Rule-based Reordering Constraints for Phrase-based SMT

Rule-based Reordering Constraints for Phrase-based SMT Chooi-Ling Goh and Takashi Onishi and Eiichiro Sumita Multilingual Translation Laboratory National Institute of Information and Communications Technology 619-0289 Kyoto, Japan {chooiling.goh,takashi.onishi,eiichiro.sumita}@nict.go.jp Abstract Translation results suffer when a standard phrase-based statistical machine transla- tion system is used for translating long sen- tences. The translation output will not pre- serve the same word order as the source, especially between a language pair that has different syntactic structures. When a sentence is long, it should be partitioned into several clauses, and the word reorder- ing during the translation should be done within clauses, not between clauses. In this paper, we propose splitting the long sentences using linguistic information, and translating the sentences with the split boundaries. In other words, we constrain the word reordering so that it cannot cross the boundaries. We propose two types of constraints: split condition and block con- dition. By doing so, word order can be pre- served and translation quality improved. Our experiments on the patent translation between Japanese and English are able to achieve better translations measured by BLEU score, NIST score and word error rate (WER). 1 Introduction Translating long and complex sentences has been a critical problem in machine translation. A stan- dard phrase-based statistical machine translation (SMT) system cannot solve the problem of word reordering in the target when the source sentence has a complex structure. A syntax-based machine translation system could solve the problem by run- ning a parser on the source sentence in order to get the syntactic structure, but when a sentence is long and complex, the parser may fail to give a correct parse tree. Klein and Manning (2003) have shown c⃝ 2011 European Association for Machine Translation. that the accuracy of parsing decreases as sentence length increases, and the parsing time increases ex- ponentially. However, in this research, we found that even when a sentence is long and complex, it is possible to split a sentence into smaller units which can be translated separately with minor con- sideration of the context. The main problem here is locating the best locations for the split. We use linguistic information such part-of-speech (POS) tags and commas as clues to determine the split positions. After splitting a sentence into small clauses, the clauses are translated almost inde- pendently. This means that word reordering can only be done within a clause, not between clauses. This constraint can be specified using “wall” tag in Moses1 (Koehn and Haddow, 2009). Furthermore, a long sentence may include some long and com- plex noun phrases. These noun phrases should also be translated individually regardless of the context. We try to locate the area of these noun phrases by looking at the sequences of nouns in the sen- tences. We then bracket these noun phrases using the “zone” constraint in Moses. We use the NTCIR-8 (Fujii et al., 2010) Patent Translation shared task data between Japanese and English in our experiment. The sentences in the patent are always long and have complex …

Content truncated to 500 words