Findings of the WMT 2020 Shared Task on Parallel Corpus Filtering and Alignment

Proceedings of the 5th Conference on Machine Translation (WMT), pages 726–742 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 726 Findings of the WMT 2020 Shared Task on Parallel Corpus Filtering and Alignment Philipp Koehn∗†, Vishrav Chaudhary†, Ahmed El-Kishky† Naman Goyal†, Peng-Jen Chen†, Francisco Guzm´an† phi@jhu.edu, vishrav@fb.com, ahelk@fb.com naman@fb.com, pipibjc@fb.com, fguzman@fb.com ∗Johns Hopkins University, Baltimore, Maryland, United States †Facebook AI, Menlo Park, California, United States Abstract Following two preceding WMT Shared Tasks on Parallel Corpus Filtering (Koehn et al., 2018, 2019), we posed again the challenge of assigning sentence-level quality scores for very noisy corpora of sentence pairs crawled from the web, with the goal of sub-selecting the highest-quality data to be used to train ma- chine translation systems. This year, the task tackled the low resource condition of Pashto– English and Khmer–English and also included the challenge of sentence alignment from doc- ument pairs. 10 participants from companies, national research labs, and universities partici- pated in this task. 1 Introduction The field of Machine Translation has experienced significant advances in recent years thanks to im- provements in neural modeling (Bahdanau et al., 2015; Gehring et al., 2016; Vaswani et al., 2017), as well as the availability of large parallel corpora for training (Tiedemann, 2012; Smith et al., 2013; Bojar et al., 2017). Unfortunately, today’s neu- ral machine translation models, perform poorly on low-resource language pairs, for which clean, high-quality training data is lacking (Koehn and Knowles, 2017). Improving performance on low resource language pairs has high impact consider- ing that these languages are spoken by a large frac- tion of the world population. This is a particular challenge for industrial machine translation sys- tems that need to support hundreds of languages in order to provide adequate services to their mul- tilingual user base. While there have been advances in using mono- lingual corpora (Lample et al., 2018; Liu et al., 2020) and parallel corpora in multiple language pairs (Aharoni et al., 2019; Fan et al., 2020), the best training data for machine translation are still parallel corpora in the targeted language pair and domain. Parallel corpora are typically gathered from any available source without much guarantees about quality. This is especially the case for parallel corpora that are extracted from the web without much control over which web sites are mined. Since noisy training data has been recognized as a challenge for neural machine translation training (Khayrallah and Koehn, 2018), an essential step in using such data is filtering or discounting noisy sentence pairs. Recently, there is increased interest in the fil- tering of noisy parallel corpora to improve the data that can be used to train translation systems. The Shared Task on Parallel Corpus Filtering and Alignment at the Conference for Machine Trans- lation (WMT 2020) was organized to promote re- search to make learning from noisy data more vi- able for low-resource languages. It is similar to the previous year’s task but tackles different lan- guages (Pashto and Khmer instead of Nepali and Sinhala) and also …

Content truncated to 500 words