featurizer

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@154 ec762483-ff6d-05da-a07a-a48fb63a330f

4 files changed, 700 insertions, 12 deletions

diff --git a/extools/Makefile.am b/extools/Makefile.am
index bce6c404..fc02f831 100644
--- a/extools/Makefile.am
+++ b/extools/Makefile.am
@@ -2,6 +2,8 @@ bin_PROGRAMS = \
   extractor \
   mr_stripe_rule_reduce \
   build_lexical_translation \
+  filter_grammar \
+  featurize_grammar \
   filter_score_grammar
 
 noinst_PROGRAMS =
@@ -10,6 +12,14 @@ filter_score_grammar_SOURCES = filter_score_grammar.cc extract.cc sentence_pair.
 filter_score_grammar_LDADD = $(top_srcdir)/decoder/libcdec.a -lz
 filter_score_grammar_LDFLAGS = -all-static
 
+filter_grammar_SOURCES = filter_grammar.cc extract.cc sentence_pair.cc
+filter_grammar_LDADD = $(top_srcdir)/decoder/libcdec.a -lz
+filter_grammar_LDFLAGS = -all-static
+
+featurize_grammar_SOURCES = featurize_grammar.cc extract.cc sentence_pair.cc
+featurize_grammar_LDADD = $(top_srcdir)/decoder/libcdec.a -lz
+featurize_grammar_LDFLAGS = -all-static
+
 build_lexical_translation_SOURCES = build_lexical_translation.cc extract.cc sentence_pair.cc
 build_lexical_translation_LDADD = $(top_srcdir)/decoder/libcdec.a -lz
 build_lexical_translation_LDFLAGS = -all-static
diff --git a/extools/featurize_grammar.cc b/extools/featurize_grammar.cc
new file mode 100644
index 00000000..1ca20a4b
--- /dev/null
+++ b/extools/featurize_grammar.cc
@@ -0,0 +1,410 @@
+/*
+ * Featurize a grammar in striped format
+ */
+#include <iostream>
+#include <string>
+#include <map>
+#include <vector>
+#include <utility>
+#include <cstdlib>
+#include <fstream>
+#include <tr1/unordered_map>
+
+#include "suffix_tree.h"
+#include "sparse_vector.h"
+#include "sentence_pair.h"
+#include "extract.h"
+#include "fdict.h"
+#include "tdict.h"
+#include "lex_trans_tbl.h"
+#include "filelib.h"
+
+#include <boost/shared_ptr.hpp>
+#include <boost/functional/hash.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+using namespace std;
+using namespace std::tr1;
+namespace po = boost::program_options;
+
+static const size_t MAX_LINE_LENGTH = 64000000;
+
+typedef unordered_map<vector<WordID>, RuleStatistics, boost::hash<vector<WordID> > > ID2RuleStatistics;
+
+void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
+  po::options_description opts("Configuration options");
+  opts.add_options()
+        ("filtered_grammar,g", po::value<string>(), "Grammar to add features to")
+        ("aligned_corpus,c", po::value<string>(), "Aligned corpus (single line format)")
+        ("help,h", "Print this help message and exit");
+  po::options_description clo("Command line options");
+  po::options_description dcmdline_options;
+  dcmdline_options.add(opts);
+  
+  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
+  po::notify(*conf);
+
+  if (conf->count("help") || conf->count("aligned_corpus")==0) {
+    cerr << "\nUsage: featurize_grammar -g FILTERED-GRAMMAR.gz -c ALIGNED_CORPUS.fr-en-al [-options] < UNFILTERED-GRAMMAR\n";
+    cerr << dcmdline_options << endl;
+    exit(1);
+  }
+}
+
+namespace {
+  inline bool IsWhitespace(char c) { return c == ' ' || c == '\t'; }
+  inline bool IsBracket(char c){return c == '[' || c == ']';}
+  inline void SkipWhitespace(const char* buf, int* ptr) {
+    while (buf[*ptr] && IsWhitespace(buf[*ptr])) { ++(*ptr); }
+  }
+}
+
+int ReadPhraseUntilDividerOrEnd(const char* buf, const int sstart, const int end, vector<WordID>* p) {
+  static const WordID kDIV = TD::Convert("|||");
+  int ptr = sstart;
+  while(ptr < end) {
+    while(ptr < end && IsWhitespace(buf[ptr])) { ++ptr; }
+    int start = ptr;
+    while(ptr < end && !IsWhitespace(buf[ptr])) { ++ptr; }
