added non-pruning intersection and a CRF tagger

- the linear-chain tagger is more of a proof of concept than a real tagger-- the context-free assumptions made in a number of places mean that the algorithms used may not be as efficient as they could be, but the model is as powerful as any CRF
- it would be easy to add latent variables or semi-CRF support (or both!)
- i've added a couple basic features that are often used for POS tagging
- non-pruning intersection is useful for lexical word alignment models and the tagger
- a sample POS tagger model will be committed later

1 files changed, 5 insertions, 0 deletions

diff --git a/decoder/trule.h b/decoder/trule.h
index d2b1babe..42edfa2c 100644
--- a/decoder/trule.h
+++ b/decoder/trule.h
@@ -39,6 +39,10 @@ class TRule {
   //    [LHS] ||| term1 [NT] term2 [OTHER_NT] [YET_ANOTHER_NT]
   static TRule* CreateRuleMonolingual(const std::string& rule);
 
+  static TRule* CreateLexicalRule(const WordID& src, const WordID& trg) {
+    return new TRule(src, trg);
+  }
+
   void ESubstitute(const std::vector<const std::vector<WordID>* >& var_values,
                    std::vector<WordID>* result) const {
     int vc = 0;
@@ -116,6 +120,7 @@ class TRule {
   short int prev_j;
 
  private:
+  TRule(const WordID& src, const WordID& trg) : e_(1, trg), f_(1, src), lhs_(), arity_(), prev_i(), prev_j() {}
   bool SanityCheck() const;
 };