summaryrefslogtreecommitdiff
path: root/decoder
diff options
context:
space:
mode:
Diffstat (limited to 'decoder')
-rw-r--r--decoder/Makefile.am6
-rw-r--r--decoder/cdec_ff.cc25
-rw-r--r--decoder/ff_parse_match.cc218
-rw-r--r--decoder/ff_parse_match.h25
-rw-r--r--decoder/ff_soft_syntax.cc201
-rw-r--r--decoder/ff_soft_syntax.h27
-rw-r--r--decoder/ff_soft_syntax2.cc234
-rw-r--r--decoder/ff_soft_syntax2.h27
-rw-r--r--decoder/ff_source_syntax2.cc159
-rw-r--r--decoder/ff_source_syntax2.h25
-rw-r--r--decoder/ff_source_syntax2_p.cc166
-rw-r--r--decoder/ff_source_syntax2_p.h25
-rw-r--r--decoder/ff_source_syntax_p.cc245
-rw-r--r--decoder/ff_source_syntax_p.h42
14 files changed, 1425 insertions, 0 deletions
diff --git a/decoder/Makefile.am b/decoder/Makefile.am
index 82b50f19..00f8b5f9 100644
--- a/decoder/Makefile.am
+++ b/decoder/Makefile.am
@@ -142,7 +142,13 @@ libcdec_a_SOURCES = \
ff_csplit.cc \
ff_tagger.cc \
ff_source_path.cc \
+ ff_parse_match.cc \
+ ff_soft_syntax.cc \
+ ff_soft_syntax2.cc \
ff_source_syntax.cc \
+ ff_source_syntax_p.cc \
+ ff_source_syntax2.cc \
+ ff_source_syntax2_p.cc \
ff_bleu.cc \
ff_factory.cc \
incremental.cc \
diff --git a/decoder/cdec_ff.cc b/decoder/cdec_ff.cc
index 0bf441d4..e7b31f50 100644
--- a/decoder/cdec_ff.cc
+++ b/decoder/cdec_ff.cc
@@ -14,8 +14,18 @@
#include "ff_rules.h"
#include "ff_ruleshape.h"
#include "ff_bleu.h"
+#include "ff_soft_syntax.h"
+#include "ff_soft_syntax2.h"
#include "ff_source_path.h"
+
+
+#include "ff_parse_match.h"
#include "ff_source_syntax.h"
+#include "ff_source_syntax_p.h"
+#include "ff_source_syntax2.h"
+#include "ff_source_syntax2_p.h"
+
+
#include "ff_register.h"
#include "ff_charset.h"
#include "ff_wordset.h"
@@ -48,8 +58,23 @@ void register_feature_functions() {
ff_registry.Register("NgramFeatures", new FFFactory<NgramDetector>());
ff_registry.Register("RuleContextFeatures", new FFFactory<RuleContextFeatures>());
ff_registry.Register("RuleIdentityFeatures", new FFFactory<RuleIdentityFeatures>());
+
+
+ ff_registry.Register("ParseMatchFeatures", new FFFactory<ParseMatchFeatures>);
+
+ ff_registry.Register("SoftSyntacticFeatures", new FFFactory<SoftSyntacticFeatures>);
+ ff_registry.Register("SoftSyntacticFeatures2", new FFFactory<SoftSyntacticFeatures2>);
+
ff_registry.Register("SourceSyntaxFeatures", new FFFactory<SourceSyntaxFeatures>);
+ ff_registry.Register("SourceSyntaxFeatures2", new FFFactory<SourceSyntaxFeatures2>);
+
ff_registry.Register("SourceSpanSizeFeatures", new FFFactory<SourceSpanSizeFeatures>);
+
+ //ff_registry.Register("PSourceSyntaxFeatures", new FFFactory<PSourceSyntaxFeatures>);
+ //ff_registry.Register("PSourceSpanSizeFeatures", new FFFactory<PSourceSpanSizeFeatures>);
+ //ff_registry.Register("PSourceSyntaxFeatures2", new FFFactory<PSourceSyntaxFeatures2>);
+
+
ff_registry.Register("CMR2008ReorderingFeatures", new FFFactory<CMR2008ReorderingFeatures>());
ff_registry.Register("RuleSourceBigramFeatures", new FFFactory<RuleSourceBigramFeatures>());
ff_registry.Register("RuleTargetBigramFeatures", new FFFactory<RuleTargetBigramFeatures>());
diff --git a/decoder/ff_parse_match.cc b/decoder/ff_parse_match.cc
new file mode 100644
index 00000000..ed556b91
--- /dev/null
+++ b/decoder/ff_parse_match.cc
@@ -0,0 +1,218 @@
+#include "ff_parse_match.h"
+
+#include <sstream>
+#include <stack>
+#include <string>
+
+#include "sentence_metadata.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// implements the parse match features as described in Vilar et al. (2008)
+// source trees must be represented in Penn Treebank format, e.g.
