handle translation from the null word

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@689 ec762483-ff6d-05da-a07a-a48fb63a330f

4 files changed, 55 insertions, 31 deletions

diff --git a/decoder/aligner.cc b/decoder/aligner.cc
index 92431be4..3f0c7347 100644
--- a/decoder/aligner.cc
+++ b/decoder/aligner.cc
@@ -24,8 +24,10 @@ void SourceEdgeCoveragesUsingParseIndices(const Hypergraph& g,
     if (edge.rule_->EWords() == 0 || edge.rule_->FWords() == 0)
       continue;
     // aligned to NULL (crf ibm variant only)
-    if (edge.prev_i_ == -1 || edge.i_ == -1)
+    if (edge.prev_i_ == -1 || edge.i_ == -1) {
+      cov.insert(-1);
       continue;
+    }
     assert(edge.j_ >= 0);
     assert(edge.prev_j_ >= 0);
     if (edge.Arity() == 0) {
@@ -211,7 +213,7 @@ void AlignerTools::WriteAlignment(const Lattice& src_lattice,
   // figure out the src and reference size;
   int src_size = src_sent.size();
   int ref_size = trg_sent.size();
-  Array2D<prob_t> align(src_size, ref_size, prob_t::Zero());
+  Array2D<prob_t> align(src_size + 1, ref_size, prob_t::Zero());
   for (int c = 0; c < g->edges_.size(); ++c) {
     const prob_t& p = edge_posteriors[c];
     const set<int>& srcs = src_cov[c];
@@ -220,7 +222,7 @@ void AlignerTools::WriteAlignment(const Lattice& src_lattice,
          si != srcs.end(); ++si) {
       for (set<int>::const_iterator ti = trgs.begin();
            ti != trgs.end(); ++ti) {
-        align(*si, *ti) += p;
+        align(*si + 1, *ti) += p;
       }
     }
   }
@@ -234,12 +236,12 @@ void AlignerTools::WriteAlignment(const Lattice& src_lattice,
   for (int j = 0; j < ref_size; ++j) {
     if (use_soft_threshold) {
       threshold = prob_t::Zero();
-      for (int i = 0; i < src_size; ++i)
+      for (int i = 0; i <= src_size; ++i)
         if (align(i, j) > threshold) threshold = align(i, j);
       //threshold *= prob_t(0.99);
     }
     for (int i = 0; i < src_size; ++i)
-      grid(i, j) = align(i, j) >= threshold;
+      grid(i, j) = align(i+1, j) >= threshold;
   }
   if (out == &cout) {
     // TODO need to do some sort of verbose flag
diff --git a/decoder/decoder.cc b/decoder/decoder.cc
index b975a5fc..eb983419 100644
--- a/decoder/decoder.cc
+++ b/decoder/decoder.cc
@@ -364,7 +364,7 @@ DecoderImpl::DecoderImpl(po::variables_map& conf, int argc, char** argv, istream
         ("beam_prune", po::value<double>(), "Prune paths from +LM forest, keep paths within exp(alpha>=0)")
     ("scale_prune_srclen", "scale beams by the input length (in # of tokens; may not be what you want for lattices")
     ("promise_power",po::value<double>()->default_value(0), "Give more beam budget to more promising previous-pass nodes when pruning - but allocate the same average beams.  0 means off, 1 means beam proportional to inside*outside prob, n means nth power (affects just --cubepruning_pop_limit).  note: for the same pop_limit, this gives more search error unless very close to 0 (recommend disabled; even 0.01 is slightly worse than 0) which is a bad sign and suggests this isn't doing a good job; further it's slightly slower to LM cube rescore with 0.01 compared to 0, as well as giving (very insignificantly) lower BLEU.  TODO: test under more conditions, or try idea with different formula, or prob. cube beams.")
-        ("lexalign_use_null", "Support source-side null words in lexical translation")
+        ("lextrans_use_null", "Support source-side null words in lexical translation")
         ("tagger_tagset,t", po::value<string>(), "(Tagger) file containing tag set")
         ("csplit_output_plf", "(Compound splitter) Output lattice in PLF format")
         ("csplit_preserve_full_word", "(Compound splitter) Always include the unsegmented form in the output lattice")
diff --git a/decoder/ff_wordalign.cc b/decoder/ff_wordalign.cc
index f2f07033..5f42b438 100644
--- a/decoder/ff_wordalign.cc
+++ b/decoder/ff_wordalign.cc
@@ -16,6 +16,8 @@
 
 static const int MAX_SENTENCE_SIZE = 100;
 
+static const int kNULL_i = 255;  // -1 as an unsigned char
+
 using namespace std;
 
 Model2BinaryFeatures::Model2BinaryFeatures(const string& ) :
@@ -149,7 +151,11 @@ void MarkovJumpFClass::FireFeature(const SentenceMetadata& smeta,
                                    int prev_src_pos,
                                    int cur_src_pos,
                                    SparseVector<double>* features) const {
+  if (prev_src_pos == kNULL_i || cur_src_pos == kNULL_i)
+    return;
+
   const int jumpsize = cur_src_pos - prev_src_pos;
+
   assert(smeta.GetSentenceID() < pos_.size());
   const WordID cur_fclass = pos_[smeta.GetSentenceID()][cur_src_pos];
   const int fid = fids_[smeta.GetSourceLength()].find(cur_fclass)->second.find(jumpsize)->second;
@@ -189,10 +195,13 @@ void MarkovJumpFClass::TraversalFeaturesImpl(const SentenceMetadata& smeta,
   }
 }
 
