Neural Machine Translation Research Based On The Semantic Vector Of Tri-lingual Parallel Corpus

In recent years, Neural Machine Translation based on deep learning has gained researchers attention and achieved comparable even better results than traditonanl Statistical Machine Translation. Traditional Neural Machine Translation system are trained based on bilingual parallel corpus, sometimes we can get a lot of three or more languages of the parallel corpus, and we can train two or more end to end translation systems in the same time. The input terminals are parallel corpora in different languages and the output terminal is the only target language. Since different input terminals are different expressions of the same meaning, we assume that the intermediate vectors obtained from different input terminals have some relation. Our research is to achieve the purpose of improving the performance of Machine Translation by exploring the relationship between semanitic vectors of different source languages. Our research is based on tri-lingual parallel corpus, two languages will be served as the source languages and the remaining one will be served as the target languages. Our research includes the following aspects.(1) We comparatively analysis the translation performance from diffe rent source languages to the same target language. We train two end to end machine translation systems based on the same machine translaion method, one is from Chinese to Japanese and the other is from English to Japanese. Considering the experimental results and the principle of NMT, we get the necessity of Machine Translation research based on the semantic vector of the tri-lingual parallel corpus.(2) Machine Translation research based on splicing of vectors and Pivot Language. We divide this research into three parts. In the first part of the research, we get a new vector c which includes information of vector c? from source language1 and vector c? from source language2. We think the new vector c is the semantic representation of parallel source language1 and source lan guage2. However the system of source language1 to target language and the system of source language2 to target language are not independent, which means two pair of parameters are not independent. Not only in the training process but also in testing process we need to input source language1 and source language2 simultaneously. In the second part of the research, we hope to help improving the worse system with better system based on Pivot Language, in the condition that parameters are independent. In the third parf of the research, we combine the first part and the second part, we still get a new vector c which includes information of vector c? from source language1 and vector c? from source language2. Differently, once the model is trained, we only need to input one language to finish the translation process. In this way, we still hope to help improving the worse system with better system based on Pivot Language, in the condition that parameters are independent.(3) Machine Translation research based on similarity of vectors. We train the NMT system translating source language1 to target language and the NMT system translating source language2 to target language synchronously. As we know, the new vector in encoder-decoder is the semantic representation of source language. In our experiment, we translate source language1 to target language, meanwhile translate parallel source language2 to target language. As the source language1 and source language2 are parallel, they express the same semantics. We conjecture the fixed-length vector c? from source language1 and the fixed-length vector c? from source language2 are similar. We take it as a constraint condition to enhance semantic representation in the training process. In the third research, two systems are independent, so parameters are independent. Only in the training process we need to input source language1 and source language2 simultaneously. Once the model is trained, we could test the performance of the system from source language1 to target language by only inputting source language1 and test the performance of the system from source language2 to target language by only inputting source language2.