+    if (ptr == start) {cerr << "Warning! empty token.\n"; return ptr; }
+    const WordID w = TD::Convert(string(buf, start, ptr - start));
+
+    if((IsBracket(buf[start]) and IsBracket(buf[ptr-1])) or( w == kDIV))
+      p->push_back(1 * w);
+    else {
+      if (w == kDIV) return ptr;
+      p->push_back(w);
+    }
+  }
+  return ptr;
+}
+
+void ParseLine(const char* buf, vector<WordID>* cur_key, ID2RuleStatistics* counts) {
+  static const WordID kDIV = TD::Convert("|||");
+  counts->clear();
+  int ptr = 0;
+  while(buf[ptr] != 0 && buf[ptr] != '\t') { ++ptr; }
+  if (buf[ptr] != '\t') {
+    cerr << "Missing tab separator between key and value!\n INPUT=" << buf << endl;
+    exit(1);
+  }
+  cur_key->clear();
+  // key is: "[X] ||| word word word"
+  int tmpp = ReadPhraseUntilDividerOrEnd(buf, 0, ptr, cur_key);
+  cur_key->push_back(kDIV);
+  ReadPhraseUntilDividerOrEnd(buf, tmpp, ptr, cur_key);
+  ++ptr;
+  int start = ptr;
+  int end = ptr;
+  int state = 0; // 0=reading label, 1=reading count
+  vector<WordID> name;
+  while(buf[ptr] != 0) {
+    while(buf[ptr] != 0 && buf[ptr] != '|') { ++ptr; }
+    if (buf[ptr] == '|') {
+      ++ptr;
+      if (buf[ptr] == '|') {
+        ++ptr;
+        if (buf[ptr] == '|') {
+          ++ptr;
+          end = ptr - 3;
+          while (end > start && IsWhitespace(buf[end-1])) { --end; }
+          if (start == end) {
+            cerr << "Got empty token!\n  LINE=" << buf << endl;
+            exit(1);
+          }
+          switch (state) {
+            case 0: ++state; name.clear(); ReadPhraseUntilDividerOrEnd(buf, start, end, &name); break;
+            case 1: --state; (*counts)[name].ParseRuleStatistics(buf, start, end); break;
+            default: cerr << "Can't happen\n"; abort();
+          }
+          SkipWhitespace(buf, &ptr);
+          start = ptr;
+        }
+      }
+    }
+  }
+  end=ptr;
+  while (end > start && IsWhitespace(buf[end-1])) { --end; }
+  if (end > start) {
+    switch (state) {
+      case 0: ++state; name.clear(); ReadPhraseUntilDividerOrEnd(buf, start, end, &name); break;
+      case 1: --state; (*counts)[name].ParseRuleStatistics(buf, start, end); break;
+      default: cerr << "Can't happen\n"; abort();
+    }
+  }
+}
+
+
+void LexTranslationTable::createTTable(const char* buf){
+  AnnotatedParallelSentence sent;
+  sent.ParseInputLine(buf);
+      
+  //iterate over the alignment to compute aligned words
+  
+  for(int i =0;i<sent.aligned.width();i++)
+    {
+      for (int j=0;j<sent.aligned.height();j++)
+        {
+          if (DEBUG) cerr << sent.aligned(i,j) << " ";
+          if( sent.aligned(i,j))
+            {
+              if (DEBUG) cerr << TD::Convert(sent.f[i])  << " aligned to " << TD::Convert(sent.e[j]);
+              ++word_translation[pair<WordID,WordID> (sent.f[i], sent.e[j])];
+              ++total_foreign[sent.f[i]];
+              ++total_english[sent.e[j]];
+            }
+        }
+      if (DEBUG)  cerr << endl;
+    }
+  if (DEBUG) cerr << endl;
+  
+  const WordID NULL_ = TD::Convert("NULL");
+  //handle unaligned words - align them to null
+  for (int j =0; j < sent.e_len; j++) {
+    if (sent.e_aligned[j]) continue;
+    ++word_translation[pair<WordID,WordID> (NULL_, sent.e[j])];
+    ++total_foreign[NULL_];
+    ++total_english[sent.e[j]];
+  }
+  
+  for (int i =0; i < sent.f_len; i++) {
+    if (sent.f_aligned[i]) continue;
+    ++word_translation[pair<WordID,WordID> (sent.f[i], NULL_)];
+    ++total_english[NULL_];
+    ++total_foreign[sent.f[i]];
+  }
+}
+
+inline float safenlog(float v) {
+  if (v == 1.0f) return 0.0f;
+  float res = -log(v);
+  if (res > 100.0f) res = 100.0f;
+  return res;
