diff options
Diffstat (limited to 'decoder')
| -rw-r--r-- | decoder/Makefile.am | 3 | ||||
| -rw-r--r-- | decoder/cdec_ff.cc | 3 | ||||
| -rw-r--r-- | decoder/ff_const_reorder.cc | 1115 | ||||
| -rw-r--r-- | decoder/ff_const_reorder.h | 43 | ||||
| -rw-r--r-- | decoder/ff_const_reorder_common.h | 1441 |
5 files changed, 2604 insertions, 1 deletions
diff --git a/decoder/Makefile.am b/decoder/Makefile.am index e46a7120..f791b3cc 100644 --- a/decoder/Makefile.am +++ b/decoder/Makefile.am @@ -45,6 +45,8 @@ 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 \ ff_external.h \ @@ -111,6 +113,7 @@ libcdec_a_SOURCES = \ ff_bleu.cc \ ff_charset.cc \ ff_context.cc \ + ff_const_reorder.cc \ ff_csplit.cc \ ff_external.cc \ ff_factory.cc \ diff --git a/decoder/cdec_ff.cc b/decoder/cdec_ff.cc index 7f7e075b..6f7227aa 100644 --- a/decoder/cdec_ff.cc +++ b/decoder/cdec_ff.cc @@ -3,6 +3,7 @@ #include "ff.h" #include "ff_basic.h" #include "ff_context.h" +#include "ff_const_reorder.h" #include "ff_spans.h" #include "ff_lm.h" #include "ff_klm.h" @@ -77,6 +78,6 @@ void register_feature_functions() { ff_registry.Register("WordPairFeatures", new FFFactory<WordPairFeatures>); ff_registry.Register("SourcePathFeatures", new FFFactory<SourcePathFeatures>); ff_registry.Register("WordSet", new FFFactory<WordSet>); + ff_registry.Register("ConstReorderFeature", new FFFactory<ConstReorderFeature>); ff_registry.Register("External", new FFFactory<ExternalFeature>); } - diff --git a/decoder/ff_const_reorder.cc b/decoder/ff_const_reorder.cc new file mode 100644 index 00000000..95546793 --- /dev/null +++ b/decoder/ff_const_reorder.cc @@ -0,0 +1,1115 @@ +#include "ff_const_reorder.h" + +#include "filelib.h" +#include "stringlib.h" +#include "hg.h" +#include "sentence_metadata.h" +#include "hash.h" +#include "ff_const_reorder_common.h" + +#include <sstream> +#include <string> +#include <vector> +#include <stdio.h> + +using namespace std; +using namespace const_reorder; + +typedef HASH_MAP<std::string, vector<double> > MapClassifier; + +inline bool is_inside(int i, int left, int right) { + if (i < left || i > right) return false; + return true; +} + +/* + * assume i <= j + * [i, j] is inside [left, right] or [i, j] equates to [left, right] + */ +inline bool is_inside(int i, int j, int left, int right) { + if (i >= left && j <= right) return true; + return false; +} + +/* + * assume i <= j + * [i, j] is inside [left, right], but [i, j] not equal to [left, right] + */ +inline bool is_proper_inside(int i, int j, int left, int right) { + if (i >= left && j <= right && right - left > j - i) return true; + return false; +} + +/* + * assume i <= j + * [i, j] is proper proper inside [left, right] + */ +inline bool is_proper_proper_inside(int i, int j, int left, int right) { + if (i > left && j < right) return true; + return false; +} + +inline bool is_overlap(int left1, int right1, int left2, int right2) { + if (is_inside(left1, left2, right2) || is_inside(left2, left1, right1)) + return true; + + return false; +} + +inline void NewAndCopyCharArray(char** p, const char* q) { + if (q != NULL) { + (*p) = new char[strlen(q) + 1]; + strcpy((*p), q); + } else + (*p) = NULL; +} + +// TODO:to make the alignment more efficient +struct TargetTranslation { + TargetTranslation(int begin_pos, int end_pos,int e_num_word) + : begin_pos_(begin_pos), + end_pos_(end_pos), + e_num_words_(e_num_word), + vec_left_most_(end_pos - begin_pos + 1, e_num_word), + vec_right_most_(end_pos - begin_pos + 1, -1), + vec_f_align_bit_array_(end_pos - begin_pos + 1), + vec_e_align_bit_array_(e_num_word) { + int len = end_pos - begin_pos + 1; + align_.reserve(1.5 * len); + } + + void InsertAlignmentPoint(int s, int t) { + int i = s - begin_pos_; + + vector<bool>& b = vec_f_align_bit_array_[i]; + if (b.empty()) b.resize(e_num_words_); + b[t] = 1; + + vector<bool>& a = vec_e_align_bit_array_[t]; + if (a.empty()) a.resize(end_pos_ - begin_pos_ + 1); + a[i] = 1; + + align_.push_back({s, t}); + + if (t > vec_right_most_[i]) vec_right_most_[i] = t; + if (t < vec_left_most_[i]) vec_left_most_[i] = t; + } + + /* + * given a source span [begin, end], whether its target side is continuous, + * return "0": the source span is translated silently + * return "1": there is at least on word inside its target span, this word + * doesn't align to any word inside [begin, end], but outside [begin, end] + * return "2": otherwise + */ + string IsTargetConstinousSpan(int begin, int end) const { + int target_begin, target_end; + FindLeftRightMostTargetSpan(begin, end, target_begin, target_end); + if (target_begin == -1) return "0"; + + for (int i = target_begin; i <= target_end; i++) { + if (vec_e_align_bit_array_[i].empty()) continue; + int j = begin; + for (; j <= end; j++) { + if (vec_e_align_bit_array_[i][j - begin_pos_]) break; + } + if (j == end + 1) // e[i] is aligned, but e[i] doesn't align to any + // source word in [begin_pos, end_pos] + return "1"; + } + return "2"; + } + + string IsTargetConstinousSpan2(int begin, int end) const { + int target_begin, target_end; + FindLeftRightMostTargetSpan(begin, end, target_begin, target_end); + if (target_begin == -1) return "Unaligned"; + + for (int i = target_begin; i <= target_end; i++) { + if (vec_e_align_bit_array_[i].empty()) continue; + int j = begin; + for (; j <= end; j++) { + if (vec_e_align_bit_array_[i][j - begin_pos_]) break; + } + if (j == end + 1) // e[i] is aligned, but e[i] doesn't align to any + // source word in [begin_pos, end_pos] + return "Discon't"; + } + return "Con't"; + } + + void FindLeftRightMostTargetSpan(int begin, int end, int& target_begin, + int& target_end) const { + int b = begin - begin_pos_; + int e = end - begin_pos_ + 1; + + target_begin = vec_left_most_[b]; + target_end = vec_right_most_[b]; + for (int i = b + 1; i < e; i++) { + if (target_begin > vec_left_most_[i]) target_begin = vec_left_most_[i]; + if (target_end < vec_right_most_[i]) target_end = vec_right_most_[i]; + } + if (target_end == -1) target_begin = -1; + return; + + target_begin = e_num_words_; + target_end = -1; + + for (int i = begin - begin_pos_; i < end - begin_pos_ + 1; i++) { + if (vec_f_align_bit_array_[i].empty()) continue; + for (int j = 0; j < target_begin; j++) + if (vec_f_align_bit_array_[i][j]) { + target_begin = j; + break; + } + } + for (int i = end - begin_pos_; i > begin - begin_pos_ - 1; i--) { + if (vec_f_align_bit_array_[i].empty()) continue; + for (int j = e_num_words_ - 1; j > target_end; j--) + if (vec_f_align_bit_array_[i][j]) { + target_end = j; + break; + } + } + + if (target_end == -1) target_begin = -1; + } + + const uint16_t begin_pos_, end_pos_; // the position in input + const uint16_t e_num_words_; + vector<AlignmentPoint> align_; + + private: + vector<short> vec_left_most_; + vector<short> vec_right_most_; + vector<vector<bool> > vec_f_align_bit_array_; + vector<vector<bool> > vec_e_align_bit_array_; +}; + +struct FocusedConstituent { + FocusedConstituent(const SParsedTree* pTree) { + if (pTree == NULL) return; + 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) { + if (true) { + + // do constituent reordering for all children of pParent + if (strcmp(pParent->m_pszTerm, "ROOT")) + focus_parents_.push_back(pParent); + } + if (pParent->m_iBrotherIndex != 0) break; + pParent = pParent->m_ptParent; + } + } + } + + ~FocusedConstituent() { // TODO + focus_parents_.clear(); + } + + vector<STreeItem*> focus_parents_; +}; + +typedef SPredicateItem FocusedPredicate; + +struct FocusedSRL { + FocusedSRL(const SSrlSentence* srl) { + if (srl == NULL) return; + for (size_t i = 0; i < srl->m_vecPred.size(); i++) { + if (strcmp(srl->m_pTree->m_vecTerminals[srl->m_vecPred[i]->m_iPosition] + ->m_ptParent->m_pszTerm, + "VA") == 0) + continue; + focus_predicates_.push_back( + new FocusedPredicate(srl->m_pTree, srl->m_vecPred[i])); + } + } + + ~FocusedSRL() { focus_predicates_.clear(); } + + vector<const FocusedPredicate*> focus_predicates_; +}; + +struct ConstReorderFeatureImpl { + ConstReorderFeatureImpl(const std::string& param) { + + b_block_feature_ = false; + b_order_feature_ = false; + b_srl_block_feature_ = false; + b_srl_order_feature_ = false; + + vector<string> terms; + SplitOnWhitespace(param, &terms); + if (terms.size() == 1) { + b_block_feature_ = true; + b_order_feature_ = true; + } else if (terms.size() >= 3) { + if (terms[1].compare("1") == 0) b_block_feature_ = true; + if (terms[2].compare("1") == 0) b_order_feature_ = true; + if (terms.size() == 6) { + if (terms[4].compare("1") == 0) b_srl_block_feature_ = true; + if (terms[5].compare("1") == 0) b_srl_order_feature_ = true; + + assert(b_srl_block_feature_ || b_srl_order_feature_); + } + + } else { + assert("ERROR"); + } + + const_reorder_classifier_left_ = NULL; + const_reorder_classifier_right_ = NULL; + + srl_reorder_classifier_left_ = NULL; + srl_reorder_classifier_right_ = NULL; + + if (b_order_feature_) { + InitializeClassifier((terms[0] + string(".left")).c_str(), + &const_reorder_classifier_left_); + InitializeClassifier((terms[0] + string(".right")).c_str(), + &const_reorder_classifier_right_); + } + + if (b_srl_order_feature_) { + InitializeClassifier((terms[3] + string(".left")).c_str(), + &srl_reorder_classifier_left_); + InitializeClassifier((terms[3] + string(".right")).c_str(), + &srl_reorder_classifier_right_); + } + + parsed_tree_ = NULL; + focused_consts_ = NULL; + + srl_sentence_ = NULL; + focused_srl_ = NULL; + + map_left_ = NULL; + map_right_ = NULL; + + map_srl_left_ = NULL; + map_srl_right_ = NULL; + + dict_block_status_ = new