Alibaba Submission to the WMT20 Parallel Corpus Filtering Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 979–984 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 979 Alibaba Submission to the WMT20 Parallel Corpus Filtering Task Jun Lu, Xin Ge, Yangbin Shi, Yuqi Zhang Machine Intelligence Technology Lab, Alibaba Group Hangzhou, China {joelu.luj, shiyi.gx, taiwu.syb, chenwei.zyq}@alibaba-inc.com Abstract This paper describes the Alibaba Machine Translation Group submissions to the WMT 2020 Shared Task on Parallel Corpus Filter- ing and Alignment. In the filtering task, three main methods are applied to evaluate the qual- ity of the parallel corpus, i.e. a) Dual Bilin- gual GPT-2 model, b) Dual Conditional Cross- Entropy Model and c) IBM word alignment model. The scores of these models are com- bined by using a positive-unlabeled (PU) learn- ing model and a brute-force search to ob- tain additional gains. Besides, a few sim- ple but efficient rules are adopted to evalu- ate the quality and the diversity of the corpus. In the alignment-filtering task, the extraction pipeline of bilingual sentence pairs includes the following steps: bilingual lexicon mining, language identification, sentence segmentation and sentence alignment. The final result shows that, in both filtering and alignment tasks, our system significantly outperforms the LASER- based system. 1 Introduction The parallel corpus is an essential resource for building a high quality machine translation(MT) system. It has been shown that, the higher the cor- pus quality, the better the performance of a MT system(Koehn and Knowles, 2017; Khayrallah and Koehn, 2018). Many successful machine transla- tion systems are built on the corpus crawled from the web. In practice, this kind of parallel corpus may be very noisy. The task of Parallel Corpus Fil- tering is aimed at tackling the problem of cleaning noisy parallel corpora. We form the bilingual sentences quality in the following aspects. Firstly, a high-quality parallel sentence pair(also called bitext) should have the property that its target sentence precisely translates the source sentence, and vice versa. In this task, we attempt to quantify the translation accuracy (also Figure 1: Framework of parallel corpus filtering called bilingual score) of bilingual sentence pairs. Secondly, the monolingual quality of the target and source sentences of a parallel corpus should also be considered. In our system, we evaluate the mono- lingual quality (also called monolingual score) of a target sentence due to its importance for the MT procedure. Finally, the bilingual and monolingual scores are combined to evaluate bilingual sentence pairs and filter out the ones with low quality. The paper is structured as follows. Section 2 de- scribes our methods which are used in the parallel corpus filtering. In Section 3, we briefly outline the pipeline of parallel sentence extraction. Sec- tion 4 specifies the experiments and results as well as the dataset for building model-based methods. Conclusions are drawn in Section 5. 2 Parallel Corpus Filtering Methods Figure 1 shows the framework of parallel corpus filtering. The raw parallel corpus is firstly filtered by heuristic rules so that the very noisy sentence pairs will be removed. Then, …

Content truncated to 500 words