Since statistical machine translation (SMT) and translation memory (TM) complement each other in TM matched and unmatched regions, a unified framework for integrating TM into phrase-based SMT is proposed in this paper. Unlike previous two-stage pipeline approaches, which directly merge TM results into the input sentences and subsequently let the SMT only translates those unmatched regions, the proposed framework refers to the corresponding TM information associated with each phrase
at the SMT decoding. Under this unified framework, several integrated models are proposed to incorporate different types of information extracted from TM to guide the SMT decoding. We thus let SMT implicitly and indirectly utilize global context with a local dependency model. Furthermore, the SMT phrase table is dynamically enhanced with TM phrase pairs when the TM database and the SMT training set are different.
On a Chinese English TM database, our experiments show that the proposed Model-I significantly improves over both SMT
and TM when the SMT training set is also adopted as the TM database and when the fuzzy match score is over 0.4 (overall 3.5
BLEU points improvement and 2.6 TER points reduction). In addition, the proposed Model-II is significantly better than the TM
and the SMT systems when the SMT training set and the TM database are different. Furthermore, the proposed Model-III outperforms both the TM and the SMT systems even when the SMT training set and the TM database are from different domains. Additionally, the proposed Model-IV further achieves significant improvements with the help of Top-N TM sentence pairs. Lastly, all our models significantly outperform those state-of-the-art approaches under all test conditions.