utils/tree.h index 8070f828..ac80f6c7 100644 --- a/utils/tree.h +++ b/utils/tree.h @@ -14,8 +14,6 @@ #include #include -using namespace std; - struct STreeItem { STreeItem(const char *pszTerm) { m_pszTerm = new char[strlen(pszTerm) + 1]; @@ -53,7 +51,7 @@ struct STreeItem { public: char *m_pszTerm; - vector m_vecChildren; // children items + std::vector m_vecChildren; // children items STreeItem *m_ptParent; // the parent item int m_iBegin; @@ -64,7 +62,7 @@ struct STreeItem { }; struct SGetHeadWord { - typedef vector CVectorStr; + typedef std::vector CVectorStr; SGetHeadWord() {} ~SGetHeadWord() {} int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { @@ -311,7 +309,7 @@ struct SParsedTree { if (strcmp(pszStr, "(())") == 0) return NULL; SParsedTree *pTree = new SParsedTree(); - vector vecSyn; + std::vector vecSyn; fnReadSyntactic(pszStr, vecSyn); int iLeft = 1, iRight = 1; //# left/right parenthesis @@ -418,13 +416,13 @@ struct SParsedTree { for (I = 0; I < ptItem->m_vecChildren.size(); I++) fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); - vector vecRight; + std::vector vecRight; if (ptItem->m_vecChildren.size() == 1) iHeadchild = 0; else { for (I = 0; I < ptItem->m_vecChildren.size(); I++) - vecRight.push_back(string(ptItem->m_vecChildren[I]->m_pszTerm)); + vecRight.push_back(std::string(ptItem->m_vecChildren[I]->m_pszTerm)); iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); } @@ -433,7 +431,8 @@ struct SParsedTree { ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; } - static void fnReadSyntactic(const char *pszSyn, vector &vec) { + static void fnReadSyntactic(const char *pszSyn, + std::vector &vec) { char *p; int I; @@ -481,29 +480,29 @@ struct SParsedTree { if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { if (pszTerm[0] == '(') { - vec.push_back(string("(")); + vec.push_back(std::string("(")); iLeftNum++; I = 1; while (pszTerm[I] == '(' && pszTerm[I] != '\0') { - vec.push_back(string("(")); + vec.push_back(std::string("(")); iLeftNum++; I++; } - if (strlen(pszTerm) > 1) vec.push_back(string(pszTerm + I)); + if (strlen(pszTerm) > 1) vec.push_back(std::string(pszTerm + I)); } else { char *pTmp; pTmp = pszTerm + strlen(pszTerm) - 1; while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; pTmp[1] = '\0'; - if (strlen(pszTerm) > 0) vec.push_back(string(pszTerm)); + if (strlen(pszTerm) > 0) vec.push_back(std::string(pszTerm)); pTmp += 2; for (I = 0; I <= (int)strlen(pTmp); I++) { - vec.push_back(string(")")); + vec.push_back(std::string(")")); iRightNum++; } } @@ -512,26 +511,26 @@ struct SParsedTree { q = strchr(pszTerm, ')'); if (q != NULL) { q[0] = '\0'; - if (pszTerm[0] != '\0') vec.push_back(string(pszTerm)); - vec.push_back(string(")")); + if (pszTerm[0] != '\0') vec.push_back(std::string(pszTerm)); + vec.push_back(std::string(")")); iRightNum++; q++; while (q[0] == ')') { - vec.push_back(string(")")); + vec.push_back(std::string(")")); q++; iRightNum++; } while (q[0] == '(') { - vec.push_back(string("(")); + vec.push_back(std::string("(")); q++; iLeftNum++; } - if (q[0] != '\0') vec.push_back(string(q)); + if (q[0] != '\0') vec.push_back(std::string(q)); } else - vec.push_back(string(pszTerm)); + vec.push_back(std::string(pszTerm)); } } @@ -547,10 +546,10 @@ struct SParsedTree { if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { //( (IP..) ) - std::vector::iterator it; + std::vector::iterator it; it = vec.begin(); it++; - vec.insert(it, string("ROOT")); + vec.insert(it, std::string("ROOT")); } break; @@ -563,7 +562,7 @@ struct SParsedTree { public: STreeItem *m_ptRoot; - vector m_vecTerminals; // the leaf nodes + std::vector m_vecTerminals; // the leaf nodes }; struct SParseReader { @@ -645,7 +644,7 @@ struct SParseReader { for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); - vector vecRight; + std::vector vecRight; int iHeadchild; @@ -658,7 +657,7 @@ struct SParseReader { if (p[0] == '*' && p[strlen(p) - 1] == '*') { iHeadchild = i; p[strlen(p) - 1] = '\0'; - string str = p + 1; + std::string str = p + 1; strcpy(p, str.c_str()); // erase the "*..*" break; } diff --git a/utils/tsuruoka_maxent.h b/utils/tsuruoka_maxent.h index e6bef232..d1b062fe 100644 --- a/utils/tsuruoka_maxent.h +++ b/utils/tsuruoka_maxent.h @@ -6,17 +6,16 @@ #ifndef TSURUOKA_MAXENT_H_ #define TSURUOKA_MAXENT_H_ -#include "synutils.h" -#include "stringlib.h" -#include "maxent.h" - #include -#include -#include #include +#include #include +#include +#include -using namespace std; +#include "synutils.h" +#include "stringlib.h" +#include "maxent.h" typedef std::unordered_map Map; typedef std::unordered_map::iterator Iterator; @@ -67,10 +66,10 @@ struct Tsuruoka_Maxent { assert(p != NULL); p[0] = '\0'; p++; - vector vecContext; - SplitOnWhitespace(string(pszLine), &vecContext); + std::vector vecContext; + SplitOnWhitespace(std::string(pszLine), &vecContext); - pmes->label = string(p); + pmes->label = std::string(p); for (size_t i = 0; i < vecContext.size(); i++) pmes->add_feature(vecContext[i]); pModel->add_training_sample((*pmes)); @@ -98,53 +97,53 @@ struct Tsuruoka_Maxent { } double fnEval(const char* pszContext, const char* pszOutcome) const { - vector vecContext; + std::vector vecContext; ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); + SplitOnWhitespace(std::string(pszContext), &vecContext); for (size_t i = 0; i < vecContext.size(); i++) pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); + std::vector vecProb = m_pModel->classify(*pmes); delete pmes; int iLableID = m_pModel->get_class_id(pszOutcome); return vecProb[iLableID]; } void fnEval(const char* pszContext, - vector >& vecOutput) const { - vector vecContext; + std::vector >& vecOutput) const { + std::vector vecContext; ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); + SplitOnWhitespace(std::string(pszContext), &vecContext); vecOutput.clear(); for (size_t i = 0; i < vecContext.size(); i++) pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); + std::vector vecProb = m_pModel->classify(*pmes); for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); + std::string label = m_pModel->get_class_label(i); vecOutput.push_back(make_pair(label, vecProb[i])); } delete pmes; } - void fnEval(const char* pszContext, vector& vecOutput) const { - vector vecContext; + void fnEval(const char* pszContext, std::vector& vecOutput) const { + std::vector vecContext; ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(string(pszContext), &vecContext); + SplitOnWhitespace(std::string(pszContext), &vecContext); vecOutput.clear(); for (size_t i = 0; i < vecContext.size(); i++) pmes->add_feature(vecContext[i]); - vector vecProb = m_pModel->classify(*pmes); + std::vector vecProb = m_pModel->classify(*pmes); for (size_t i = 0; i < vecProb.size(); i++) { - string label = m_pModel->get_class_label(i); + std::string label = m_pModel->get_class_label(i); vecOutput.push_back(vecProb[i]); } delete pmes; } - int fnGetClassId(const string& strLabel) const { + int fnGetClassId(const std::string& strLabel) const { return m_pModel->get_class_id(strLabel); } -- cgit v1.2.3 From 9b87efca9cef49a76866ff419bb914a433d89830 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 16:58:44 -0400 Subject: Run clang-format --- decoder/apply_models.cc | 9 ++++--- decoder/ff_const_reorder.cc | 2 +- utils/argument_reorder_model.cc | 38 ++++++++++++++--------------- utils/constituent_reorder_model.cc | 50 +++++++++++++++++++------------------- utils/srl_sentence.h | 5 ++-- utils/tree.h | 8 +++--- 6 files changed, 56 insertions(+), 56 deletions(-) (limited to 'utils') diff --git a/decoder/apply_models.cc b/decoder/apply_models.cc index 278516c1..18c83fd4 100644 --- a/decoder/apply_models.cc +++ b/decoder/apply_models.cc @@ -276,7 +276,7 @@ public: models.EraseIgnoredBytes(&o_item_state); assert(item_state == o_item_state); } else { - assert(o_item->state_ == item->state_); // sanity check! + assert(o_item->state_ == item->state_); // sanity check! } o_item->est_prob_ = item->est_prob_; @@ -623,9 +623,10 @@ void ApplyModelSet(const Hypergraph& in, if (models.stateless() || config.algorithm == IntersectionConfiguration::FULL) { NoPruningRescorer ma(models, smeta, in, out); // avoid overhead of best-first when no state ma.Apply(); - } else if (config.algorithm == IntersectionConfiguration::CUBE - || config.algorithm == IntersectionConfiguration::FAST_CUBE_PRUNING - || config.algorithm == IntersectionConfiguration::FAST_CUBE_PRUNING_2) { + } else if (config.algorithm == IntersectionConfiguration::CUBE || + config.algorithm == IntersectionConfiguration::FAST_CUBE_PRUNING || + config.algorithm == + IntersectionConfiguration::FAST_CUBE_PRUNING_2) { int pl = config.pop_limit; const int max_pl_for_large=50; if (pl > max_pl_for_large && in.nodes_.size() > 80000) { diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc index fc774c3a..e51cc6cc 100644 --- a/decoder/ff_const_reorder.cc +++ b/decoder/ff_const_reorder.cc @@ -304,7 +304,7 @@ struct TargetTranslation { if (target_end == -1) target_begin = -1; } - const uint16_t begin_pos_, end_pos_; // the position in parse + const uint16_t begin_pos_, end_pos_; // the position in parse const uint16_t input_begin_pos_, input_end_pos_; // the position in input const uint16_t e_num_words_; vector align_; diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc index 58886251..537dfb20 100644 --- a/utils/argument_reorder_model.cc +++ b/utils/argument_reorder_model.cc @@ -64,13 +64,13 @@ inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, struct SArgumentReorderTrainer { SArgumentReorderTrainer( - const char* pszSRLFname, // source-side srl tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name + const char* pszSRLFname, // source-side srl tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name const char* pszTopPredicateFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelFname, // classifier model file name + const char* pszInstanceFname, // training instance file name + const char* pszModelFname, // classifier model file name int iCutoff) { fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, pszTargetFname, pszTopPredicateFname, @@ -110,14 +110,14 @@ struct SArgumentReorderTrainer { } void fnGenerateInstanceFiles( - const char* pszSRLFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename + const char* pszSRLFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename const char* pszSourceFname, // source file name const char* pszTargetFname, // target file name const char* pszTopPredicateFname, // top predicate file name (we only // consider predicates with 100+ // occurrences - const char* pszInstanceFname // training instance file name + const char* pszInstanceFname // training instance file name ) { SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); @@ -264,17 +264,17 @@ inline string str(char const* name, po::variables_map const& conf) { } //--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file -///scratch0/mt_exp/gale-align/gale-align.nw.al --source_file -///scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file -///scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file -///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix -///scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 +/// scratch0/mt_exp/gale-align/gale-align.nw.al --source_file +/// scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file +/// scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file +/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix +/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 //--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file -///scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file -///scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file -///scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file -///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix -///scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 +/// scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix +/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 int main(int argc, char** argv) { po::options_description opts("Configuration options"); diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc index 042c751b..d5c1cf5b 100644 --- a/utils/constituent_reorder_model.cc +++ b/utils/constituent_reorder_model.cc @@ -73,14 +73,14 @@ inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, struct SConstReorderTrainer { SConstReorderTrainer( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename const char* pszSourceFname, // source file name const char* pszTargetFname, // target file name const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type - int iCutoff, // feature count threshold + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold const char* pszOption // other classifier parameters (for svmlight) ) { fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, @@ -401,8 +401,8 @@ delete pZhangleMaxent;*/ } void fnGenerateInstanceFile( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename const char* pszSourceFname, // source file name const char* pszTargetFname, // target file name const char* pszInstanceFname // training instance file name @@ -507,8 +507,8 @@ delete pZhangleMaxent;*/ } void fnGenerateInstanceFile2( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename const char* pszSourceFname, // source file name const char* pszTargetFname, // target file name const char* pszInstanceFname // training instance file name @@ -582,13 +582,13 @@ struct SConstContTrainer { SConstContTrainer( const char* pszFlattenedSynFname, // source-side flattened parse tree // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type - int iCutoff, // feature count threshold + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iClassifierType, // classifier type + int iCutoff, // feature count threshold const char* pszOption // other classifier parameters (for svmlight) ) { fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, @@ -699,10 +699,10 @@ struct SConstContTrainer { void fnGenerateInstanceFile( const char* pszFlattenedSynFname, // source-side flattened parse tree // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name ) { SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); SParseReader* pParseReader = new SParseReader(pszFlattenedSynFname, true); @@ -781,11 +781,11 @@ inline string str(char const* name, po::variables_map const& conf) { } //--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file -///scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file -///scratch0/mt_exp/gq-ctb/data/train.cn --target_file -///scratch0/mt_exp/gq-ctb/data/train.en --instance_file -///scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix -///scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 +/// scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file +/// scratch0/mt_exp/gq-ctb/data/train.cn --target_file +/// scratch0/mt_exp/gq-ctb/data/train.en --instance_file +/// scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix +/// scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 //--classifier_type 1 int main(int argc, char** argv) { diff --git a/utils/srl_sentence.h b/utils/srl_sentence.h index 10f0414b..9d509600 100644 --- a/utils/srl_sentence.h +++ b/utils/srl_sentence.h @@ -8,7 +8,6 @@ #ifndef SRL_SENTENCE_H_ #define SRL_SENTENCE_H_ - #include #include @@ -36,7 +35,7 @@ struct SArgument { char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP int m_iBegin; - int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] + int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] float m_fProb; // the probability of this role, STreeItem* m_pTreeItem; }; @@ -66,7 +65,7 @@ struct SPredicate { char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same // as the predicate itself - int m_iPosition; // the position in sentence + int m_iPosition; // the position in sentence std::vector m_vecArgt; // arguments associated to the predicate }; diff --git a/utils/tree.h b/utils/tree.h index ac80f6c7..6c3406d6 100644 --- a/utils/tree.h +++ b/utils/tree.h @@ -52,12 +52,12 @@ struct STreeItem { char *m_pszTerm; std::vector m_vecChildren; // children items - STreeItem *m_ptParent; // the parent item + STreeItem *m_ptParent; // the parent item int m_iBegin; - int m_iEnd; // the node span words[m_iBegin, m_iEnd] - int m_iHeadChild; // the index of its head child - int m_iHeadWord; // the index of its head word + int m_iEnd; // the node span words[m_iBegin, m_iEnd] + int m_iHeadChild; // the index of its head child + int m_iHeadWord; // the index of its head word int m_iBrotherIndex; // the index in his brothers }; -- cgit v1.2.3 From 37cfe3758a07f414dfd28c36890567db2f2224c4 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 17:12:03 -0400 Subject: Trainers for reordering constraint models --- .gitignore | 1 + utils/Makefile.am | 4 +- utils/argument_reorder_model.cc | 6 ++ utils/constituent_reorder_model.cc | 197 +------------------------------------ 4 files changed, 13 insertions(+), 195 deletions(-) (limited to 'utils') diff --git a/.gitignore b/.gitignore index 4efcdd8d..c75328dc 100644 --- a/.gitignore +++ b/.gitignore @@ -181,6 +181,7 @@ training/mr_reduce_to_weights training/optimize_test training/plftools training/test_ngram +utils/argument_reorder_model_trainer utils/atools utils/bin/ utils/const_reorder_model_trainer diff --git a/utils/Makefile.am b/utils/Makefile.am index 0bd21b2b..53967561 100644 --- a/utils/Makefile.am +++ b/utils/Makefile.am @@ -1,4 +1,4 @@ -bin_PROGRAMS = reconstruct_weights atools const_reorder_model_trainer +bin_PROGRAMS = reconstruct_weights atools const_reorder_model_trainer argument_reorder_model_trainer noinst_PROGRAMS = \ ts \ @@ -108,6 +108,8 @@ atools_LDADD = libutils.a atools_LDFLAGS = $(STATIC_FLAGS) const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc const_reorder_model_trainer_LDADD = libutils.a +argument_reorder_model_trainer_SOURCES = argument_reorder_model.cc +argument_reorder_model_trainer_LDADD = libutils.a phmt_SOURCES = phmt.cc phmt_LDADD = libutils.a $(BOOST_UNIT_TEST_FRAMEWORK_LDFLAGS) $(BOOST_UNIT_TEST_FRAMEWORK_LIBS) diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc index 537dfb20..40647f8a 100644 --- a/utils/argument_reorder_model.cc +++ b/utils/argument_reorder_model.cc @@ -6,12 +6,18 @@ */ #include +#include #include +#include +#include +#include #include "argument_reorder_model.h" #include "synutils.h" #include "tsuruoka_maxent.h" +using namespace std; + inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { SFReader* pFReader = new STxtFileReader(pszFName); diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc index d5c1cf5b..5a5c5546 100644 --- a/utils/constituent_reorder_model.cc +++ b/utils/constituent_reorder_model.cc @@ -79,7 +79,6 @@ struct SConstReorderTrainer { const char* pszTargetFname, // target file name const char* pszInstanceFname, // training instance file name const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type int iCutoff, // feature count threshold const char* pszOption // other classifier parameters (for svmlight) ) { @@ -578,193 +577,6 @@ delete pZhangleMaxent;*/ } }; -struct SConstContTrainer { - SConstContTrainer( - const char* pszFlattenedSynFname, // source-side flattened parse tree - // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iClassifierType, // classifier type - int iCutoff, // feature count threshold - const char* pszOption // other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszFlattenedSynFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszInstanceFname); - // fnTraining(pszInstanceFname, pszModelPrefix, iClassifierType, iCutoff, - // pszOption); - fnTraining(pszInstanceFname, pszModelPrefix, iCutoff); - } - ~SConstContTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); - pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, - 2.0); - delete pZhangleMaxent;*/ - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszInstanceFname, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - void fnGetFocusedParentNodes(const SParsedTree* pTree, - vector& vecFocused) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - - // pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - // do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - inline void fnGetOutcome(int iL1, int iR1, const SAlignment* pAlign, - string& strOutcome) { - strOutcome = pAlign->fnIsContinuous(iL1, iR1); - } - - inline string fnGetLengthType(int iLen) { - if (iLen == 1) return string("1"); - if (iLen == 2) return string("2"); - if (iLen == 3) return string("3"); - if (iLen < 6) return string("4"); - if (iLen < 11) return string("6"); - return string("11"); - } - - /* - * Source side (11 features): - * f1: the syntactic category - * f2: the syntactic category of its parent - * f3: the head word's pos - * f4: =1 if it's the head of its parent node - * or - * the head of its parent node - * f5: length type - */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pCon1, - const SAlignment* pAlign, - const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, - ostringstream& ostr) { - - fnGetOutcome(pCon1->m_iBegin, pCon1->m_iEnd, pAlign, strOutcome); - - // generate features - // f1 - ostr << "f1=" << pCon1->m_pszTerm; - // f2 - ostr << " f2=" << pCon1->m_ptParent->m_pszTerm; - // f3 - ostr << " f3=" << pTree->m_vecTerminals[pCon1->m_iHeadWord] - ->m_ptParent->m_pszTerm; - // f4 - if (pCon1->m_iBrotherIndex == pCon1->m_ptParent->m_iHeadChild) { - ostr << " f4=1"; - } else { - ostr << " f4=" - << pCon1->m_ptParent->m_vecChildren[pCon1->m_ptParent->m_iHeadChild] - ->m_pszTerm; - } - // f5 - ostr << " f5=" << fnGetLengthType(pCon1->m_iEnd - pCon1->m_iBegin + 1); - } - - void fnGenerateInstanceFile( - const char* pszFlattenedSynFname, // source-side flattened parse tree - // file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszFlattenedSynFname, true); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - - FILE* fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - char* pszLine = new char[50001]; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(pTree != NULL); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); - vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); - vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; - i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem* pParent = vecFocused[i]; - - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent->m_vecChildren[j], pAlign, - vecSTerms, vecTTerms, strOutcome, ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); - } - } - - delete pAlign; - delete pTree; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - fclose(fpOut); - delete pAlignReader; - delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; - } -}; - inline void print_options(std::ostream& out, po::options_description const& opts) { typedef std::vector > Ds; @@ -786,7 +598,6 @@ inline string str(char const* name, po::variables_map const& conf) { /// scratch0/mt_exp/gq-ctb/data/train.en --instance_file /// scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix /// scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 -//--classifier_type 1 int main(int argc, char** argv) { po::options_description opts("Configuration options"); @@ -798,9 +609,7 @@ int main(int argc, char** argv) { "instance_file", po::value(), "Instance file path (output)")( "model_prefix", po::value(), "Model file path prefix (output): three files will be generated")( - "classifier_type", po::value()->default_value(1), - "Classifier type: 1 for openNLP maxent; 2 for Zhangle maxent; and 3 for " - "SVMLight")("feature_cutoff", po::value()->default_value(100), + "feature_cutoff", po::value()->default_value(100), "Feature cutoff threshold")( "svm_option", po::value(), "Parameters for SVMLight classifier")( "help", "produce help message"); @@ -839,8 +648,8 @@ int main(int argc, char** argv) { str("parse_file", vm).c_str(), str("align_file", vm).c_str(), str("source_file", vm).c_str(), str("target_file", vm).c_str(), str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), - vm["classifier_type"].as(), vm["feature_cutoff"].as(), pOption); + vm["feature_cutoff"].as(), pOption); delete pTrainer; - return 1; + return 0; } -- cgit v1.2.3 From 74521aa976e4deafd68c773b9230a1ddda0ca81e Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 17:12:33 -0400 Subject: Run clang-format --- utils/constituent_reorder_model.cc | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'utils') diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc index 5a5c5546..2ff8413f 100644 --- a/utils/constituent_reorder_model.cc +++ b/utils/constituent_reorder_model.cc @@ -609,9 +609,9 @@ int main(int argc, char** argv) { "instance_file", po::value(), "Instance file path (output)")( "model_prefix", po::value(), "Model file path prefix (output): three files will be generated")( - "feature_cutoff", po::value()->default_value(100), - "Feature cutoff threshold")( - "svm_option", po::value(), "Parameters for SVMLight classifier")( + "feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")("svm_option", po::value(), + "Parameters for SVMLight classifier")( "help", "produce help message"); po::variables_map vm; -- cgit v1.2.3 From 73abcb3a73ef5f24f29a3c3f9fda50bc431e9175 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Fri, 10 Oct 2014 17:56:53 -0400 Subject: Eliminate compiler warnings --- decoder/ff_const_reorder.cc | 2 +- utils/argument_reorder_model.cc | 2 +- utils/constituent_reorder_model.cc | 12 ++++++------ utils/lbfgs.h | 3 ++- utils/tsuruoka_maxent.h | 2 +- 5 files changed, 11 insertions(+), 10 deletions(-) (limited to 'utils') diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc index d29820ef..c0269125 100644 --- a/decoder/ff_const_reorder.cc +++ b/decoder/ff_const_reorder.cc @@ -1339,7 +1339,7 @@ void ConstReorderFeature::PrepareForInput(const SentenceMetadata& smeta) { void ConstReorderFeature::TraversalFeaturesImpl( const SentenceMetadata& /* smeta */, const Hypergraph::Edge& edge, const vector& ant_states, SparseVector* features, - SparseVector* estimated_features, void* state) const { + SparseVector* /*estimated_features*/, void* state) const { pimpl_->SetConstReorderFeature(edge, features, ant_states, state); } diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc index 40647f8a..5caf318f 100644 --- a/utils/argument_reorder_model.cc +++ b/utils/argument_reorder_model.cc @@ -263,7 +263,7 @@ inline void print_options(std::ostream& out, if (i) out << ' '; out << "--" << ds[i]->long_name(); } - out << '"\n'; + out << '\n'; } inline string str(char const* name, po::variables_map const& conf) { return conf[name].as(); diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc index 2ff8413f..df75a1a0 100644 --- a/utils/constituent_reorder_model.cc +++ b/utils/constituent_reorder_model.cc @@ -80,7 +80,7 @@ struct SConstReorderTrainer { const char* pszInstanceFname, // training instance file name const char* pszModelPrefix, // classifier model file name prefix int iCutoff, // feature count threshold - const char* pszOption // other classifier parameters (for svmlight) + const char* /*pszOption*/ // other classifier parameters (for svmlight) ) { fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, pszTargetFname, pszInstanceFname); @@ -134,14 +134,14 @@ delete pZhangleMaxent;*/ } inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, - const SAlignment* pAlign, string& strOutcome) { + const SAlignment* /*pAlign*/, string& strOutcome) { if (iL1 == -1 && iL2 == -1) strOutcome = "BU"; // 1. both are untranslated else if (iL1 == -1) strOutcome = "1U"; // 2. XP1 is untranslated else if (iL2 == -1) strOutcome = "2U"; // 3. XP2 is untranslated - else if (iL1 == iL2 && iR2 == iR2) + else if (iL1 == iL2 && iR1 == iR2) strOutcome = "SS"; // 4. Have same scope else if (iL1 <= iL2 && iR1 >= iR2) strOutcome = "1C2"; // 5. XP1's translation covers XP2's @@ -240,7 +240,7 @@ delete pZhangleMaxent;*/ int iPos, const vector& vecChunkStatus, const vector& vecPosition, const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, + const vector& /*vecTTerms*/, string& strOutcome, ostringstream& ostr) { STreeItem* pCon1, *pCon2; pCon1 = pParent->m_vecChildren[iPos - 1]; @@ -313,11 +313,11 @@ delete pZhangleMaxent;*/ * f8: the first and the last word of XP2's translation (f8_f, f8_l) * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + void fnGenerateInstance(const SParsedTree* /*pTree*/, const STreeItem* pParent, const STreeItem* pCon1, const STreeItem* pCon2, const SAlignment* pAlign, const vector& vecSTerms, - const vector& vecTTerms, string& strOutcome, + const vector& /*vecTTerms*/, string& strOutcome, ostringstream& ostr) { int iLeft1, iRight1, iLeft2, iRight2; diff --git a/utils/lbfgs.h b/utils/lbfgs.h index ed5cd944..4d706f7a 100644 --- a/utils/lbfgs.h +++ b/utils/lbfgs.h @@ -1,6 +1,8 @@ #ifndef _LBFGS_H_ #define _LBFGS_H_ +#include + // template // std::vector // perform_LBFGS(FuncGrad func_grad, const std::vector & x0); @@ -13,7 +15,6 @@ std::vector perform_OWLQN( double (*func_grad)(const std::vector &, std::vector &), const std::vector &x0, const double C); -// const int LBFGS_M = 7; const int LBFGS_M = 10; #endif diff --git a/utils/tsuruoka_maxent.h b/utils/tsuruoka_maxent.h index d1b062fe..550a4b7f 100644 --- a/utils/tsuruoka_maxent.h +++ b/utils/tsuruoka_maxent.h @@ -34,7 +34,7 @@ struct Tsuruoka_Maxent { } void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, - const char* pszModelFName, int iNumIteration) { + const char* pszModelFName, int /*iNumIteration*/) { assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || strcmp(pszAlgorithm, "sgd") == 0 || strcmp(pszAlgorithm, "SGD") == 0); -- cgit v1.2.3 From e39742355186ebf59a369b2e5b369648047ca216 Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Sat, 6 Dec 2014 10:37:48 -0500 Subject: Remove synutils.h --- decoder/ff_const_reorder.cc | 13 ++-- utils/argument_reorder_model.cc | 126 ++++++++++++++------------------- utils/constituent_reorder_model.cc | 138 ++++++++++++++++--------------------- utils/synutils.h | 123 --------------------------------- utils/tsuruoka_maxent.h | 1 - 5 files changed, 119 insertions(+), 282 deletions(-) delete mode 100644 utils/synutils.h (limited to 'utils') diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc index c0269125..8da551c6 100644 --- a/decoder/ff_const_reorder.cc +++ b/decoder/ff_const_reorder.cc @@ -1,5 +1,6 @@ #include "ff_const_reorder.h" +#include "filelib.h" #include "stringlib.h" #include "hg.h" #include "sentence_metadata.h" @@ -383,13 +384,13 @@ struct SIndexMap { index_map_parse_2_input = NULL; return; } - STxtFileReader* reader = new STxtFileReader(index_map_file.c_str()); - char szLine[10001]; - szLine[0] = '\0'; - reader->fnReadNextLine(szLine, NULL); - delete reader; vector terms; - SplitOnWhitespace(string(szLine), &terms); + { + ReadFile file(index_map_file); + string line; + assert(getline(*file.stream(), line)); + SplitOnWhitespace(line, &terms); + } index_map_input_2_parse = new short int[terms.size() + 1]; int ix = 0; diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc index 5caf318f..c4e90cba 100644 --- a/utils/argument_reorder_model.cc +++ b/utils/argument_reorder_model.cc @@ -12,60 +12,49 @@ #include #include +#include "filelib.h" + #include "argument_reorder_model.h" -#include "synutils.h" #include "tsuruoka_maxent.h" using namespace std; inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { - SFReader* pFReader = new STxtFileReader(pszFName); - char* pszLine = new char[100001]; - int iLen; Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; + { + ReadFile in(pszFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + for (const auto& i : terms) { + ++hashPredicate[i]; } } } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE* fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; + + { + ReadFile in(pszFName); + WriteFile out(pszNewFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + bool written = false; + for (const auto& i : terms) { + if (hashPredicate[i] >= iCutoff) { + (*out.stream()) << i << " "; + written = true; + } + } + if (written) { + (*out.stream()) << "\n"; } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); } } - fclose(fpOut); - delete pFReader; - - delete[] pszLine; } struct SArgumentReorderTrainer { @@ -127,8 +116,8 @@ struct SArgumentReorderTrainer { ) { SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); Map* pMapPredicate; if (pszTopPredicateFname != NULL) @@ -136,13 +125,10 @@ struct SArgumentReorderTrainer { else pMapPredicate = NULL; - char* pszLine = new char[50001]; + string line; - FILE* fpLeftOut, *fpRightOut; - sprintf(pszLine, "%s.left", pszInstanceFname); - fpLeftOut = fopen(pszLine, "w"); - sprintf(pszLine, "%s.right", pszInstanceFname); - fpRightOut = fopen(pszLine, "w"); + WriteFile left_file(pszInstanceFname + string(".left")); + WriteFile right_file(pszInstanceFname + string(".right")); // read sentence by sentence SAlignment* pAlign; @@ -153,12 +139,12 @@ struct SArgumentReorderTrainer { pSRL = pSRLReader->fnReadNextSrlSentence(); assert(pSRL != NULL); pTree = pSRL->m_pTree; - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + assert(getline(*source_file.stream(), line)); vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + SplitOnWhitespace(line, &vecSTerms); + assert(getline(*target_file.stream(), line)); vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); + SplitOnWhitespace(line, &vecTTerms); // vecTPOSTerms.size() == 0, given the case when an english sentence fails // parsing @@ -204,10 +190,10 @@ struct SArgumentReorderTrainer { // strOutcome.c_str()); // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), // strOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str()); - fprintf(fpRightOut, "%s %s\n", ostr.str().c_str(), - strRightOutcome.c_str()); + (*left_file.stream()) << ostr.str() << " " << strLeftOutcome + << "\n"; + (*right_file.stream()) << ostr.str() << " " << strRightOutcome + << "\n"; } } } @@ -218,36 +204,28 @@ struct SArgumentReorderTrainer { if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); } - delete[] pszLine; - - fclose(fpLeftOut); - fclose(fpRightOut); delete pAlignReader; delete pSRLReader; - delete pTxtSReader; - delete pTxtTReader; } Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { if (pszTopPredicateFname == NULL) return NULL; Map* pMapPredicate = new Map(); - STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); - char* pszLine = new char[50001]; + // STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); + ReadFile in(pszTopPredicateFname); + // char* pszLine = new char[50001]; + string line; int iNumCount = 0; - while (pReader->fnReadNextLine(pszLine, NULL)) { - if (pszLine[0] == '#') continue; - char* p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - int iCount = atoi(p); + while (getline(*in.stream(), line)) { + if (line.size() && line[0] == '#') continue; + auto p = line.find(' '); + assert(p != string::npos); + int iCount = atoi(line.substr(p + 1).c_str()); if (iCount < 100) break; - (*pMapPredicate)[string(pszLine)] = iNumCount++; + (*pMapPredicate)[line] = iNumCount++; } - delete pReader; - delete[] pszLine; return pMapPredicate; } }; diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc index df75a1a0..bdb7c5d1 100644 --- a/utils/constituent_reorder_model.cc +++ b/utils/constituent_reorder_model.cc @@ -5,15 +5,17 @@ * Author: junhuili */ +#include +#include + #include +#include "filelib.h" + #include "alignment.h" #include "tree.h" -#include "synutils.h" #include "tsuruoka_maxent.h" -#include - using namespace std; typedef std::unordered_map Map; @@ -23,52 +25,40 @@ namespace po = boost::program_options; inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, const char* pszNewFName) { - SFReader* pFReader = new STxtFileReader(pszFName); - char* pszLine = new char[100001]; - int iLen; Map hashPredicate; - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - if (iter == hashPredicate.end()) { - hashPredicate[vecTerms[i]] = 1; - - } else { - iter->second++; + { + ReadFile f(pszFName); + string line; + while (getline(*f.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + for (const auto& i : terms) { + ++hashPredicate[i]; } } } - delete pFReader; - - pFReader = new STxtFileReader(pszFName); - FILE* fpOut = fopen(pszNewFName, "w"); - while (pFReader->fnReadNextLine(pszLine, &iLen)) { - if (iLen == 0) continue; - - vector vecTerms; - SplitOnWhitespace(string(pszLine), &vecTerms); - ostringstream ostr; - for (size_t i = 0; i < vecTerms.size() - 1; i++) { - Iterator iter = hashPredicate.find(vecTerms[i]); - assert(iter != hashPredicate.end()); - if (iter->second >= iCutoff) { - ostr << vecTerms[i] << " "; + + { + ReadFile in(pszFName); + WriteFile out(pszNewFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + bool written = false; + for (const auto& i : terms) { + if (hashPredicate[i] >= iCutoff) { + (*out.stream()) << i << " "; + written = true; + } + } + if (written) { + (*out.stream()) << "\n"; } - } - if (ostr.str().length() > 0) { - ostr << vecTerms[vecTerms.size() - 1]; - fprintf(fpOut, "%s\n", ostr.str().c_str()); } } - fclose(fpOut); - delete pFReader; - - delete[] pszLine; } struct SConstReorderTrainer { @@ -408,31 +398,29 @@ delete pZhangleMaxent;*/ ) { SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); SParseReader* pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - FILE* fpLeftOut = fopen(strInstanceLeftFname.c_str(), "w"); - assert(fpLeftOut != NULL); - - FILE* fpRightOut = fopen(strInstanceRightFname.c_str(), "w"); - assert(fpRightOut != NULL); + WriteFile left_file(strInstanceLeftFname); + WriteFile right_file(strInstanceRightFname); // read sentence by sentence SAlignment* pAlign; SParsedTree* pTree; - char* pszLine = new char[50001]; + string line; int iSentNum = 0; while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + + assert(getline(*source_file.stream(), line)); vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + SplitOnWhitespace(line, &vecSTerms); + + assert(getline(*target_file.stream(), line)); vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); + SplitOnWhitespace(line, &vecTTerms); if (pTree != NULL) { @@ -475,16 +463,18 @@ delete pZhangleMaxent;*/ vecLeftPosition, vecSTerms, vecTTerms, strLeftOutcome, ostr); + string ostr_str = ostr.str(); + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), // strLeftOutcome.c_str()); - fprintf(fpLeftOut, "%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str()); + (*left_file.stream()) << ostr_str << " " << strLeftOutcome << "\n"; string strRightOutcome; fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], strRightOutcome); - fprintf(fpRightOut, "%s LeftOrder=%s %s\n", ostr.str().c_str(), - strLeftOutcome.c_str(), strRightOutcome.c_str()); + (*right_file.stream()) << ostr_str + << " LeftOrder=" << strLeftOutcome << " " + << strRightOutcome << "\n"; } } delete pTree; @@ -496,13 +486,8 @@ delete pZhangleMaxent;*/ if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); } - fclose(fpLeftOut); - fclose(fpRightOut); delete pAlignReader; delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; } void fnGenerateInstanceFile2( @@ -514,25 +499,26 @@ delete pZhangleMaxent;*/ ) { SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); SParseReader* pParseReader = new SParseReader(pszSynFname, false); - STxtFileReader* pTxtSReader = new STxtFileReader(pszSourceFname); - STxtFileReader* pTxtTReader = new STxtFileReader(pszTargetFname); - FILE* fpOut = fopen(pszInstanceFname, "w"); - assert(fpOut != NULL); + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); + + WriteFile output_file(pszInstanceFname); // read sentence by sentence SAlignment* pAlign; SParsedTree* pTree; - char* pszLine = new char[50001]; + string line; int iSentNum = 0; while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { pTree = pParseReader->fnReadNextParseTree(); - assert(pTxtSReader->fnReadNextLine(pszLine, NULL)); + assert(getline(*source_file.stream(), line)); vector vecSTerms; - SplitOnWhitespace(string(pszLine), &vecSTerms); - assert(pTxtTReader->fnReadNextLine(pszLine, NULL)); + SplitOnWhitespace(line, &vecSTerms); + + assert(getline(*target_file.stream(), line)); vector vecTTerms; - SplitOnWhitespace(string(pszLine), &vecTTerms); + SplitOnWhitespace(line, &vecTTerms); if (pTree != NULL) { @@ -556,7 +542,7 @@ delete pZhangleMaxent;*/ // fprintf(stderr, "%s %s\n", ostr.str().c_str(), // strOutcome.c_str()); - fprintf(fpOut, "%s %s\n", ostr.str().c_str(), strOutcome.c_str()); + (*output_file.stream()) << ostr.str() << " " << strOutcome << "\n"; } } delete pTree; @@ -568,12 +554,8 @@ delete pZhangleMaxent;*/ if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); } - fclose(fpOut); delete pAlignReader; delete pParseReader; - delete pTxtSReader; - delete pTxtTReader; - delete[] pszLine; } }; diff --git a/utils/synutils.h b/utils/synutils.h deleted file mode 100644 index f611553e..00000000 --- a/utils/synutils.h +++ /dev/null @@ -1,123 +0,0 @@ -/* - * utility.h - * - * Created on: Jun 24, 2013 - * Author: lijunhui - */ - -#ifndef UTILITY_H_ -#define UTILITY_H_ - -#include -#include -#include -#include -#include - -#include -#include - -typedef std::unordered_map MapString2Int; -typedef std::unordered_map MapString2Float; -typedef std::unordered_map::iterator - MapString2FloatIterator; - -struct SFReader { - SFReader() {} - virtual ~SFReader() {} - - virtual bool fnReadNextLine(char* pszLine, int* piLength) = 0; - virtual bool fnReadNextLine(std::string& strLine) = 0; -}; - -struct STxtFileReader : public SFReader { - STxtFileReader(const char* pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~STxtFileReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - bool fnReadNextLine(char* pszLine, int* piLength) { - if (feof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - if (iLen == 0) return false; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(std::string& strLine) { - char* pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) - strLine = std::string(pszLine); - else - strLine = std::string(""); - delete[] pszLine; - - return bOut; - } - - private: - FILE* m_fpIn; -}; - -struct SGZFileReader : public SFReader { - SGZFileReader(const char* pszFname) { - m_fpIn = gzopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~SGZFileReader() { - if (m_fpIn != NULL) gzclose(m_fpIn); - } - - bool fnReadNextLine(char* pszLine, int* piLength) { - if (m_fpIn == NULL) exit(0); - if (gzeof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - gzgets(m_fpIn, pszLine, 10001); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - bool fnReadNextLine(std::string& strLine) { - char* pszLine = new char[10001]; - bool bOut = fnReadNextLine(pszLine, NULL); - if (bOut) - strLine = std::string(pszLine); - else - strLine = std::string(""); - delete[] pszLine; - - return bOut; - } - - private: - gzFile m_fpIn; -}; - -#endif /* UTILITY_H_ */ diff --git a/utils/tsuruoka_maxent.h b/utils/tsuruoka_maxent.h index 550a4b7f..82da44ff 100644 --- a/utils/tsuruoka_maxent.h +++ b/utils/tsuruoka_maxent.h @@ -13,7 +13,6 @@ #include #include -#include "synutils.h" #include "stringlib.h" #include "maxent.h" -- cgit v1.2.3 From f74345e54b9f5f5f894d81df61e85fddc53a520a Mon Sep 17 00:00:00 2001 From: "Wu, Ke" Date: Sat, 6 Dec 2014 12:17:27 -0500 Subject: Move non-MaxEnt code out of utils 1. alignment.h, argument_reorder_model.h, src_sentence.h, tree.h, tsuruoka_maxent.h -> decoder/ff_const_reorder_common.h. 2. Trainers source files (argument_reorder_model.cc and constituent_reorder_model.cc) are moved to training/const_reorder. --- .gitignore | 4 +- configure.ac | 1 + decoder/Makefile.am | 1 + decoder/ff_const_reorder.cc | 6 +- decoder/ff_const_reorder_common.h | 1441 ++++++++++++++++++++ training/Makefile.am | 4 +- training/const_reorder/Makefile.am | 8 + training/const_reorder/argument_reorder_model.cc | 307 +++++ .../const_reorder/constituent_reorder_model.cc | 636 +++++++++ utils/Makefile.am | 12 +- utils/alignment.h | 200 --- utils/argument_reorder_model.cc | 307 ----- utils/argument_reorder_model.h | 246 ---- utils/constituent_reorder_model.cc | 637 --------- utils/srl_sentence.h | 211 --- utils/tree.h | 699 ---------- utils/tsuruoka_maxent.h | 153 --- 17 files changed, 2401 insertions(+), 2472 deletions(-) create mode 100644 decoder/ff_const_reorder_common.h create mode 100644 training/const_reorder/Makefile.am create mode 100644 training/const_reorder/argument_reorder_model.cc create mode 100644 training/const_reorder/constituent_reorder_model.cc delete mode 100644 utils/alignment.h delete mode 100644 utils/argument_reorder_model.cc delete mode 100644 utils/argument_reorder_model.h delete mode 100644 utils/constituent_reorder_model.cc delete mode 100644 utils/srl_sentence.h delete mode 100644 utils/tree.h delete mode 100644 utils/tsuruoka_maxent.h (limited to 'utils') diff --git a/.gitignore b/.gitignore index c75328dc..fccfcce7 100644 --- a/.gitignore +++ b/.gitignore @@ -181,10 +181,10 @@ training/mr_reduce_to_weights training/optimize_test training/plftools training/test_ngram -utils/argument_reorder_model_trainer +training/const_reorder/argument_reorder_model_trainer +training/const_reorder/const_reorder_model_trainer utils/atools utils/bin/ -utils/const_reorder_model_trainer utils/crp_test utils/dict_test utils/logval_test diff --git a/configure.ac b/configure.ac index 0b1f04f0..aaad9a78 100644 --- a/configure.ac +++ b/configure.ac @@ -235,6 +235,7 @@ AC_CONFIG_FILES([training/minrisk/Makefile]) AC_CONFIG_FILES([training/mira/Makefile]) AC_CONFIG_FILES([training/latent_svm/Makefile]) AC_CONFIG_FILES([training/dtrain/Makefile]) +AC_CONFIG_FILES([training/const_reorder/Makefile]) # external feature function example code AC_CONFIG_FILES([example_extff/Makefile]) diff --git a/decoder/Makefile.am b/decoder/Makefile.am index cc52ccd4..f791b3cc 100644 --- a/decoder/Makefile.am +++ b/decoder/Makefile.am @@ -45,6 +45,7 @@ libcdec_a_SOURCES = \ ff_basic.h \ ff_bleu.h \ ff_charset.h \ + ff_const_reorder_common.h \ ff_const_reorder.h \ ff_context.h \ ff_csplit.h \ diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc index 8da551c6..e4fb4725 100644 --- a/decoder/ff_const_reorder.cc +++ b/decoder/ff_const_reorder.cc @@ -4,11 +4,8 @@ #include "stringlib.h" #include "hg.h" #include "sentence_metadata.h" -#include "tree.h" -#include "srl_sentence.h" -#include "tsuruoka_maxent.h" #include "hash.h" -#include "argument_reorder_model.h" +#include "ff_const_reorder_common.h" #include #include @@ -16,6 +13,7 @@ #include using namespace std; +using namespace const_reorder; typedef HASH_MAP > MapClassifier; diff --git a/decoder/ff_const_reorder_common.h b/decoder/ff_const_reorder_common.h new file mode 100644 index 00000000..7c111de3 --- /dev/null +++ b/decoder/ff_const_reorder_common.h @@ -0,0 +1,1441 @@ +#ifndef _FF_CONST_REORDER_COMMON_H +#define _FF_CONST_REORDER_COMMON_H + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "maxent.h" +#include "stringlib.h" + +namespace const_reorder { + +struct STreeItem { + STreeItem(const char *pszTerm) { + m_pszTerm = new char[strlen(pszTerm) + 1]; + strcpy(m_pszTerm, pszTerm); + + m_ptParent = NULL; + m_iBegin = -1; + m_iEnd = -1; + m_iHeadChild = -1; + m_iHeadWord = -1; + m_iBrotherIndex = -1; + } + ~STreeItem() { + delete[] m_pszTerm; + for (size_t i = 0; i < m_vecChildren.size(); i++) delete m_vecChildren[i]; + } + int fnAppend(STreeItem *ptChild) { + m_vecChildren.push_back(ptChild); + ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; + ptChild->m_ptParent = this; + return m_vecChildren.size() - 1; + } + int fnGetChildrenNum() { return m_vecChildren.size(); } + + bool fnIsPreTerminal(void) { + int I; + if (this == NULL || m_vecChildren.size() == 0) return false; + + for (I = 0; I < m_vecChildren.size(); I++) + if (m_vecChildren[I]->m_vecChildren.size() > 0) return false; + + return true; + } + + public: + char *m_pszTerm; + + std::vector m_vecChildren; // children items + STreeItem *m_ptParent; // the parent item + + int m_iBegin; + int m_iEnd; // the node span words[m_iBegin, m_iEnd] + int m_iHeadChild; // the index of its head child + int m_iHeadWord; // the index of its head word + int m_iBrotherIndex; // the index in his brothers +}; + +struct SGetHeadWord { + typedef std::vector CVectorStr; + SGetHeadWord() {} + ~SGetHeadWord() {} + int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { + // 0 indicating from right to left while 1 indicating from left to right + char szaHeadLists[201] = "0"; + + /* //head rules for Egnlish + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + if (strcmp(pszCFGLeft, "ADJP") == 0) + strcpy(szaHeadLists, "0ADJP JJ 0AD NN CS 0"); + else if (strcmp(pszCFGLeft, "ADVP") == 0) + strcpy(szaHeadLists, "0ADVP AD 0"); + else if (strcmp(pszCFGLeft, "CLP") == 0) + strcpy(szaHeadLists, "0CLP M 0"); + else if (strcmp(pszCFGLeft, "CP") == 0) + strcpy(szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0"); + else if (strcmp(pszCFGLeft, "DNP") == 0) + strcpy(szaHeadLists, "0DNP DEG 0DEC 0"); + else if (strcmp(pszCFGLeft, "DVP") == 0) + strcpy(szaHeadLists, "0DVP DEV 0"); + else if (strcmp(pszCFGLeft, "DP") == 0) + strcpy(szaHeadLists, "1DP DT 1"); + else if (strcmp(pszCFGLeft, "FRAG") == 0) + strcpy(szaHeadLists, "0VV NR NN 0"); + else if (strcmp(pszCFGLeft, "INTJ") == 0) + strcpy(szaHeadLists, "0INTJ IJ 0"); + else if (strcmp(pszCFGLeft, "LST") == 0) + strcpy(szaHeadLists, "1LST CD OD 1"); + else if (strcmp(pszCFGLeft, "IP") == 0) + strcpy(szaHeadLists, "0IP VP 0VV 0"); + // strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); + else if (strcmp(pszCFGLeft, "LCP") == 0) + strcpy(szaHeadLists, "0LCP LC 0"); + else if (strcmp(pszCFGLeft, "NP") == 0) + strcpy(szaHeadLists, "0NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "PP") == 0) + strcpy(szaHeadLists, "1PP P 1"); + else if (strcmp(pszCFGLeft, "PRN") == 0) + strcpy(szaHeadLists, "0 NP IP VP NT NR NN 0"); + else if (strcmp(pszCFGLeft, "QP") == 0) + strcpy(szaHeadLists, "0QP CLP CD OD 0"); + else if (strcmp(pszCFGLeft, "VP") == 0) + strcpy(szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1"); + else if (strcmp(pszCFGLeft, "VCD") == 0) + strcpy(szaHeadLists, "0VCD VV VA VC VE 0"); + if (strcmp(pszCFGLeft, "VRD") == 0) + strcpy(szaHeadLists, "0VRD VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VSB") == 0) + strcpy(szaHeadLists, "0VSB VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VCP") == 0) + strcpy(szaHeadLists, "0VCP VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VNV") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "VPT") == 0) + strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); + else if (strcmp(pszCFGLeft, "UCP") == 0) + strcpy(szaHeadLists, "0"); + else if (strcmp(pszCFGLeft, "WHNP") == 0) + strcpy(szaHeadLists, "0WHNP NP NN NT NR QP 0"); + else if (strcmp(pszCFGLeft, "WHPP") == 0) + strcpy(szaHeadLists, "1WHPP PP P 1"); + + /* //head rules for GENIA corpus + if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) + strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP + 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); + else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) + strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN + 1NP 1JJS 1NN 1" ); + else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); + else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) + strcpy( szaHeadLists, "0" ); + else if( strcmp( pszCFGLeft, "LST" ) == 0 ) + strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); + else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) + strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP + 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); + else if( strcmp( pszCFGLeft, "PP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); + else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) + strcpy( szaHeadLists, "1RP 1" ); + else if( strcmp( pszCFGLeft, "QP" ) == 0 ) + strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR + 0JJS 0" ); + else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) + strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); + else if( strcmp( pszCFGLeft, "S" ) == 0 ) + strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); + else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) + strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ + 0SINV 0SBAR 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) + strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); + else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP + 0" ); + else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) + strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); + else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) + strcpy( szaHeadLists, "1" ); + else if( strcmp( pszCFGLeft, "VP" ) == 0 ) + strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP + 0ADJP 0NN 0NNS 0NP 0" ); + else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) + strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); + else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) + strcpy( szaHeadLists, "1CC 1WRB 1" ); + else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) + strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); + else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) + strcpy( szaHeadLists, "1IN 1TO FW 1" ); + else if( strcmp( pszCFGLeft, "NP" ) == 0 ) + strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN + 0CD 0JJ JJS RB QP 0" ); + */ + + return fnMyOwnHeadWordRule(szaHeadLists, vectRight); + } + + private: + int fnMyOwnHeadWordRule(char *pszaHeadLists, CVectorStr vectRight) { + char szHeadList[201], *p; + char szTerm[101]; + int J; + + p = pszaHeadLists; + + int iCountRight; + + iCountRight = vectRight.size(); + + szHeadList[0] = '\0'; + while (1) { + szTerm[0] = '\0'; + sscanf(p, "%s", szTerm); + if (strlen(szHeadList) == 0) { + if (strcmp(szTerm, "0") == 0) { + return iCountRight - 1; + } + if (strcmp(szTerm, "1") == 0) { + return 0; + } + + sprintf(szHeadList, "%c %s ", szTerm[0], szTerm + 1); + p = strstr(p, szTerm); + p += strlen(szTerm); + } else { + if ((szTerm[0] == '0') || (szTerm[0] == '1')) { + if (szHeadList[0] == '0') { + for (J = iCountRight - 1; J >= 0; J--) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } else { + for (J = 0; J < iCountRight; J++) { + sprintf(szTerm, " %s ", vectRight.at(J).c_str()); + if (strstr(szHeadList, szTerm) != NULL) return J; + } + } + + szHeadList[0] = '\0'; + } else { + strcat(szHeadList, szTerm); + strcat(szHeadList, " "); + + p = strstr(p, szTerm); + p += strlen(szTerm); + } + } + } + + return 0; + } +}; + +struct SParsedTree { + SParsedTree() { m_ptRoot = NULL; } + ~SParsedTree() { + if (m_ptRoot != NULL) delete m_ptRoot; + } + static SParsedTree *fnConvertFromString(const char *pszStr) { + if (strcmp(pszStr, "(())") == 0) return NULL; + SParsedTree *pTree = new SParsedTree(); + + std::vector vecSyn; + fnReadSyntactic(pszStr, vecSyn); + + int iLeft = 1, iRight = 1; //# left/right parenthesis + + STreeItem *pcurrent; + + pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); + + pcurrent = pTree->m_ptRoot; + + for (size_t i = 2; i < vecSyn.size() - 1; i++) { + if (strcmp(vecSyn[i].c_str(), "(") == 0) + iLeft++; + else if (strcmp(vecSyn[i].c_str(), ")") == 0) { + iRight++; + if (pcurrent == NULL) { + // error + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + pcurrent = pcurrent->m_ptParent; + } else { + STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); + pcurrent->fnAppend(ptNewItem); + pcurrent = ptNewItem; + + if (strcmp(vecSyn[i - 1].c_str(), "(") != 0 && + strcmp(vecSyn[i - 1].c_str(), ")") != 0) { + pTree->m_vecTerminals.push_back(ptNewItem); + pcurrent = pcurrent->m_ptParent; + } + } + } + + if (iLeft != iRight) { + // error + fprintf(stderr, "the left and right parentheses are not matched!"); + fprintf(stderr, "ERROR in ConvertFromString\n"); + fprintf(stderr, "%s\n", pszStr); + return NULL; + } + + return pTree; + } + + int fnGetNumWord() { return m_vecTerminals.size(); } + + void fnSetSpanInfo() { + int iNextNum = 0; + fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); + } + + void fnSetHeadWord() { + for (size_t i = 0; i < m_vecTerminals.size(); i++) + m_vecTerminals[i]->m_iHeadWord = i; + SGetHeadWord *pGetHeadWord = new SGetHeadWord(); + fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); + delete pGetHeadWord; + } + + STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { + STreeItem *pTreeItem = m_vecTerminals[iLeft]; + + while (pTreeItem->m_iEnd < iRight) { + pTreeItem = pTreeItem->m_ptParent; + if (pTreeItem == NULL) break; + } + if (pTreeItem == NULL) return NULL; + if (pTreeItem->m_iEnd > iRight) return NULL; + + assert(pTreeItem->m_iEnd == iRight); + if (bLowest) return pTreeItem; + + while (pTreeItem->m_ptParent != NULL && + pTreeItem->m_ptParent->fnGetChildrenNum() == 1) + pTreeItem = pTreeItem->m_ptParent; + + return pTreeItem; + } + + private: + void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int &iNextNum) { + int I; + int iNumChildren = ptItem->fnGetChildrenNum(); + for (I = 0; I < iNumChildren; I++) + fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[I], iNextNum); + + if (I == 0) { + ptItem->m_iBegin = iNextNum; + ptItem->m_iEnd = iNextNum++; + } else { + ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; + ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; + } + } + + void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, + SGetHeadWord *pGetHeadWord) { + int I, iHeadchild; + + if (ptItem->m_vecChildren.size() == 0) return; + + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); + + std::vector vecRight; + + if (ptItem->m_vecChildren.size() == 1) + iHeadchild = 0; + else { + for (I = 0; I < ptItem->m_vecChildren.size(); I++) + vecRight.push_back(std::string(ptItem->m_vecChildren[I]->m_pszTerm)); + + iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); + } + + ptItem->m_iHeadChild = iHeadchild; + ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + + static void fnReadSyntactic(const char *pszSyn, + std::vector &vec) { + char *p; + int I; + + int iLeftNum, iRightNum; + char *pszTmp, *pszTerm; + pszTmp = new char[strlen(pszSyn)]; + pszTerm = new char[strlen(pszSyn)]; + pszTmp[0] = pszTerm[0] = '\0'; + + vec.clear(); + + char *pszLine; + pszLine = new char[strlen(pszSyn) + 1]; + strcpy(pszLine, pszSyn); + + char *pszLine2; + + while (1) { + while ((strlen(pszLine) > 0) && (pszLine[strlen(pszLine) - 1] > 0) && + (pszLine[strlen(pszLine) - 1] <= ' ')) + pszLine[strlen(pszLine) - 1] = '\0'; + + if (strlen(pszLine) == 0) break; + + // printf( "%s\n", pszLine ); + pszLine2 = pszLine; + while (pszLine2[0] <= ' ') pszLine2++; + if (pszLine2[0] == '<') continue; + + sscanf(pszLine2 + 1, "%s", pszTmp); + + if (pszLine2[0] == '(') { + iLeftNum = 0; + iRightNum = 0; + } + + p = pszLine2; + while (1) { + pszTerm[0] = '\0'; + sscanf(p, "%s", pszTerm); + + if (strlen(pszTerm) == 0) break; + p = strstr(p, pszTerm); + p += strlen(pszTerm); + + if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { + if (pszTerm[0] == '(') { + vec.push_back(std::string("(")); + iLeftNum++; + + I = 1; + while (pszTerm[I] == '(' && pszTerm[I] != '\0') { + vec.push_back(std::string("(")); + iLeftNum++; + + I++; + } + + if (strlen(pszTerm) > 1) vec.push_back(std::string(pszTerm + I)); + } else { + char *pTmp; + pTmp = pszTerm + strlen(pszTerm) - 1; + while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; + pTmp[1] = '\0'; + + if (strlen(pszTerm) > 0) vec.push_back(std::string(pszTerm)); + pTmp += 2; + + for (I = 0; I <= (int)strlen(pTmp); I++) { + vec.push_back(std::string(")")); + iRightNum++; + } + } + } else { + char *q; + q = strchr(pszTerm, ')'); + if (q != NULL) { + q[0] = '\0'; + if (pszTerm[0] != '\0') vec.push_back(std::string(pszTerm)); + vec.push_back(std::string(")")); + iRightNum++; + + q++; + while (q[0] == ')') { + vec.push_back(std::string(")")); + q++; + iRightNum++; + } + + while (q[0] == '(') { + vec.push_back(std::string("(")); + q++; + iLeftNum++; + } + + if (q[0] != '\0') vec.push_back(std::string(q)); + } else + vec.push_back(std::string(pszTerm)); + } + } + + if (iLeftNum != iRightNum) { + fprintf(stderr, "%s\n", pszSyn); + assert(iLeftNum == iRightNum); + } + /*if ( iLeftNum != iRightNum ) { + printf( "ERROR: left( and right ) is not matched, %d ( and %d + )\n", iLeftNum, iRightNum ); + return; + }*/ + + if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { + //( (IP..) ) + std::vector::iterator it; + it = vec.begin(); + it++; + vec.insert(it, std::string("ROOT")); + } + + break; + } + + delete[] pszLine; + delete[] pszTmp; + delete[] pszTerm; + } + + public: + STreeItem *m_ptRoot; + std::vector m_vecTerminals; // the leaf nodes +}; + +struct SParseReader { + SParseReader(const char *pszParse_Fname, bool bFlattened = false) + : m_bFlattened(bFlattened) { + m_fpIn = fopen(pszParse_Fname, "r"); + assert(m_fpIn != NULL); + } + ~SParseReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + SParsedTree *fnReadNextParseTree() { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + pTree = SParsedTree::fnConvertFromString(pszLine); + if (pTree == NULL) break; + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + SParsedTree *fnReadNextParseTreeWithProb(double *pProb) { + SParsedTree *pTree = NULL; + char *pszLine = new char[100001]; + int iLen; + + while (fnReadNextSentence(pszLine, &iLen) == true) { + if (iLen == 0) continue; + + char *p = strchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + if (pProb) (*pProb) = atof(pszLine); + + pTree = SParsedTree::fnConvertFromString(p); + if (m_bFlattened) + fnPostProcessingFlattenedParse(pTree); + else { + pTree->fnSetSpanInfo(); + pTree->fnSetHeadWord(); + } + break; + } + + delete[] pszLine; + return pTree; + } + + private: + /* + * since to the parse tree is a flattened tree, use the head mark to identify + * head info. + * the head node will be marked as "*XP*" + */ + void fnSetParseTreeHeadInfo(SParsedTree *pTree) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) + pTree->m_vecTerminals[i]->m_iHeadWord = i; + fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); + } + + void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { + if (pTreeItem->m_vecChildren.size() == 0) return; + + for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) + fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); + + std::vector vecRight; + + int iHeadchild; + + if (pTreeItem->fnIsPreTerminal()) { + iHeadchild = 0; + } else { + size_t i; + for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { + char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; + if (p[0] == '*' && p[strlen(p) - 1] == '*') { + iHeadchild = i; + p[strlen(p) - 1] = '\0'; + std::string str = p + 1; + strcpy(p, str.c_str()); // erase the "*..