Abstract
Over the last decade, an increasing amount of work has been done to advance the phrase-based statistical machine translation model in which the method of extracting phrase pairs consists of word alignment and phrase extraction. In this paper, we show that, for Japanese-English and Chinese-English statistical machine translation systems, this method is indeed missing potentially useful phrase pairs which could lead to better translation scores. These potentially useful phrase pairs can be detected by looking at the segmentation traces after decoding. We choose to see the problem of extracting potentially useful phrase pairs as a two-class classification problem: among all the possible phrase pairs, distinguish the useful ones from the not-useful ones. As for any classification problem, the question is to discover the relevant features which contribute the most. Extracting potentially useful phrase pairs resulted in a statistically significant improvement of 7.65 BLEU points in English-Chinese and 7.61 BLEU points in Chinese-English experiments. A slight increase of 0.94 BLEU points and 0.4 BLEU points is also observed for English-Japanese system and Japanese-English system, respectively.
