Splitting Long Input Sentences for Phrase-based Statistical Machine Translation

Translation results suffer when a standard phrasebased statistical machine translation system is used for translating long sentences. The translation output will not have the same word order as the source. When a sentence is long, it should be partitioned into several clauses, and the word reordering during the translation done within these clauses, not between the clauses. In this paper, we propose splitting the long sentences using linguistic information, and translating the sentence piece by piece. In other words, we constrain the word reordering so that it can only be done within the pieces but not between the pieces. We then apply a language model to join the pieces back together in the original sequence in order to reduce disfluencies in the connection. By doing so, word order can be preserved and translation quality improved. Our experiments on the patent translation from Japanese to English are able to achieve better translations measured by both BLEU score and word error rate (WER).