基于深度可分离卷积的汉越神经机器翻译

(昆明理工大学信息工程与自动化学院,云南省人工智能重点实验室,云南 昆明 650500)

汉越神经机器翻译; 数据稀疏; 粒度; 深度可分离卷积

Chinese-Vietnamese neural machine translation based on deep separable convolution
XU Yu,LAI Hua*,YU Zhengtao,GAO Shengxiang,WEN Yonghua

(Yunnan Key Laboratory of Artificial Intelligence,College of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

DOI: 10.6043/j.issn.0438-0479.201908038

备注

在汉越神经机器翻译中,由于汉越平行语料稀少,使得数据稀疏问题十分严重,极大地影响了模型的翻译效果.为了提升数据稀疏情况下的汉越神经机器翻译性能,提出一种基于深度可分离卷积的汉越神经机器翻译方法.该方法根据越南语的语言特点,将越南语切分为词、音节、字符、子词4种不同的粒度并利用深度可分离卷积改进神经机器翻译模型,通过增加深度可分离卷积神经网络,对模型输入的不同粒度序列进行卷积运算,提取更多的特征数据,相比传统卷积降低了模型的理论计算量.实验结果表明,该方法在越南语4种不同翻译粒度上均取得最佳效果,一定程度上提升了汉越神经机器翻译性能.

In the Chinese-Vietnamese neural machine translation,the lack of Chinese-Vietnamese parallel corpus has created a serious problem of sparse data,which greatly affects the translation performance of the model.In order to improve the performance of Chinese-Vietnamese neural machine translation with sparse date,a deep separable convolution based Chinese-Vietnamese neural machine translation method is proposed.First,according to the linguistic characteristics of Vietnamese,the method divides Vietnamese into four different granularities:words,syllables,characters and subwords.Second,the neural machine translation model is improved by using the depth separable convolution.By adding depth separable convolution in traditional neural networks,more features can thus be extracted with convolution operation performed on sequence inputs of different granularities.Compared with the traditional convolution,the improved method reduces the theoretical calculation amount.Finally, experimental results show that the proposed method achieves considerably in four different translation granularities of Vietnamese,improving the neural machine translation ability in Chinese-Vietnamese to a certain extent.