+// (S (NP John) (VP (V left)))
+
+struct ParseMatchFeaturesImpl {
+ ParseMatchFeaturesImpl(const string& param) {
+ if (param.compare("") != 0) {
+ char score_param = (char) param[0];
+ switch(score_param) {
+ case 'b':
+ scoring_method = 0;
+ break;
+ case 'l':
+ scoring_method = 1;
+ break;
+ case 'e':
+ scoring_method = 2;
+ break;
+ case 'r':
+ scoring_method = 3;
+ break;
+ default:
+ scoring_method = 0;
+ }
+ }
+ else {
+ scoring_method = 0;
+ }
+ }
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ //cerr << "TREE: " << tree << endl;
+ src_sent_len = src_len;
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ // cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ // cur_cat.second spans from cur_cat.first to i
+ // cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl;
+ // NOTE: unary rule chains end up being labeled with the top-most category
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ int FireFeatures(const TRule& rule, const int i, const int j, int* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ //cerr << rule << endl;
+ //cerr << "span: " << i << " " << j << endl;
+ const WordID lhs = src_tree(i,j);
+ int fid_ef = FD::Convert("PM");
+ int min_dist; // minimal distance to next syntactic constituent of this rule's LHS
+ int summed_min_dists; // minimal distances of LHS and NTs summed up
+ if (TD::Convert(lhs).compare("XX") != 0)
+ min_dist= 0;
+ // compute the distance to the next syntactical constituent
+ else {
+ int ok = 0;
+ for (unsigned int k = 1; k < (j - i); k++) {
+ min_dist = k;
+ for (unsigned int l = 0; l <= k; l++) {
+ // check if adding k words to the rule span will
+ // lead to a syntactical constituent
+ int l_add = i-l;
+ int r_add = j+(k-l);
+ //cerr << "Adding: " << l_add << " " << r_add << endl;
+ if ((l_add < src_tree.width() && r_add < src_tree.height()) && (TD::Convert(src_tree(l_add, r_add)).compare("XX") != 0)) {
+ //cerr << TD::Convert(src_tree(i-l,j+(k-l))) << endl;
+ //cerr << "span_add: " << l_add << " " << r_add << endl;
+ ok = 1;
+ break;
+ }
+ // check if removing k words from the rule span will
+ // lead to a syntactical constituent
+ else {
+ //cerr << "Hilfe...!" << endl;
+ int l_rem= i+l;
+ int r_rem = j-(k-l);
+ //cerr << "Removing: " << l_rem << " " << r_rem << endl;
+ if ((l_rem < src_tree.width() && r_rem < src_tree.height()) && TD::Convert(src_tree(l_rem, r_rem)).compare("XX") != 0) {
+ //cerr << TD::Convert(src_tree(i+l,j-(k-l))) << endl;
+ //cerr << "span_rem: " << l_rem << " " << r_rem << endl;
+ ok = 1;
+ break;
+ }
+ }
+ }
+ if (ok) break;
+ }
+ }
+ summed_min_dists = min_dist;
+ //cerr << min_dist << endl;
+ unsigned ntc = 0;
+ for (unsigned k = 0; k < rule.f_.size(); ++k) {
+ int fj = rule.f_[k];
+ if (fj <= 0)
+ summed_min_dists += ants[ntc++];
+ }
+ switch(scoring_method) {
+ case 0:
+ // binary scoring
+ feats->set_value(fid_ef, (summed_min_dists == 0));
+ break;
+ // CHECK: for the remaining scoring methods, the question remains if
+ // min_dist or summed_min_dists should be used
+ case 1:
+ // linear scoring
+ feats->set_value(fid_ef, 1.0/(min_dist+1));
+ break;
+ case 2:
+ // exponential scoring
+ feats->set_value(fid_ef, 1.0/exp(min_dist));
+ break;
+ case 3:
+ // relative scoring
+ feats->set_value(fid_ef, (j-i)/((j-i) + min_dist));
+ break;
+ default:
+ // binary scoring in case nothing is defined
+ feats->set_value(fid_ef, (summed_min_dists == 0));
+ }
+ return min_dist;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ unsigned int src_sent_len;
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ int scoring_method;
+};
+
+ParseMatchFeatures::ParseMatchFeatures(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new ParseMatchFeaturesImpl(param);
+}
+
+ParseMatchFeatures::~ParseMatchFeatures() {
+ delete impl;
+ impl = NULL;
+}
+
+void ParseMatchFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ int ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const int*>(ant_contexts[i]);
+
+ *static_cast<int*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void ParseMatchFeatures::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
diff --git a/decoder/ff_parse_match.h b/decoder/ff_parse_match.h
new file mode 100644
index 00000000..fa73481a
--- /dev/null
+++ b/decoder/ff_parse_match.h
@@ -0,0 +1,25 @@
+#ifndef _FF_PARSE_MATCH_H_
+#define _FF_PARSE_MATCH_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct ParseMatchFeaturesImpl;
+
+class ParseMatchFeatures : public FeatureFunction {
+ public:
+ ParseMatchFeatures(const std::string& param);
+ ~ParseMatchFeatures();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ ParseMatchFeaturesImpl* impl;
+};
+
+#endif
diff --git a/decoder/ff_soft_syntax.cc b/decoder/ff_soft_syntax.cc
new file mode 100644
index 00000000..9981fa45
--- /dev/null
+++ b/decoder/ff_soft_syntax.cc
@@ -0,0 +1,201 @@
+#include "ff_soft_syntax.h"
+
+#include <cstdio>
+#include <sstream>
+#include <stack>
+#include <string>
+#include <vector>
+
+#include "sentence_metadata.h"
+#include "stringlib.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// Implements the soft syntactic features described in
+// Marton and Resnik (2008): "Soft Syntacitc Constraints for Hierarchical Phrase-Based Translation".
