Phrase-Boundary Translation Model Using Shallow Syntactic Labels

Phrase-boundary model for statistical machine translation labels the rules with classes of boundary words on the target side phrases of training corpus. In this paper, we extend the phrase-boundary model using shallow syntactic labels including POS tags and chunk labels. With the priority of chunk labels, the proposed model names non-terminals with shallow syntactic labels on the boundaries of the target side phrases. In comparison to the base phrase-boundary model, our variant uses phrase labels in addition to word classes. In other words, if there is no chunk label in one boundary, the labeler uses the word POS tag. The boundary labels are concatenated where there is no label for the whole target span. Using chunks as phrase labels, the proposed model generalizes the rules to decrease the model sparseness. The sparseness has more importance in the language pairs with a lot of differences in the word order because they have less number of aligned phrase pairs for extraction of rules. Compared with Syntax Augmented Machine Translation (SAMT) that labels rules with the syntax trees of the target side sentences, the proposed model does not need deep syntactic parsing. Thus, it is applicable even for low-resource languages having no syntactic parser. Some translation experiments are performed from Persian and German to English as the source and target languages with different word orders. In the experiments, our model achieved improvements of about 0.5 point of BLEU over a variant of SAMT.