Improve Neural Machine Translation by Syntax Tree

Abstract

Most of the neural machine translation methods are devoted to using syntactic information at one end of the Encoder-Decoder framework. They didn't use syntactic information at both ends, so that the syntactic information cannot be fully utilized to improve the translation effect. There are still errors due to insufficient grammar information. In order to solve this problem and make full use of syntactic information, we proposed a new tree-to-tree neural machine translation model. Syntax tree is added to the encoder as priori knowledge, and the bidirectional tree is used to obtain the information of the syntax tree, thereby generating a high-quality representation. At the same time, syntax structure is also added to the decoder to guide sentence generation. The experiment was carried out on Chinese-English language pairs, which proved to improve the effect of the neural machine translation.