+}
+
+static bool IsZero(float f) { return (f > 0.999 && f < 1.001); }
+
+struct FeatureExtractor {
+  // create any keys necessary
+  virtual void ObserveFilteredRule(const WordID lhs,
+                                   const vector<WordID>& src,
+                                   const vector<WordID>& trg) {}
+
+  // compute statistics over keys, the same lhs-src-trg tuple may be seen
+  // more than once
+  virtual void ObserveUnfilteredRule(const WordID lhs,
+                                     const vector<WordID>& src,
+                                     const vector<WordID>& trg,
+                                     const RuleStatistics& info) {}
+
+  // compute features, a unique lhs-src-trg tuple will be seen exactly once
+  virtual void ExtractFeatures(const WordID lhs,
+                               const vector<WordID>& src,
+                               const vector<WordID>& trg,
+                               const RuleStatistics& info,
+                               SparseVector<float>* result) const = 0;
+
+  virtual ~FeatureExtractor() {}
+};
+
+struct LogRuleCount : public FeatureExtractor {
+  LogRuleCount() :
+    fid_(FD::Convert("LogRuleCount")),
+    sfid_(FD::Convert("SingletonRule")),
+    kCFE(FD::Convert("CFE")) {}
+  virtual void ExtractFeatures(const WordID lhs,
+                               const vector<WordID>& src,
+                               const vector<WordID>& trg,
+                               const RuleStatistics& info,
+                               SparseVector<float>* result) const {
+    (void) lhs; (void) src; (void) trg;
+    result->set_value(fid_, log(info.counts.value(kCFE)));
+    if (IsZero(info.counts.value(kCFE)))
+      result->set_value(sfid_, 1);
+  }
+  const int fid_;
+  const int sfid_;
+  const int kCFE;
+};
+
+// this extracts the lexical translation prob features
+// in BOTH directions.
+struct LexProbExtractor : public FeatureExtractor {
+  LexProbExtractor(const std::string& corpus) :
+      e2f_(FD::Convert("LexE2F")), f2e_(FD::Convert("LexF2E")) {
+    ReadFile rf(corpus);
+    //create lexical translation table
+    cerr << "Computing lexical translation probabilities from " << corpus << "..." << endl;
+    char* buf = new char[MAX_LINE_LENGTH];
+    istream& alignment = *rf.stream();
+    while(alignment) {
+      alignment.getline(buf, MAX_LINE_LENGTH);
+      if (buf[0] == 0) continue;
+      table.createTTable(buf);              
+    }
+    delete[] buf;
+  }
+
+  virtual void ExtractFeatures(const WordID lhs,
+                               const vector<WordID>& src,
+                               const vector<WordID>& trg,
+                               const RuleStatistics& info,
+                               SparseVector<float>* result) const {
+    map <WordID, pair<int, float> > foreign_aligned;
+    map <WordID, pair<int, float> > english_aligned;
+
+    //Loop over all the alignment points to compute lexical translation probability
+    const vector< pair<short,short> >& al = info.aligns;
+    vector< pair<short,short> >::const_iterator ita;
+    for (ita = al.begin(); ita != al.end(); ++ita) {
+            if (DEBUG) {
+              cerr << "\nA:" << ita->first << "," << ita->second << "::";
+              cerr <<  TD::Convert(src[ita->first]) << "-" << TD::Convert(trg[ita->second]);
+            }
+
+            //Lookup this alignment probability in the table
+            int temp = table.word_translation[pair<WordID,WordID> (src[ita->first],trg[ita->second])];
+            float f2e=0, e2f=0;
+            if ( table.total_foreign[src[ita->first]] != 0)
+              f2e = (float) temp / table.total_foreign[src[ita->first]];
