Research On Domain Adaptation For Neural Machine Translation

Currently,neural machine translation(NMT)has become dominant in the community of machine translation due to its excellent performance.With the development of NMT,prevailing NMT models become more and more complex with large numbers of parameters,which often require large-scale,high-quality corpora for effective training.However,two difficulties are encountered in the practical applications of NMT.On the one hand,training an NMT model for a specific domain requires a large quantity of parallel sentence in such domain,which is often not readily available.On the other hand,the translated sentences often belong to multiple domains,thus requiring an NMT model general to different domains.To deal with this issue,many researchers have conducted studies on the domain adaptation for NMT.Domain adaptation aims to fully explore the effective translation information in existing,large-scale domain training corpora to improve the performance of low-resource domain neural machine translation models.Specifically,studies of domain adaptation can be classified into two general categories.One is to use the mixed-domain training corpus to construct a unified NMT model for all domains.The other is to transfer the rich-resource domain(out-of-domain)translation knowledge to benefit the low-resource(in-domain)NMT model.Here,we propose two approaches for two categories.1.For multi-domain neural machine translation,we propose an improved multidomain NMT model with domain context discrimination,which jointly models NMT and monolingual attention-based domain classifications.We distinguish and simultaneously model source-side domain-specific and domain-shared information,and strengthen the training loss of target-side domain-related words in the sentence.By this way,we can not only exploit domain-shared translation information,but also retain the unique text style and syntax of each domain.2.For N-to-one domain transfer,we propose a novel iterative dual domain adaptation framework for NMT.Under this framework,out-of-domain NMT model and in-domain NMT model iteratively perform bi-directional translation knowledge transfer(from in-domain to out-of-domain and then vice versa).In this way,both indomain and out-of-domain NMT models are expected to constantly reinforce each other,which is likely to achieve better NMT domain adaptation.We have compared these two approaches with the contrast models which widely used in domain adaptation for NMT,and demonstrated the effectiveness of the proposed approaches through detailed experimental analysis.