-//  std::vector<std::map<int, int> > flen2jump2fid_;
 MarkovJump::MarkovJump(const string& param) :
     FeatureFunction(1),
     fid_(FD::Convert("MarkovJump")),
+    fid_lex_null_(FD::Convert("JumpLexNull")),
+    fid_null_lex_(FD::Convert("JumpNullLex")),
+    fid_null_null_(FD::Convert("JumpNullNull")),
+    fid_lex_lex_(FD::Convert("JumpLexLex")),
     binary_params_(false) {
   cerr << "    MarkovJump";
   vector<string> argv;
@@ -218,7 +227,7 @@ MarkovJump::MarkovJump(const string& param) :
   cerr << endl;
 }
 
-// TODO handle NULLs according to Och 2000
+// TODO handle NULLs according to Och 2000?
 void MarkovJump::TraversalFeaturesImpl(const SentenceMetadata& smeta,
                                        const Hypergraph::Edge& edge,
                                        const vector<const void*>& ant_states,
@@ -229,19 +238,20 @@ void MarkovJump::TraversalFeaturesImpl(const SentenceMetadata& smeta,
   const int flen = smeta.GetSourceLength();
   if (edge.Arity() == 0) {
     dpstate = static_cast<unsigned int>(edge.i_);
-    if (edge.prev_i_ == 0) {
-      if (binary_params_) {
-        // NULL will be tricky
-        // TODO initial state distribution, not normal jumps
+    if (edge.prev_i_ == 0) {     // first word in sentence
+      if (edge.i_ >= 0 && binary_params_) {
         const int fid = flen2jump2fid_[flen].find(edge.i_ + 1)->second;
         features->set_value(fid, 1.0);
+      } else if (edge.i_ < 0 && binary_params_) {
+        // handled by bigram features
       }
     } else if (edge.prev_i_ == smeta.GetTargetLength() - 1) {
-        // NULL will be tricky
-      if (binary_params_) {
+      if (edge.i_ >= 0 && binary_params_) {
         int jumpsize = flen - edge.i_;
         const int fid = flen2jump2fid_[flen].find(jumpsize)->second;
         features->set_value(fid, 1.0);
+      } else if (edge.i_ < 0 && binary_params_) {
+        // handled by bigram features
       }
     }
   } else if (edge.Arity() == 1) {
@@ -253,13 +263,24 @@ void MarkovJump::TraversalFeaturesImpl(const SentenceMetadata& smeta,
       dpstate = static_cast<unsigned int>(left_index);
     else
       dpstate = static_cast<unsigned int>(right_index);
-    const int jumpsize = right_index - left_index;
+    if (left_index == kNULL_i || right_index == kNULL_i) {
+      if (left_index == kNULL_i && right_index == kNULL_i)
+        features->set_value(fid_null_null_, 1.0);
+      else if (left_index == kNULL_i)
+        features->set_value(fid_null_lex_, 1.0);
+      else
+        features->set_value(fid_lex_null_, 1.0);
 
-    if (binary_params_) {
-      const int fid = flen2jump2fid_[flen].find(jumpsize)->second;
-      features->set_value(fid, 1.0);
     } else {
-      features->set_value(fid_, fabs(jumpsize - 1));  // Blunsom and Cohn def
+      features->set_value(fid_lex_lex_, 1.0); // TODO should only use if NULL is enabled
+      const int jumpsize = right_index - left_index;
+
+      if (binary_params_) {
+        const int fid = flen2jump2fid_[flen].find(jumpsize)->second;
+        features->set_value(fid, 1.0);
+      } else {
+        features->set_value(fid_, fabs(jumpsize - 1));  // Blunsom and Cohn def
+      }
     }
   } else {
     assert(!"something really unexpected is happening");
@@ -294,15 +315,6 @@ void SourceBigram::FireFeature(WordID left,
     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,
@@ -386,8 +398,14 @@ void SourcePOSBigram::TraversalFeaturesImpl(const SentenceMetadata& smeta,
   if (arity == 0) {
     assert(smeta.GetSentenceID() < pos_.size());
     const vector<WordID>& pos_sent = pos_[smeta.GetSentenceID()];
-    assert(edge.i_ < pos_sent.size());
-    out_context = pos_sent[edge.i_];
+    if (edge.i_ >= 0) {  // non-NULL source
+      assert(edge.i_ < pos_sent.size());
+      out_context = pos_sent[edge.i_];
+    } else { // NULL source
+      // should assert that source is kNULL?
+      static const WordID kNULL = TD::Convert("<eps>");
+      out_context = kNULL;
+    }
     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
diff --git a/decoder/ff_wordalign.h b/decoder/ff_wordalign.h
index 30ddf7a1..0714229c 100644
--- a/decoder/ff_wordalign.h
+++ b/decoder/ff_wordalign.h
@@ -49,6 +49,11 @@ class MarkovJump : public FeatureFunction {
                                      void* out_context) const;
  private:
   const int fid_;
+  const int fid_lex_null_;
+  const int fid_null_lex_;
+  const int fid_null_null_;
+  const int fid_lex_lex_;
+
   bool binary_params_;
   std::vector<std::map<int, int> > flen2jump2fid_;
 };
@@ -96,7 +101,6 @@ class SourceBigram : public FeatureFunction {
                    WordID trg,
                    SparseVector<double>* features) const;
   mutable Class2Class2FID fmap_;
-  mutable Class2FID ufmap_;
 };
 
 class SourcePOSBigram : public FeatureFunction {