Dict(); + dict_block_status_->Convert("Unaligned", false); + dict_block_status_->Convert("Discon't", false); + dict_block_status_->Convert("Con't", false); + } + + ~ConstReorderFeatureImpl() { + if (const_reorder_classifier_left_) delete const_reorder_classifier_left_; + if (const_reorder_classifier_right_) delete const_reorder_classifier_right_; + if (srl_reorder_classifier_left_) delete srl_reorder_classifier_left_; + if (srl_reorder_classifier_right_) delete srl_reorder_classifier_right_; + FreeSentenceVariables(); + + delete dict_block_status_; + } + + static int ReserveStateSize() { return 1 * sizeof(TargetTranslation*); } + + void InitializeInputSentence(const std::string& parse_file, + const std::string& srl_file) { + FreeSentenceVariables(); + if (b_srl_block_feature_ || b_srl_order_feature_) { + assert(srl_file != ""); + srl_sentence_ = ReadSRLSentence(srl_file); + parsed_tree_ = srl_sentence_->m_pTree; + } else { + assert(parse_file != ""); + srl_sentence_ = NULL; + parsed_tree_ = ReadParseTree(parse_file); + } + + if (b_block_feature_ || b_order_feature_) { + focused_consts_ = new FocusedConstituent(parsed_tree_); + + if (b_order_feature_) { + // we can do the classifier "off-line" + map_left_ = new MapClassifier(); + map_right_ = new MapClassifier(); + InitializeConstReorderClassifierOutput(); + } + } + + if (b_srl_block_feature_ || b_srl_order_feature_) { + focused_srl_ = new FocusedSRL(srl_sentence_); + + if (b_srl_order_feature_) { + map_srl_left_ = new MapClassifier(); + map_srl_right_ = new MapClassifier(); + InitializeSRLReorderClassifierOutput(); + } + } + + if (parsed_tree_ != NULL) { + size_t i = parsed_tree_->m_vecTerminals.size(); + vec_target_tran_.reserve(20 * i * i * i); + } else + vec_target_tran_.reserve(1000000); + } + + void SetConstReorderFeature(const Hypergraph::Edge& edge, + SparseVector<double>* features, + const vector<const void*>& ant_states, + void* state) { + if (parsed_tree_ == NULL) return; + + short int begin = edge.i_, end = edge.j_ - 1; + + typedef TargetTranslation* PtrTargetTranslation; + PtrTargetTranslation* remnant = + reinterpret_cast<PtrTargetTranslation*>(state); + + vector<const TargetTranslation*> vec_node; + vec_node.reserve(edge.tail_nodes_.size()); + for (size_t i = 0; i < edge.tail_nodes_.size(); i++) { + const PtrTargetTranslation* astate = + reinterpret_cast<const PtrTargetTranslation*>(ant_states[i]); + vec_node.push_back(astate[0]); + } + + int e_num_word = edge.rule_->e_.size(); + for (size_t i = 0; i < vec_node.size(); i++) { + e_num_word += vec_node[i]->e_num_words_; + e_num_word--; + } + + remnant[0] = new TargetTranslation(begin, end, e_num_word); + vec_target_tran_.push_back(remnant[0]); + + // reset the alignment + // for the source side, we know its position in source sentence + // for the target side, we always assume its starting position is 0 + unsigned vc = 0; + const TRulePtr rule = edge.rule_; + std::vector<int> f_index(rule->f_.size()); + int index = edge.i_; + for (unsigned i = 0; i < rule->f_.size(); i++) { + f_index[i] = index; + const WordID& c = rule->f_[i]; + if (c < 1) + index = vec_node[vc++]->end_pos_ + 1; + else + index++; + } + assert(vc == vec_node.size()); + assert(index == edge.j_); + + std::vector<int> e_index(rule->e_.size()); + index = 0; + vc = 0; + for (unsigned i = 0; i < rule->e_.size(); i++) { + e_index[i] = index; + const WordID& c = rule->e_[i]; + if (c < 1) { + index += vec_node[-c]->e_num_words_; + vc++; + } else + index++; + } + assert(vc == vec_node.size()); + + size_t nt_pos = 0; + for (size_t i = 0; i < edge.rule_->f_.size(); i++) { + if (edge.rule_->f_[i] > 0) continue; + + // it's an NT + size_t j; + for (j = 0; j < edge.rule_->e_.size(); j++) + if (edge.rule_->e_[j] * -1 == nt_pos) break; + assert(j != edge.rule_->e_.size()); + nt_pos++; + + // i aligns j + int eindex = e_index[j]; + const vector<AlignmentPoint>& align = + vec_node[-1 * edge.rule_->e_[j]]->align_; + for (size_t k = 0; k < align.size(); k++) { + remnant[0]->InsertAlignmentPoint(align[k].s_, eindex + align[k].t_); + } + } + for (size_t i = 0; i < edge.rule_->a_.size(); i++) { + int findex = f_index[edge.rule_->a_[i].s_]; + int eindex = e_index[edge.rule_->a_[i].t_]; + remnant[0]->InsertAlignmentPoint(findex, eindex); + } + + // till now, we finished setting state values + // next, use the state values to calculate constituent reorder feature + SetConstReorderFeature(begin, end, features, remnant[0], + vec_node, f_index); + } + + void SetConstReorderFeature(short int begin, short int end, + SparseVector<double>* features, + const TargetTranslation* target_translation, + const vector<const TargetTranslation*>& vec_node, + std::vector<int>& /*findex*/) { + if (b_srl_block_feature_ || b_srl_order_feature_) { + double logprob_srl_reorder_left = 