+// Source trees must be represented in Penn Treebank format,
+// e.g. (S (NP John) (VP (V left))).
+
+struct SoftSyntacticFeaturesImpl {
+ SoftSyntacticFeaturesImpl(const string& param) {
+ vector<string> labels = SplitOnWhitespace(param);
+ for (unsigned int i = 0; i < labels.size(); i++)
+ //cerr << "Labels: " << labels.at(i) << endl;
+ for (unsigned int i = 0; i < labels.size(); i++) {
+ string label = labels.at(i);
+ pair<string, string> feat_label;
+ feat_label.first = label.substr(0, label.size() - 1);
+ feat_label.second = label.at(label.size() - 1);
+ feat_labels.push_back(feat_label);
+ }
+}
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ //cerr << "String " << tree << endl;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ //cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ // cur_cat.second spans from cur_cat.first to i
+ //cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl;
+ // NOTE: unary rule chains end up being labeled with the top-most category
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ const WordID lhs = src_tree(i,j);
+ string lhs_str = TD::Convert(lhs);
+ //cerr << "LHS: " << lhs_str << " from " << i << " to " << j << endl;
+ //cerr << "RULE :"<< rule << endl;
+ int& fid_ef = fids_ef(i,j)[&rule];
+ for (unsigned int i = 0; i < feat_labels.size(); i++) {
+ ostringstream os;
+ string label = feat_labels.at(i).first;
+ //cerr << "This Label: " << label << endl;
+ char feat_type = (char) feat_labels.at(i).second.c_str()[0];
+ //cerr << "feat_type: " << feat_type << endl;
+ switch(feat_type) {
+ case '2':
+ if (lhs_str.compare(label) == 0) {
+ os << "SYN:" << label << "_conform";
+ }
+ else {
+ os << "SYN:" << label << "_cross";
+ }
+ fid_ef = FD::Convert(os.str());
+ if (fid_ef > 0) {
+ //cerr << "Feature :" << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ break;
+ case '_':
+ os << "SYN:" << label;
+ fid_ef = FD::Convert(os.str());
+ if (lhs_str.compare(label) == 0) {
+ if (fid_ef > 0) {
+ //cerr << "Feature: " << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ }
+ else {
+ if (fid_ef > 0) {
+ //cerr << "Feature: " << os.str() << endl;
+ feats->set_value(fid_ef, -1.0);
+ }
+ }
+ break;
+ case '+':
+ if (lhs_str.compare(label) == 0) {
+ os << "SYN:" << label << "_conform";
+ fid_ef = FD::Convert(os.str());
+ if (fid_ef > 0) {
+ //cerr << "Feature: " << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ }
+ break;
+ case '-':
+ //cerr << "-" << endl;
+ if (lhs_str.compare(label) != 0) {
+ os << "SYN:" << label << "_cross";
+ fid_ef = FD::Convert(os.str());
+ if (fid_ef > 0) {
+ //cerr << "Feature :" << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ }
+ break;
+ os.clear();
+ os.str("");
+ }
+ //cerr << "Feature: " << os.str() << endl;
+ //cerr << endl;
+ }
+ return lhs;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ vector<pair<string, string> > feat_labels;
+};
+
+SoftSyntacticFeatures::SoftSyntacticFeatures(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new SoftSyntacticFeaturesImpl(param);
+}
+
+SoftSyntacticFeatures::~SoftSyntacticFeatures() {
+ delete impl;
+ impl = NULL;
+}
+
+void SoftSyntacticFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ WordID ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
+
+ *static_cast<WordID*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void SoftSyntacticFeatures::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
diff --git a/decoder/ff_soft_syntax.h b/decoder/ff_soft_syntax.h
new file mode 100644
index 00000000..79352f49
--- /dev/null
+++ b/decoder/ff_soft_syntax.h
@@ -0,0 +1,27 @@
+#ifndef _FF_SOFTSYNTAX_H_
+#define _FF_SOFTSYNTAX_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct SoftSyntacticFeaturesImpl;
+
+class SoftSyntacticFeatures : public FeatureFunction {
+ public:
+ SoftSyntacticFeatures(const std::string& param);
+ ~SoftSyntacticFeatures();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ SoftSyntacticFeaturesImpl* impl;
+};
+
+
+
+#endif
diff --git a/decoder/ff_soft_syntax2.cc b/decoder/ff_soft_syntax2.cc
new file mode 100644
index 00000000..121bc39b
--- /dev/null
+++ b/decoder/ff_soft_syntax2.cc
@@ -0,0 +1,234 @@
+#include "ff_soft_syntax2.h"
+
+#include <cstdio>
+#include <sstream>
+#include <stack>
+#include <string>
+#include <vector>
+
+#include "sentence_metadata.h"
+#include "stringlib.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// Implements the soft syntactic features described in
+// Marton and Resnik (2008): "Soft Syntacitc Constraints for Hierarchical Phrase-Based Translation".
