A Hybrid Word Alignment Model for Phrase-Based Statistical Machine Translation

摘要

This paper proposes a hybrid word alignment model for Phrase-Based Statistical Machine translation (PB-SMT). The proposed hybrid alignment model provides most informative alignment links which are offered by both un-supervised and semi-supervised word alignment models. Two unsupervised word alignment models (GIZA++ and Berkeley aligner) and a rule based aligner are combined together. The rule based aligner only aligns named entities (NEs) and chunks. The NEs are aligned through transliteration using a joint source-channel model. Chunks are aligned employing a bootstrapping approach by translating the source chunks into the target language using a baseline PB-SMT system and subsequently validating the target chunks using a fuzzy matching technique against the target corpus. All the experiments are carried out after single-tokenizing the multi-word NEs. Our best system provided significant improvements over the baseline as measured by BLEU.