0.0, logprob_srl_reorder_right = 0.0; + for (size_t i = 0; i < focused_srl_->focus_predicates_.size(); i++) { + const FocusedPredicate* pred = focused_srl_->focus_predicates_[i]; + if (!is_overlap(begin, end, pred->begin_, pred->end_)) + continue; // have no overlap between this predicate (with its + // argument) and the current edge + + size_t j; + for (j = 0; j < vec_node.size(); j++) { + if (is_inside(pred->begin_, pred->end_, vec_node[j]->begin_pos_, + vec_node[j]->end_pos_)) + break; + } + if (j < vec_node.size()) continue; + + vector<int> vecBlockStatus; + vecBlockStatus.reserve(pred->vec_items_.size()); + for (j = 0; j < pred->vec_items_.size(); j++) { + const STreeItem* con1 = pred->vec_items_[j]->tree_item_; + if (con1->m_iBegin < begin || con1->m_iEnd > end) { + vecBlockStatus.push_back(0); + continue; + } // the node is partially outside the current edge + + string type = target_translation->IsTargetConstinousSpan2( + con1->m_iBegin, con1->m_iEnd); + vecBlockStatus.push_back(dict_block_status_->Convert(type, false)); + + if (!b_srl_block_feature_) continue; + // see if the node is covered by an NT + size_t k; + for (k = 0; k < vec_node.size(); k++) { + if (is_inside(con1->m_iBegin, con1->m_iEnd, vec_node[k]->begin_pos_, + vec_node[k]->end_pos_)) + break; + } + if (k < vec_node.size()) continue; + int f_id = FD::Convert(string(pred->vec_items_[j]->role_) + type); + if (f_id) features->add_value(f_id, 1); + } + + if (!b_srl_order_feature_) continue; + + vector<int> vecPosition, vecRelativePosition; + vector<int> vecRightPosition, vecRelativeRightPosition; + vecPosition.reserve(pred->vec_items_.size()); + vecRelativePosition.reserve(pred->vec_items_.size()); + vecRightPosition.reserve(pred->vec_items_.size()); + vecRelativeRightPosition.reserve(pred->vec_items_.size()); + for (j = 0; j < pred->vec_items_.size(); j++) { + const STreeItem* con1 = pred->vec_items_[j]->tree_item_; + if (con1->m_iBegin < begin || con1->m_iEnd > end) { + vecPosition.push_back(-1); + vecRightPosition.push_back(-1); + continue; + } // the node is partially outside the current edge + int left1 = -1, right1 = -1; + target_translation->FindLeftRightMostTargetSpan( + con1->m_iBegin, con1->m_iEnd, left1, right1); + vecPosition.push_back(left1); + vecRightPosition.push_back(right1); + } + fnGetRelativePosition(vecPosition, vecRelativePosition); + fnGetRelativePosition(vecRightPosition, vecRelativeRightPosition); + + for (j = 1; j < pred->vec_items_.size(); j++) { + const STreeItem* con1 = pred->vec_items_[j - 1]->tree_item_; + const STreeItem* con2 = pred->vec_items_[j]->tree_item_; + + if (con1->m_iBegin < begin || con2->m_iEnd > end) + continue; // one of the two nodes is partially outside the current + // edge + + // both con1 and con2 are covered, need to check if they are covered + // by the same NT + size_t k; + for (k = 0; k < vec_node.size(); k++) { + if (is_inside(con1->m_iBegin, con2->m_iEnd, vec_node[k]->begin_pos_, + vec_node[k]->end_pos_)) + break; + } + if (k < vec_node.size()) continue; + + // they are not covered bye the same NT + string outcome; + string key; + GenerateKey(pred->vec_items_[j - 1]->tree_item_, + pred->vec_items_[j]->tree_item_, vecBlockStatus[j - 1], + vecBlockStatus[j], key); + + fnGetOutcome(vecRelativePosition[j - 1], vecRelativePosition[j], + outcome); + double prob = CalculateConstReorderProb(srl_reorder_classifier_left_, + map_srl_left_, key, outcome); + // printf("%s %s %f\n", ostr.str().c_str(), outcome.c_str(), prob); + logprob_srl_reorder_left += log10(prob); + + fnGetOutcome(vecRelativeRightPosition[j - 1], + vecRelativeRightPosition[j], outcome); + prob = CalculateConstReorderProb(srl_reorder_classifier_right_, + map_srl_right_, key, outcome); + logprob_srl_reorder_right += log10(prob); + } + } + + if (b_srl_order_feature_) { + int f_id = FD::Convert("SRLReorderFeatureLeft"); + if (f_id && logprob_srl_reorder_left != 0.0) + features->set_value(f_id, logprob_srl_reorder_left); + f_id = FD::Convert("SRLReorderFeatureRight"); + if (f_id && logprob_srl_reorder_right != 0.0) + features->set_value(f_id, logprob_srl_reorder_right); + } + } + + if (b_block_feature_ || b_order_feature_) { + double logprob_const_reorder_left = 0.0, + logprob_const_reorder_right = 0.0; + + for (size_t i = 0; i < focused_consts_->focus_parents_.size(); i++) { + STreeItem* parent = focused_consts_->focus_parents_[i]; + if (!is_overlap(begin, end, parent->m_iBegin, + parent->m_iEnd)) + continue; // have no overlap between this parent node and the current + // edge + + size_t j; + for (j = 0; j < vec_node.size(); j++) { + if (is_inside(parent->m_iBegin, parent->m_iEnd, + vec_node[j]->begin_pos_, vec_node[j]->end_pos_)) + break; + } + if (j < vec_node.size()) continue; + + if (b_block_feature_) { + if (parent->m_iBegin >= begin && + parent->m_iEnd <= end) { + string type = target_translation->IsTargetConstinousSpan2( + parent->m_iBegin, parent->m_iEnd); + int f_id = FD::Convert(string(parent->m_pszTerm) + type); + if (f_id) features->add_value(f_id, 1); + } + } + + if (parent->m_vecChildren.size() == 1 || !b_order_feature_) continue; + + vector<int> vecChunkBlock; + vecChunkBlock.reserve(parent->m_vecChildren.size()); + + for (j = 0; j < parent->m_vecChildren.size(); j++) { + STreeItem* con1 = parent->m_vecChildren[j]; + if (con1->m_iBegin < begin || con1->m_iEnd > end) { + vecChunkBlock.push_back(0); + continue; + } // the node is partially outside the current edge + + string type = target_translation->IsTargetConstinousSpan2( + con1->m_iBegin, con1->m_iEnd); + vecChunkBlock.push_back(dict_block_status_->Convert(type, false)); + + /*if (!b_block_feature_) continue; + //see if the node is covered by an NT + size_t k; + for (k = 0; k < vec_node.size(); k++) { + if (is_inside(con1->m_iBegin, con1->m_iEnd, + vec_node[k]->begin_pos_, vec_node[k]->end_pos_)) + break; + } + if (k < vec_node.size()) continue; + int f_id = FD::Convert(string(con1->m_pszTerm) + type); + if (f_id) + features->add_value(f_id, 1);*/ + } + + if (!b_order_feature_) continue; + + vector<int> vecPosition, vecRelativePosition; + vector<int> vecRightPosition, vecRelativeRightPosition; + vecPosition.reserve(parent->m_vecChildren.size()); + vecRelativePosition.reserve(parent->m_vecChildren.size()); + vecRightPosition.reserve(parent->m_vecChildren.size()); + vecRelativeRightPosition.reserve(parent->m_vecChildren.size()); + for (j = 0; j < parent->m_vecChildren.size(); j++) { + STreeItem* con1 = parent->m_vecChildren[j]; + if (con1->m_iBegin < begin || con1->m_iEnd > end) { + vecPosition.push_back(-1); + vecRightPosition.push_back(-1); + continue; + } // the node is partially outside the current edge + int left1 = -1, right1 = -1; + target_translation->FindLeftRightMostTargetSpan( + con1->m_iBegin, con1->m_iEnd, left1, right1); + vecPosition.push_back(left1); + vecRightPosition.push_back(right1); + } + fnGetRelativePosition(vecPosition, vecRelativePosition); + fnGetRelativePosition(vecRightPosition, vecRelativeRightPosition); + + for (j = 1; j < parent->m_vecChildren.size(); j++) { + STreeItem* con1 = parent->m_vecChildren[j - 1]; + STreeItem* con2 = parent->m_vecChildren[j]; + + if (con1->m_iBegin < begin || con2->m_iEnd > end) + continue; // one of the two nodes is partially outside the current + // edge + + // both con1 and con2 are covered, need to check if they are covered + // by the same NT + size_t k; + for (k = 0; k < vec_node.size(); k++) { + if (is_inside(con1->m_iBegin, con2->m_iEnd, vec_node[k]->begin_pos_, + vec_node[k]->end_pos_)) + break; + } + if (k < vec_node.size()) continue; + + // they are not covered bye the same NT + string outcome; + string key; + GenerateKey(parent->m_vecChildren[j - 1], parent->m_vecChildren[j], + vecChunkBlock[j - 1], vecChunkBlock[j], key); + + fnGetOutcome(vecRelativePosition[j - 1], vecRelativePosition[j], + outcome); + double prob = CalculateConstReorderProb( + const_reorder_classifier_left_, map_left_, key, outcome); + // printf("%s %s %f\n", ostr.str().c_str(), outcome.c_str(), prob); + logprob_const_reorder_left += log10(prob); + + fnGetOutcome(vecRelativeRightPosition[j - 1], + vecRelativeRightPosition[j], outcome); + prob = CalculateConstReorderProb(const_reorder_classifier_right_, + map_right_, key, outcome); + logprob_const_reorder_right += log10(prob); + } + } + + if (b_order_feature_) { + int f_id = FD::Convert("ConstReorderFeatureLeft"); + if (f_id && logprob_const_reorder_left != 0.0) + features->set_value(f_id, logprob_const_reorder_left); + f_id = FD::Convert("ConstReorderFeatureRight"); + if (f_id && logprob_const_reorder_right != 0.0) + features->set_value(f_id, logprob_const_reorder_right); + } + } + } + + private: + void Byte_to_Char(unsigned char* str, int n) { + str[0] = (n & 255); + str[1] = n / 256; + } + void GenerateKey(const STreeItem* pCon1, const STreeItem* pCon2, + int iBlockStatus1, int iBlockStatus2, string& key) { + assert(iBlockStatus1 != 0); + assert(iBlockStatus2 != 0); + unsigned char szTerm[1001]; + Byte_to_Char(szTerm, pCon1->m_iBegin); + Byte_to_Char(szTerm + 2, pCon2->m_iEnd); + szTerm[4] = (char)iBlockStatus1; + szTerm[5] = (char)iBlockStatus2; + szTerm[6] = '\0'; + // sprintf(szTerm, "%d|%d|%d|%d|%s|%s", pCon1->m_iBegin, pCon1->m_iEnd, + // pCon2->m_iBegin, pCon2->m_iEnd, strBlockStatus1.c_str(), + // strBlockStatus2.c_str()); + key = string(szTerm, szTerm + 6); + } + void InitializeConstReorderClassifierOutput() { + if (!b_order_feature_) return; + int size_block_status = dict_block_status_->max(); + + for (size_t i = 0; i < focused_consts_->focus_parents_.size(); i++) { + STreeItem* parent = focused_consts_->focus_parents_[i]; + + for (size_t j = 1; j < parent->m_vecChildren.size(); j++) { + for (size_t k = 1; k <= size_block_status; k++) { + for (size_t l = 1; l <= size_block_status; l++) { + ostringstream ostr; + GenerateFeature(parsed_tree_, parent, j, + dict_block_status_->Convert(k), + dict_block_status_->Convert(l), ostr); + + string strKey; + GenerateKey(parent->m_vecChildren[j - 1], parent->m_vecChildren[j], + k, l, strKey); + + vector<double> vecOutput; + const_reorder_classifier_left_->fnEval(ostr.str().c_str(), + vecOutput); + (*map_left_)[strKey] = vecOutput; + + const_reorder_classifier_right_->fnEval(ostr.str().c_str(), + vecOutput); + (*map_right_)[strKey] = vecOutput; + } + } + } + } + } + + void InitializeSRLReorderClassifierOutput() { + if (!b_srl_order_feature_) return; + int size_block_status = dict_block_status_->max(); + + for (size_t i = 0; i < focused_srl_->focus_predicates_.size(); i++) { + const FocusedPredicate* pred = focused_srl_->focus_predicates_[i]; + + for (size_t j = 1; j < pred->vec_items_.size(); j++) { + for (size_t k = 1; k <= size_block_status; k++) { + for (size_t l = 1; l <= size_block_status; l++) { + ostringstream ostr; + + SArgumentReorderModel::fnGenerateFeature( + parsed_tree_, pred->pred_, pred, j, + dict_block_status_->Convert(k), dict_block_status_->Convert(l), + ostr); + + string strKey; + GenerateKey(pred->vec_items_[j - 1]->tree_item_, + pred->vec_items_[j]->tree_item_, k, l, strKey); + + vector<double> vecOutput; + srl_reorder_classifier_left_->fnEval(ostr.str().c_str(), vecOutput); + (*map_srl_left_)[strKey] = vecOutput; + + srl_reorder_classifier_right_->fnEval(ostr.str().c_str(), + vecOutput); + (*map_srl_right_)[strKey] = vecOutput; + } + } + } + } + } + + double CalculateConstReorderProb( + const Tsuruoka_Maxent* const_reorder_classifier, const MapClassifier* map, + const string& key, const string& outcome) { + MapClassifier::const_iterator iter = (*map).find(key); + assert(iter != map->end()); + int id = const_reorder_classifier->fnGetClassId(outcome); + return iter->second[id]; + } + + void FreeSentenceVariables() { + if (srl_sentence_ != NULL) { + delete srl_sentence_; + srl_sentence_ = NULL; + } else { + if (parsed_tree_ != NULL) delete parsed_tree_; + parsed_tree_ = NULL; + } + + if (focused_consts_ != NULL) delete focused_consts_; + focused_consts_ = NULL; + + for (size_t i = 0; i < vec_target_tran_.size(); i++) + delete vec_target_tran_[i]; + vec_target_tran_.clear(); + + if (map_left_ != NULL) delete map_left_; + map_left_ = NULL; + if (map_right_ != NULL) delete map_right_; + map_right_ = NULL; + + if (map_srl_left_ != NULL) delete map_srl_left_; + map_srl_left_ = NULL; + if (map_srl_right_ != NULL) delete map_srl_right_; + map_srl_right_ = NULL; + } + + void InitializeClassifier(const char* pszFname, + Tsuruoka_Maxent** ppClassifier) { + (*ppClassifier) = new Tsuruoka_Maxent(pszFname); + } + + void GenerateOutcome(const vector<int>& vecPos, vector<string>& vecOutcome) { + vecOutcome.clear(); + + for (size_t i = 1; i < vecPos.size(); i++) { + if (vecPos[i] == -1 || vecPos[i] == vecPos[i - 1]) { + vecOutcome.push_back("M"); // monotone + continue; + } + + if (vecPos[i - 1] == -1) { + // vecPos[i] is not -1 + size_t j = i - 2; + while (j > -1 && vecPos[j] == -1) j--; + + size_t k; + for (k = 0; k < j; k++) { + if (vecPos[k] > vecPos[j] || vecPos[k] <= vecPos[i]) break; + } + if (k < j) { + vecOutcome.push_back("DM"); + continue; + } + + for (k = i + 1; k < vecPos.size(); k++) + if (vecPos[k] < vecPos[i] && (j == -1 && vecPos[k] >= vecPos[j])) + break; + if (k < vecPos.size()) { + vecOutcome.push_back("DM"); + continue; + } + vecOutcome.push_back("M"); + } else { + // neither of vecPos[i-1] and vecPos[i] is -1 + if (vecPos[i - 1] < vecPos[i]) { + // monotone or discon't monotone + size_t j; + for (j = 0; j < i - 1; j++) + if (vecPos[j] > vecPos[i - 1] && vecPos[j] <= vecPos[i]) break; + if (j < i - 1) { + vecOutcome.push_back("DM"); + continue; + } + for (j = i + 1; j < vecPos.size(); j++) + if (vecPos[j] >= vecPos[i - 1] && vecPos[j] < vecPos[i]) break; + if (j < vecPos.size()) { + vecOutcome.push_back("DM"); + continue; + } + vecOutcome.push_back("M"); + } else { + // swap or discon't swap + size_t j; + for (j = 0; j < i - 1; j++) + if (vecPos[j] > vecPos[i] && vecPos[j] <= vecPos[i - 1]) break; + if (j < i - 1) { + vecOutcome.push_back("DS"); + continue; + } + for (j = i + 1; j < vecPos.size(); j++) + if (vecPos[j] >= vecPos[i] && vecPos[j] < vecPos[i - 1]) break; + if (j < vecPos.size()) { + vecOutcome.push_back("DS"); + continue; + } + vecOutcome.push_back("S"); + } + } + } + + assert(vecOutcome.size() == vecPos.size() - 1); + } + + void fnGetRelativePosition(const