+// Source trees must be represented in Penn Treebank format,
+// e.g. (S (NP John) (VP (V left))).
+
+struct SoftSyntacticFeatures2Impl {
+ SoftSyntacticFeatures2Impl(const string& param) {
+ vector<string> labels = SplitOnWhitespace(param);
+ //for (unsigned int i = 0; i < labels.size(); i++)
+ //cerr << "Labels: " << labels.at(i) << endl;
+ for (unsigned int i = 0; i < labels.size(); i++) {
+ string label = labels.at(i);
+ pair<string, string> feat_label;
+ feat_label.first = label.substr(0, label.size() - 1);
+ feat_label.second = label.at(label.size() - 1);
+ feat_labels.push_back(feat_label);
+ }
+ }
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ //cerr << "String " << tree << endl;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ //cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ // cur_cat.second spans from cur_cat.first to i
+ //cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl;
+ // NOTE: unary rule chains end up being labeled with the top-most category
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ const WordID lhs = src_tree(i,j);
+ string lhs_str = TD::Convert(lhs);
+ //cerr << "LHS: " << lhs_str << " from " << i << " to " << j << endl;
+ //cerr << "RULE :"<< rule << endl;
+ int& fid_ef = fids_ef(i,j)[&rule];
+ string lhs_to_str = TD::Convert(lhs);
+ int min_dist;
+ string min_dist_label;
+ if (lhs_to_str.compare("XX") != 0) {
+ min_dist = 0;
+ min_dist_label = lhs_to_str;
+ }
+ else {
+ int ok = 0;
+ for (unsigned int k = 1; k < (j - i); k++) {
+ min_dist = k;
+ for (unsigned int l = 0; l <= k; l++) {
+ int l_add = i-l;
+ int r_add = j+(k-l);
+ if ((l_add < src_tree.width() && r_add < src_tree.height()) && (TD::Convert(src_tree(l_add, r_add)).compare("XX") != 0)) {
+ ok = 1;
+ min_dist_label = (TD::Convert(src_tree(l_add, r_add)));
+ break;
+ }
+ else {
+ int l_rem= i+l;
+ int r_rem = j-(k-l);
+ if ((l_rem < src_tree.width() && r_rem < src_tree.height()) && TD::Convert(src_tree(l_rem, r_rem)).compare("XX") != 0) {
+ ok = 1;
+ min_dist_label = (TD::Convert(src_tree(l_rem, r_rem)));
+ break;
+ }
+ }
+ }
+ if (ok) break;
+ }
+ }
+ //cerr << "SPAN: " << i << " " << j << endl;
+ //cerr << "MINDIST: " << min_dist << endl;
+ //cerr << "MINDISTLABEL: " << min_dist_label << endl;
+ for (unsigned int i = 0; i < feat_labels.size(); i++) {
+ ostringstream os;
+ string label = feat_labels.at(i).first;
+ //cerr << "This Label: " << label << endl;
+ char feat_type = (char) feat_labels.at(i).second.c_str()[0];
+ //cerr << "feat_type: " << feat_type << endl;
+ switch(feat_type) {
+ case '2':
+ if (min_dist_label.compare(label) == 0) {
+ if (min_dist == 0) {
+ os << "SYN:" << label << "_conform";
+ }
+ else {
+ os << "SYN:" << label << "_cross";
+ }
+ fid_ef = FD::Convert(os.str());
+ //cerr << "Feature :" << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ break;
+ case '_':
+ os << "SYN:" << label;
+ fid_ef = FD::Convert(os.str());
+ if (min_dist_label.compare(label) == 0) {
+ //cerr << "Feature: " << os.str() << endl;
+ if (min_dist == 0) {
+ feats->set_value(fid_ef, 1.0);
+ }
+ else {
+ //cerr << "Feature: " << os.str() << endl;
+ feats->set_value(fid_ef, -1.0);
+ }
+ }
+ break;
+ case '+':
+ if (min_dist_label.compare(label) == 0) {
+ os << "SYN:" << label << "_conform";
+ fid_ef = FD::Convert(os.str());
+ if (min_dist == 0) {
+ //cerr << "Feature: " << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ }
+ break;
+ case '-':
+ //cerr << "-" << endl;
+ if (min_dist_label.compare(label) != 0) {
+ os << "SYN:" << label << "_cross";
+ fid_ef = FD::Convert(os.str());
+ if (min_dist > 0) {
+ //cerr << "Feature :" << os.str() << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ }
+ break;
+ os.clear();
+ os.str("");
+ }
+ //cerr << "FEATURE: " << os.str() << endl;
+ //cerr << endl;
+ }
+ return lhs;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ vector<pair<string, string> > feat_labels;
+};
+
+SoftSyntacticFeatures2::SoftSyntacticFeatures2(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new SoftSyntacticFeatures2Impl(param);
+}
+
+SoftSyntacticFeatures2::~SoftSyntacticFeatures2() {
+ delete impl;
+ impl = NULL;
+}
+
+void SoftSyntacticFeatures2::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ WordID ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
+
+ *static_cast<WordID*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void SoftSyntacticFeatures2::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
