【摘要】 神经机器翻译现已成为翻译技术主流，在诸多大语种上已取得了极佳的进展，但对于老挝语这种低资源语言的机器翻译技术尚存在欠缺。该文使用弱监督的方法将汉老双语映射在同一向量空间中，减小了因跨语言而带来的嵌入差异，接着使用反向翻译策略缓解了老挝语低资源语言问题，运用汉老平行句对筛选策略得到了扩充的高质量平行句对。在此基础上，对Transformer模型进行改进，实现了编码器-解码器层间的信息增强，使模型翻译性能得到提升，同时引入平均注意力机制，使模型解码速度得到提升。实验表明，该文模型在老汉和汉老翻译任务中BLEU值较基线系统提升了1.36和1.22。更多还原

【Abstract】 Neural machine translation has become the mainstream of translation technology and has made great progress in many major languages. However, there is still a lack of machine translation technology for low resource languages such as Lao. This paper adopts a weakly supervised learning method to map Chinese and Lao languages into the same vector space, which reduces the embedding differences between the two languages. Then, the back translation strategy is used to alleviate the lack of corpus in Lao, and the parallel sentence pair filtering strategy is used to obtain high-quality parallel sentence pairs. On this basis, the Transformer model is improved to enhance the information transfer between the encoder and decoder layers. At the same time, the average attention network is used to improve the decoding speed. Experiments show that the model achieves BLEU score improvements of 1.36 and 1.22 over the baseline system in the translation tasks of Lao-Chinese and Chinese-Lao.更多还原