A LSTM-Based Bidirectional Translation Model for Optimizing Rare Words and Terminologies

Neural translation model has greatly grown in recent years. Many researches have come up with very good solutions to deficiencies in Neural translation model. However, it is difficult to get best effect for rare words and terminologies what are marked as unknown words because of the limit of the dictionary's size. This paper presents a bidirectional translation model what can be used to translate between bilinguals and optimize rare words and terminologies. At first we use word2vec to get a word similarity model. By replacing the rare words to be trained and tested by similarity model, we solve the problems caused by rare words. In addition, all terminologies are treated as a rare word to join this model, so that there is a good performance in translating terminologies. Then, by introducing mutual learning in the symmetric LSTM, the translation accuracy between bilinguals has been improved. As experimental results show, this method achieves expected goal in effectiveness and accuracy.