+            if ( table.total_english[trg[ita->second]] !=0 )
+              e2f = (float) temp / table.total_english[trg[ita->second]];
+            if (DEBUG) printf (" %d %E %E\n", temp, f2e, e2f);
+              
+            //local counts to keep track of which things haven't been aligned, to later compute their null alignment
+            if (foreign_aligned.count(src[ita->first])) {
+              foreign_aligned[ src[ita->first] ].first++;
+              foreign_aligned[ src[ita->first] ].second += e2f;
+            } else {
+              foreign_aligned[ src[ita->first] ] = pair<int,float> (1,e2f);
+            }
+  
+            if (english_aligned.count( trg[ ita->second] )) {
+               english_aligned[ trg[ ita->second] ].first++;
+               english_aligned[ trg[ ita->second] ].second += f2e;
+            } else {
+               english_aligned[ trg[ ita->second] ] = pair<int,float> (1,f2e);
+            }
+          }
+
+          float final_lex_f2e=1, final_lex_e2f=1;
+          static const WordID NULL_ = TD::Convert("NULL");
+
+          //compute lexical weight P(F|E) and include unaligned foreign words
+           for(int i=0;i<src.size(); i++) {
+               if (!table.total_foreign.count(src[i])) continue;      //if we dont have it in the translation table, we won't know its lexical weight
+               
+               if (foreign_aligned.count(src[i])) 
+                 {
+                   pair<int, float> temp_lex_prob = foreign_aligned[src[i]];
+                   final_lex_e2f *= temp_lex_prob.second / temp_lex_prob.first;
+                 }
+               else //dealing with null alignment
+                 {
+                   int temp_count = table.word_translation[pair<WordID,WordID> (src[i],NULL_)];
+                   float temp_e2f = (float) temp_count / table.total_english[NULL_];
+                   final_lex_e2f *= temp_e2f;
+                 }                              
+
+             }
+
+           //compute P(E|F) unaligned english words
+           for(int j=0; j< trg.size(); j++) {
+               if (!table.total_english.count(trg[j])) continue;
+               
+               if (english_aligned.count(trg[j]))
+                 {
+                   pair<int, float> temp_lex_prob = english_aligned[trg[j]];
+                   final_lex_f2e *= temp_lex_prob.second / temp_lex_prob.first;
+                 }
+               else //dealing with null
+                 {
+                   int temp_count = table.word_translation[pair<WordID,WordID> (NULL_,trg[j])];
+                   float temp_f2e = (float) temp_count / table.total_foreign[NULL_];
+                   final_lex_f2e *= temp_f2e;
+                 }
+           }
+     result->set_value(e2f_, safenlog(final_lex_e2f));
+     result->set_value(f2e_, safenlog(final_lex_f2e));
+  }
+  const int e2f_, f2e_;
+  mutable LexTranslationTable table;
+};
+
+int main(int argc, char** argv){
+  po::variables_map conf;
+  InitCommandLine(argc, argv, &conf);
+  ifstream alignment (conf["aligned_corpus"].as<string>().c_str());
+  ReadFile fg1(conf["filtered_grammar"].as<string>());
+
+  istream& fs1 = *fg1.stream();
+
+  // TODO make this list configurable
+  vector<boost::shared_ptr<FeatureExtractor> > extractors;
+  extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogRuleCount));
+  extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LexProbExtractor(conf["aligned_corpus"].as<string>())));
+
+  //score unscored grammar
+  cerr << "Reading filtered grammar to detect keys..." << endl;
+  char* buf = new char[MAX_LINE_LENGTH];
+
+  ID2RuleStatistics acc, cur_counts;
+  vector<WordID> key, cur_key,temp_key;
