面向平行语料库和多层次语言特征的英语翻译系统研究

An English translation system for parallel corpus and multi-levellanguage characteristics

针对传统英语翻译系统对于平行语料库和多层次语言特征的提取准确率低、翻译效果不佳的问题,提出基于深度可分离卷积的英语神经机器翻译方法。此方法根据英语的语言特征,将英语切分为词、音节、字符、子词四种不同层次的语言粒度,以此降低英语低频词数量;然后通过深度可分离卷积对基于注意力机制的神经机器翻译模型进行改进,得到深度可分离卷积的英语神经机器翻译模型。实验结果表明,在对汉语~汉语翻译的切分结果中,本模型的在英汉翻译的切分语粒度BLEU分数均保持在21%及以上,均高于传统的CNN模型和Transformer机器翻译模型。且对平行语料和多层次语言特征进行测试发现,本模型的训练时间仅为16 h, CNN模型和Transformer机器翻译模型的训练时间分别为18 h和24 h,训练时长比本模型高出11%左右。由此可知,本模型可提升英语翻译系统计算效率,模型训练和学习能力明显增强,计算量减少,特征提取效果显著提升。

For the problem of low extraction accuracy and poor translation effect of parallel corpus and multi-level language features, a new English neural machine translation method based on deeply separable convolution is proposed. According to the language characteristics of English, English is divided into the language granularity of words, syllables, characters, and subwords to reduce the number of English low-frequency words;then improve the neural machine translation model based on attention mechanism through deeply separable convolution, and obtain the deeply separable convolution model. The experimental results show that the segmentation score of Chinese Chinese translation in BLEU scores is kept at 21% or above, which is higher than the traditional CNN model and Transformer machine translation model. Moreover, testing the parallel corpus and multi-level language features found that the training time of this model was only 16h, and the training time of CNN model and Transformer machine translation model was 18h and 24h, respectively, and the training time was about 11% higher than that of this model. Therefore, this model can improve the computing efficiency of the English translation system, significantly enhance the model training and learning ability, reduce the computation amount, and significantly improve the feature extraction effect.