Phrase extraction based on constraints of word similarities

Aimed at the problem that the traditional phrase extraction method is strictly dependent on word alignments, and is not pruned to alignment errors, a loose phrase extraction method, which does not strictly depend on word alignments. In this method, constraints are posed on alignment points to avoid ill-formed phrase pairs. Three constraint strategies are proposed based on word similarities: Dice coefficient, Phi-square coefficient and log-likelihood ratio. Experiments were carried out on the corpus of IWSLT 2004. Results show that the BLEU scores of the best results of loose phrase extraction can be improved by 15.14%, compared with the baseline system Pharaoh.