+  WordID lhs = 0;
+  vector<WordID> src;
+
+#if 0
+  int line = 0;
+  while(fs1) {
+    fs1.getline(buf, MAX_LINE_LENGTH);
+    if (buf[0] == 0) continue;
+    ParseLine(buf, &cur_key, &cur_counts);
+    src.resize(cur_key.size() - 2);
+    for (int i = 0; i < src.size(); ++i) src[i] = cur_key[i+2];
+    lhs = cur_key[0];
+    for (ID2RuleStatistics::const_iterator it = cur_counts.begin(); it != cur_counts.end(); ++it) {
+      for (int i = 0; i < extractors.size(); ++i)
+        extractors[i]->ObserveFilteredRule(lhs, src, it->first);
+    }
+  }
+
+  cerr << "Reading unfiltered grammar..." << endl;
+  while(cin) {
+    cin.getline(buf, MAX_LINE_LENGTH);
+    if (buf[0] == 0) continue;
+    ParseLine(buf, &cur_key, &cur_counts);
+    src.resize(cur_key.size() - 2);
+    for (int i = 0; i < src.size(); ++i) src[i] = cur_key[i+2];
+    lhs = cur_key[0];
+    for (ID2RuleStatistics::const_iterator it = cur_counts.begin(); it != cur_counts.end(); ++it) {
+      // TODO set lhs, src, trg
+      for (int i = 0; i < extractors.size(); ++i)
+        extractors[i]->ObserveUnfilteredRule(lhs, src, it->first, it->second);
+    }
+  }
+#endif
+
+  ReadFile fg2(conf["filtered_grammar"].as<string>());
+  istream& fs2 = *fg2.stream();
+  cerr << "Reading filtered grammar and adding features..." << endl;
+  while(fs2) {
+    fs2.getline(buf, MAX_LINE_LENGTH);
+    if (buf[0] == 0) continue;
+    ParseLine(buf, &cur_key, &cur_counts);
+    src.resize(cur_key.size() - 2);
+    for (int i = 0; i < src.size(); ++i) src[i] = cur_key[i+2];
+    lhs = cur_key[0];
+
+    //loop over all the Target side phrases that this source aligns to
+    for (ID2RuleStatistics::const_iterator it = cur_counts.begin(); it != cur_counts.end(); ++it) {
+      SparseVector<float> feats;
+      for (int i = 0; i < extractors.size(); ++i)
+        extractors[i]->ExtractFeatures(lhs, src, it->first, it->second, &feats);
+      cout << TD::GetString(cur_key) << " ||| " << TD::GetString(it->first) << " ||| ";
+      feats.Write(false, &cout);
+      cout << endl;
+    }
+  }
+}
+
diff --git a/extools/filter_grammar.cc b/extools/filter_grammar.cc
new file mode 100644
index 00000000..a2992f7d
--- /dev/null
+++ b/extools/filter_grammar.cc
@@ -0,0 +1,235 @@
+/*
+ * Filter a grammar in striped format
+ */
+#include <iostream>
+#include <string>
+#include <map>
+#include <vector>
+#include <utility>
+#include <cstdlib>
+#include <fstream>
+#include <tr1/unordered_map>
+
+#include "suffix_tree.h"
+#include "sparse_vector.h"
+#include "sentence_pair.h"
+#include "extract.h"
+#include "fdict.h"
+#include "tdict.h"
+#include "lex_trans_tbl.h"
+#include "filelib.h"
+
+#include <boost/shared_ptr.hpp>
+#include <boost/functional/hash.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+using namespace std;
+using namespace std::tr1;
+namespace po = boost::program_options;
+
+static const size_t MAX_LINE_LENGTH = 64000000;
+
+typedef unordered_map<vector<WordID>, RuleStatistics, boost::hash<vector<WordID> > > ID2RuleStatistics;
+
+void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
+  po::options_description opts("Configuration options");
+  opts.add_options()
+        ("test_set,t", po::value<string>(), "Filter for this test set")
+        ("top_e_given_f,n", po::value<size_t>()->default_value(30), "Keep top N rules, according to p(e|f). 0 for all")
+        ("help,h", "Print this help message and exit");
+  po::options_description clo("Command line options");
+  po::options_description dcmdline_options;
+  dcmdline_options.add(opts);
+  
+  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
+  po::notify(*conf);
+
+  if (conf->count("help") || conf->count("test_set")==0) {
