Research On Neural Machine Translation Incorporating Phrase Knowledge

With the rapid development of informatization and networking,the process of global integration has been accelerated dramatically and the need for practical language conversion technology has become increasingly urgent,therefore the machine translation technology attracted the interest of more and more researchers.At present,the most advanced machine translation technology is neural machine translation(NMT).Compared with the traditional statistical machine translation(SMT)methods,NMT directly uses neural networks to map source language text to target language text.It does not need manually designing features,and can make full use of the context information in sentences,so the translation performance was significantly improved.However,the mainstream NMT methods take words as its basic processing units,and their ability to model phrases is limited.Phrase-based SMT methods can automatically learn phrase translation knowledge between source language and target language from large-scale bilingual corpus and establish relevant probability model,which has certain advantages in phrase translation.How to combine the two kinds of method and make their advantages complementary is an important research issue.Focusing on the above issue,this paper studies the NMT technology which incorporates phrase knowledge.We proposed a suffix-based phrase knowledge representation method.First the source sentence is parsed to obtain its syntactic structure,and then the phrases in the sentence are identified.They are matched with the phrase table generated by the SMT system to acquire bilingual phrase segments.Such information is added to the end of the source sentence in the form of suffix to guide the phrase translation in NMT.The source sentence with suffix is input to the encoder,so that the phrase knowledge is integrated into the encoded sentence vector and improves the translation performance.Experimental results on Chinese-English corpora show that that the accuracy of machine translation can be improved by combining phrase knowledge with source sentences,which indicates that the integration of appropriate amount of phrase translation information in NMT encoder can effectively guide the generation of translation.