NJU’s submission to the WMT20 QE Shared Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 1004–1009 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 1004 NJU’s submission to the WMT20 QE Shared Task Qu Cui and Xiang Geng and Shujian Huang∗and Jiajun Chen National Key Laboratory for Novel Software Technology, Nanjing University {cuiq,gx}@smail.nju.edu.cn, {huangsj,chenjj}@nju.edu.cn Abstract This paper describes our system of the sentence-level and word-level Quality Estima- tion Shared Task of WMT20. Our system is based on the QE Brain, and we simply enhance it by injecting noise at the target side. And to obtain the deep bi-directional information, we use a masked language model at the target side instead of two single directional decoders. Meanwhile, we try to use the extra QE data from the WMT17 and WMT19 to improve our system’s performance. Finally, we ensemble the features or the results from different mod- els to get our best results. Our system finished fifth in the end at sentence-level on both EN- ZH and EN-DE language pairs. 1 Introduction Quality Estimation (QE) is a task to predict the quality of translations without relying on any ref- erences. QE plays a critical role in machine trans- lation to reduce human efforts, such as deciding whether a translation is good enough for post- editing and indicating what edits are needed. This paper describes our system of the Shared Task on Word and Sentence-Level (QE Tasks 2) at WMT20. With the post-edited translations, all the quality scores can be computed automatically by TER- COM (Snover et al., 2006). Traditional QE models (Kozlova et al., 2016) use some time-consuming and expensive hand-craft features to represent the translation pairs. With the great success of deep neural networks in natural language processing (NLP), some researches have begun to apply automatic neural features to do QE tasks (Chen et al., 2017; Shah et al., 2016). How- ever, the rare QE data can’t fully release the power of deep neural networks. To address this prob- lem, researchers try to transfer bilingual knowledge ∗ Corresponding Author. from parallel data to QE tasks (Fan et al., 2018). These works usually follow a predictor-estimator framework (Kim et al., 2017). This framework first trains the predictor to predict each token of the target sentence given the source and the context of the target sentence on parallel data. Then, the estimator is trained using the features of QE data produced by the predictor. However, existing predictor-estimator frame- works cannot fully use the information from paral- lel data because of the discrepancy of data quality between the predictor and the estimator. The pre- dictor is trained on parallel data, which are nearly no errors in translations. While the translations in QE data is generated by a real machine translation system and may have some errors. When the es- timator is training on the QE data, the predictor needs to extract the features of translations with some errors, which is quite different from the par- allel data. Thus, the predictor can’t extract features well. To fix this problem, …

Content truncated to 500 words