Structural Phrase Alignment Based on Consistency Criteria

In this paper, we propose a new method for phrase alignment using a dependency type distance and a distance-score function. With this method, appropriate correspondences can be selected among correspondence candidates that often include ambiguous or incorrect ones. Furthermore, this method makes it possible to measure the overall alignment consistency. We conduct an alignment experiment using 500 parallel sentences on newspaper domain, and achieve an F-measure improvement of 35 points over the simple statistical method (GIZA++), and 3.0 points over a baseline system. We also conducted a translation experiment and achieved a BLEU score improvement of 0.4 points over a baseline system.