*" + break; + } + } + assert(i < pTreeItem->m_vecChildren.size()); + } + + pTreeItem->m_iHeadChild = iHeadchild; + pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; + } + void fnPostProcessingFlattenedParse(SParsedTree *pTree) { + pTree->fnSetSpanInfo(); + fnSetParseTreeHeadInfo(pTree); + } + bool fnReadNextSentence(char *pszLine, int *piLength) { + if (feof(m_fpIn) == true) return false; + + int iLen; + + pszLine[0] = '\0'; + + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + if (piLength != NULL) (*piLength) = iLen; + + return true; + } + + private: + FILE *m_fpIn; + const bool m_bFlattened; +}; + +/* + * Note: + * m_vec_s_align.size() may not be equal to the length of source side + *sentence + * due to the last words may not be aligned + * + */ +struct SAlignment { + typedef std::vector SingleAlign; + SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } + ~SAlignment() {} + + bool fnIsAligned(int i, bool s) const { + const std::vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + if ((*palign)[i].size() == 0) return false; + return true; + } + + /* + * return true if [b, e] is aligned phrases on source side (if s==true) or on + * the target side (if s==false); + * return false, otherwise. + */ + bool fnIsAlignedPhrase(int b, int e, bool s, int* pob, int* poe) const { + int ob, oe; //[b, e] on the other side + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + + if (ob == -1) { + if (pob != NULL) (*pob) = -1; + if (poe != NULL) (*poe) = -1; + return false; // no aligned word among [b, e] + } + if (pob != NULL) (*pob) = ob; + if (poe != NULL) (*poe) = oe; + + int bb, be; //[b, e] back given [ob, oe] on the other side + if (s == true) + fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); + else + fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); + + if (bb < b || be > e) return false; + return true; + } + + bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { + const std::vector* palign; + if (s == true) + palign = &m_vec_s_align; + else + palign = &m_vec_t_align; + + if ((*palign).size() <= e || (*palign)[b].size() == 0 || + (*palign)[e].size() == 0) + return false; + + return fnIsAlignedPhrase(b, e, s, pob, poe); + } + + void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { + if (s == true) + fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); + else + fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); + } + + /* + * look the translation of source[b, e] is continuous or not + * 1) return "Unaligned": if the source[b, e] is translated silently; + * 2) return "Con't": if none of target words in target[.., ..] is exclusively + * aligned to any word outside source[b, e] + * 3) return "Discon't": otherwise; + */ + std::string fnIsContinuous(int b, int e) const { + int ob, oe; + fnGetLeftRightMost(b, e, true, ob, oe); + if (ob == -1) return "Unaligned"; + + for (int i = ob; i <= oe; i++) { + if (!fnIsAligned(i, false)) continue; + const SingleAlign& a = m_vec_t_align[i]; + int j; + for (j = 0; j < a.size(); j++) + if (a[j] >= b && a[j] <= e) break; + if (j == a.size()) return "Discon't"; + } + return "Con't"; + } + + const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { + if (s == true) { + if (i >= m_vec_s_align.size()) return NULL; + return &(m_vec_s_align[i]); + } else { + if (i >= m_vec_t_align.size()) return NULL; + return &(m_vec_t_align[i]); + } + } + + private: + void fnGetLeftRightMost(int b, int e, const std::vector& align, + int& ob, int& oe) const { + ob = oe = -1; + for (int i = b; i <= e && i < align.size(); i++) { + if (align[i].size() > 0) { + if (align[i][0] < ob || ob == -1) ob = align[i][0]; + if (oe < align[i][align[i].size() - 1]) + oe = align[i][align[i].size() - 1]; + } + } + } + void fnInitializeAlignment(const char* pszAlign) { + m_vec_s_align.clear(); + m_vec_t_align.clear(); + + std::vector terms = SplitOnWhitespace(std::string(pszAlign)); + int si, ti; + for (size_t i = 0; i < terms.size(); i++) { + sscanf(terms[i].c_str(), "%d-%d", &si, &ti); + + while (m_vec_s_align.size() <= si) { + SingleAlign sa; + m_vec_s_align.push_back(sa); + } + while (m_vec_t_align.size() <= ti) { + SingleAlign sa; + m_vec_t_align.push_back(sa); + } + + m_vec_s_align[si].push_back(ti); + m_vec_t_align[ti].push_back(si); + } + + // sort + for (size_t i = 0; i < m_vec_s_align.size(); i++) { + std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); + } + for (size_t i = 0; i < m_vec_t_align.size(); i++) { + std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); + } + } + + private: + std::vector m_vec_s_align; // source side words' alignment + std::vector m_vec_t_align; // target side words' alignment +}; + +struct SAlignmentReader { + SAlignmentReader(const char* pszFname) { + m_fpIn = fopen(pszFname, "r"); + assert(m_fpIn != NULL); + } + ~SAlignmentReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + SAlignment* fnReadNextAlignment() { + if (feof(m_fpIn) == true) return NULL; + char* pszLine = new char[100001]; + pszLine[0] = '\0'; + fgets(pszLine, 10001, m_fpIn); + int iLen = strlen(pszLine); + if (iLen == 0) return NULL; + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + SAlignment* pAlign = new SAlignment(pszLine); + delete[] pszLine; + return pAlign; + } + + private: + FILE* m_fpIn; +}; + +struct SArgument { + SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { + m_pszRole = new char[strlen(pszRole) + 1]; + strcpy(m_pszRole, pszRole); + m_iBegin = iBegin; + m_iEnd = iEnd; + m_fProb = fProb; + m_pTreeItem = NULL; + } + ~SArgument() { delete[] m_pszRole; } + + void fnSetTreeItem(STreeItem* pTreeItem) { + m_pTreeItem = pTreeItem; + if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { + assert(m_pTreeItem->m_iBegin == m_iBegin); + assert(m_pTreeItem->m_iEnd == m_iEnd); + } + } + + char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP + int m_iBegin; + int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] + float m_fProb; // the probability of this role, + STreeItem* m_pTreeItem; +}; + +struct SPredicate { + SPredicate(const char* pszLemma, int iPosition) { + if (pszLemma != NULL) { + m_pszLemma = new char[strlen(pszLemma) + 1]; + strcpy(m_pszLemma, pszLemma); + } else + m_pszLemma = NULL; + m_iPosition = iPosition; + } + ~SPredicate() { + if (m_pszLemma != NULL) delete[] m_pszLemma; + for (size_t i = 0; i < m_vecArgt.size(); i++) delete m_vecArgt[i]; + } + int fnAppend(const char* pszRole, int iBegin, int iEnd) { + SArgument* pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); + return fnAppend(pArgt); + } + int fnAppend(SArgument* pArgt) { + m_vecArgt.push_back(pArgt); + int iPosition = m_vecArgt.size() - 1; + return iPosition; + } + + char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same + // as the predicate itself + int m_iPosition; // the position in sentence + std::vector m_vecArgt; // arguments associated to the predicate +}; + +struct SSrlSentence { + SSrlSentence() { m_pTree = NULL; } + ~SSrlSentence() { + if (m_pTree != NULL) delete m_pTree; + + for (size_t i = 0; i < m_vecPred.size(); i++) delete m_vecPred[i]; + } + int fnAppend(const char* pszLemma, int iPosition) { + SPredicate* pPred = new SPredicate(pszLemma, iPosition); + return fnAppend(pPred); + } + int fnAppend(SPredicate* pPred) { + m_vecPred.push_back(pPred); + int iPosition = m_vecPred.size() - 1; + return iPosition; + } + int GetPredicateNum() { return m_vecPred.size(); } + + SParsedTree* m_pTree; + std::vector m_vecPred; +}; + +struct SSrlSentenceReader { + SSrlSentenceReader(const char* pszSrlFname) { + m_fpIn = fopen(pszSrlFname, "r"); + assert(m_fpIn != NULL); + } + ~SSrlSentenceReader() { + if (m_fpIn != NULL) fclose(m_fpIn); + } + + inline void fnReplaceAll(std::string& str, const std::string& from, + const std::string& to) { + size_t start_pos = 0; + while ((start_pos = str.find(from, start_pos)) != std::string::npos) { + str.replace(start_pos, from.length(), to); + start_pos += to.length(); // In case 'to' contains 'from', like replacing + // 'x' with 'yx' + } + } + + // TODO: here only considers flat predicate-argument structure + // i.e., no overlap among them + SSrlSentence* fnReadNextSrlSentence() { + std::vector > vecContent; + if (fnReadNextContent(vecContent) == false) return NULL; + + SSrlSentence* pSrlSentence = new SSrlSentence(); + int iSize = vecContent.size(); + // put together syntactic text + std::ostringstream ostr; + for (int i = 0; i < iSize; i++) { + std::string strSynSeg = + vecContent[i][5]; // the 5th column is the syntactic segment + size_t iPosition = strSynSeg.find_first_of('*'); + assert(iPosition != std::string::npos); + std::ostringstream ostrTmp; + ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] + << ")"; // the 2th column is POS-tag, and the 0th column is word + strSynSeg.replace(iPosition, 1, ostrTmp.str()); + fnReplaceAll(strSynSeg, "(", " ("); + ostr << strSynSeg; + } + std::string strSyn = ostr.str(); + pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); + pSrlSentence->m_pTree->fnSetHeadWord(); + pSrlSentence->m_pTree->fnSetSpanInfo(); + + // read predicate-argument structure + int iNumPred = vecContent[0].size() - 8; + for (int i = 0; i < iNumPred; i++) { + std::vector vecRole; + std::vector vecBegin; + std::vector vecEnd; + int iPred = -1; + for (int j = 0; j < iSize; j++) { + const char* p = vecContent[j][i + 8].c_str(); + const char* q; + if (p[0] == '(') { + // starting position of an argument(or predicate) + vecBegin.push_back(j); + q = strchr(p, '*'); + assert(q != NULL); + vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); + if (vecRole.back().compare("V") == 0) { + assert(iPred == -1); + iPred = vecRole.size() - 1; + } + } + if (p[strlen(p) - 1] == ')') { + // end position of an argument(or predicate) + vecEnd.push_back(j); + assert(vecBegin.size() == vecEnd.size()); + } + } + assert(iPred != -1); + SPredicate* pPred = new SPredicate( + pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, + vecBegin[iPred]); + pSrlSentence->fnAppend(pPred); + for (size_t j = 0; j < vecBegin.size(); j++) { + if (j == iPred) continue; + pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); + pPred->m_vecArgt.back()->fnSetTreeItem( + pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], + false)); + } + } + return pSrlSentence; + } + + private: + bool fnReadNextContent(std::vector >& vecContent) { + vecContent.clear(); + if (feof(m_fpIn) == true) return false; + char* pszLine; + pszLine = new char[100001]; + pszLine[0] = '\0'; + int iLen; + while (!feof(m_fpIn)) { + fgets(pszLine, 10001, m_fpIn); + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + if (iLen == 0) break; // end of this sentence + + std::vector terms = SplitOnWhitespace(std::string(pszLine)); + assert(terms.size() > 7); + vecContent.push_back(terms); + } + delete[] pszLine; + return true; + } + + private: + FILE* m_fpIn; +}; + +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; + +struct Tsuruoka_Maxent { + Tsuruoka_Maxent(const char* pszModelFName) { + if (pszModelFName != NULL) { + m_pModel = new ME_Model(); + m_pModel->load_from_file(pszModelFName); + } else + m_pModel = NULL; + } + + ~Tsuruoka_Maxent() { + if (m_pModel != NULL) delete m_pModel; + } + + void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, + const char* pszModelFName, int /*iNumIteration*/) { + assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || + strcmp(pszAlgorithm, "sgd") == 0 || + strcmp(pszAlgorithm, "SGD") == 0); + FILE* fpIn = fopen(pszInstanceFName, "r"); + + ME_Model* pModel = new ME_Model(); + + char* pszLine = new char[100001]; + int iNumInstances = 0; + int iLen; + while (!feof(fpIn)) { + pszLine[0] = '\0'; + fgets(pszLine, 20000, fpIn); + if (strlen(pszLine) == 0) { + continue; + } + + iLen = strlen(pszLine); + while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { + pszLine[iLen - 1] = '\0'; + iLen--; + } + + iNumInstances++; + + ME_Sample* pmes = new ME_Sample(); + + char* p = strrchr(pszLine, ' '); + assert(p != NULL); + p[0] = '\0'; + p++; + std::vector vecContext; + SplitOnWhitespace(std::string(pszLine), &vecContext); + + pmes->label = std::string(p); + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + pModel->add_training_sample((*pmes)); + if (iNumInstances % 100000 == 0) + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + delete pmes; + } + fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); + fclose(fpIn); + + if (strcmp(pszAlgorithm, "l1") == 0) + pModel->use_l1_regularizer(1.0); + else if (strcmp(pszAlgorithm, "l2") == 0) + pModel->use_l2_regularizer(1.0); + else + pModel->use_SGD(); + + pModel->train(); + pModel->save_to_file(pszModelFName); + + delete pModel; + fprintf(stdout, "......Finished Training\n"); + fprintf(stdout, "......Model saved as %s\n", pszModelFName); + delete[] pszLine; + } + + double fnEval(const char* pszContext, const char* pszOutcome) const { + std::vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(std::string(pszContext), &vecContext); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + std::vector vecProb = m_pModel->classify(*pmes); + delete pmes; + int iLableID = m_pModel->get_class_id(pszOutcome); + return vecProb[iLableID]; + } + void fnEval(const char* pszContext, + std::vector >& vecOutput) const { + std::vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(std::string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + std::vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + std::string label = m_pModel->get_class_label(i); + vecOutput.push_back(make_pair(label, vecProb[i])); + } + delete pmes; + } + void fnEval(const char* pszContext, std::vector& vecOutput) const { + std::vector vecContext; + ME_Sample* pmes = new ME_Sample(); + SplitOnWhitespace(std::string(pszContext), &vecContext); + + vecOutput.clear(); + + for (size_t i = 0; i < vecContext.size(); i++) + pmes->add_feature(vecContext[i]); + std::vector vecProb = m_pModel->classify(*pmes); + + for (size_t i = 0; i < vecProb.size(); i++) { + std::string label = m_pModel->get_class_label(i); + vecOutput.push_back(vecProb[i]); + } + delete pmes; + } + int fnGetClassId(const std::string& strLabel) const { + return m_pModel->get_class_id(strLabel); + } + + private: + ME_Model* m_pModel; +}; + +// an argument item or a predicate item (the verb itself) +struct SSRLItem { + SSRLItem(const STreeItem *tree_item, std::string role) + : tree_item_(tree_item), role_(role) {} + ~SSRLItem() {} + const STreeItem *tree_item_; + const std::string role_; +}; + +struct SPredicateItem { + SPredicateItem(const SParsedTree *tree, const SPredicate *pred) + : pred_(pred) { + vec_items_.reserve(pred->m_vecArgt.size() + 1); + for (int i = 0; i < pred->m_vecArgt.size(); i++) { + vec_items_.push_back( + new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, + std::string(pred->m_vecArgt[i]->m_pszRole))); + } + vec_items_.push_back( + new SSRLItem(tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, + std::string("Pred"))); + sort(vec_items_.begin(), vec_items_.end(), SortFunction); + + begin_ = vec_items_[0]->tree_item_->m_iBegin; + end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; + } + + ~SPredicateItem() { vec_items_.clear(); } + + static bool SortFunction(SSRLItem *i, SSRLItem *j) { + return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); + } + + std::vector vec_items_; + int begin_; + int end_; + const SPredicate *pred_; +}; + +struct SArgumentReorderModel { + public: + static std::string fnGetBlockOutcome(int iBegin, int iEnd, + SAlignment *pAlign) { + return pAlign->fnIsContinuous(iBegin, iEnd); + } + static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, + std::vector &vecStrLeftReorder, + std::vector &vecStrRightReorder) { + std::vector vecLeft, vecRight; + for (int i = 0; i < pPredItem->vec_items_.size(); i++) { + const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + std::vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + std::vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vecStrLeftReorder.clear(); + vecStrRightReorder.clear(); + for (int i = 1; i < vecLeftPosition.size(); i++) { + std::string strOutcome; + fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); + vecStrLeftReorder.push_back(strOutcome); + fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); + vecStrRightReorder.push_back(strOutcome); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + * + * f1: (left_role, right_role, predicate_term) + * f2: (left_role, right_role, predicate_term, other_right_role) + * f3: (left_role, right_role, predicate_term, other_left_role) + * f4: (left_role, right_role, left_head_pos) + * f5: (left_role, right_role, left_head_word) + * f6: (left_role, right_role, left_syntactic_label) + * f7: (left_role, right_role, right_head_pos) + * f8: (left_role, right_role, right_head_word) + * f8: (left_role, right_role, right_syntactic_label) + * f8: (left_role, right_role, left_chunk_status) + * f9: (left_role, right_role, right_chunk_status) + * f10: (left_role, right_role, left_chunk_status) + * f11: (left_role, right_role, right_chunk_status) + * f12: (left_label, parent_label) + * f13: (right_label, parent_label) + */ + static void fnGenerateFeature(const SParsedTree *pTree, + const SPredicate *pPred, + const SPredicateItem *pPredItem, int iPos, + const std::string &strBlock1, + const std::string &strBlock2, + std::ostringstream &ostr) { + SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; + SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; + const STreeItem *pCon1 = pSRLItem1->tree_item_; + const STreeItem *pCon2 = pSRLItem2->tree_item_; + + std::string left_role = pSRLItem1->role_; + std::string right_role = pSRLItem2->role_; + + std::string predicate_term = + pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; + + std::vector vec_other_right_sibling; + for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) + vec_other_right_sibling.push_back( + std::string(pPredItem->vec_items_[i]->role_)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(std::string("NULL")); + + std::vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_right_sibling.push_back( + std::string(pPredItem->vec_items_[i]->role_)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(std::string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; + ostr << "f1=" << left_role << "_" << right_role; + + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) { + ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_right_sibling[i]; + ostr << " f2=" << left_role << "_" << right_role << "_" + << vec_other_right_sibling[i]; + } + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) { + ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term + << "_" << vec_other_left_sibling[i]; + ostr << " f3=" << left_role << "_" << right_role << "_" + << vec_other_left_sibling[i]; + } + // f4 + ostr << " f4=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; + // f6 + ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f7 + ostr << " f7=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f8 + ostr << " f8=" << left_role << "_" << right_role << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; + // f9 + ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; + // f10 + ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; + // f11 + ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; + // f12 + ostr << " f12=" << left_role << "_" << predicate_term; + ostr << " f12=" << left_role; + // f13 + ostr << " f13=" << right_role << "_" << predicate_term; + ostr << " f13=" << right_role; + } + + private: + static void fnGetOutcome(int i1, int i2, std::string &strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = std::string("DM"); + else + strOutcome = std::string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = std::string("DS"); + else + strOutcome = std::string("S"); + } + } + + static void fnGetRelativePosition(const std::vector &vecLeft, + std::vector &vecPosition) { + vecPosition.clear(); + + std::vector vec; + for (int i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (int i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (int j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } +}; +} // namespace const_reorder + +#endif // _FF_CONST_REORDER_COMMON_H diff --git a/training/Makefile.am b/training/Makefile.am index 8ef3c939..2812a9be 100644 --- a/training/Makefile.am +++ b/training/Makefile.am @@ -8,5 +8,5 @@ SUBDIRS = \ dtrain \ latent_svm \ mira \ - rampion - + rampion \ + const_reorder diff --git a/training/const_reorder/Makefile.am b/training/const_reorder/Makefile.am new file mode 100644 index 00000000..2e81e588 --- /dev/null +++ b/training/const_reorder/Makefile.am @@ -0,0 +1,8 @@ +bin_PROGRAMS = const_reorder_model_trainer argument_reorder_model_trainer + +AM_CPPFLAGS = -I$(top_srcdir) -I$(top_srcdir)/utils -I$(top_srcdir)/decoder + +const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc +const_reorder_model_trainer_LDADD = ../../utils/libutils.a +argument_reorder_model_trainer_SOURCES = argument_reorder_model.cc +argument_reorder_model_trainer_LDADD = ../../utils/libutils.a diff --git a/training/const_reorder/argument_reorder_model.cc b/training/const_reorder/argument_reorder_model.cc new file mode 100644 index 00000000..54402436 --- /dev/null +++ b/training/const_reorder/argument_reorder_model.cc @@ -0,0 +1,307 @@ +/* + * argument_reorder_model.cc + * + * Created on: Dec 15, 2013 + * Author: lijunhui + */ + +#include +#include +#include +#include +#include +#include + +#include "utils/filelib.h" + +#include "decoder/ff_const_reorder_common.h" + +using namespace std; +using namespace const_reorder; + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + Map hashPredicate; + { + ReadFile in(pszFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + for (const auto& i : terms) { + ++hashPredicate[i]; + } + } + } + + { + ReadFile in(pszFName); + WriteFile out(pszNewFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + bool written = false; + for (const auto& i : terms) { + if (hashPredicate[i] >= iCutoff) { + (*out.stream()) << i << " "; + written = true; + } + } + if (written) { + (*out.stream()) << "\n"; + } + } + } +} + +struct SArgumentReorderTrainer { + SArgumentReorderTrainer( + const char* pszSRLFname, // source-side srl tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelFname, // classifier model file name + int iCutoff) { + fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszTopPredicateFname, + pszInstanceFname); + + string strInstanceFname, strModelFname; + strInstanceFname = string(pszInstanceFname) + string(".left"); + strModelFname = string(pszModelFname) + string(".left"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + strInstanceFname = string(pszInstanceFname) + string(".right"); + strModelFname = string(pszModelFname) + string(".right"); + fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); + } + + ~SArgumentReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + void fnGenerateInstanceFiles( + const char* pszSRLFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszTopPredicateFname, // top predicate file name (we only + // consider predicates with 100+ + // occurrences + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); + + Map* pMapPredicate; + if (pszTopPredicateFname != NULL) + pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); + else + pMapPredicate = NULL; + + string line; + + WriteFile left_file(pszInstanceFname + string(".left")); + WriteFile right_file(pszInstanceFname + string(".right")); + + // read sentence by sentence + SAlignment* pAlign; + SSrlSentence* pSRL; + SParsedTree* pTree; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pSRL = pSRLReader->fnReadNextSrlSentence(); + assert(pSRL != NULL); + pTree = pSRL->m_pTree; + assert(getline(*source_file.stream(), line)); + vector vecSTerms; + SplitOnWhitespace(line, &vecSTerms); + assert(getline(*target_file.stream(), line)); + vector vecTTerms; + SplitOnWhitespace(line, &vecTTerms); + // vecTPOSTerms.size() == 0, given the case when an english sentence fails + // parsing + + if (pTree != NULL) { + for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { + SPredicate* pPred = pSRL->m_vecPred[i]; + if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition] + ->m_ptParent->m_pszTerm, + "VA") == 0) + continue; + string strPred = + string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); + if (pMapPredicate != NULL) { + Map::iterator iter_map = pMapPredicate->find(strPred); + if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) + continue; + } + + SPredicateItem* pPredItem = new SPredicateItem(pTree, pPred); + + vector vecStrBlock; + for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { + SSRLItem* pItem1 = pPredItem->vec_items_[j]; + vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome( + pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, + pAlign)); + } + + vector vecStrLeftReorderType; + vector vecStrRightReorderType; + SArgumentReorderModel::fnGetReorderType( + pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); + for (int j = 1; j < pPredItem->vec_items_.size(); j++) { + string strLeftOutcome, strRightOutcome; + strLeftOutcome = vecStrLeftReorderType[j - 1]; + strRightOutcome = vecStrRightReorderType[j - 1]; + ostringstream ostr; + SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, + vecStrBlock[j - 1], + vecStrBlock[j], ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), + // strOutcome.c_str()); + (*left_file.stream()) << ostr.str() << " " << strLeftOutcome + << "\n"; + (*right_file.stream()) << ostr.str() << " " << strRightOutcome + << "\n"; + } + } + } + delete pSRL; + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + delete pAlignReader; + delete pSRLReader; + } + + Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { + if (pszTopPredicateFname == NULL) return NULL; + + Map* pMapPredicate = new Map(); + // STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); + ReadFile in(pszTopPredicateFname); + // char* pszLine = new char[50001]; + string line; + int iNumCount = 0; + while (getline(*in.stream(), line)) { + if (line.size() && line[0] == '#') continue; + auto p = line.find(' '); + assert(p != string::npos); + int iCount = atoi(line.substr(p + 1).c_str()); + if (iCount < 100) break; + (*pMapPredicate)[line] = iNumCount++; + } + return pMapPredicate; + } +}; + +namespace po = boost::program_options; + +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); + out << '"'; + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); + } + out << '\n'; +} +inline string str(char const* name, po::variables_map const& conf) { + return conf[name].as(); +} + +//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file +/// scratch0/mt_exp/gale-align/gale-align.nw.al --source_file +/// scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file +/// scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file +/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix +/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 +//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file +/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix +/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options()("srl_file", po::value(), "srl file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")("help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("srl_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + SArgumentReorderTrainer* pTrainer = new SArgumentReorderTrainer( + str("srl_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), NULL, + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["feature_cutoff"].as()); + delete pTrainer; + + return 1; +} diff --git a/training/const_reorder/constituent_reorder_model.cc b/training/const_reorder/constituent_reorder_model.cc new file mode 100644 index 00000000..6bec3f0b --- /dev/null +++ b/training/const_reorder/constituent_reorder_model.cc @@ -0,0 +1,636 @@ +/* + * constituent_reorder_model.cc + * + * Created on: Jul 10, 2013 + * Author: junhuili + */ + +#include +#include + +#include + +#include "utils/filelib.h" + +#include "decoder/ff_const_reorder_common.h" + +using namespace std; +using namespace const_reorder; + +typedef std::unordered_map Map; +typedef std::unordered_map::iterator Iterator; + +namespace po = boost::program_options; + +inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, + const char* pszNewFName) { + Map hashPredicate; + { + ReadFile f(pszFName); + string line; + while (getline(*f.