Statistical Phrase-Based Translation

4.5 Phrase Extraction Heuristic

1. Why do IBM-model alignments need symmetrization for phrase extraction, and what are the extremes?
   IBM models align each foreign word to at most one English word, so one direction can’t express many-to-many links. Aligning both directions gives two alignments: their intersection is high-precision, their union high-recall.
   • Koehn et al., 2003, Statistical Phrase-Based Translation (Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of ACL (HLT-NAACL 2003)), §4.5, pp. 52–53
2. Describe the base symmetrization heuristic (grow) of Och et al. (1999) used by Koehn et al.
   Start from the intersection; add only points from the union that connect at least one still unaligned word. First iteratively add points directly adjacent to existing ones until nothing changes, then in a final step add non-adjacent points under the same conditions. Variants: diag also allows diagonal neighbours; diag-and requires both words unaligned in the final step.
   • Koehn et al., 2003, Statistical Phrase-Based Translation (Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of ACL (HLT-NAACL 2003)), §4.5, pp. 52–53
3. How much does the choice of symmetrization heuristic matter?
   A lot: about 0.02 BLEU between the best and worst, and rankings change with corpus size. diag-and was best for almost all sizes; f2e alone started near the bottom but became competitive at 320k sentence pairs.
   • Koehn et al., 2003, Statistical Phrase-Based Translation (Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of ACL (HLT-NAACL 2003)), §4.5, pp. 52–53