SEDAR: a Large Scale French-English Financial Domain Parallel Corpus

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3595–3602 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3595 SEDAR: a Large Scale French-English Financial Domain Parallel Corpus Abbas Ghaddar, Philippe Langlais RALI-DIRO Montreal, Canada abbas.ghaddar@umontreal.ca, felipe@iro.umontreal.ca Abstract This paper describes the acquisition, preprocessing and characteristics of SEDAR, a large scale English-French parallel corpus for the financial domain. Our extensive experiments on machine translation show that SEDAR is essential to obtain good performance on finance. We observe a large gain in performance of machine translation systems trained on SEDAR when tested on finance, which makes SEDAR suitable to study domain adaptation for neural machine translation. The first release of the corpus comprises 8.6 million high quality sentence pairs that is publicly available for research at https://github.com/autorite/sedar-bitext. Keywords: Parallel Corpus, Financial Data, Machine Translation 1. Introduction Neural machine translation (NMT) (Cho et al., 2014; Bah- danau et al., 2014; Vaswani et al., 2017; Edunov et al., 2018) has become the most popular approach in recent years for machine translation. These models require the use of large scale parallel corpora to train millions of inter- nal parameters (Ott et al., 2018). The quantity and quality of training data is crucial for systems performances, as well as train/test domain matching. Koehn and Knowles (2017) show that NMT performs poorly in two scenarios: lack of large amount of training data, and out-of-domain transla- tion (domain mismatch). Unfortunately, large parallel corpora are only available for a limited set of domains such as news and political dis- courses. Several approaches for NMT domain adaption have been proposed 1, but these techniques require a small amount of parallel data or large quantity of monolingual data. This can be problematic for domains such as finance, where even monolingual data is scarce or non-existent due to the commercial value and privacy issue of such data. The most commonly known and well studied English- French bilingual data is the train portion provided within the WMT’14 shared task (Bojar et al., 2014). It con- tains 40.8M sentence pairs extracted from five datasets that cover various domains: EUROPARL V7 (Koehn, 2005), UNITED NATIONS CORPUS (Eisele and Chen, 2010), COMMON CRAWL CORPUS, NEWS COMMENTARY, and 109 FRENCH-ENGLISH corpus. Lison and Tiedemann (2016) present OPENSUBTITLES, a parallel corpus of 2.6 billion sentences across 60 languages originally mined from movies and Tv episode subtitles. The English-French portion of OPUS (Tiedemann, 2012) corpus contains a small collection (ECB) of parallel data in finance domain, that is collected from documents published by the European Central Bank. PARACRAWL2 is an ongoing project aiming to collect par- allel data from the web for all 24 official European Union Work done while the first author was interning at Autorit´e des march´es financiers (Qu´ebec). 1See (Chu and Wang, 2018) for a review 2https://paracrawl.eu languages. The dataset is extremely large and noisy, therefore there have been endeavors to filter high quality pairs (Koehn et al., 2018). For example, even if the English- French …

Content truncated to 500 words