基于Transformer的蒙汉神经机器翻译研究

针对传统循环神经网络和卷积神经网络的缺点,搭建完全基于多头自注意力机制的Transformer蒙汉神经机器翻译模型。实验结果表明,该模型比基于LSTM的蒙汉翻译模型提高了9个BLEU值左右。这说明Transformer翻译模型在句子语义提取和语义表达方面优于LSTM翻译模型。同时在语料预处理阶段,还对中蒙文语料进行了不同粒度的切分。通过实验对比分析,蒙文进行BPE处理后的翻译结果优于对中文单独使用分词处理的结果;在较小语料库中,对中文进行分字处理效果优于分词效果。

MONGOLIAN-CHINESE NEURAL MACHINE TRANSI ATION BASED ON TRANSFORMER

Aiming at disadvantages of traditional recurrent neural network and convolutional neural network,the Mongolian-Chinese machine translation model based on Transformer is constructed completely by multi-head self-attention mechanism.The experimental results show that the model improves about 9 BLEU values compared with the model based on LSTM.It indicates that the translation model based on Transformer is superior to the LSTM in terms of sentence semantic extraction and semantic expression.The Chinese and Mongolian corpora are segmented at different granularity in the pre-processing.Through the experimental comparison and analysis,the translation results of Mongolian after BPE processing are better than that of Chinese word segmentation.Moreover,in the small corpus,the effect of word segmentation on Chinese is better than that of word segmentation.