+    cerr << "\nUsage: filter_grammar -t TEST-SET.fr [-options] < grammar\n";
+    cerr << dcmdline_options << endl;
+    exit(1);
+  }
+}   
+namespace {
+  inline bool IsWhitespace(char c) { return c == ' ' || c == '\t'; }
+  inline bool IsBracket(char c){return c == '[' || c == ']';}
+  inline void SkipWhitespace(const char* buf, int* ptr) {
+    while (buf[*ptr] && IsWhitespace(buf[*ptr])) { ++(*ptr); }
+  }
+}
+
+int ReadPhraseUntilDividerOrEnd(const char* buf, const int sstart, const int end, vector<WordID>* p) {
+  static const WordID kDIV = TD::Convert("|||");
+  int ptr = sstart;
+  while(ptr < end) {
+    while(ptr < end && IsWhitespace(buf[ptr])) { ++ptr; }
+    int start = ptr;
+    while(ptr < end && !IsWhitespace(buf[ptr])) { ++ptr; }
+    if (ptr == start) {cerr << "Warning! empty token.\n"; return ptr; }
+    const WordID w = TD::Convert(string(buf, start, ptr - start));
+
+    if((IsBracket(buf[start]) and IsBracket(buf[ptr-1])) or( w == kDIV))
+      p->push_back(1 * w);
+    else {
+      if (w == kDIV) return ptr;
+      p->push_back(w);
+    }
+  }
+  return ptr;
+}
+
+
+void ParseLine(const char* buf, vector<WordID>* cur_key, ID2RuleStatistics* counts) {
+  static const WordID kDIV = TD::Convert("|||");
+  counts->clear();
+  int ptr = 0;
+  while(buf[ptr] != 0 && buf[ptr] != '\t') { ++ptr; }
+  if (buf[ptr] != '\t') {
+    cerr << "Missing tab separator between key and value!\n INPUT=" << buf << endl;
+    exit(1);
+  }
+  cur_key->clear();
+  // key is: "[X] ||| word word word"
+  int tmpp = ReadPhraseUntilDividerOrEnd(buf, 0, ptr, cur_key);
+  cur_key->push_back(kDIV);
+  ReadPhraseUntilDividerOrEnd(buf, tmpp, ptr, cur_key);
+  ++ptr;
+  int start = ptr;
+  int end = ptr;
+  int state = 0; // 0=reading label, 1=reading count
+  vector<WordID> name;
+  while(buf[ptr] != 0) {
+    while(buf[ptr] != 0 && buf[ptr] != '|') { ++ptr; }
+    if (buf[ptr] == '|') {
+      ++ptr;
+      if (buf[ptr] == '|') {
+        ++ptr;
+        if (buf[ptr] == '|') {
+          ++ptr;
+          end = ptr - 3;
+          while (end > start && IsWhitespace(buf[end-1])) { --end; }
+          if (start == end) {
+            cerr << "Got empty token!\n  LINE=" << buf << endl;
+            exit(1);
+          }
+          switch (state) {
+            case 0: ++state; name.clear(); ReadPhraseUntilDividerOrEnd(buf, start, end, &name); break;
+            case 1: --state; (*counts)[name].ParseRuleStatistics(buf, start, end); break;
+            default: cerr << "Can't happen\n"; abort();
+          }
+          SkipWhitespace(buf, &ptr);
+          start = ptr;
+        }
+      }
+    }
+  }
+  end=ptr;
+  while (end > start && IsWhitespace(buf[end-1])) { --end; }
+  if (end > start) {
+    switch (state) {
+      case 0: ++state; name.clear(); ReadPhraseUntilDividerOrEnd(buf, start, end, &name); break;
+      case 1: --state; (*counts)[name].ParseRuleStatistics(buf, start, end); break;
+      default: cerr << "Can't happen\n"; abort();
+    }
+  }
+}
+
+
+struct SourceFilter {
+  // return true to keep the rule, otherwise false
+  virtual bool Matches(const vector<WordID>& key) const = 0;
+  virtual ~SourceFilter() {}
+};
+
+struct DumbSuffixTreeFilter : SourceFilter {
+  DumbSuffixTreeFilter(const string& corpus) :
+      kDIV(TD::Convert("|||")) {
+    cerr << "Build suffix tree from test set in " << corpus << endl;
+    assert(FileExists(corpus));
+    ReadFile rfts(corpus);
+    istream& testSet = *rfts.stream();
+    char* buf = new char[MAX_LINE_LENGTH];
+    AnnotatedParallelSentence sent;
+
+    /* process the data set to build suffix tree
+     */
+    while(!testSet.eof()) {