vector<int>& vecLeft, + vector<int>& vecPosition) { + vecPosition.clear(); + + vector<float> vec; + vec.reserve(vecLeft.size()); + 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); + } + + for (size_t i = 1; i < vecPosition.size(); i++) { + if (vecPosition[i - 1] == vecPosition[i]) { + for (size_t j = 0; j < vecLeft.size(); j++) cout << vecLeft[j] << " "; + cout << "\n"; + 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"); + } + } + + // features in constituent_reorder_model.cc + void GenerateFeature(const SParsedTree* pTree, const STreeItem* pParent, + int iPos, const string& strBlockStatus1, + const string& strBlockStatus2, ostringstream& ostr) { + STreeItem* pCon1, *pCon2; + pCon1 = pParent->m_vecChildren[iPos - 1]; + pCon2 = pParent->m_vecChildren[iPos]; + + string left_label = string(pCon1->m_pszTerm); + string right_label = string(pCon2->m_pszTerm); + string parent_label = string(pParent->m_pszTerm); + + vector<string> 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<string> 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 << "_" + << pTree->m_vecTerminals[pCon1->m_iHeadWord]->m_pszTerm; + // f6 + ostr << " f6=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_ptParent->m_pszTerm; + // f7 + ostr << " f7=" << left_label << "_" << right_label << "_" + << pTree->m_vecTerminals[pCon2->m_iHeadWord]->m_pszTerm; + // f8 + ostr << " f8=" << left_label << "_" << right_label << "_" + << strBlockStatus1; + // f9 + ostr << " f9=" << left_label << "_" << right_label << "_" + << strBlockStatus2; + + // f10 + ostr << " f10=" << left_label << "_" << parent_label; + // f11 + ostr << " f11=" << right_label << "_" << parent_label; + } + + SParsedTree* ReadParseTree(const std::string& parse_file) { + SParseReader* reader = new SParseReader(parse_file.c_str(), false); + SParsedTree* tree = reader->fnReadNextParseTree(); + // assert(tree != NULL); + delete reader; + return tree; + } + + SSrlSentence* ReadSRLSentence(const std::string& srl_file) { + SSrlSentenceReader* reader = new SSrlSentenceReader(srl_file.c_str()); + SSrlSentence* srl = reader->fnReadNextSrlSentence(); + // assert(srl != NULL); + delete reader; + return srl; + } + + private: + Tsuruoka_Maxent* const_reorder_classifier_left_; + Tsuruoka_Maxent* const_reorder_classifier_right_; + + Tsuruoka_Maxent* srl_reorder_classifier_left_; + Tsuruoka_Maxent* srl_reorder_classifier_right_; + + MapClassifier* map_left_; + MapClassifier* map_right_; + + MapClassifier* map_srl_left_; + MapClassifier* map_srl_right_; + + SParsedTree* parsed_tree_; + FocusedConstituent* focused_consts_; + vector<TargetTranslation*> vec_target_tran_; + + bool b_order_feature_; + bool b_block_feature_; + + bool b_srl_block_feature_; + bool b_srl_order_feature_; + SSrlSentence* srl_sentence_; + FocusedSRL* focused_srl_; + + Dict* dict_block_status_; +}; + +ConstReorderFeature::ConstReorderFeature(const std::string& param) { + pimpl_ = new ConstReorderFeatureImpl(param); + SetStateSize(ConstReorderFeatureImpl::ReserveStateSize()); + SetIgnoredStateSize(ConstReorderFeatureImpl::ReserveStateSize()); + name_ = "ConstReorderFeature"; +} + +ConstReorderFeature::~ConstReorderFeature() { // TODO + delete pimpl_; +} + +void ConstReorderFeature::PrepareForInput(const SentenceMetadata& smeta) { + string parse_file = smeta.GetSGMLValue("parse"); + string srl_file = smeta.GetSGMLValue("srl"); + assert(!(parse_file == "" && srl_file == "")); + + pimpl_->InitializeInputSentence(parse_file, srl_file); +} + +void ConstReorderFeature::TraversalFeaturesImpl( + const SentenceMetadata& /* smeta */, const Hypergraph::Edge& edge, + const vector<const void*>& ant_states, SparseVector<double>* features, + SparseVector<double>* /*estimated_features*/, void* state) const { + pimpl_->SetConstReorderFeature(edge, features, ant_states, state); +} + +string ConstReorderFeature::usage(bool show_params, bool show_details) { + ostringstream out; + out << "ConstReorderFeature"; + if (show_params) { + out << " model_file_prefix [const_block=1 const_order=1] [srl_block=0 " + "srl_order=0]" + << "\nParameters:\n" + << " const_{block,order}: enable/disable constituency constraints.\n" + << " src_{block,order}: enable/disable semantic role labeling " + "constraints.\n"; + } + if (show_details) { + out << "\n" + << "Soft reordering constraint features from " + "http://www.aclweb.org/anthology/P14-1106. To train the classifers, " + "use utils/const_reorder_model_trainer for constituency reordering " + "constraints and utils/argument_reorder_model_trainer for semantic " + "role labeling reordering constraints.\n" + << "Input segments should provide path to parse tree (resp. SRL parse) " + "as \"parse\" (resp. \"srl\") properties.