Head Finalization Reordering for Chinese-to-Japanese Machine Translation

In Statistical Machine Translation, reordering rules have proved useful in extracting bilingual phrases and in decoding during translation between languages that are structurally different. Linguistically motivated rules have been incorporated into Chineseto-English (Wang et al., 2007) and Englishto-Japanese (Isozaki et al., 2010b) translation with significant gains to the statistical translation system. Here, we carry out a linguistic analysis of the Chinese-to-Japanese translation problem and propose one of the first reordering rules for this language pair. Experimental results show substantially improvements (from 20.70 to 23.17 BLEU) when head-finalization rules based on HPSG parses are used, and further gains (to 24.14 BLEU) were obtained using more refined rules.