another feature, another POS

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@664 ec762483-ff6d-05da-a07a-a48fb63a330f

4 files changed, 90 insertions, 1 deletions

diff --git a/decoder/cdec_ff.cc b/decoder/cdec_ff.cc
index 84ba19fa..c0c595a5 100644
--- a/decoder/cdec_ff.cc
+++ b/decoder/cdec_ff.cc
@@ -46,6 +46,7 @@ void register_feature_functions() {
   ff_registry.Register("Model2BinaryFeatures", new FFFactory<Model2BinaryFeatures>);
   ff_registry.Register("MarkovJump", new FFFactory<MarkovJump>);
   ff_registry.Register("MarkovJumpFClass", new FFFactory<MarkovJumpFClass>);
+  ff_registry.Register("SourceBigram", new FFFactory<SourceBigram>);
   ff_registry.Register("SourcePOSBigram", new FFFactory<SourcePOSBigram>);
   ff_registry.Register("BlunsomSynchronousParseHack", new FFFactory<BlunsomSynchronousParseHack>);
   ff_registry.Register("AlignerResults", new FFFactory<AlignerResults>);
diff --git a/decoder/ff.cc b/decoder/ff.cc
index 7bdd21e3..a32c0dcb 100644
--- a/decoder/ff.cc
+++ b/decoder/ff.cc
@@ -171,7 +171,9 @@ void ModelSet::AddFeaturesToEdge(const SentenceMetadata& smeta,
                                  prob_t* combination_cost_estimate) const {
   edge->reset_info();
   context->resize(state_size_);
-  memset(&(*context)[0], 0, state_size_);
+  if (state_size_ > 0) {
+    memset(&(*context)[0], 0, state_size_);
+  }
   SparseVector<double> est_vals;  // only computed if combination_cost_estimate is non-NULL
   if (combination_cost_estimate) *combination_cost_estimate = prob_t::One();
   for (int i = 0; i < models_.size(); ++i) {
diff --git a/decoder/ff_wordalign.cc b/decoder/ff_wordalign.cc
index a1968159..da86b714 100644
--- a/decoder/ff_wordalign.cc
+++ b/decoder/ff_wordalign.cc
@@ -266,6 +266,72 @@ void MarkovJump::TraversalFeaturesImpl(const SentenceMetadata& smeta,
   }
 }
 
+// state: src word used, number of trg words generated
+SourceBigram::SourceBigram(const std::string& param) :
+    FeatureFunction(sizeof(WordID) + sizeof(int)) {
+}
+
+void SourceBigram::FinalTraversalFeatures(const void* context,
+                                      SparseVector<double>* features) const {
+  WordID left = *static_cast<const WordID*>(context);
+  int left_wc = *(static_cast<const int*>(context) + 1);
+  if (left_wc == 1)
+    FireFeature(-1, left, features);
+  FireFeature(left, -1, features);
+}
+
+void SourceBigram::FireFeature(WordID left,
+                   WordID right,
+                   SparseVector<double>* features) const {
+  int& fid = fmap_[left][right];
+  // TODO important important !!! escape strings !!!
+  if (!fid) {
+    ostringstream os;
+    os << "SB:";
+    if (left < 0) { os << "BOS"; } else { os << TD::Convert(left); }
+    os << '_';
+    if (right < 0) { os << "EOS"; } else { os << TD::Convert(right); }
+    fid = FD::Convert(os.str());
+    if (fid == 0) fid = -1;
+  }
+  if (fid > 0) features->set_value(fid, 1.0);
+  int& ufid = ufmap_[left];
+  if (!ufid) {
+    ostringstream os;
+    os << "SU:";
+    if (left < 0) { os << "BOS"; } else { os << TD::Convert(left); }
+    ufid = FD::Convert(os.str());
+    if (ufid == 0) fid = -1;
+  }
+  if (ufid > 0) features->set_value(ufid, 1.0);
+}
+
+void SourceBigram::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 {
+  WordID& out_context = *static_cast<WordID*>(context);
+  int& out_word_count = *(static_cast<int*>(context) + 1);
+  const int arity = edge.Arity();
+  if (arity == 0) {
+    out_context = edge.rule_->f()[0];
+    out_word_count = edge.rule_->EWords();
+    assert(out_word_count == 1); // this is only defined for lex translation!
+    // revisit this if you want to translate into null words
+  } else if (arity == 2) {
+    WordID left = *static_cast<const WordID*>(ant_contexts[0]);
+    WordID right = *static_cast<const WordID*>(ant_contexts[1]);
+    int left_wc = *(static_cast<const int*>(ant_contexts[0]) + 1);
+    int right_wc = *(static_cast<const int*>(ant_contexts[0]) + 1);
+    if (left_wc == 1 && right_wc == 1)
+      FireFeature(-1, left, features);
+    FireFeature(left, right, features);
+    out_word_count = left_wc + right_wc;
+    out_context = right;
+  }
+}
 // state: POS of src word used, number of trg words generated
 SourcePOSBigram::SourcePOSBigram(const std::string& param) :
     FeatureFunction(sizeof(WordID) + sizeof(int)) {
diff --git a/decoder/ff_wordalign.h b/decoder/ff_wordalign.h
index c44ad26b..ebbecfea 100644
--- a/decoder/ff_wordalign.h
+++ b/decoder/ff_wordalign.h
@@ -78,6 +78,26 @@ class MarkovJumpFClass : public FeatureFunction {
 
 typedef std::map<WordID, int> Class2FID;
 typedef std::map<WordID, Class2FID> Class2Class2FID;
+class SourceBigram : public FeatureFunction {
+ public:
+  SourceBigram(const std::string& param);
+  virtual void FinalTraversalFeatures(const void* context,
+                                      SparseVector<double>* features) const;
+ 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;
+ private:
+  void FireFeature(WordID src,
+                   WordID trg,
+                   SparseVector<double>* features) const;
+  mutable Class2Class2FID fmap_;
+  mutable Class2FID ufmap_;
+};
+
 class SourcePOSBigram : public FeatureFunction {
  public:
   SourcePOSBigram(const std::string& param);