\n"; + } + return out.str(); +} + +boost::shared_ptr<FeatureFunction> CreateConstReorderModel( + const std::string& param) { + ConstReorderFeature* ret = new ConstReorderFeature(param); + return boost::shared_ptr<FeatureFunction>(ret); +} diff --git a/decoder/ff_const_reorder.h b/decoder/ff_const_reorder.h new file mode 100644 index 00000000..a5be02d0 --- /dev/null +++ b/decoder/ff_const_reorder.h @@ -0,0 +1,43 @@ +/* + * ff_const_reorder.h + * + * Created on: Jul 11, 2013 + * Author: junhuili + */ + +#ifndef FF_CONST_REORDER_H_ +#define FF_CONST_REORDER_H_ + +#include "ff_factory.h" +#include "ff.h" + +struct ConstReorderFeatureImpl; + +// Soft reordering constraint features from +// http://www.aclweb.org/anthology/P14-1106. To train the classifers, +// use utils/const_reorder_model_trainer for constituency reordering +// constraints and utils/argument_reorder_model_trainer for SRL +// reordering constraints. +// +// Input segments should provide path to parse tree (resp. SRL parse) +// as "parse" (resp. "srl") properties. +class ConstReorderFeature : public FeatureFunction { + public: + ConstReorderFeature(const std::string& param); + ~ConstReorderFeature(); + static std::string usage(bool param, bool verbose); + + protected: + virtual void PrepareForInput(const SentenceMetadata& smeta); + + virtual void TraversalFeaturesImpl( + const SentenceMetadata& smeta, const HG::Edge& edge, + const std::vector<const void*>& ant_contexts, + SparseVector<double>* features, SparseVector<double>* estimated_features, + void* out_context) const; + + private: + ConstReorderFeatureImpl* pimpl_; +}; + +#endif /* FF_CONST_REORDER_H_ */ 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 <string> +#include <assert.h> +#include <stdio.h> +#include <string.h> +#include <string> +#include <sstream> +#include <unordered_map> +#include <utility> +#include <vector> + +#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<STreeItem *> 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<std::string> 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<std::string> 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<std::string> 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<std::string> &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<std::string>::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<STreeItem *> 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<std::string> 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<int> SingleAlign; + SAlignment(const char* pszAlign) { fnInitializeAlignment(pszAlign); } + ~SAlignment() {} + + bool fnIsAligned(int i, bool s) const { + const std::vector<SingleAlign>* 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<SingleAlign>* 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<SingleAlign>& 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<std::string> 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<SingleAlign> m_vec_s_align; // source side words' alignment + std::vector<SingleAlign> 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<SArgument*> 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<SPredicate*> 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<std::vector<std::string> > 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<std::string> vecRole; + std::vector<int> vecBegin; + std::vector<int> 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<std::vector<std::string> >& 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<std::string> 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<std::string, int> Map; +typedef std::unordered_map<std::string, int>::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<std::string> 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<std::string> 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<double> 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<std::pair<std::string, double> >& vecOutput) const { + std::vector<std::string> 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<double> 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<double>& vecOutput) const { + std::vector<std::string> 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<double> 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<SSRLItem *> 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<std::string> &vecStrLeftReorder, + std::vector<std::string> &vecStrRightReorder) { + std::vector<int> 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<int> vecLeftPosition; + fnGetRelativePosition(vecLeft, vecLeftPosition); + std::vector<int> 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<std::string> 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<std::string> 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<int> &vecLeft, + std::vector<int> &vecPosition) { + vecPosition.clear(); + + std::vector<float> 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 |