TransQuest: Translation Quality Estimation with Cross-lingual Transformers

Recent years have seen big advances in the field of sentence-level quality\nestimation (QE), largely as a result of using neural-based architectures.\nHowever, the majority of these methods work only on the language pair they are\ntrained on and need retraining for new language pairs. This process can prove\ndifficult from a technical point of view and is usually computationally\nexpensive. In this paper we propose a simple QE framework based on\ncross-lingual transformers, and we use it to implement and evaluate two\ndifferent neural architectures. Our evaluation shows that the proposed methods\nachieve state-of-the-art results outperforming current open-source quality\nestimation frameworks when trained on datasets from WMT. In addition, the\nframework proves very useful in transfer learning settings, especially when\ndealing with low-resourced languages, allowing us to obtain very competitive\nresults.\n