Research on Neural Network Machine Translation Model Based on Entity Tagging Improvement

Abstract

Machine translation, as an efficient tool, can achieve equivalent conversion between different languages while preserving the original semantics. At present, machine translation models based on deep neural networks have become a hot research topic in the fields of natural language processing and image processing. However, the randomness of neural networks leads to the existing neural network machine translation models unable to effectively reflect the linguistic dependencies and having unsatisfactory results when dealing with long sentence sequences. To solve these two problems, a new neural network machine translation model with entity tagging improvement is proposed. First, for the low-frequency word translation problem, UNK entity tags replacement is used to compensate for the weakness of the randomness of neural networks and the encoding/decoding strategy of entity tagging is improved. Then, on the basis of the LSTM translation model, an attention mechanism is introduced to dynamically adjust the degree of influence of the context at the source language end on the target language sequence to improve the feature learning ability of the translation model in processing long sentences. The analysis of the experimental results shows that the translation evaluation index BLEU of the proposed translation model is significantly improved compared with various translation models, which verifies its effectiveness.