Tencent Submissions for the CCMT 2020 Quality Estimation Task

This paper presents our submissions to CCMT 2020 Quality Estimation (QE) sentence-level task for both Chinese-to-English (ZH-EN) and English-to-Chinese (EN-ZH). We propose new methods based on the predictor-estimator architecture. For the predictor, we propose XLM-predictor and Transformer-predictor. XLM-predictor novelly produces two kinds of contextual token representation, i.e., maskXLM and non-mask-XLM. For the estimator, both RNN-estimator and Transformer-estimator are conducted and two novel strategies, i.e. top-K strategy and multi-head attention strategy, are proposed to enhance the sentence feature representation. We also propose new effective ensemble technique for sentence-level predictions.