diff --git a/decoder/ff_soft_syntax2.h b/decoder/ff_soft_syntax2.h
new file mode 100644
index 00000000..4de91d86
--- /dev/null
+++ b/decoder/ff_soft_syntax2.h
@@ -0,0 +1,27 @@
+#ifndef _FF_SOFTSYNTAX2_H_
+#define _FF_SOFTSYNTAX2_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct SoftSyntacticFeatures2Impl;
+
+class SoftSyntacticFeatures2 : public FeatureFunction {
+ public:
+ SoftSyntacticFeatures2(const std::string& param);
+ ~SoftSyntacticFeatures2();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ SoftSyntacticFeatures2Impl* impl;
+};
+
+
+
+#endif
diff --git a/decoder/ff_source_syntax2.cc b/decoder/ff_source_syntax2.cc
new file mode 100644
index 00000000..08ece917
--- /dev/null
+++ b/decoder/ff_source_syntax2.cc
@@ -0,0 +1,159 @@
+#include "ff_source_syntax2.h"
+
+#include <sstream>
+#include <stack>
+#include <string>
+#include <tr1/unordered_set>
+
+#include "sentence_metadata.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// implements the source side syntax features described in Blunsom et al. (EMNLP 2008)
+// source trees must be represented in Penn Treebank format, e.g.
+// (S (NP John) (VP (V left)))
+
+struct SourceSyntaxFeatures2Impl {
+ SourceSyntaxFeatures2Impl(const string& param) {
+ if (!(param.compare("") == 0)) {
+ string triggered_features_fn = param;
+ ReadFile triggered_features(triggered_features_fn);
+ string in;
+ while(getline(*triggered_features, in)) {
+ feature_filter.insert(FD::Convert(in));
+ }
+ }
+ }
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ //cerr << "TREE: " << tree << endl;
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ // cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ const WordID lhs = src_tree(i,j);
+ int& fid_ef = fids_ef(i,j)[&rule];
+ ostringstream os;
+ os << "SYN:" << TD::Convert(lhs);
+ os << ':';
+ unsigned ntc = 0;
+ for (unsigned k = 0; k < rule.f_.size(); ++k) {
+ int fj = rule.f_[k];
+ if (k > 0 && fj <= 0) os << '_';
+ if (fj <= 0) {
+ os << '[' << TD::Convert(ants[ntc++]) << ']';
+ } /*else {
+ os << TD::Convert(fj);
+ }*/
+ }
+ os << ':';
+ for (unsigned k = 0; k < rule.e_.size(); ++k) {
+ const int ei = rule.e_[k];
+ if (k > 0) os << '_';
+ if (ei <= 0)
+ os << '[' << (1-ei) << ']';
+ else
+ os << TD::Convert(ei);
+ }
+ fid_ef = FD::Convert(os.str());
+ //cerr << "FEATURE: " << os.str() << endl;
+ //cerr << "FID_EF: " << fid_ef << endl;
+ if (feature_filter.find(fid_ef) != feature_filter.end()) {
+ cerr << "SYN-Feature was trigger more than once on training set." << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ else cerr << "SYN-Feature was triggered less than once on training set." << endl;
+ return lhs;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ tr1::unordered_set<int> feature_filter;
+
+};
+
+SourceSyntaxFeatures2::SourceSyntaxFeatures2(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new SourceSyntaxFeatures2Impl(param);
+}
+
+SourceSyntaxFeatures2::~SourceSyntaxFeatures2() {
+ delete impl;
+ impl = NULL;
+}
+
+void SourceSyntaxFeatures2::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ WordID ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
+
+ *static_cast<WordID*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void SourceSyntaxFeatures2::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
diff --git a/decoder/ff_source_syntax2.h b/decoder/ff_source_syntax2.h
new file mode 100644
index 00000000..b6b7dc3d
--- /dev/null
+++ b/decoder/ff_source_syntax2.h
@@ -0,0 +1,25 @@
+#ifndef _FF_SOURCE_TOOLS2_H_
+#define _FF_SOURCE_TOOLS2_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct SourceSyntaxFeatures2Impl;
+
+class SourceSyntaxFeatures2 : public FeatureFunction {
+ public:
+ SourceSyntaxFeatures2(const std::string& param);
+ ~SourceSyntaxFeatures2();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ SourceSyntaxFeatures2Impl* impl;
+};
+
+#endif
diff --git a/decoder/ff_source_syntax2_p.cc b/decoder/ff_source_syntax2_p.cc
new file mode 100644
index 00000000..dfa791ea
--- /dev/null
+++ b/decoder/ff_source_syntax2_p.cc
@@ -0,0 +1,166 @@
+#include "ff_source_syntax2_p.h"
+
+#include <sstream>
+#include <stack>
+#include <string>
+#include <tr1/unordered_set>
+
+#include "sentence_metadata.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// implements the source side syntax features described in Blunsom et al. (EMNLP 2008)
+// source trees must be represented in Penn Treebank format, e.g.