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + for (const auto& i : terms) { + ++hashPredicate[i]; + } + } + } + + { + ReadFile in(pszFName); + WriteFile out(pszNewFName); + string line; + while (getline(*in.stream(), line)) { + if (!line.size()) continue; + vector terms; + SplitOnWhitespace(line, &terms); + bool written = false; + for (const auto& i : terms) { + if (hashPredicate[i] >= iCutoff) { + (*out.stream()) << i << " "; + written = true; + } + } + if (written) { + (*out.stream()) << "\n"; + } + } + } +} + +struct SConstReorderTrainer { + SConstReorderTrainer( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname, // training instance file name + const char* pszModelPrefix, // classifier model file name prefix + int iCutoff, // feature count threshold + const char* /*pszOption*/ // other classifier parameters (for svmlight) + ) { + fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, + pszTargetFname, pszInstanceFname); + + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + + string strModelLeftFname = string(pszModelPrefix) + string(".left"); + string strModelRightFname = string(pszModelPrefix) + string(".right"); + + fprintf(stdout, "...Training the left ordering model\n"); + fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), + iCutoff); + fprintf(stdout, "...Training the right ordering model\n"); + fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), + iCutoff); + } + ~SConstReorderTrainer() {} + + private: + void fnTraining(const char* pszInstanceFname, const char* pszModelFname, + int iCutoff) { + char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; + if (iCutoff > 0) { + sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); + fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); + } else { + strcpy(pszNewInstanceFName, pszInstanceFname); + } + + /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); +pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); +delete pZhangleMaxent;*/ + + Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); + pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); + delete pMaxent; + + if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { + sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); + system(pszNewInstanceFName); + } + delete[] pszNewInstanceFName; + } + + inline bool fnIsVerbPOS(const char* pszTerm) { + if (strcmp(pszTerm, "VV") == 0 || strcmp(pszTerm, "VA") == 0 || + strcmp(pszTerm, "VC") == 0 || strcmp(pszTerm, "VE") == 0) + return true; + return false; + } + + inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, + const SAlignment* /*pAlign*/, string& strOutcome) { + if (iL1 == -1 && iL2 == -1) + strOutcome = "BU"; // 1. both are untranslated + else if (iL1 == -1) + strOutcome = "1U"; // 2. XP1 is untranslated + else if (iL2 == -1) + strOutcome = "2U"; // 3. XP2 is untranslated + else if (iL1 == iL2 && iR1 == iR2) + strOutcome = "SS"; // 4. Have same scope + else if (iL1 <= iL2 && iR1 >= iR2) + strOutcome = "1C2"; // 5. XP1's translation covers XP2's + else if (iL1 >= iL2 && iR1 <= iR2) + strOutcome = "2C1"; // 6. XP2's translation covers XP1's + else if (iR1 < iL2) { + int i = iR1 + 1; + /*while (i < iL2) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL2) + strOutcome = "M"; // 7. Monotone + else + strOutcome = "DM"; // 8. Discontinuous monotone + } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) + strOutcome = "OM"; // 9. Overlap monotone + else if (iR2 < iL1) { + int i = iR2 + 1; + /*while (i < iL1) { + if (pAlign->fnIsAligned(i, false)) + break; + i++; + }*/ + if (i == iL1) + strOutcome = "S"; // 10. Swap + else + strOutcome = "DS"; // 11. Discontinuous swap + } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) + strOutcome = "OS"; // 12. Overlap swap + else + assert(false); + } + + inline void fnGetOutcome(int i1, int i2, string& strOutcome) { + assert(i1 != i2); + if (i1 < i2) { + if (i2 > i1 + 1) + strOutcome = string("DM"); + else + strOutcome = string("M"); + } else { + if (i1 > i2 + 1) + strOutcome = string("DS"); + else + strOutcome = string("S"); + } + } + + inline void fnGetRelativePosition(const vector& vecLeft, + vector& vecPosition) { + vecPosition.clear(); + + vector vec; + for (size_t i = 0; i < vecLeft.size(); i++) { + if (vecLeft[i] == -1) { + if (i == 0) + vec.push_back(-1); + else + vec.push_back(vecLeft[i - 1] + 0.1); + } else + vec.push_back(vecLeft[i]); + } + + for (size_t i = 0; i < vecLeft.size(); i++) { + int count = 0; + + for (size_t j = 0; j < vecLeft.size(); j++) { + if (j == i) continue; + if (vec[j] < vec[i]) { + count++; + } else if (vec[j] == vec[i] && j < i) { + count++; + } + } + vecPosition.push_back(count); + } + } + + /* + * features: + * f1: (left_label, right_label, parent_label) + * f2: (left_label, right_label, parent_label, other_right_sibling_label) + * f3: (left_label, right_label, parent_label, other_left_sibling_label) + * f4: (left_label, right_label, left_head_pos) + * f5: (left_label, right_label, left_head_word) + * f6: (left_label, right_label, right_head_pos) + * f7: (left_label, right_label, right_head_word) + * f8: (left_label, right_label, left_chunk_status) + * f9: (left_label, right_label, right_chunk_status) + * f10: (left_label, parent_label) + * f11: (right_label, parent_label) + */ + void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, + int iPos, const vector& vecChunkStatus, + const vector& vecPosition, + const vector& vecSTerms, + const vector& /*vecTTerms*/, string& strOutcome, + ostringstream& ostr) { + STreeItem* pCon1, *pCon2; + pCon1 = pParent->m_vecChildren[iPos - 1]; + pCon2 = pParent->m_vecChildren[iPos]; + + fnGetOutcome(vecPosition[iPos - 1], vecPosition[iPos], strOutcome); + + string left_label = string(pCon1->m_pszTerm); + string right_label = string(pCon2->m_pszTerm); + string parent_label = string(pParent->m_pszTerm); + + vector vec_other_right_sibling; + for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) + vec_other_right_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_right_sibling.size() == 0) + vec_other_right_sibling.push_back(string("NULL")); + vector vec_other_left_sibling; + for (int i = 0; i < iPos - 1; i++) + vec_other_left_sibling.push_back( + string(pParent->m_vecChildren[i]->m_pszTerm)); + if (vec_other_left_sibling.size() == 0) + vec_other_left_sibling.push_back(string("NULL")); + + // generate features + // f1 + ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; + // f2 + for (int i = 0; i < vec_other_right_sibling.size(); i++) + ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_right_sibling[i]; + // f3 + for (int i = 0; i < vec_other_left_sibling.size(); i++) + ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label + << "_" << vec_other_left_sibling[i]; + // f4 + ostr << " f4=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; + // f5 + ostr << " f5=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon1->m_iHeadWord]; + // f6 + ostr << " f6=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f7 + ostr << " f7=" << left_label << "_" << right_label << "_" + << vecSTerms[pCon2->m_iHeadWord]; + // f8 + ostr << " f8=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos - 1]; + // f9 + ostr << " f9=" << left_label << "_" << right_label << "_" + << vecChunkStatus[iPos]; + // f10 + ostr << " f10=" << left_label << "_" << parent_label; + // f11 + ostr << " f11=" << right_label << "_" << parent_label; + } + + /* + * Source side (11 features): + * f1: the categories of XP1 and XP2 (f1_1, f1_2) + * f2: the head words of XP1 and XP2 (f2_1, f2_2) + * f3: the first and last word of XP1 (f3_f, f3_l) + * f4: the first and last word of XP2 (f4_f, f4_l) + * f5: is XP1 or XP2 the head node (f5_1, f5_2) + * f6: the category of the common parent + * Target side (6 features): + * f7: the first and the last word of XP1's translation (f7_f, f7_l) + * f8: the first and the last word of XP2's translation (f8_f, f8_l) + * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) + */ + void fnGenerateInstance(const SParsedTree* /*pTree*/, const STreeItem* pParent, + const STreeItem* pCon1, const STreeItem* pCon2, + const SAlignment* pAlign, + const vector& vecSTerms, + const vector& /*vecTTerms*/, string& strOutcome, + ostringstream& ostr) { + + int iLeft1, iRight1, iLeft2, iRight2; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, + iRight1); + pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, + iRight2); + + fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); + + // generate features + // f1 + ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; + // f2 + ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" + << vecSTerms[pCon2->m_iHeadWord]; + // f3 + ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] + << " f3_l=" << vecSTerms[pCon1->m_iEnd]; + // f4 + ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] + << " f4_l=" << vecSTerms[pCon2->m_iEnd]; + // f5 + if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) + ostr << " f5_1=1"; + else + ostr << " f5_1=0"; + if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) + ostr << " f5_2=1"; + else + ostr << " f5_2=0"; + // f6 + ostr << " f6=" << pParent->m_pszTerm; + + /*//f7 + if (iLeft1 != -1) { + ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << + vecTTerms[iRight1]; + } + if (iLeft2 != -1) { + ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << + vecTTerms[iRight2]; + } + + const vector* pvecTarget = + pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); + string str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_1=" << str.substr(0, str.size()-1); + } + pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); + str = ""; + for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { + str += vecTTerms[(*pvecTarget)[i]] + "_"; + } + if (str.length() > 0) { + ostr << " f9_2=" << str.substr(0, str.size()-1); + } */ + } + + void fnGetFocusedParentNodes(const SParsedTree* pTree, + vector& vecFocused) { + for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { + STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; + + while (pParent != NULL) { + // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - + // pParent->m_iBegin > 5) { + if (pParent->m_vecChildren.size() > 1) { + // do constituent reordering for all children of pParent + vecFocused.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + void fnGenerateInstanceFile( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); + string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); + string strInstanceRightFname = string(pszInstanceFname) + string(".right"); + WriteFile left_file(strInstanceLeftFname); + WriteFile right_file(strInstanceRightFname); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + string line; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + + assert(getline(*source_file.stream(), line)); + vector vecSTerms; + SplitOnWhitespace(line, &vecSTerms); + + assert(getline(*target_file.stream(), line)); + vector vecTTerms; + SplitOnWhitespace(line, &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; i < vecFocused.size(); i++) { + + STreeItem* pParent = vecFocused[i]; + + vector vecLeft, vecRight; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + STreeItem* pCon1 = pParent->m_vecChildren[j]; + int iLeft1, iRight1; + pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, + iLeft1, iRight1); + vecLeft.push_back(iLeft1); + vecRight.push_back(iRight1); + } + vector vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + vector vecRightPosition; + fnGetRelativePosition(vecRight, vecRightPosition); + + vector vecChunkStatus; + for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { + string strOutcome = + pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, + pParent->m_vecChildren[j]->m_iEnd); + vecChunkStatus.push_back(strOutcome); + } + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strLeftOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, j, vecChunkStatus, + vecLeftPosition, vecSTerms, vecTTerms, + strLeftOutcome, ostr); + + string ostr_str = ostr.str(); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strLeftOutcome.c_str()); + (*left_file.stream()) << ostr_str << " " << strLeftOutcome << "\n"; + + string strRightOutcome; + fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], + strRightOutcome); + (*right_file.stream()) << ostr_str + << " LeftOrder=" << strLeftOutcome << " " + << strRightOutcome << "\n"; + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + delete pAlignReader; + delete pParseReader; + } + + void fnGenerateInstanceFile2( + const char* pszSynFname, // source-side flattened parse tree file name + const char* pszAlignFname, // alignment filename + const char* pszSourceFname, // source file name + const char* pszTargetFname, // target file name + const char* pszInstanceFname // training instance file name + ) { + SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); + SParseReader* pParseReader = new SParseReader(pszSynFname, false); + + ReadFile source_file(pszSourceFname); + ReadFile target_file(pszTargetFname); + + WriteFile output_file(pszInstanceFname); + + // read sentence by sentence + SAlignment* pAlign; + SParsedTree* pTree; + string line; + int iSentNum = 0; + while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { + pTree = pParseReader->fnReadNextParseTree(); + assert(getline(*source_file.stream(), line)); + vector vecSTerms; + SplitOnWhitespace(line, &vecSTerms); + + assert(getline(*target_file.stream(), line)); + vector vecTTerms; + SplitOnWhitespace(line, &vecTTerms); + + if (pTree != NULL) { + + vector vecFocused; + fnGetFocusedParentNodes(pTree, vecFocused); + + for (size_t i = 0; + i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { + + STreeItem* pParent = vecFocused[i]; + + for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { + // children[j-1] vs. children[j] reordering + + string strOutcome; + ostringstream ostr; + + fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j - 1], + pParent->m_vecChildren[j], pAlign, vecSTerms, + vecTTerms, strOutcome, ostr); + + // fprintf(stderr, "%s %s\n", ostr.str().c_str(), + // strOutcome.c_str()); + (*output_file.stream()) << ostr.str() << " " << strOutcome << "\n"; + } + } + delete pTree; + } + + delete pAlign; + iSentNum++; + + if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); + } + + delete pAlignReader; + delete pParseReader; + } +}; + +inline void print_options(std::ostream& out, + po::options_description const& opts) { + typedef std::vector > Ds; + Ds const& ds = opts.options(); + out << '"'; + for (unsigned i = 0; i < ds.size(); ++i) { + if (i) out << ' '; + out << "--" << ds[i]->long_name(); + } + out << '\n'; +} +inline string str(char const* name, po::variables_map const& conf) { + return conf[name].as(); +} + +//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file +/// scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file +/// scratch0/mt_exp/gq-ctb/data/train.cn --target_file +/// scratch0/mt_exp/gq-ctb/data/train.en --instance_file +/// scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix +/// scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 +int main(int argc, char** argv) { + + po::options_description opts("Configuration options"); + opts.add_options()("parse_file", po::value(), + "parse file path (input)")( + "align_file", po::value(), "Alignment file path (input)")( + "source_file", po::value(), "Source text file path (input)")( + "target_file", po::value(), "Target text file path (input)")( + "instance_file", po::value(), "Instance file path (output)")( + "model_prefix", po::value(), + "Model file path prefix (output): three files will be generated")( + "feature_cutoff", po::value()->default_value(100), + "Feature cutoff threshold")("svm_option", po::value(), + "Parameters for SVMLight classifier")( + "help", "produce help message"); + + po::variables_map vm; + if (argc) { + po::store(po::parse_command_line(argc, argv, opts), vm); + po::notify(vm); + } + + if (vm.count("help")) { + print_options(cout, opts); + return 1; + } + + if (!vm.count("parse_file") || !vm.count("align_file") || + !vm.count("source_file") || !vm.count("target_file") || + !vm.count("instance_file") || !vm.count("model_prefix")) { + print_options(cout, opts); + if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; + if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; + if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; + if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; + if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; + if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; + exit(0); + } + + const char* pOption; + if (vm.count("svm_option")) + pOption = str("svm_option", vm).c_str(); + else + pOption = NULL; + + SConstReorderTrainer* pTrainer = new SConstReorderTrainer( + str("parse_file", vm).c_str(), str("align_file", vm).c_str(), + str("source_file", vm).c_str(), str("target_file", vm).c_str(), + str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), + vm["feature_cutoff"].as(), pOption); + delete pTrainer; + + return 0; +} diff --git a/utils/Makefile.am b/utils/Makefile.am index 53967561..fabb4454 100644 --- a/utils/Makefile.am +++ b/utils/Makefile.am @@ -1,4 +1,4 @@ -bin_PROGRAMS = reconstruct_weights atools const_reorder_model_trainer argument_reorder_model_trainer +bin_PROGRAMS = reconstruct_weights atools noinst_PROGRAMS = \ ts \ @@ -20,9 +20,7 @@ noinst_LIBRARIES = libutils.a libutils_a_SOURCES = \ test_data \ alias_sampler.h \ - alignment.h \ alignment_io.h \ - argument_reorder_model.h \ array2d.h \ b64tools.h \ batched_append.h \ @@ -61,15 +59,12 @@ libutils_a_SOURCES = \ show.h \ small_vector.h \ sparse_vector.h \ - srl_sentence.h \ star.h \ static_utoa.h \ stringlib.h \ string_piece.hh \ tdict.h \ timing_stats.h \ - tree.h \ - tsuruoka_maxent.h \ utoa.h \ value_array.h \ verbose.h \ @@ -91,7 +86,6 @@ libutils_a_SOURCES = \ filelib.cc \ stringlib.cc \ string_piece.cc \ - synutils.h \ sparse_vector.cc \ timing_stats.cc \ verbose.cc \ @@ -106,10 +100,6 @@ reconstruct_weights_LDADD = libutils.a atools_SOURCES = atools.cc atools_LDADD = libutils.a atools_LDFLAGS = $(STATIC_FLAGS) -const_reorder_model_trainer_SOURCES = constituent_reorder_model.cc -const_reorder_model_trainer_LDADD = libutils.a -argument_reorder_model_trainer_SOURCES = argument_reorder_model.cc -argument_reorder_model_trainer_LDADD = libutils.a phmt_SOURCES = phmt.cc phmt_LDADD = libutils.a $(BOOST_UNIT_TEST_FRAMEWORK_LDFLAGS) $(BOOST_UNIT_TEST_FRAMEWORK_LIBS) diff --git a/utils/alignment.h b/utils/alignment.h deleted file mode 100644 index 456577ca..00000000 --- a/utils/alignment.h +++ /dev/null @@ -1,200 +0,0 @@ -/* - * alignment.h - * - * Created on: May 23, 2013 - * Author: lijunhui - */ - -#ifndef ALIGNMENT_H_ -#define ALIGNMENT_H_ - -#include -#include -#include -#include - -#include "stringlib.h" - -/* - * Note: - * m_vec_s_align.size() may not be equal to the length of source side - *sentence - * due to the last words may not be aligned - * - */ -struct SAlignment { - typedef std::vector SingleAlign; - SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } - ~SAlignment() {} - - bool fnIsAligned(int i, bool s) const { - const std::vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - if ((*palign)[i].size() == 0) return false; - return true; - } - - /* - * return true if [b, e] is aligned phrases on source side (if s==true) or on - * the target side (if s==false); - * return false, otherwise. - */ - bool fnIsAlignedPhrase(int b, int e, bool s, int* pob, int* poe) const { - int ob, oe; //[b, e] on the other side - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - - if (ob == -1) { - if (pob != NULL) (*pob) = -1; - if (poe != NULL) (*poe) = -1; - return false; // no aligned word among [b, e] - } - if (pob != NULL) (*pob) = ob; - if (poe != NULL) (*poe) = oe; - - int bb, be; //[b, e] back given [ob, oe] on the other side - if (s == true) - fnGetLeftRightMost(ob, oe, m_vec_t_align, bb, be); - else - fnGetLeftRightMost(ob, oe, m_vec_s_align, bb, be); - - if (bb < b || be > e) return false; - return true; - } - - bool fnIsAlignedTightPhrase(int b, int e, bool s, int* pob, int* poe) const { - const std::vector* palign; - if (s == true) - palign = &m_vec_s_align; - else - palign = &m_vec_t_align; - - if ((*palign).size() <= e || (*palign)[b].size() == 0 || - (*palign)[e].size() == 0) - return false; - - return fnIsAlignedPhrase(b, e, s, pob, poe); - } - - void fnGetLeftRightMost(int b, int e, bool s, int& ob, int& oe) const { - if (s == true) - fnGetLeftRightMost(b, e, m_vec_s_align, ob, oe); - else - fnGetLeftRightMost(b, e, m_vec_t_align, ob, oe); - } - - /* - * look the translation of source[b, e] is continuous or not - * 1) return "Unaligned": if the source[b, e] is translated silently; - * 2) return "Con't": if none of target words in target[.., ..] is exclusively - * aligned to any word outside source[b, e] - * 3) return "Discon't": otherwise; - */ - std::string fnIsContinuous(int b, int e) const { - int ob, oe; - fnGetLeftRightMost(b, e, true, ob, oe); - if (ob == -1) return "Unaligned"; - - for (int i = ob; i <= oe; i++) { - if (!fnIsAligned(i, false)) continue; - const SingleAlign& a = m_vec_t_align[i]; - int j; - for (j = 0; j < a.size(); j++) - if (a[j] >= b && a[j] <= e) break; - if (j == a.size()) return "Discon't"; - } - return "Con't"; - } - - const SingleAlign* fnGetSingleWordAlign(int i, bool s) const { - if (s == true) { - if (i >= m_vec_s_align.size()) return NULL; - return &(m_vec_s_align[i]); - } else { - if (i >= m_vec_t_align.size()) return NULL; - return &(m_vec_t_align[i]); - } - } - - private: - void fnGetLeftRightMost(int b, int e, const std::vector& align, - int& ob, int& oe) const { - ob = oe = -1; - for (int i = b; i <= e && i < align.size(); i++) { - if (align[i].size() > 0) { - if (align[i][0] < ob || ob == -1) ob = align[i][0]; - if (oe < align[i][align[i].size() - 1]) - oe = align[i][align[i].size() - 1]; - } - } - } - void fnInitializeAlignment(const char* pszAlign) { - m_vec_s_align.clear(); - m_vec_t_align.clear(); - - std::vector terms = SplitOnWhitespace(std::string(pszAlign)); - int si, ti; - for (size_t i = 0; i < terms.size(); i++) { - sscanf(terms[i].c_str(), "%d-%d", &si, &ti); - - while (m_vec_s_align.size() <= si) { - SingleAlign sa; - m_vec_s_align.push_back(sa); - } - while (m_vec_t_align.size() <= ti) { - SingleAlign sa; - m_vec_t_align.push_back(sa); - } - - m_vec_s_align[si].push_back(ti); - m_vec_t_align[ti].push_back(si); - } - - // sort - for (size_t i = 0; i < m_vec_s_align.size(); i++) { - std::sort(m_vec_s_align[i].begin(), m_vec_s_align[i].end()); - } - for (size_t i = 0; i < m_vec_t_align.size(); i++) { - std::sort(m_vec_t_align[i].begin(), m_vec_t_align[i].end()); - } - } - - private: - std::vector m_vec_s_align; // source side words' alignment - std::vector m_vec_t_align; // target side words' alignment -}; - -struct SAlignmentReader { - SAlignmentReader(const char* pszFname) { - m_fpIn = fopen(pszFname, "r"); - assert(m_fpIn != NULL); - } - ~SAlignmentReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - SAlignment* fnReadNextAlignment() { - if (feof(m_fpIn) == true) return NULL; - char* pszLine = new char[100001]; - pszLine[0] = '\0'; - fgets(pszLine, 10001, m_fpIn); - int iLen = strlen(pszLine); - if (iLen == 0) return NULL; - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - SAlignment* pAlign = new SAlignment(pszLine); - delete[] pszLine; - return pAlign; - } - - private: - FILE* m_fpIn; -}; - -#endif /* ALIGNMENT_H_ */ diff --git a/utils/argument_reorder_model.cc b/utils/argument_reorder_model.cc deleted file mode 100644 index c4e90cba..00000000 --- a/utils/argument_reorder_model.cc +++ /dev/null @@ -1,307 +0,0 @@ -/* - * argument_reorder_model.cc - * - * Created on: Dec 15, 2013 - * Author: lijunhui - */ - -#include -#include -#include -#include -#include -#include - -#include "filelib.h" - -#include "argument_reorder_model.h" -#include "tsuruoka_maxent.h" - -using namespace std; - -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, - const char* pszNewFName) { - Map hashPredicate; - { - ReadFile in(pszFName); - string line; - while (getline(*in.stream(), line)) { - if (!line.size()) continue; - vector terms; - SplitOnWhitespace(line, &terms); - for (const auto& i : terms) { - ++hashPredicate[i]; - } - } - } - - { - ReadFile in(pszFName); - WriteFile out(pszNewFName); - string line; - while (getline(*in.stream(), line)) { - if (!line.size()) continue; - vector terms; - SplitOnWhitespace(line, &terms); - bool written = false; - for (const auto& i : terms) { - if (hashPredicate[i] >= iCutoff) { - (*out.stream()) << i << " "; - written = true; - } - } - if (written) { - (*out.stream()) << "\n"; - } - } - } -} - -struct SArgumentReorderTrainer { - SArgumentReorderTrainer( - const char* pszSRLFname, // source-side srl tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszTopPredicateFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelFname, // classifier model file name - int iCutoff) { - fnGenerateInstanceFiles(pszSRLFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszTopPredicateFname, - pszInstanceFname); - - string strInstanceFname, strModelFname; - strInstanceFname = string(pszInstanceFname) + string(".left"); - strModelFname = string(pszModelFname) + string(".left"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - strInstanceFname = string(pszInstanceFname) + string(".right"); - strModelFname = string(pszModelFname) + string(".right"); - fnTraining(strInstanceFname.c_str(), strModelFname.c_str(), iCutoff); - } - - ~SArgumentReorderTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - void fnGenerateInstanceFiles( - const char* pszSRLFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszTopPredicateFname, // top predicate file name (we only - // consider predicates with 100+ - // occurrences - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SSrlSentenceReader* pSRLReader = new SSrlSentenceReader(pszSRLFname); - ReadFile source_file(pszSourceFname); - ReadFile target_file(pszTargetFname); - - Map* pMapPredicate; - if (pszTopPredicateFname != NULL) - pMapPredicate = fnLoadTopPredicates(pszTopPredicateFname); - else - pMapPredicate = NULL; - - string line; - - WriteFile left_file(pszInstanceFname + string(".left")); - WriteFile right_file(pszInstanceFname + string(".right")); - - // read sentence by sentence - SAlignment* pAlign; - SSrlSentence* pSRL; - SParsedTree* pTree; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pSRL = pSRLReader->fnReadNextSrlSentence(); - assert(pSRL != NULL); - pTree = pSRL->m_pTree; - assert(getline(*source_file.stream(), line)); - vector vecSTerms; - SplitOnWhitespace(line, &vecSTerms); - assert(getline(*target_file.stream(), line)); - vector vecTTerms; - SplitOnWhitespace(line, &vecTTerms); - // vecTPOSTerms.size() == 0, given the case when an english sentence fails - // parsing - - if (pTree != NULL) { - for (size_t i = 0; i < pSRL->m_vecPred.size(); i++) { - SPredicate* pPred = pSRL->m_vecPred[i]; - if (strcmp(pTree->m_vecTerminals[pPred->m_iPosition] - ->m_ptParent->m_pszTerm, - "VA") == 0) - continue; - string strPred = - string(pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm); - if (pMapPredicate != NULL) { - Map::iterator iter_map = pMapPredicate->find(strPred); - if (pMapPredicate != NULL && iter_map == pMapPredicate->end()) - continue; - } - - SPredicateItem* pPredItem = new SPredicateItem(pTree, pPred); - - vector vecStrBlock; - for (size_t j = 0; j < pPredItem->vec_items_.size(); j++) { - SSRLItem* pItem1 = pPredItem->vec_items_[j]; - vecStrBlock.push_back(SArgumentReorderModel::fnGetBlockOutcome( - pItem1->tree_item_->m_iBegin, pItem1->tree_item_->m_iEnd, - pAlign)); - } - - vector vecStrLeftReorderType; - vector vecStrRightReorderType; - SArgumentReorderModel::fnGetReorderType( - pPredItem, pAlign, vecStrLeftReorderType, vecStrRightReorderType); - for (int j = 1; j < pPredItem->vec_items_.size(); j++) { - string strLeftOutcome, strRightOutcome; - strLeftOutcome = vecStrLeftReorderType[j - 1]; - strRightOutcome = vecStrRightReorderType[j - 1]; - ostringstream ostr; - SArgumentReorderModel::fnGenerateFeature(pTree, pPred, pPredItem, j, - vecStrBlock[j - 1], - vecStrBlock[j], ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strOutcome.c_str()); - // fprintf(fpOut, "sentid=%d %s %s\n", iSentNum, ostr.str().c_str(), - // strOutcome.c_str()); - (*left_file.stream()) << ostr.str() << " " << strLeftOutcome - << "\n"; - (*right_file.stream()) << ostr.str() << " " << strRightOutcome - << "\n"; - } - } - } - delete pSRL; - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - delete pAlignReader; - delete pSRLReader; - } - - Map* fnLoadTopPredicates(const char* pszTopPredicateFname) { - if (pszTopPredicateFname == NULL) return NULL; - - Map* pMapPredicate = new Map(); - // STxtFileReader* pReader = new STxtFileReader(pszTopPredicateFname); - ReadFile in(pszTopPredicateFname); - // char* pszLine = new char[50001]; - string line; - int iNumCount = 0; - while (getline(*in.stream(), line)) { - if (line.size() && line[0] == '#') continue; - auto p = line.find(' '); - assert(p != string::npos); - int iCount = atoi(line.substr(p + 1).c_str()); - if (iCount < 100) break; - (*pMapPredicate)[line] = iNumCount++; - } - return pMapPredicate; - } -}; - -namespace po = boost::program_options; - -inline void print_options(std::ostream& out, - po::options_description const& opts) { - typedef std::vector > Ds; - Ds const& ds = opts.options(); - out << '"'; - for (unsigned i = 0; i < ds.size(); ++i) { - if (i) out << ' '; - out << "--" << ds[i]->long_name(); - } - out << '\n'; -} -inline string str(char const* name, po::variables_map const& conf) { - return conf[name].as(); -} - -//--srl_file /scratch0/mt_exp/gale-align/gale-align.nw.srl.cn --align_file -/// scratch0/mt_exp/gale-align/gale-align.nw.al --source_file -/// scratch0/mt_exp/gale-align/gale-align.nw.cn --target_file -/// scratch0/mt_exp/gale-align/gale-align.nw.en --instance_file -/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.instance --model_prefix -/// scratch0/mt_exp/gale-align/gale-align.nw.argreorder.model --feature_cutoff 2 -//--srl_file /scratch0/mt_exp/gale-ctb/gale-ctb.srl.cn --align_file -/// scratch0/mt_exp/gale-ctb/gale-ctb.align --source_file -/// scratch0/mt_exp/gale-ctb/gale-ctb.cn --target_file -/// scratch0/mt_exp/gale-ctb/gale-ctb.en0 --instance_file -/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.instance --model_prefix -/// scratch0/mt_exp/gale-ctb/gale-ctb.argreorder.model --feature_cutoff 2 -int main(int argc, char** argv) { - - po::options_description opts("Configuration options"); - opts.add_options()("srl_file", po::value(), "srl file path (input)")( - "align_file", po::value(), "Alignment file path (input)")( - "source_file", po::value(), "Source text file path (input)")( - "target_file", po::value(), "Target text file path (input)")( - "instance_file", po::value(), "Instance file path (output)")( - "model_prefix", po::value(), - "Model file path prefix (output): three files will be generated")( - "feature_cutoff", po::value()->default_value(100), - "Feature cutoff threshold")("help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("srl_file") || !vm.count("align_file") || - !vm.count("source_file") || !vm.count("target_file") || - !vm.count("instance_file") || !vm.count("model_prefix")) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - SArgumentReorderTrainer* pTrainer = new SArgumentReorderTrainer( - str("srl_file", vm).c_str(), str("align_file", vm).c_str(), - str("source_file", vm).c_str(), str("target_file", vm).c_str(), NULL, - str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), - vm["feature_cutoff"].as()); - delete pTrainer; - - return 1; -} diff --git a/utils/argument_reorder_model.h b/utils/argument_reorder_model.h deleted file mode 100644 index 077fa5ba..00000000 --- a/utils/argument_reorder_model.h +++ /dev/null @@ -1,246 +0,0 @@ -/* - * argument_reorder_model.h - * - * Created on: Dec 15, 2013 - * Author: lijunhui - */ - -#ifndef ARGUMENT_REORDER_MODEL_H_ -#define ARGUMENT_REORDER_MODEL_H_ - -#include -#include - -#include "alignment.h" -#include "tree.h" -#include "srl_sentence.h" - -// an argument item or a predicate item (the verb itself) -struct SSRLItem { - SSRLItem(const STreeItem *tree_item, std::string role) - : tree_item_(tree_item), role_(role) {} - ~SSRLItem() {} - const STreeItem *tree_item_; - const std::string role_; -}; - -struct SPredicateItem { - SPredicateItem(const SParsedTree *tree, const SPredicate *pred) - : pred_(pred) { - vec_items_.reserve(pred->m_vecArgt.size() + 1); - for (int i = 0; i < pred->m_vecArgt.size(); i++) { - vec_items_.push_back( - new SSRLItem(pred->m_vecArgt[i]->m_pTreeItem, - std::string(pred->m_vecArgt[i]->m_pszRole))); - } - vec_items_.push_back( - new SSRLItem(tree->m_vecTerminals[pred->m_iPosition]->m_ptParent, - std::string("Pred"))); - sort(vec_items_.begin(), vec_items_.end(), SortFunction); - - begin_ = vec_items_[0]->tree_item_->m_iBegin; - end_ = vec_items_[vec_items_.size() - 1]->tree_item_->m_iEnd; - } - - ~SPredicateItem() { vec_items_.clear(); } - - static bool SortFunction(SSRLItem *i, SSRLItem *j) { - return (i->tree_item_->m_iBegin < j->tree_item_->m_iBegin); - } - - std::vector vec_items_; - int begin_; - int end_; - const SPredicate *pred_; -}; - -struct SArgumentReorderModel { - public: - static std::string fnGetBlockOutcome(int iBegin, int iEnd, - SAlignment *pAlign) { - return pAlign->fnIsContinuous(iBegin, iEnd); - } - static void fnGetReorderType(SPredicateItem *pPredItem, SAlignment *pAlign, - std::vector &vecStrLeftReorder, - std::vector &vecStrRightReorder) { - std::vector vecLeft, vecRight; - for (int i = 0; i < pPredItem->vec_items_.size(); i++) { - const STreeItem *pCon1 = pPredItem->vec_items_[i]->tree_item_; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, - iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - std::vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - std::vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vecStrLeftReorder.clear(); - vecStrRightReorder.clear(); - for (int i = 1; i < vecLeftPosition.size(); i++) { - std::string strOutcome; - fnGetOutcome(vecLeftPosition[i - 1], vecLeftPosition[i], strOutcome); - vecStrLeftReorder.push_back(strOutcome); - fnGetOutcome(vecRightPosition[i - 1], vecRightPosition[i], strOutcome); - vecStrRightReorder.push_back(strOutcome); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - * - * f1: (left_role, right_role, predicate_term) - * f2: (left_role, right_role, predicate_term, other_right_role) - * f3: (left_role, right_role, predicate_term, other_left_role) - * f4: (left_role, right_role, left_head_pos) - * f5: (left_role, right_role, left_head_word) - * f6: (left_role, right_role, left_syntactic_label) - * f7: (left_role, right_role, right_head_pos) - * f8: (left_role, right_role, right_head_word) - * f8: (left_role, right_role, right_syntactic_label) - * f8: (left_role, right_role, left_chunk_status) - * f9: (left_role, right_role, right_chunk_status) - * f10: (left_role, right_role, left_chunk_status) - * f11: (left_role, right_role, right_chunk_status) - * f12: (left_label, parent_label) - * f13: (right_label, parent_label) - */ - static void fnGenerateFeature(const SParsedTree *pTree, - const SPredicate *pPred, - const SPredicateItem *pPredItem, int iPos, - const std::string &strBlock1, - const std::string &strBlock2, - std::ostringstream &ostr) { - SSRLItem *pSRLItem1 = pPredItem->vec_items_[iPos - 1]; - SSRLItem *pSRLItem2 = pPredItem->vec_items_[iPos]; - const STreeItem *pCon1 = pSRLItem1->tree_item_; - const STreeItem *pCon2 = pSRLItem2->tree_item_; - - std::string left_role = pSRLItem1->role_; - std::string right_role = pSRLItem2->role_; - - std::string predicate_term = - pTree->m_vecTerminals[pPred->m_iPosition]->m_pszTerm; - - std::vector vec_other_right_sibling; - for (int i = iPos + 1; i < pPredItem->vec_items_.size(); i++) - vec_other_right_sibling.push_back( - std::string(pPredItem->vec_items_[i]->role_)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(std::string("NULL")); - - std::vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_right_sibling.push_back( - std::string(pPredItem->vec_items_[i]->role_)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(std::string("NULL")); - - // generate features - // f1 - ostr << "f1=" << left_role << "_" << right_role << "_" << predicate_term; - ostr << "f1=" << left_role << "_" << right_role; - - // f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) { - ostr << " f2=" << left_role << "_" << right_role << "_" << predicate_term - << "_" << vec_other_right_sibling[i]; - ostr << " f2=" << left_role << "_" << right_role << "_" - << vec_other_right_sibling[i]; - } - // f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) { - ostr << " f3=" << left_role << "_" << right_role << "_" << predicate_term - << "_" << vec_other_left_sibling[i]; - ostr << " f3=" << left_role << "_" << right_role << "_" - << vec_other_left_sibling[i]; - } - // f4 - ostr << " f4=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - // f5 - ostr << " f5=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; - // f6 - ostr << " f6=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - // f7 - ostr << " f7=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - // f8 - ostr << " f8=" << left_role << "_" << right_role << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; - // f9 - ostr << " f9=" << left_role << "_" << right_role << "_" << pCon2->m_pszTerm; - // f10 - ostr << " f10=" << left_role << "_" << right_role << "_" << strBlock1; - // f11 - ostr << " f11=" << left_role << "_" << right_role << "_" << strBlock2; - // f12 - ostr << " f12=" << left_role << "_" << predicate_term; - ostr << " f12=" << left_role; - // f13 - ostr << " f13=" << right_role << "_" << predicate_term; - ostr << " f13=" << right_role; - } - - private: - static void fnGetOutcome(int i1, int i2, std::string &strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) - strOutcome = std::string("DM"); - else - strOutcome = std::string("M"); - } else { - if (i1 > i2 + 1) - strOutcome = std::string("DS"); - else - strOutcome = std::string("S"); - } - } - - static void fnGetRelativePosition(const std::vector &vecLeft, - std::vector &vecPosition) { - vecPosition.clear(); - - std::vector vec; - for (int i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i - 1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (int i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (int j = 0; j < vecLeft.size(); j++) { - if (j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } -}; - -#endif /* ARGUMENT_REORDER_MODEL_H_ */ diff --git a/utils/constituent_reorder_model.cc b/utils/constituent_reorder_model.cc deleted file mode 100644 index bdb7c5d1..00000000 --- a/utils/constituent_reorder_model.cc +++ /dev/null @@ -1,637 +0,0 @@ -/* - * constituent_reorder_model.cc - * - * Created on: Jul 10, 2013 - * Author: junhuili - */ - -#include -#include - -#include - -#include "filelib.h" - -#include "alignment.h" -#include "tree.h" -#include "tsuruoka_maxent.h" - -using namespace std; - -typedef std::unordered_map Map; -typedef std::unordered_map::iterator Iterator; - -namespace po = boost::program_options; - -inline void fnPreparingTrainingdata(const char* pszFName, int iCutoff, - const char* pszNewFName) { - Map hashPredicate; - { - ReadFile f(pszFName); - string line; - while (getline(*f.stream(), line)) { - if (!line.size()) continue; - vector terms; - SplitOnWhitespace(line, &terms); - for (const auto& i : terms) { - ++hashPredicate[i]; - } - } - } - - { - ReadFile in(pszFName); - WriteFile out(pszNewFName); - string line; - while (getline(*in.stream(), line)) { - if (!line.size()) continue; - vector terms; - SplitOnWhitespace(line, &terms); - bool written = false; - for (const auto& i : terms) { - if (hashPredicate[i] >= iCutoff) { - (*out.stream()) << i << " "; - written = true; - } - } - if (written) { - (*out.stream()) << "\n"; - } - } - } -} - -struct SConstReorderTrainer { - SConstReorderTrainer( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname, // training instance file name - const char* pszModelPrefix, // classifier model file name prefix - int iCutoff, // feature count threshold - const char* /*pszOption*/ // other classifier parameters (for svmlight) - ) { - fnGenerateInstanceFile(pszSynFname, pszAlignFname, pszSourceFname, - pszTargetFname, pszInstanceFname); - - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - - string strModelLeftFname = string(pszModelPrefix) + string(".left"); - string strModelRightFname = string(pszModelPrefix) + string(".right"); - - fprintf(stdout, "...Training