Learning Phrase Boundaries for Hierarchical Phrase-based Translation

Hierarchical phrase-based models provide a powerful mechanism to capture non-local phrase reorderings for statistical machine translation (SMT). However, many phrase reorderings are arbitrary because the models are weak on determining phrase boundaries for patternmatching. This paper presents a novel approach to learn phrase boundaries directly from word-aligned corpus without using any syntactical information. We use phrase boundaries, which indicate the beginning/ending of phrase reordering, as soft constraints for decoding. Experimental results and analysis show that the approach yields significant improvements over the baseline on large-scale Chineseto-English translation.