A Bilingual Word Alignment Method of Chinese-English based on Recurrent Neural Network

Abstract

Word alignment is an important step in statistical machine translation. Chinese-English bilingual language has a large difference in language characteristics, which may lead to some inconsistent results in word alignment. In this paper, a word alignment method based on recurrent neural network (RNN) is proposed. Firstly, Chinese-English bilingual words are transformed into word embedding, which are input to RNN model and incorporate context information. RNN uses internal memory to process input sequences of arbitrary time series. The experimental results show that compared with DNN and IBM4 models, this method improves the accuracy of word alignment and the quality of machine translation.