Syntactic Constraints on Phrase Extraction for Phrase-Based Machine Translation

A typical phrase-based machine translation (PBMT) system uses phrase pairs extracted from word-aligned parallel corpora. All phrase pairs that are consistent with word alignments are collected. The resulting phrase table is very large and includes many non-syntactic phrases which may not be necessary. We propose to filter the phrase table based on source language syntactic constraints. Rather than filter out all non-syntactic phrases, we only apply syntactic constraints when there is phrase segmentation ambiguity arising from unaligned words. Our method is very simple and yields a 24.38% phrase pair reduction and a 0.52 BLEU point improvement when compared to a baseline PBMT system with full-size tables.