+      testSet.getline(buf, MAX_LINE_LENGTH);
+      if (buf[0] == 0) continue;
+
+      //hack to read in the test set using AnnotatedParallelSentence
+      strcat(buf," ||| fake ||| 0-0");
+      sent.ParseInputLine(buf);
+
+      //add each successive suffix to the tree
+      for(int i=0; i<sent.f_len; i++)
+        root.InsertPath(sent.f, i, sent.f_len - 1);
+    }
+    delete[] buf;
+  }
+  virtual bool Matches(const vector<WordID>& key) const {
+    const Node<int>* curnode = &root;
+    const int ks = key.size() - 1;
+    for(int i=0; i < ks; i++) {
+      const string& word = TD::Convert(key[i]);
+      if (key[i] == kDIV || (word[0] == '[' && word[word.size() - 1] == ']')) { // non-terminal
+        curnode = &root;
+      } else if (curnode) {
+        curnode = curnode->Extend(key[i]);
+        if (!curnode) return false;
+      }
+    }
+    return true;
+  }
+  const WordID kDIV;
+  Node<int> root;
+};
+
+int main(int argc, char** argv){
+  po::variables_map conf;
+  InitCommandLine(argc, argv, &conf);
+  const int max_options = conf["top_e_given_f"].as<size_t>();;
+  istream& unscored_grammar = cin;
+
+  cerr << "Loading test set " << conf["test_set"].as<string>() << "...\n";
+  boost::shared_ptr<SourceFilter> filter;
+  filter.reset(new DumbSuffixTreeFilter(conf["test_set"].as<string>()));
+
+  cerr << "Filtering...\n";
+  //score unscored grammar
+  char* buf = new char[MAX_LINE_LENGTH];
+
+  ID2RuleStatistics acc, cur_counts;
+  vector<WordID> key, cur_key,temp_key;
+  int line = 0;
+
+  multimap<float, ID2RuleStatistics::const_iterator> options; 
+  const int kCOUNT = FD::Convert("CFE");
+  while(!unscored_grammar.eof())
+    {
+      ++line;
+      options.clear();
+      unscored_grammar.getline(buf, MAX_LINE_LENGTH);
+      if (buf[0] == 0) continue;
+      ParseLine(buf, &cur_key, &cur_counts);
+      if (!filter || filter->Matches(cur_key)) {
+        // sort by counts
+        for (ID2RuleStatistics::const_iterator it = cur_counts.begin(); it != cur_counts.end(); ++it) {
+          options.insert(make_pair(-it->second.counts.value(kCOUNT), it));
+        }
+        int ocount = 0;
+        cout << TD::GetString(cur_key) << '\t';
+
+        bool first = true;
+        for (multimap<float,ID2RuleStatistics::const_iterator>::iterator it = options.begin(); it != options.end(); ++it) {
+          if (first) { first = false; } else { cout << " ||| "; }
+          cout << TD::GetString(it->second->first) << " ||| " << it->second->second;
+          ++ocount;
+          if (ocount == max_options) break;
+        }
+        cout << endl;
+      }
+    }
+}
+
diff --git a/extools/filter_score_grammar.cc b/extools/filter_score_grammar.cc
index f34b240d..fe9a2a07 100644
--- a/extools/filter_score_grammar.cc
+++ b/extools/filter_score_grammar.cc
@@ -37,7 +37,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   opts.add_options()
         ("test_set,t", po::value<string>(), "Filter for this test set (not specified = no filtering)")
         ("top_e_given_f,n", po::value<size_t>()->default_value(30), "Keep top N rules, according to p(e|f). 0 for all")
-        ("hiero_features", "Use 'Hiero' features")
+        ("backoff_features", "Extract backoff X-features, assumes E, F, EF counts")
 //        ("feature,f", po::value<vector<string> >()->composing(), "List of features to compute")
         ("aligned_corpus,c", po::value<string>(), "Aligned corpus (single line format)")
         ("help,h", "Print this help message and exit");
@@ -247,36 +247,66 @@ struct FeatureExtractor {
   const string extractor_name;
 };
 
+static bool IsZero(float f) { return (f > 0.999 && f < 1.001); }
+
 struct LogRuleCount : public FeatureExtractor {
