Zero-Shot Translation Quality Estimation with Explicit Cross-Lingual Patterns

This paper describes our submission of the WMT 2020 Shared Task on Sentence\nLevel Direct Assessment, Quality Estimation (QE). In this study, we empirically\nreveal the \\textit{mismatching issue} when directly adopting BERTScore to QE.\nSpecifically, there exist lots of mismatching errors between the source\nsentence and translated candidate sentence with token pairwise similarity. In\nresponse to this issue, we propose to expose explicit cross-lingual patterns,\n\\textit{e.g.} word alignments and generation score, to our proposed zero-shot\nmodels. Experiments show that our proposed QE model with explicit cross-lingual\npatterns could alleviate the mismatching issue, thereby improving the\nperformance. Encouragingly, our zero-shot QE method could achieve comparable\nperformance with supervised QE method, and even outperforms the supervised\ncounterpart on 2 out of 6 directions. We expect our work could shed light on\nthe zero-shot QE model improvement.\n