+// (S (NP John) (VP (V left)))
+
+struct PSourceSyntaxFeatures2Impl {
+ PSourceSyntaxFeatures2Impl(const string& param) {
+ if (param.compare("") != 0) {
+ string triggered_features_fn = param;
+ ReadFile triggered_features(triggered_features_fn);
+ string in;
+ while(getline(*triggered_features, in)) {
+ feature_filter.insert(FD::Convert(in));
+ }
+ }
+ /*cerr << "find(\"One\") == " << boolalpha << (table.find("One") != table.end()) << endl;
+ cerr << "find(\"Three\") == " << boolalpha << (table.find("Three") != table.end()) << endl;*/
+ }
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ //cerr << "TREE: " << tree << endl;
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ // cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ }
+ //cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ const WordID lhs = src_tree(i,j);
+ int& fid_ef = fids_ef(i,j)[&rule];
+ ostringstream os;
+ os << "SYN:" << TD::Convert(lhs);
+ os << ':';
+ unsigned ntc = 0;
+ for (unsigned k = 0; k < rule.f_.size(); ++k) {
+ int fj = rule.f_[k];
+ if (k > 0 && fj <= 0) os << '_';
+ if (fj <= 0) {
+ os << '[' << TD::Convert(ants[ntc++]) << ']';
+ } /*else {
+ os << TD::Convert(fj);
+ }*/
+ }
+ os << ':';
+ for (unsigned k = 0; k < rule.e_.size(); ++k) {
+ const int ei = rule.e_[k];
+ if (k > 0) os << '_';
+ if (ei <= 0)
+ os << '[' << (1-ei) << ']';
+ else
+ os << TD::Convert(ei);
+ }
+ fid_ef = FD::Convert(os.str());
+ //cerr << "FEATURE: " << os.str() << endl;
+ //cerr << "FID_EF: " << fid_ef << endl;
+ if (feature_filter.size() > 0) {
+ if (feature_filter.find(fid_ef) != feature_filter.end()) {
+ //cerr << "SYN-Feature was trigger more than once on training set." << endl;
+ feats->set_value(fid_ef, 1.0);
+ }
+ //else cerr << "SYN-Feature was triggered less than once on training set." << endli;
+ }
+ else {
+ feats->set_value(fid_ef, 1.0);
+ }
+ return lhs;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ tr1::unordered_set<int> feature_filter;
+
+};
+
+PSourceSyntaxFeatures2::PSourceSyntaxFeatures2(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new PSourceSyntaxFeatures2Impl(param);
+}
+
+PSourceSyntaxFeatures2::~PSourceSyntaxFeatures2() {
+ delete impl;
+ impl = NULL;
+}
+
+void PSourceSyntaxFeatures2::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ WordID ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
+
+ *static_cast<WordID*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void PSourceSyntaxFeatures2::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
diff --git a/decoder/ff_source_syntax2_p.h b/decoder/ff_source_syntax2_p.h
new file mode 100644
index 00000000..d56ecab0
--- /dev/null
+++ b/decoder/ff_source_syntax2_p.h
@@ -0,0 +1,25 @@
+#ifndef _FF_SOURCE_TOOLS2_H_
+#define _FF_SOURCE_TOOLS2_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct PSourceSyntaxFeatures2Impl;
+
+class PSourceSyntaxFeatures2 : public FeatureFunction {
+ public:
+ PSourceSyntaxFeatures2(const std::string& param);
+ ~PSourceSyntaxFeatures2();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ PSourceSyntaxFeatures2Impl* impl;
+};
+
+#endif
diff --git a/decoder/ff_source_syntax_p.cc b/decoder/ff_source_syntax_p.cc
new file mode 100644
index 00000000..cd081544
--- /dev/null
+++ b/decoder/ff_source_syntax_p.cc
@@ -0,0 +1,245 @@
+#include "ff_source_syntax_p.h"
+
+#include <sstream>
+#include <stack>
+#include <tr1/unordered_set>
+
+#include "sentence_metadata.h"
+#include "array2d.h"
+#include "filelib.h"
+
+using namespace std;
+
+// implements the source side syntax features described in Blunsom et al. (EMNLP 2008)
+// source trees must be represented in Penn Treebank format, e.g.
