CCMT 2022 Translation Quality Estimation Task

Chang Su, Miaomiao Ma, Hao Yang, Shimin Tao, Jiaxin Guo, Minghan Wang, Min Zhang, Xiaosong Qiao

Abstract: This paper presents the method used by Huawei Translation Services Center (HW-TSC) in the quality estimation (QE) task: sentence-level post-editing effort estimation in the 18th China Conference on Machine Translation (CCMT) 2022. This method is based on a predictor-estimator model. The predictor is an XLM-RoBERTa model pre-trained on a large-scale parallel corpus and extracts features from the source language text and machine-translated text. The estimator is a fully connected layer that is used to regress the post-editing distance scores using the extracted features. In the experiment, it is found that pre-training the predictor with the semantic textual similarity (STS) task in the parallel corpus and using augmented training data constructed by different machine translation (MT) engines can improve the prediction effect of the Human-targeted Translation Edit Rate (HTER) in both Chinese-English and English-Chinese tasks.