   LogRuleCount() :
     FeatureExtractor("LogRuleCount"),
-    fid_(FD::Convert("LogRuleCount")), kCFE(FD::Convert("CFE")) {}
+    fid_(FD::Convert("LogRuleCount")),
+    sfid_(FD::Convert("SingletonRule")),
+    kCFE(FD::Convert("CFE")) {}
   virtual void ExtractFeatures(const vector<WordID>& lhs_src,
                                const vector<WordID>& trg,
                                const RuleStatistics& info,
                                SparseVector<float>* result) const {
     (void) lhs_src; (void) trg;
     result->set_value(fid_, log(info.counts.value(kCFE)));
+    if (IsZero(info.counts.value(kCFE)))
+      result->set_value(sfid_, 1);
   }
   const int fid_;
+  const int sfid_;
   const int kCFE;
 };
 
-struct SingletonRule : public FeatureExtractor {
-  SingletonRule() :
-    FeatureExtractor("SingletonRule"),
-    fid_(FD::Convert("SingletonRule")), kCFE(FD::Convert("CFE")) {}
+struct LogECount : public FeatureExtractor {
+  LogECount() :
+    FeatureExtractor("LogECount"),
+    sfid_(FD::Convert("SingletonE")),
+    fid_(FD::Convert("LogECount")), kCE(FD::Convert("CE")) {}
   virtual void ExtractFeatures(const vector<WordID>& lhs_src,
                                const vector<WordID>& trg,
                                const RuleStatistics& info,
                                SparseVector<float>* result) const {
     (void) lhs_src; (void) trg;
-    if (info.counts.value(kCFE) > 0.999 && info.counts.value(kCFE) < 1.001) {
-      result->set_value(fid_, 1.0);
-    }
+    assert(info.counts.value(kCE) > 0);
+    result->set_value(fid_, log(info.counts.value(kCE)));
+    if (IsZero(info.counts.value(kCE)))
+      result->set_value(sfid_, 1);
   }
+  const int sfid_;
   const int fid_;
-  const int kCFE;
+  const int kCE;
+};
+
+struct LogFCount : public FeatureExtractor {
+  LogFCount() :
+    FeatureExtractor("LogFCount"),
+    sfid_(FD::Convert("SingletonF")),
+    fid_(FD::Convert("LogFCount")), kCF(FD::Convert("CF")) {}
+  virtual void ExtractFeatures(const vector<WordID>& lhs_src,
+                               const vector<WordID>& trg,
+                               const RuleStatistics& info,
+                               SparseVector<float>* result) const {
+    (void) lhs_src; (void) trg;
+    assert(info.counts.value(kCF) > 0);
+    result->set_value(fid_, log(info.counts.value(kCF)));
+    if (IsZero(info.counts.value(kCF)))
+      result->set_value(sfid_, 1);
+  }
+  const int sfid_;
+  const int fid_;
+  const int kCF;
 };
 
 struct EGivenFExtractor : public FeatureExtractor {
@@ -437,13 +467,16 @@ int main(int argc, char** argv){
 
   // TODO make this list configurable
   vector<boost::shared_ptr<FeatureExtractor> > extractors;
-  if (conf.count("hiero_features")) {
+  if (conf.count("backoff_features")) {
+    extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogRuleCount));
+    extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogECount));
+    extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogFCount));
     extractors.push_back(boost::shared_ptr<FeatureExtractor>(new EGivenFExtractor));
     extractors.push_back(boost::shared_ptr<FeatureExtractor>(new FGivenEExtractor));
     extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LexProbExtractor(conf["aligned_corpus"].as<string>())));
   } else {
     extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogRuleCount));
-    extractors.push_back(boost::shared_ptr<FeatureExtractor>(new SingletonRule));
+    extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LogFCount));
     extractors.push_back(boost::shared_ptr<FeatureExtractor>(new LexProbExtractor(conf["aligned_corpus"].as<string>())));
   }