+// (S (NP John) (VP (V left)))
+
+// log transform to make long spans cluster together
+// but preserve differences
+inline int SpanSizeTransform(unsigned span_size) {
+ if (!span_size) return 0;
+ return static_cast<int>(log(span_size+1) / log(1.39)) - 1;
+}
+
+struct PSourceSyntaxFeaturesImpl {
+ PSourceSyntaxFeaturesImpl() {}
+
+ PSourceSyntaxFeaturesImpl(const string& param) {
+ if (!(param.compare("") == 0)) {
+ string triggered_features_fn = param;
+ ReadFile triggered_features(triggered_features_fn);
+ string in;
+ while(getline(*triggered_features, in)) {
+ feature_filter.insert(FD::Convert(in));
+ }
+ }
+ }
+
+ void InitializeGrids(const string& tree, unsigned src_len) {
+ assert(tree.size() > 0);
+ //fids_cat.clear();
+ fids_ef.clear();
+ src_tree.clear();
+ //fids_cat.resize(src_len, src_len + 1);
+ fids_ef.resize(src_len, src_len + 1);
+ src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
+ ParseTreeString(tree, src_len);
+ }
+
+ void ParseTreeString(const string& tree, unsigned src_len) {
+ stack<pair<int, WordID> > stk; // first = i, second = category
+ pair<int, WordID> cur_cat; cur_cat.first = -1;
+ unsigned i = 0;
+ unsigned p = 0;
+ while(p < tree.size()) {
+ const char cur = tree[p];
+ if (cur == '(') {
+ stk.push(cur_cat);
+ ++p;
+ unsigned k = p + 1;
+ while (k < tree.size() && tree[k] != ' ') { ++k; }
+ cur_cat.first = i;
+ cur_cat.second = TD::Convert(tree.substr(p, k - p));
+ // cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ p = k + 1;
+ } else if (cur == ')') {
+ unsigned k = p;
+ while (k < tree.size() && tree[k] == ')') { ++k; }
+ const unsigned num_closes = k - p;
+ for (unsigned ci = 0; ci < num_closes; ++ci) {
+ // cur_cat.second spans from cur_cat.first to i
+ // cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl;
+ // NOTE: unary rule chains end up being labeled with the top-most category
+ src_tree(cur_cat.first, i) = cur_cat.second;
+ cur_cat = stk.top();
+ stk.pop();
+ }
+ p = k;
+ while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
+ } else if (cur == ' ' || cur == '\t') {
+ cerr << "Unexpected whitespace in: " << tree << endl;
+ abort();
+ } else { // terminal symbol
+ unsigned k = p + 1;
+ do {
+ while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
+ // cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
+ ++i;
+ assert(i <= src_len);
+ while (k < tree.size() && tree[k] == ' ') { ++k; }
+ p = k;
+ } while (p < tree.size() && tree[p] != ')');
+ }
+ }
+ // cerr << "i=" << i << " src_len=" << src_len << endl;
+ assert(i == src_len); // make sure tree specified in src_tree is
+ // the same length as the source sentence
+ }
+
+ WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ //cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
+ const WordID lhs = src_tree(i,j);
+ //int& fid_cat = fids_cat(i,j);
+ int& fid_ef = fids_ef(i,j)[&rule];
+ if (fid_ef <= 0) {
+ ostringstream os;
+ //ostringstream os2;
+ os << "SYN:" << TD::Convert(lhs);
+ //os2 << "SYN:" << TD::Convert(lhs) << '_' << SpanSizeTransform(j - i);
+ //fid_cat = FD::Convert(os2.str());
+ os << ':';
+ unsigned ntc = 0;
+ for (unsigned k = 0; k < rule.f_.size(); ++k) {
+ if (k > 0) os << '_';
+ int fj = rule.f_[k];
+ if (fj <= 0) {
+ os << '[' << TD::Convert(ants[ntc++]) << ']';
+ } else {
+ os << TD::Convert(fj);
+ }
+ }
+ os << ':';
+ for (unsigned k = 0; k < rule.e_.size(); ++k) {
+ const int ei = rule.e_[k];
+ if (k > 0) os << '_';
+ if (ei <= 0)
+ os << '[' << (1-ei) << ']';
+ else
+ os << TD::Convert(ei);
+ }
+ fid_ef = FD::Convert(os.str());
+ }
+ //if (fid_cat > 0)
+ // feats->set_value(fid_cat, 1.0);
+ if (fid_ef > 0 && (feature_filter.find(fid_ef) != feature_filter.end()))