the left ordering model\n"); - fnTraining(strInstanceLeftFname.c_str(), strModelLeftFname.c_str(), - iCutoff); - fprintf(stdout, "...Training the right ordering model\n"); - fnTraining(strInstanceRightFname.c_str(), strModelRightFname.c_str(), - iCutoff); - } - ~SConstReorderTrainer() {} - - private: - void fnTraining(const char* pszInstanceFname, const char* pszModelFname, - int iCutoff) { - char* pszNewInstanceFName = new char[strlen(pszInstanceFname) + 50]; - if (iCutoff > 0) { - sprintf(pszNewInstanceFName, "%s.tmp", pszInstanceFname); - fnPreparingTrainingdata(pszInstanceFname, iCutoff, pszNewInstanceFName); - } else { - strcpy(pszNewInstanceFName, pszInstanceFname); - } - - /*Zhangle_Maxent *pZhangleMaxent = new Zhangle_Maxent(NULL); -pZhangleMaxent->fnTrain(pszInstanceFname, "lbfgs", pszModelFname, 100, 2.0); -delete pZhangleMaxent;*/ - - Tsuruoka_Maxent* pMaxent = new Tsuruoka_Maxent(NULL); - pMaxent->fnTrain(pszNewInstanceFName, "l1", pszModelFname, 300); - delete pMaxent; - - if (strcmp(pszNewInstanceFName, pszInstanceFname) != 0) { - sprintf(pszNewInstanceFName, "rm %s.tmp", pszInstanceFname); - system(pszNewInstanceFName); - } - delete[] pszNewInstanceFName; - } - - inline bool fnIsVerbPOS(const char* pszTerm) { - if (strcmp(pszTerm, "VV") == 0 || strcmp(pszTerm, "VA") == 0 || - strcmp(pszTerm, "VC") == 0 || strcmp(pszTerm, "VE") == 0) - return true; - return false; - } - - inline void fnGetOutcome(int iL1, int iR1, int iL2, int iR2, - const SAlignment* /*pAlign*/, string& strOutcome) { - if (iL1 == -1 && iL2 == -1) - strOutcome = "BU"; // 1. both are untranslated - else if (iL1 == -1) - strOutcome = "1U"; // 2. XP1 is untranslated - else if (iL2 == -1) - strOutcome = "2U"; // 3. XP2 is untranslated - else if (iL1 == iL2 && iR1 == iR2) - strOutcome = "SS"; // 4. Have same scope - else if (iL1 <= iL2 && iR1 >= iR2) - strOutcome = "1C2"; // 5. XP1's translation covers XP2's - else if (iL1 >= iL2 && iR1 <= iR2) - strOutcome = "2C1"; // 6. XP2's translation covers XP1's - else if (iR1 < iL2) { - int i = iR1 + 1; - /*while (i < iL2) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL2) - strOutcome = "M"; // 7. Monotone - else - strOutcome = "DM"; // 8. Discontinuous monotone - } else if (iL1 < iL2 && iL2 <= iR1 && iR1 < iR2) - strOutcome = "OM"; // 9. Overlap monotone - else if (iR2 < iL1) { - int i = iR2 + 1; - /*while (i < iL1) { - if (pAlign->fnIsAligned(i, false)) - break; - i++; - }*/ - if (i == iL1) - strOutcome = "S"; // 10. Swap - else - strOutcome = "DS"; // 11. Discontinuous swap - } else if (iL2 < iL1 && iL1 <= iR2 && iR2 < iR1) - strOutcome = "OS"; // 12. Overlap swap - else - assert(false); - } - - inline void fnGetOutcome(int i1, int i2, string& strOutcome) { - assert(i1 != i2); - if (i1 < i2) { - if (i2 > i1 + 1) - strOutcome = string("DM"); - else - strOutcome = string("M"); - } else { - if (i1 > i2 + 1) - strOutcome = string("DS"); - else - strOutcome = string("S"); - } - } - - inline void fnGetRelativePosition(const vector& vecLeft, - vector& vecPosition) { - vecPosition.clear(); - - vector vec; - for (size_t i = 0; i < vecLeft.size(); i++) { - if (vecLeft[i] == -1) { - if (i == 0) - vec.push_back(-1); - else - vec.push_back(vecLeft[i - 1] + 0.1); - } else - vec.push_back(vecLeft[i]); - } - - for (size_t i = 0; i < vecLeft.size(); i++) { - int count = 0; - - for (size_t j = 0; j < vecLeft.size(); j++) { - if (j == i) continue; - if (vec[j] < vec[i]) { - count++; - } else if (vec[j] == vec[i] && j < i) { - count++; - } - } - vecPosition.push_back(count); - } - } - - /* - * features: - * f1: (left_label, right_label, parent_label) - * f2: (left_label, right_label, parent_label, other_right_sibling_label) - * f3: (left_label, right_label, parent_label, other_left_sibling_label) - * f4: (left_label, right_label, left_head_pos) - * f5: (left_label, right_label, left_head_word) - * f6: (left_label, right_label, right_head_pos) - * f7: (left_label, right_label, right_head_word) - * f8: (left_label, right_label, left_chunk_status) - * f9: (left_label, right_label, right_chunk_status) - * f10: (left_label, parent_label) - * f11: (right_label, parent_label) - */ - void fnGenerateInstance(const SParsedTree* pTree, const STreeItem* pParent, - int iPos, const vector& vecChunkStatus, - const vector& vecPosition, - const vector& vecSTerms, - const vector& /*vecTTerms*/, string& strOutcome, - ostringstream& ostr) { - STreeItem* pCon1, *pCon2; - pCon1 = pParent->m_vecChildren[iPos - 1]; - pCon2 = pParent->m_vecChildren[iPos]; - - fnGetOutcome(vecPosition[iPos - 1], vecPosition[iPos], strOutcome); - - string left_label = string(pCon1->m_pszTerm); - string right_label = string(pCon2->m_pszTerm); - string parent_label = string(pParent->m_pszTerm); - - vector vec_other_right_sibling; - for (int i = iPos + 1; i < pParent->m_vecChildren.size(); i++) - vec_other_right_sibling.push_back( - string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_right_sibling.size() == 0) - vec_other_right_sibling.push_back(string("NULL")); - vector vec_other_left_sibling; - for (int i = 0; i < iPos - 1; i++) - vec_other_left_sibling.push_back( - string(pParent->m_vecChildren[i]->m_pszTerm)); - if (vec_other_left_sibling.size() == 0) - vec_other_left_sibling.push_back(string("NULL")); - - // generate features - // f1 - ostr << "f1=" << left_label << "_" << right_label << "_" << parent_label; - // f2 - for (int i = 0; i < vec_other_right_sibling.size(); i++) - ostr << " f2=" << left_label << "_" << right_label << "_" << parent_label - << "_" << vec_other_right_sibling[i]; - // f3 - for (int i = 0; i < vec_other_left_sibling.size(); i++) - ostr << " f3=" << left_label << "_" << right_label << "_" << parent_label - << "_" << vec_other_left_sibling[i]; - // f4 - ostr << " f4=" << left_label << "_" << right_label << "_" - << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_ptParent->m_pszTerm; - // f5 - ostr << " f5=" << left_label << "_" << right_label << "_" - << vecSTerms[pCon1->m_iHeadWord]; - // f6 - ostr << " f6=" << left_label << "_" << right_label << "_" - << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; - // f7 - ostr << " f7=" << left_label << "_" << right_label << "_" - << vecSTerms[pCon2->m_iHeadWord]; - // f8 - ostr << " f8=" << left_label << "_" << right_label << "_" - << vecChunkStatus[iPos - 1]; - // f9 - ostr << " f9=" << left_label << "_" << right_label << "_" - << vecChunkStatus[iPos]; - // f10 - ostr << " f10=" << left_label << "_" << parent_label; - // f11 - ostr << " f11=" << right_label << "_" << parent_label; - } - - /* - * Source side (11 features): - * f1: the categories of XP1 and XP2 (f1_1, f1_2) - * f2: the head words of XP1 and XP2 (f2_1, f2_2) - * f3: the first and last word of XP1 (f3_f, f3_l) - * f4: the first and last word of XP2 (f4_f, f4_l) - * f5: is XP1 or XP2 the head node (f5_1, f5_2) - * f6: the category of the common parent - * Target side (6 features): - * f7: the first and the last word of XP1's translation (f7_f, f7_l) - * f8: the first and the last word of XP2's translation (f8_f, f8_l) - * f9: the translation of XP1's and XP2's head word (f9_1, f9_2) - */ - void fnGenerateInstance(const SParsedTree* /*pTree*/, const STreeItem* pParent, - const STreeItem* pCon1, const STreeItem* pCon2, - const SAlignment* pAlign, - const vector& vecSTerms, - const vector& /*vecTTerms*/, string& strOutcome, - ostringstream& ostr) { - - int iLeft1, iRight1, iLeft2, iRight2; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, iLeft1, - iRight1); - pAlign->fnGetLeftRightMost(pCon2->m_iBegin, pCon2->m_iEnd, true, iLeft2, - iRight2); - - fnGetOutcome(iLeft1, iRight1, iLeft2, iRight2, pAlign, strOutcome); - - // generate features - // f1 - ostr << "f1_1=" << pCon1->m_pszTerm << " f1_2=" << pCon2->m_pszTerm; - // f2 - ostr << " f2_1=" << vecSTerms[pCon1->m_iHeadWord] << " f2_2" - << vecSTerms[pCon2->m_iHeadWord]; - // f3 - ostr << " f3_f=" << vecSTerms[pCon1->m_iBegin] - << " f3_l=" << vecSTerms[pCon1->m_iEnd]; - // f4 - ostr << " f4_f=" << vecSTerms[pCon2->m_iBegin] - << " f4_l=" << vecSTerms[pCon2->m_iEnd]; - // f5 - if (pParent->m_iHeadChild == pCon1->m_iBrotherIndex) - ostr << " f5_1=1"; - else - ostr << " f5_1=0"; - if (pParent->m_iHeadChild == pCon2->m_iBrotherIndex) - ostr << " f5_2=1"; - else - ostr << " f5_2=0"; - // f6 - ostr << " f6=" << pParent->m_pszTerm; - - /*//f7 - if (iLeft1 != -1) { - ostr << " f7_f=" << vecTTerms[iLeft1] << " f7_l=" << - vecTTerms[iRight1]; - } - if (iLeft2 != -1) { - ostr << " f8_f=" << vecTTerms[iLeft2] << " f8_l=" << - vecTTerms[iRight2]; - } - - const vector* pvecTarget = - pAlign->fnGetSingleWordAlign(pCon1->m_iHeadWord, true); - string str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_1=" << str.substr(0, str.size()-1); - } - pvecTarget = pAlign->fnGetSingleWordAlign(pCon2->m_iHeadWord, true); - str = ""; - for (size_t i = 0; pvecTarget != NULL && i < pvecTarget->size(); i++) { - str += vecTTerms[(*pvecTarget)[i]] + "_"; - } - if (str.length() > 0) { - ostr << " f9_2=" << str.substr(0, str.size()-1); - } */ - } - - void fnGetFocusedParentNodes(const SParsedTree* pTree, - vector& vecFocused) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) { - STreeItem* pParent = pTree->m_vecTerminals[i]->m_ptParent; - - while (pParent != NULL) { - // if (pParent->m_vecChildren.size() > 1 && pParent->m_iEnd - - // pParent->m_iBegin > 5) { - if (pParent->m_vecChildren.size() > 1) { - // do constituent reordering for all children of pParent - vecFocused.push_back(pParent); - } - if (pParent->m_iBrotherIndex != 0) break; - pParent = pParent->m_ptParent; - } - } - } - - void fnGenerateInstanceFile( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszSynFname, false); - - ReadFile source_file(pszSourceFname); - ReadFile target_file(pszTargetFname); - string strInstanceLeftFname = string(pszInstanceFname) + string(".left"); - string strInstanceRightFname = string(pszInstanceFname) + string(".right"); - WriteFile left_file(strInstanceLeftFname); - WriteFile right_file(strInstanceRightFname); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - string line; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - - assert(getline(*source_file.stream(), line)); - vector vecSTerms; - SplitOnWhitespace(line, &vecSTerms); - - assert(getline(*target_file.stream(), line)); - vector vecTTerms; - SplitOnWhitespace(line, &vecTTerms); - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; i < vecFocused.size(); i++) { - - STreeItem* pParent = vecFocused[i]; - - vector vecLeft, vecRight; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - STreeItem* pCon1 = pParent->m_vecChildren[j]; - int iLeft1, iRight1; - pAlign->fnGetLeftRightMost(pCon1->m_iBegin, pCon1->m_iEnd, true, - iLeft1, iRight1); - vecLeft.push_back(iLeft1); - vecRight.push_back(iRight1); - } - vector vecLeftPosition; - fnGetRelativePosition(vecLeft, vecLeftPosition); - vector vecRightPosition; - fnGetRelativePosition(vecRight, vecRightPosition); - - vector vecChunkStatus; - for (size_t j = 0; j < pParent->m_vecChildren.size(); j++) { - string strOutcome = - pAlign->fnIsContinuous(pParent->m_vecChildren[j]->m_iBegin, - pParent->m_vecChildren[j]->m_iEnd); - vecChunkStatus.push_back(strOutcome); - } - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strLeftOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, j, vecChunkStatus, - vecLeftPosition, vecSTerms, vecTTerms, - strLeftOutcome, ostr); - - string ostr_str = ostr.str(); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strLeftOutcome.c_str()); - (*left_file.stream()) << ostr_str << " " << strLeftOutcome << "\n"; - - string strRightOutcome; - fnGetOutcome(vecRightPosition[j - 1], vecRightPosition[j], - strRightOutcome); - (*right_file.stream()) << ostr_str - << " LeftOrder=" << strLeftOutcome << " " - << strRightOutcome << "\n"; - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - delete pAlignReader; - delete pParseReader; - } - - void fnGenerateInstanceFile2( - const char* pszSynFname, // source-side flattened parse tree file name - const char* pszAlignFname, // alignment filename - const char* pszSourceFname, // source file name - const char* pszTargetFname, // target file name - const char* pszInstanceFname // training instance file name - ) { - SAlignmentReader* pAlignReader = new SAlignmentReader(pszAlignFname); - SParseReader* pParseReader = new SParseReader(pszSynFname, false); - - ReadFile source_file(pszSourceFname); - ReadFile target_file(pszTargetFname); - - WriteFile output_file(pszInstanceFname); - - // read sentence by sentence - SAlignment* pAlign; - SParsedTree* pTree; - string line; - int iSentNum = 0; - while ((pAlign = pAlignReader->fnReadNextAlignment()) != NULL) { - pTree = pParseReader->fnReadNextParseTree(); - assert(getline(*source_file.stream(), line)); - vector vecSTerms; - SplitOnWhitespace(line, &vecSTerms); - - assert(getline(*target_file.stream(), line)); - vector vecTTerms; - SplitOnWhitespace(line, &vecTTerms); - - if (pTree != NULL) { - - vector vecFocused; - fnGetFocusedParentNodes(pTree, vecFocused); - - for (size_t i = 0; - i < vecFocused.size() && pTree->m_vecTerminals.size() > 10; i++) { - - STreeItem* pParent = vecFocused[i]; - - for (size_t j = 1; j < pParent->m_vecChildren.size(); j++) { - // children[j-1] vs. children[j] reordering - - string strOutcome; - ostringstream ostr; - - fnGenerateInstance(pTree, pParent, pParent->m_vecChildren[j - 1], - pParent->m_vecChildren[j], pAlign, vecSTerms, - vecTTerms, strOutcome, ostr); - - // fprintf(stderr, "%s %s\n", ostr.str().c_str(), - // strOutcome.c_str()); - (*output_file.stream()) << ostr.str() << " " << strOutcome << "\n"; - } - } - delete pTree; - } - - delete pAlign; - iSentNum++; - - if (iSentNum % 100000 == 0) fprintf(stderr, "#%d\n", iSentNum); - } - - delete pAlignReader; - delete pParseReader; - } -}; - -inline void print_options(std::ostream& out, - po::options_description const& opts) { - typedef std::vector > Ds; - Ds const& ds = opts.options(); - out << '"'; - for (unsigned i = 0; i < ds.size(); ++i) { - if (i) out << ' '; - out << "--" << ds[i]->long_name(); - } - out << '\n'; -} -inline string str(char const* name, po::variables_map const& conf) { - return conf[name].as(); -} - -//--parse_file /scratch0/mt_exp/gq-ctb/data/train.srl.cn --align_file -/// scratch0/mt_exp/gq-ctb/data/aligned.grow-diag-final-and --source_file -/// scratch0/mt_exp/gq-ctb/data/train.cn --target_file -/// scratch0/mt_exp/gq-ctb/data/train.en --instance_file -/// scratch0/mt_exp/gq-ctb/data/srl-instance --model_prefix -/// scratch0/mt_exp/gq-ctb/data/srl-instance --feature_cutoff 10 -int main(int argc, char** argv) { - - po::options_description opts("Configuration options"); - opts.add_options()("parse_file", po::value(), - "parse file path (input)")( - "align_file", po::value(), "Alignment file path (input)")( - "source_file", po::value(), "Source text file path (input)")( - "target_file", po::value(), "Target text file path (input)")( - "instance_file", po::value(), "Instance file path (output)")( - "model_prefix", po::value(), - "Model file path prefix (output): three files will be generated")( - "feature_cutoff", po::value()->default_value(100), - "Feature cutoff threshold")("svm_option", po::value(), - "Parameters for SVMLight classifier")( - "help", "produce help message"); - - po::variables_map vm; - if (argc) { - po::store(po::parse_command_line(argc, argv, opts), vm); - po::notify(vm); - } - - if (vm.count("help")) { - print_options(cout, opts); - return 1; - } - - if (!vm.count("parse_file") || !vm.count("align_file") || - !vm.count("source_file") || !vm.count("target_file") || - !vm.count("instance_file") || !vm.count("model_prefix")) { - print_options(cout, opts); - if (!vm.count("parse_file")) cout << "--parse_file NOT FOUND\n"; - if (!vm.count("align_file")) cout << "--align_file NOT FOUND\n"; - if (!vm.count("source_file")) cout << "--source_file NOT FOUND\n"; - if (!vm.count("target_file")) cout << "--target_file NOT FOUND\n"; - if (!vm.count("instance_file")) cout << "--instance_file NOT FOUND\n"; - if (!vm.count("model_prefix")) cout << "--model_prefix NOT FOUND\n"; - exit(0); - } - - const char* pOption; - if (vm.count("svm_option")) - pOption = str("svm_option", vm).c_str(); - else - pOption = NULL; - - SConstReorderTrainer* pTrainer = new SConstReorderTrainer( - str("parse_file", vm).c_str(), str("align_file", vm).c_str(), - str("source_file", vm).c_str(), str("target_file", vm).c_str(), - str("instance_file", vm).c_str(), str("model_prefix", vm).c_str(), - vm["feature_cutoff"].as(), pOption); - delete pTrainer; - - return 0; -} diff --git a/utils/srl_sentence.h b/utils/srl_sentence.h deleted file mode 100644 index 9d509600..00000000 --- a/utils/srl_sentence.h +++ /dev/null @@ -1,211 +0,0 @@ -/* - * srl_sentence.h - * - * Created on: May 26, 2013 - * Author: junhuili - */ - -#ifndef SRL_SENTENCE_H_ -#define SRL_SENTENCE_H_ - -#include -#include - -#include "tree.h" -#include "stringlib.h" - -struct SArgument { - SArgument(const char* pszRole, int iBegin, int iEnd, float fProb) { - m_pszRole = new char[strlen(pszRole) + 1]; - strcpy(m_pszRole, pszRole); - m_iBegin = iBegin; - m_iEnd = iEnd; - m_fProb = fProb; - m_pTreeItem = NULL; - } - ~SArgument() { delete[] m_pszRole; } - - void fnSetTreeItem(STreeItem* pTreeItem) { - m_pTreeItem = pTreeItem; - if (m_pTreeItem != NULL && m_pTreeItem->m_iBegin != -1) { - assert(m_pTreeItem->m_iBegin == m_iBegin); - assert(m_pTreeItem->m_iEnd == m_iEnd); - } - } - - char* m_pszRole; // argument rule, e.g., ARG0, ARGM-TMP - int m_iBegin; - int m_iEnd; // the span of the argument, [m_iBegin, m_iEnd] - float m_fProb; // the probability of this role, - STreeItem* m_pTreeItem; -}; - -struct SPredicate { - SPredicate(const char* pszLemma, int iPosition) { - if (pszLemma != NULL) { - m_pszLemma = new char[strlen(pszLemma) + 1]; - strcpy(m_pszLemma, pszLemma); - } else - m_pszLemma = NULL; - m_iPosition = iPosition; - } - ~SPredicate() { - if (m_pszLemma != NULL) delete[] m_pszLemma; - for (size_t i = 0; i < m_vecArgt.size(); i++) delete m_vecArgt[i]; - } - int fnAppend(const char* pszRole, int iBegin, int iEnd) { - SArgument* pArgt = new SArgument(pszRole, iBegin, iEnd, 1.0); - return fnAppend(pArgt); - } - int fnAppend(SArgument* pArgt) { - m_vecArgt.push_back(pArgt); - int iPosition = m_vecArgt.size() - 1; - return iPosition; - } - - char* m_pszLemma; // lemma of the predicate, for Chinese, it's always as same - // as the predicate itself - int m_iPosition; // the position in sentence - std::vector m_vecArgt; // arguments associated to the predicate -}; - -struct SSrlSentence { - SSrlSentence() { m_pTree = NULL; } - ~SSrlSentence() { - if (m_pTree != NULL) delete m_pTree; - - for (size_t i = 0; i < m_vecPred.size(); i++) delete m_vecPred[i]; - } - int fnAppend(const char* pszLemma, int iPosition) { - SPredicate* pPred = new SPredicate(pszLemma, iPosition); - return fnAppend(pPred); - } - int fnAppend(SPredicate* pPred) { - m_vecPred.push_back(pPred); - int iPosition = m_vecPred.size() - 1; - return iPosition; - } - int GetPredicateNum() { return m_vecPred.size(); } - - SParsedTree* m_pTree; - std::vector m_vecPred; -}; - -struct SSrlSentenceReader { - SSrlSentenceReader(const char* pszSrlFname) { - m_fpIn = fopen(pszSrlFname, "r"); - assert(m_fpIn != NULL); - } - ~SSrlSentenceReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - inline void fnReplaceAll(std::string& str, const std::string& from, - const std::string& to) { - size_t start_pos = 0; - while ((start_pos = str.find(from, start_pos)) != std::string::npos) { - str.replace(start_pos, from.length(), to); - start_pos += to.length(); // In case 'to' contains 'from', like replacing - // 'x' with 'yx' - } - } - - // TODO: here only considers flat predicate-argument structure - // i.e., no overlap among them - SSrlSentence* fnReadNextSrlSentence() { - std::vector > vecContent; - if (fnReadNextContent(vecContent) == false) return NULL; - - SSrlSentence* pSrlSentence = new SSrlSentence(); - int iSize = vecContent.size(); - // put together syntactic text - std::ostringstream ostr; - for (int i = 0; i < iSize; i++) { - std::string strSynSeg = - vecContent[i][5]; // the 5th column is the syntactic segment - size_t iPosition = strSynSeg.find_first_of('*'); - assert(iPosition != std::string::npos); - std::ostringstream ostrTmp; - ostrTmp << "(" << vecContent[i][2] << " " << vecContent[i][0] - << ")"; // the 2th column is POS-tag, and the 0th column is word - strSynSeg.replace(iPosition, 1, ostrTmp.str()); - fnReplaceAll(strSynSeg, "(", " ("); - ostr << strSynSeg; - } - std::string strSyn = ostr.str(); - pSrlSentence->m_pTree = SParsedTree::fnConvertFromString(strSyn.c_str()); - pSrlSentence->m_pTree->fnSetHeadWord(); - pSrlSentence->m_pTree->fnSetSpanInfo(); - - // read predicate-argument structure - int iNumPred = vecContent[0].size() - 8; - for (int i = 0; i < iNumPred; i++) { - std::vector vecRole; - std::vector vecBegin; - std::vector vecEnd; - int iPred = -1; - for (int j = 0; j < iSize; j++) { - const char* p = vecContent[j][i + 8].c_str(); - const char* q; - if (p[0] == '(') { - // starting position of an argument(or predicate) - vecBegin.push_back(j); - q = strchr(p, '*'); - assert(q != NULL); - vecRole.push_back(vecContent[j][i + 8].substr(1, q - p - 1)); - if (vecRole.back().compare("V") == 0) { - assert(iPred == -1); - iPred = vecRole.size() - 1; - } - } - if (p[strlen(p) - 1] == ')') { - // end position of an argument(or predicate) - vecEnd.push_back(j); - assert(vecBegin.size() == vecEnd.size()); - } - } - assert(iPred != -1); - SPredicate* pPred = new SPredicate( - pSrlSentence->m_pTree->m_vecTerminals[vecBegin[iPred]]->m_pszTerm, - vecBegin[iPred]); - pSrlSentence->fnAppend(pPred); - for (size_t j = 0; j < vecBegin.size(); j++) { - if (j == iPred) continue; - pPred->fnAppend(vecRole[j].c_str(), vecBegin[j], vecEnd[j]); - pPred->m_vecArgt.back()->fnSetTreeItem( - pSrlSentence->m_pTree->fnFindNodeForSpan(vecBegin[j], vecEnd[j], - false)); - } - } - return