融合XLM词语表示的神经机器译文自动评价方法

Neural Automatic Evaluation of Machine Translation Method Combinedwith XLM Word Representation

机器译文自动评价对机器翻译的发展和应用起着重要的促进作用,其一般通过计算机器译文和人工参考译文的相似度来度量机器译文的质量。该文通过跨语种预训练语言模型XLM将源语言句子、机器译文和人工参考译文映射到相同的语义空间,结合分层注意力和内部注意力提取源语言句子与机器译文、机器译文与人工参考译文以及源语言句子与人工参考译文之间的差异特征,并将其融入基于Bi-LSTM神经译文自动评价方法中。在WMT 19译文自动评价数据集上的实验结果表明,融合XLM词语表示的神经机器译文自动评价方法显著提高了其与人工评价的相关性。

The automatic evaluation of machine translation plays an important role in promoting the development and application of machine translation.It generally measures the quality of machine translation through calculating the similarity between machine translation and its reference.This paper uses the cross-lingual language model XLM to map source sentences,machine translations and references to the same semantic space,and combines layer-wise attention and intra attention to extract the difference features from source sentences and machine translations,machine translations and its references,source sentences and its references,then integrates them into the automatic evaluation method based on Bi-LSTM neural network.The experimental results on the dataset of WMT 19 Metrics task show that the neural automatic evaluation method of machine translation combined with XLM word representation significantly improves its correlation with human judgments.