+ feats->set_value(fid_ef, 1.0);
+ return lhs;
+ }
+
+ Array2D<WordID> src_tree; // src_tree(i,j) NT = type
+ // mutable Array2D<int> fids_cat; // this tends to overfit baddly
+ mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
+ tr1::unordered_set<int> feature_filter;
+};
+
+PSourceSyntaxFeatures::PSourceSyntaxFeatures(const string& param) :
+ FeatureFunction(sizeof(WordID)) {
+ impl = new PSourceSyntaxFeaturesImpl(param);
+}
+
+PSourceSyntaxFeatures::~PSourceSyntaxFeatures() {
+ delete impl;
+ impl = NULL;
+}
+
+void PSourceSyntaxFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ WordID ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
+
+ *static_cast<WordID*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void PSourceSyntaxFeatures::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
+}
+
+struct PSourceSpanSizeFeaturesImpl {
+ PSourceSpanSizeFeaturesImpl() {}
+
+ void InitializeGrids(unsigned src_len) {
+ fids.clear();
+ fids.resize(src_len, src_len + 1);
+ }
+
+ int FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
+ if (rule.Arity() > 0) {
+ int& fid = fids(i,j)[&rule];
+ if (fid <= 0) {
+ ostringstream os;
+ os << "SSS:";
+ unsigned ntc = 0;
+ for (unsigned k = 0; k < rule.f_.size(); ++k) {
+ if (k > 0) os << '_';
+ int fj = rule.f_[k];
+ if (fj <= 0) {
+ os << '[' << TD::Convert(-fj) << ants[ntc++] << ']';
+ } else {
+ os << TD::Convert(fj);
+ }
+ }
+ os << ':';
+ for (unsigned k = 0; k < rule.e_.size(); ++k) {
+ const int ei = rule.e_[k];
+ if (k > 0) os << '_';
+ if (ei <= 0)
+ os << '[' << (1-ei) << ']';
+ else
+ os << TD::Convert(ei);
+ }
+ fid = FD::Convert(os.str());
+ }
+ if (fid > 0)
+ feats->set_value(fid, 1.0);
+ }
+ return SpanSizeTransform(j - i);
+ }
+
+ mutable Array2D<map<const TRule*, int> > fids;
+};
+
+PSourceSpanSizeFeatures::PSourceSpanSizeFeatures(const string& param) :
+ FeatureFunction(sizeof(char)) {
+ impl = new PSourceSpanSizeFeaturesImpl;
+}
+
+PSourceSpanSizeFeatures::~PSourceSpanSizeFeatures() {
+ delete impl;
+ impl = NULL;
+}
+
+void PSourceSpanSizeFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const {
+ int ants[8];
+ for (unsigned i = 0; i < ant_contexts.size(); ++i)
+ ants[i] = *static_cast<const char*>(ant_contexts[i]);
+
+ *static_cast<char*>(context) =
+ impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
+}
+
+void PSourceSpanSizeFeatures::PrepareForInput(const SentenceMetadata& smeta) {
+ impl->InitializeGrids(smeta.GetSourceLength());
+}
+
+
diff --git a/decoder/ff_source_syntax_p.h b/decoder/ff_source_syntax_p.h
new file mode 100644
index 00000000..2dd9094a
--- /dev/null
+++ b/decoder/ff_source_syntax_p.h
@@ -0,0 +1,42 @@
+#ifndef _FF_SOURCE_TOOLS_H_
+#define _FF_SOURCE_TOOLS_H_
+
+#include "ff.h"
+#include "hg.h"
+
+struct PSourceSyntaxFeaturesImpl;
+
+class PSourceSyntaxFeatures : public FeatureFunction {
+ public:
+ PSourceSyntaxFeatures(const std::string& param);
+ ~PSourceSyntaxFeatures();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ PSourceSyntaxFeaturesImpl* impl;
+};
+
+struct PSourceSpanSizeFeaturesImpl;
+class PSourceSpanSizeFeatures : public FeatureFunction {
+ public:
+ PSourceSpanSizeFeatures(const std::string& param);
+ ~PSourceSpanSizeFeatures();
+ protected:
+ virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+ const Hypergraph::Edge& edge,
+ const std::vector<const void*>& ant_contexts,
+ SparseVector<double>* features,
+ SparseVector<double>* estimated_features,
+ void* context) const;
+ virtual void PrepareForInput(const SentenceMetadata& smeta);
+ private:
+ PSourceSpanSizeFeaturesImpl* impl;
+};
+
+#endif
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-16 04:33:29 +0000