pSrlSentence; - } - - private: - bool fnReadNextContent(std::vector >& vecContent) { - vecContent.clear(); - if (feof(m_fpIn) == true) return false; - char* pszLine; - pszLine = new char[100001]; - pszLine[0] = '\0'; - int iLen; - while (!feof(m_fpIn)) { - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - if (iLen == 0) break; // end of this sentence - - std::vector terms = SplitOnWhitespace(std::string(pszLine)); - assert(terms.size() > 7); - vecContent.push_back(terms); - } - delete[] pszLine; - return true; - } - - private: - FILE* m_fpIn; -}; -#endif /* SRL_SENTENCE_H_ */ diff --git a/utils/tree.h b/utils/tree.h deleted file mode 100644 index 6c3406d6..00000000 --- a/utils/tree.h +++ /dev/null @@ -1,699 +0,0 @@ -/* - * tree.h - * - * Created on: May 23, 2013 - * Author: lijunhui - */ - -#ifndef TREE_H_ -#define TREE_H_ - -#include -#include - -#include -#include - -struct STreeItem { - STreeItem(const char *pszTerm) { - m_pszTerm = new char[strlen(pszTerm) + 1]; - strcpy(m_pszTerm, pszTerm); - - m_ptParent = NULL; - m_iBegin = -1; - m_iEnd = -1; - m_iHeadChild = -1; - m_iHeadWord = -1; - m_iBrotherIndex = -1; - } - ~STreeItem() { - delete[] m_pszTerm; - for (size_t i = 0; i < m_vecChildren.size(); i++) delete m_vecChildren[i]; - } - int fnAppend(STreeItem *ptChild) { - m_vecChildren.push_back(ptChild); - ptChild->m_iBrotherIndex = m_vecChildren.size() - 1; - ptChild->m_ptParent = this; - return m_vecChildren.size() - 1; - } - int fnGetChildrenNum() { return m_vecChildren.size(); } - - bool fnIsPreTerminal(void) { - int I; - if (this == NULL || m_vecChildren.size() == 0) return false; - - for (I = 0; I < m_vecChildren.size(); I++) - if (m_vecChildren[I]->m_vecChildren.size() > 0) return false; - - return true; - } - - public: - char *m_pszTerm; - - std::vector m_vecChildren; // children items - STreeItem *m_ptParent; // the parent item - - int m_iBegin; - int m_iEnd; // the node span words[m_iBegin, m_iEnd] - int m_iHeadChild; // the index of its head child - int m_iHeadWord; // the index of its head word - int m_iBrotherIndex; // the index in his brothers -}; - -struct SGetHeadWord { - typedef std::vector CVectorStr; - SGetHeadWord() {} - ~SGetHeadWord() {} - int fnGetHeadWord(char *pszCFGLeft, CVectorStr vectRight) { - // 0 indicating from right to left while 1 indicating from left to right - char szaHeadLists[201] = "0"; - - /* //head rules for Egnlish - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP - 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN - 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP - 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR - 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ - 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP - 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP - 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN - 0CD 0JJ JJS RB QP 0" ); - */ - - if (strcmp(pszCFGLeft, "ADJP") == 0) - strcpy(szaHeadLists, "0ADJP JJ 0AD NN CS 0"); - else if (strcmp(pszCFGLeft, "ADVP") == 0) - strcpy(szaHeadLists, "0ADVP AD 0"); - else if (strcmp(pszCFGLeft, "CLP") == 0) - strcpy(szaHeadLists, "0CLP M 0"); - else if (strcmp(pszCFGLeft, "CP") == 0) - strcpy(szaHeadLists, "0DEC SP 1ADVP CS 0CP IP 0"); - else if (strcmp(pszCFGLeft, "DNP") == 0) - strcpy(szaHeadLists, "0DNP DEG 0DEC 0"); - else if (strcmp(pszCFGLeft, "DVP") == 0) - strcpy(szaHeadLists, "0DVP DEV 0"); - else if (strcmp(pszCFGLeft, "DP") == 0) - strcpy(szaHeadLists, "1DP DT 1"); - else if (strcmp(pszCFGLeft, "FRAG") == 0) - strcpy(szaHeadLists, "0VV NR NN 0"); - else if (strcmp(pszCFGLeft, "INTJ") == 0) - strcpy(szaHeadLists, "0INTJ IJ 0"); - else if (strcmp(pszCFGLeft, "LST") == 0) - strcpy(szaHeadLists, "1LST CD OD 1"); - else if (strcmp(pszCFGLeft, "IP") == 0) - strcpy(szaHeadLists, "0IP VP 0VV 0"); - // strcpy( szaHeadLists, "0VP 0VV 1IP 0" ); - else if (strcmp(pszCFGLeft, "LCP") == 0) - strcpy(szaHeadLists, "0LCP LC 0"); - else if (strcmp(pszCFGLeft, "NP") == 0) - strcpy(szaHeadLists, "0NP NN NT NR QP 0"); - else if (strcmp(pszCFGLeft, "PP") == 0) - strcpy(szaHeadLists, "1PP P 1"); - else if (strcmp(pszCFGLeft, "PRN") == 0) - strcpy(szaHeadLists, "0 NP IP VP NT NR NN 0"); - else if (strcmp(pszCFGLeft, "QP") == 0) - strcpy(szaHeadLists, "0QP CLP CD OD 0"); - else if (strcmp(pszCFGLeft, "VP") == 0) - strcpy(szaHeadLists, "1VP VA VC VE VV BA LB VCD VSB VRD VNV VCP 1"); - else if (strcmp(pszCFGLeft, "VCD") == 0) - strcpy(szaHeadLists, "0VCD VV VA VC VE 0"); - if (strcmp(pszCFGLeft, "VRD") == 0) - strcpy(szaHeadLists, "0VRD VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VSB") == 0) - strcpy(szaHeadLists, "0VSB VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VCP") == 0) - strcpy(szaHeadLists, "0VCP VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VNV") == 0) - strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "VPT") == 0) - strcpy(szaHeadLists, "0VNV VV VA VC VE 0"); - else if (strcmp(pszCFGLeft, "UCP") == 0) - strcpy(szaHeadLists, "0"); - else if (strcmp(pszCFGLeft, "WHNP") == 0) - strcpy(szaHeadLists, "0WHNP NP NN NT NR QP 0"); - else if (strcmp(pszCFGLeft, "WHPP") == 0) - strcpy(szaHeadLists, "1WHPP PP P 1"); - - /* //head rules for GENIA corpus - if( strcmp( pszCFGLeft, "ADJP" ) == 0 ) - strcpy( szaHeadLists, "0NNS 0QP 0NN 0$ 0ADVP 0JJ 0VBN 0VBG 0ADJP - 0JJR 0NP 0JJS 0DT 0FW 0RBR 0RBS 0SBAR 0RB 0" ); - else if( strcmp( pszCFGLeft, "ADVP" ) == 0 ) - strcpy( szaHeadLists, "1RB 1RBR 1RBS 1FW 1ADVP 1TO 1CD 1JJR 1JJ 1IN - 1NP 1JJS 1NN 1" ); - else if( strcmp( pszCFGLeft, "CONJP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1RB 1IN 1" ); - else if( strcmp( pszCFGLeft, "FRAG" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "INTJ" ) == 0 ) - strcpy( szaHeadLists, "0" ); - else if( strcmp( pszCFGLeft, "LST" ) == 0 ) - strcpy( szaHeadLists, "1LS 1: 1CLN 1" ); - else if( strcmp( pszCFGLeft, "NAC" ) == 0 ) - strcpy( szaHeadLists, "0NN 0NNS 0NNP 0NNPS 0NP 0NAC 0EX 0$ 0CD 0QP - 0PRP 0VBG 0JJ 0JJS 0JJR 0ADJP 0FW 0" ); - else if( strcmp( pszCFGLeft, "PP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO 1VBG 1VBN 1RP 1FW 1" ); - else if( strcmp( pszCFGLeft, "PRN" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "PRT" ) == 0 ) - strcpy( szaHeadLists, "1RP 1" ); - else if( strcmp( pszCFGLeft, "QP" ) == 0 ) - strcpy( szaHeadLists, "0$ 0IN 0NNS 0NN 0JJ 0RB 0DT 0CD 0NCD 0QP 0JJR - 0JJS 0" ); - else if( strcmp( pszCFGLeft, "RRC" ) == 0 ) - strcpy( szaHeadLists, "1VP 1NP 1ADVP 1ADJP 1PP 1" ); - else if( strcmp( pszCFGLeft, "S" ) == 0 ) - strcpy( szaHeadLists, "0TO 0IN 0VP 0S 0SBAR 0ADJP 0UCP 0NP 0" ); - else if( strcmp( pszCFGLeft, "SBAR" ) == 0 ) - strcpy( szaHeadLists, "0WHNP 0WHPP 0WHADVP 0WHADJP 0IN 0DT 0S 0SQ - 0SINV 0SBAR 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SBARQ" ) == 0 ) - strcpy( szaHeadLists, "0SQ 0S 0SINV 0SBARQ 0FRAG 0" ); - else if( strcmp( pszCFGLeft, "SINV" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0S 0SINV 0ADJP 0NP - 0" ); - else if( strcmp( pszCFGLeft, "SQ" ) == 0 ) - strcpy( szaHeadLists, "0VBZ 0VBD 0VBP 0VB 0MD 0VP 0SQ 0" ); - else if( strcmp( pszCFGLeft, "UCP" ) == 0 ) - strcpy( szaHeadLists, "1" ); - else if( strcmp( pszCFGLeft, "VP" ) == 0 ) - strcpy( szaHeadLists, "0TO 0VBD 0VBN 0MD 0VBZ 0VB 0VBG 0VBP 0VP - 0ADJP 0NN 0NNS 0NP 0" ); - else if( strcmp( pszCFGLeft, "WHADJP" ) == 0 ) - strcpy( szaHeadLists, "0CC 0WRB 0JJ 0ADJP 0" ); - else if( strcmp( pszCFGLeft, "WHADVP" ) == 0 ) - strcpy( szaHeadLists, "1CC 1WRB 1" ); - else if( strcmp( pszCFGLeft, "WHNP" ) == 0 ) - strcpy( szaHeadLists, "0WDT 0WP 0WP$ 0WHADJP 0WHPP 0WHNP 0" ); - else if( strcmp( pszCFGLeft, "WHPP" ) == 0 ) - strcpy( szaHeadLists, "1IN 1TO FW 1" ); - else if( strcmp( pszCFGLeft, "NP" ) == 0 ) - strcpy( szaHeadLists, "0NN NNP NNS NNPS NX POS JJR 0NP 0$ ADJP PRN - 0CD 0JJ JJS RB QP 0" ); - */ - - return fnMyOwnHeadWordRule(szaHeadLists, vectRight); - } - - private: - int fnMyOwnHeadWordRule(char *pszaHeadLists, CVectorStr vectRight) { - char szHeadList[201], *p; - char szTerm[101]; - int J; - - p = pszaHeadLists; - - int iCountRight; - - iCountRight = vectRight.size(); - - szHeadList[0] = '\0'; - while (1) { - szTerm[0] = '\0'; - sscanf(p, "%s", szTerm); - if (strlen(szHeadList) == 0) { - if (strcmp(szTerm, "0") == 0) { - return iCountRight - 1; - } - if (strcmp(szTerm, "1") == 0) { - return 0; - } - - sprintf(szHeadList, "%c %s ", szTerm[0], szTerm + 1); - p = strstr(p, szTerm); - p += strlen(szTerm); - } else { - if ((szTerm[0] == '0') || (szTerm[0] == '1')) { - if (szHeadList[0] == '0') { - for (J = iCountRight - 1; J >= 0; J--) { - sprintf(szTerm, " %s ", vectRight.at(J).c_str()); - if (strstr(szHeadList, szTerm) != NULL) return J; - } - } else { - for (J = 0; J < iCountRight; J++) { - sprintf(szTerm, " %s ", vectRight.at(J).c_str()); - if (strstr(szHeadList, szTerm) != NULL) return J; - } - } - - szHeadList[0] = '\0'; - } else { - strcat(szHeadList, szTerm); - strcat(szHeadList, " "); - - p = strstr(p, szTerm); - p += strlen(szTerm); - } - } - } - - return 0; - } -}; - -struct SParsedTree { - SParsedTree() { m_ptRoot = NULL; } - ~SParsedTree() { - if (m_ptRoot != NULL) delete m_ptRoot; - } - static SParsedTree *fnConvertFromString(const char *pszStr) { - if (strcmp(pszStr, "(())") == 0) return NULL; - SParsedTree *pTree = new SParsedTree(); - - std::vector vecSyn; - fnReadSyntactic(pszStr, vecSyn); - - int iLeft = 1, iRight = 1; //# left/right parenthesis - - STreeItem *pcurrent; - - pTree->m_ptRoot = new STreeItem(vecSyn[1].c_str()); - - pcurrent = pTree->m_ptRoot; - - for (size_t i = 2; i < vecSyn.size() - 1; i++) { - if (strcmp(vecSyn[i].c_str(), "(") == 0) - iLeft++; - else if (strcmp(vecSyn[i].c_str(), ")") == 0) { - iRight++; - if (pcurrent == NULL) { - // error - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - pcurrent = pcurrent->m_ptParent; - } else { - STreeItem *ptNewItem = new STreeItem(vecSyn[i].c_str()); - pcurrent->fnAppend(ptNewItem); - pcurrent = ptNewItem; - - if (strcmp(vecSyn[i - 1].c_str(), "(") != 0 && - strcmp(vecSyn[i - 1].c_str(), ")") != 0) { - pTree->m_vecTerminals.push_back(ptNewItem); - pcurrent = pcurrent->m_ptParent; - } - } - } - - if (iLeft != iRight) { - // error - fprintf(stderr, "the left and right parentheses are not matched!"); - fprintf(stderr, "ERROR in ConvertFromString\n"); - fprintf(stderr, "%s\n", pszStr); - return NULL; - } - - return pTree; - } - - int fnGetNumWord() { return m_vecTerminals.size(); } - - void fnSetSpanInfo() { - int iNextNum = 0; - fnSuffixTraverseSetSpanInfo(m_ptRoot, iNextNum); - } - - void fnSetHeadWord() { - for (size_t i = 0; i < m_vecTerminals.size(); i++) - m_vecTerminals[i]->m_iHeadWord = i; - SGetHeadWord *pGetHeadWord = new SGetHeadWord(); - fnSuffixTraverseSetHeadWord(m_ptRoot, pGetHeadWord); - delete pGetHeadWord; - } - - STreeItem *fnFindNodeForSpan(int iLeft, int iRight, bool bLowest) { - STreeItem *pTreeItem = m_vecTerminals[iLeft]; - - while (pTreeItem->m_iEnd < iRight) { - pTreeItem = pTreeItem->m_ptParent; - if (pTreeItem == NULL) break; - } - if (pTreeItem == NULL) return NULL; - if (pTreeItem->m_iEnd > iRight) return NULL; - - assert(pTreeItem->m_iEnd == iRight); - if (bLowest) return pTreeItem; - - while (pTreeItem->m_ptParent != NULL && - pTreeItem->m_ptParent->fnGetChildrenNum() == 1) - pTreeItem = pTreeItem->m_ptParent; - - return pTreeItem; - } - - private: - void fnSuffixTraverseSetSpanInfo(STreeItem *ptItem, int &iNextNum) { - int I; - int iNumChildren = ptItem->fnGetChildrenNum(); - for (I = 0; I < iNumChildren; I++) - fnSuffixTraverseSetSpanInfo(ptItem->m_vecChildren[I], iNextNum); - - if (I == 0) { - ptItem->m_iBegin = iNextNum; - ptItem->m_iEnd = iNextNum++; - } else { - ptItem->m_iBegin = ptItem->m_vecChildren[0]->m_iBegin; - ptItem->m_iEnd = ptItem->m_vecChildren[I - 1]->m_iEnd; - } - } - - void fnSuffixTraverseSetHeadWord(STreeItem *ptItem, - SGetHeadWord *pGetHeadWord) { - int I, iHeadchild; - - if (ptItem->m_vecChildren.size() == 0) return; - - for (I = 0; I < ptItem->m_vecChildren.size(); I++) - fnSuffixTraverseSetHeadWord(ptItem->m_vecChildren[I], pGetHeadWord); - - std::vector vecRight; - - if (ptItem->m_vecChildren.size() == 1) - iHeadchild = 0; - else { - for (I = 0; I < ptItem->m_vecChildren.size(); I++) - vecRight.push_back(std::string(ptItem->m_vecChildren[I]->m_pszTerm)); - - iHeadchild = pGetHeadWord->fnGetHeadWord(ptItem->m_pszTerm, vecRight); - } - - ptItem->m_iHeadChild = iHeadchild; - ptItem->m_iHeadWord = ptItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - - static void fnReadSyntactic(const char *pszSyn, - std::vector &vec) { - char *p; - int I; - - int iLeftNum, iRightNum; - char *pszTmp, *pszTerm; - pszTmp = new char[strlen(pszSyn)]; - pszTerm = new char[strlen(pszSyn)]; - pszTmp[0] = pszTerm[0] = '\0'; - - vec.clear(); - - char *pszLine; - pszLine = new char[strlen(pszSyn) + 1]; - strcpy(pszLine, pszSyn); - - char *pszLine2; - - while (1) { - while ((strlen(pszLine) > 0) && (pszLine[strlen(pszLine) - 1] > 0) && - (pszLine[strlen(pszLine) - 1] <= ' ')) - pszLine[strlen(pszLine) - 1] = '\0'; - - if (strlen(pszLine) == 0) break; - - // printf( "%s\n", pszLine ); - pszLine2 = pszLine; - while (pszLine2[0] <= ' ') pszLine2++; - if (pszLine2[0] == '<') continue; - - sscanf(pszLine2 + 1, "%s", pszTmp); - - if (pszLine2[0] == '(') { - iLeftNum = 0; - iRightNum = 0; - } - - p = pszLine2; - while (1) { - pszTerm[0] = '\0'; - sscanf(p, "%s", pszTerm); - - if (strlen(pszTerm) == 0) break; - p = strstr(p, pszTerm); - p += strlen(pszTerm); - - if ((pszTerm[0] == '(') || (pszTerm[strlen(pszTerm) - 1] == ')')) { - if (pszTerm[0] == '(') { - vec.push_back(std::string("(")); - iLeftNum++; - - I = 1; - while (pszTerm[I] == '(' && pszTerm[I] != '\0') { - vec.push_back(std::string("(")); - iLeftNum++; - - I++; - } - - if (strlen(pszTerm) > 1) vec.push_back(std::string(pszTerm + I)); - } else { - char *pTmp; - pTmp = pszTerm + strlen(pszTerm) - 1; - while ((pTmp[0] == ')') && (pTmp >= pszTerm)) pTmp--; - pTmp[1] = '\0'; - - if (strlen(pszTerm) > 0) vec.push_back(std::string(pszTerm)); - pTmp += 2; - - for (I = 0; I <= (int)strlen(pTmp); I++) { - vec.push_back(std::string(")")); - iRightNum++; - } - } - } else { - char *q; - q = strchr(pszTerm, ')'); - if (q != NULL) { - q[0] = '\0'; - if (pszTerm[0] != '\0') vec.push_back(std::string(pszTerm)); - vec.push_back(std::string(")")); - iRightNum++; - - q++; - while (q[0] == ')') { - vec.push_back(std::string(")")); - q++; - iRightNum++; - } - - while (q[0] == '(') { - vec.push_back(std::string("(")); - q++; - iLeftNum++; - } - - if (q[0] != '\0') vec.push_back(std::string(q)); - } else - vec.push_back(std::string(pszTerm)); - } - } - - if (iLeftNum != iRightNum) { - fprintf(stderr, "%s\n", pszSyn); - assert(iLeftNum == iRightNum); - } - /*if ( iLeftNum != iRightNum ) { - printf( "ERROR: left( and right ) is not matched, %d ( and %d - )\n", iLeftNum, iRightNum ); - return; - }*/ - - if (vec.size() >= 2 && strcmp(vec[1].c_str(), "(") == 0) { - //( (IP..) ) - std::vector::iterator it; - it = vec.begin(); - it++; - vec.insert(it, std::string("ROOT")); - } - - break; - } - - delete[] pszLine; - delete[] pszTmp; - delete[] pszTerm; - } - - public: - STreeItem *m_ptRoot; - std::vector m_vecTerminals; // the leaf nodes -}; - -struct SParseReader { - SParseReader(const char *pszParse_Fname, bool bFlattened = false) - : m_bFlattened(bFlattened) { - m_fpIn = fopen(pszParse_Fname, "r"); - assert(m_fpIn != NULL); - } - ~SParseReader() { - if (m_fpIn != NULL) fclose(m_fpIn); - } - - SParsedTree *fnReadNextParseTree() { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true) { - if (iLen == 0) continue; - - pTree = SParsedTree::fnConvertFromString(pszLine); - if (pTree == NULL) break; - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete[] pszLine; - return pTree; - } - - SParsedTree *fnReadNextParseTreeWithProb(double *pProb) { - SParsedTree *pTree = NULL; - char *pszLine = new char[100001]; - int iLen; - - while (fnReadNextSentence(pszLine, &iLen) == true) { - if (iLen == 0) continue; - - char *p = strchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - if (pProb) (*pProb) = atof(pszLine); - - pTree = SParsedTree::fnConvertFromString(p); - if (m_bFlattened) - fnPostProcessingFlattenedParse(pTree); - else { - pTree->fnSetSpanInfo(); - pTree->fnSetHeadWord(); - } - break; - } - - delete[] pszLine; - return pTree; - } - - private: - /* - * since to the parse tree is a flattened tree, use the head mark to identify - * head info. - * the head node will be marked as "*XP*" - */ - void fnSetParseTreeHeadInfo(SParsedTree *pTree) { - for (size_t i = 0; i < pTree->m_vecTerminals.size(); i++) - pTree->m_vecTerminals[i]->m_iHeadWord = i; - fnSuffixTraverseSetHeadWord(pTree->m_ptRoot); - } - - void fnSuffixTraverseSetHeadWord(STreeItem *pTreeItem) { - if (pTreeItem->m_vecChildren.size() == 0) return; - - for (size_t i = 0; i < pTreeItem->m_vecChildren.size(); i++) - fnSuffixTraverseSetHeadWord(pTreeItem->m_vecChildren[i]); - - std::vector vecRight; - - int iHeadchild; - - if (pTreeItem->fnIsPreTerminal()) { - iHeadchild = 0; - } else { - size_t i; - for (i = 0; i < pTreeItem->m_vecChildren.size(); i++) { - char *p = pTreeItem->m_vecChildren[i]->m_pszTerm; - if (p[0] == '*' && p[strlen(p) - 1] == '*') { - iHeadchild = i; - p[strlen(p) - 1] = '\0'; - std::string str = p + 1; - strcpy(p, str.c_str()); // erase the "*..*" - break; - } - } - assert(i < pTreeItem->m_vecChildren.size()); - } - - pTreeItem->m_iHeadChild = iHeadchild; - pTreeItem->m_iHeadWord = pTreeItem->m_vecChildren[iHeadchild]->m_iHeadWord; - } - void fnPostProcessingFlattenedParse(SParsedTree *pTree) { - pTree->fnSetSpanInfo(); - fnSetParseTreeHeadInfo(pTree); - } - bool fnReadNextSentence(char *pszLine, int *piLength) { - if (feof(m_fpIn) == true) return false; - - int iLen; - - pszLine[0] = '\0'; - - fgets(pszLine, 10001, m_fpIn); - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - if (piLength != NULL) (*piLength) = iLen; - - return true; - } - - private: - FILE *m_fpIn; - const bool m_bFlattened; -}; - -#endif /* TREE_H_ */ diff --git a/utils/tsuruoka_maxent.h b/utils/tsuruoka_maxent.h deleted file mode 100644 index 82da44ff..00000000 --- a/utils/tsuruoka_maxent.h +++ /dev/null @@ -1,153 +0,0 @@ -/* - * tsuruoka_maxent.h - * - */ - -#ifndef TSURUOKA_MAXENT_H_ -#define TSURUOKA_MAXENT_H_ - -#include -#include -#include -#include -#include -#include - -#include "stringlib.h" -#include "maxent.h" - -typedef std::unordered_map Map; -typedef std::unordered_map::iterator Iterator; - -struct Tsuruoka_Maxent { - Tsuruoka_Maxent(const char* pszModelFName) { - if (pszModelFName != NULL) { - m_pModel = new ME_Model(); - m_pModel->load_from_file(pszModelFName); - } else - m_pModel = NULL; - } - - ~Tsuruoka_Maxent() { - if (m_pModel != NULL) delete m_pModel; - } - - void fnTrain(const char* pszInstanceFName, const char* pszAlgorithm, - const char* pszModelFName, int /*iNumIteration*/) { - assert(strcmp(pszAlgorithm, "l1") == 0 || strcmp(pszAlgorithm, "l2") == 0 || - strcmp(pszAlgorithm, "sgd") == 0 || - strcmp(pszAlgorithm, "SGD") == 0); - FILE* fpIn = fopen(pszInstanceFName, "r"); - - ME_Model* pModel = new ME_Model(); - - char* pszLine = new char[100001]; - int iNumInstances = 0; - int iLen; - while (!feof(fpIn)) { - pszLine[0] = '\0'; - fgets(pszLine, 20000, fpIn); - if (strlen(pszLine) == 0) { - continue; - } - - iLen = strlen(pszLine); - while (iLen > 0 && pszLine[iLen - 1] > 0 && pszLine[iLen - 1] < 33) { - pszLine[iLen - 1] = '\0'; - iLen--; - } - - iNumInstances++; - - ME_Sample* pmes = new ME_Sample(); - - char* p = strrchr(pszLine, ' '); - assert(p != NULL); - p[0] = '\0'; - p++; - std::vector vecContext; - SplitOnWhitespace(std::string(pszLine), &vecContext); - - pmes->label = std::string(p); - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - pModel->add_training_sample((*pmes)); - if (iNumInstances % 100000 == 0) - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - delete pmes; - } - fprintf(stdout, "......Reading #Instances: %1d\n", iNumInstances); - fclose(fpIn); - - if (strcmp(pszAlgorithm, "l1") == 0) - pModel->use_l1_regularizer(1.0); - else if (strcmp(pszAlgorithm, "l2") == 0) - pModel->use_l2_regularizer(1.0); - else - pModel->use_SGD(); - - pModel->train(); - pModel->save_to_file(pszModelFName); - - delete pModel; - fprintf(stdout, "......Finished Training\n"); - fprintf(stdout, "......Model saved as %s\n", pszModelFName); - delete[] pszLine; - } - - double fnEval(const char* pszContext, const char* pszOutcome) const { - std::vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(std::string(pszContext), &vecContext); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - std::vector vecProb = m_pModel->classify(*pmes); - delete pmes; - int iLableID = m_pModel->get_class_id(pszOutcome); - return vecProb[iLableID]; - } - void fnEval(const char* pszContext, - std::vector >& vecOutput) const { - std::vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(std::string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - std::vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - std::string label = m_pModel->get_class_label(i); - vecOutput.push_back(make_pair(label, vecProb[i])); - } - delete pmes; - } - void fnEval(const char* pszContext, std::vector& vecOutput) const { - std::vector vecContext; - ME_Sample* pmes = new ME_Sample(); - SplitOnWhitespace(std::string(pszContext), &vecContext); - - vecOutput.clear(); - - for (size_t i = 0; i < vecContext.size(); i++) - pmes->add_feature(vecContext[i]); - std::vector vecProb = m_pModel->classify(*pmes); - - for (size_t i = 0; i < vecProb.size(); i++) { - std::string label = m_pModel->get_class_label(i); - vecOutput.push_back(vecProb[i]); - } - delete pmes; - } - int fnGetClassId(const std::string& strLabel) const { - return m_pModel->get_class_id(strLabel); - } - - private: - ME_Model* m_pModel; -}; - -#endif /* TSURUOKA_MAXENT_H_ */ -- cgit v1.2.3