ParaPat: The Multi-Million Sentences Parallel Corpus of Patents Abstracts

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3769–3774 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3769 ParaPat: The Multi-Million Sentences Parallel Corpus of Patents Abstracts Felipe Soares1,2, Mark Stevenson1, Diego Bartolome2 and Anna Zaretskaya2 (1) University of Sheffield, Computer Science Department, 211 Portobello Road, Sheffield, S1 4DP, UK (2) TransPerfect Translations, Passeig de Gr`acia, 11, Barcelona, Spain fs@felipesoares.net, mark.stevenson@sheffield.ac.uk, dbartolome@translations.com, azaretskaya@translations.com Abstract The Google Patents is one of the main important sources of patents information. A striking characteristic is that many of its abstracts are presented in more than one language, thus making it a potential source of parallel corpora. This article presents the development of a parallel corpus from the open access Google Patents dataset in 74 language pairs, comprising more than 68 million sentences and 800 million tokens. Sentences were automatically aligned using the Hunalign algorithm for the largest 22 language pairs, while the others were abstract (i.e. paragraph) aligned. We demonstrate the capabilities of our corpus by training Neural Machine Translation (NMT) models for the main 9 language pairs, with a total of 18 models. Our parallel corpus is freely available in TSV format and with a SQLite database, with complementary information regarding patent metadata. Keywords: Parallel Corpus, Patents, Machine Translation Models 1. Introduction The availability of parallel corpora is required by current Statistical and Neural Machine Translation systems (SMT and NMT). Acquiring a high-quality parallel corpus that is large enough to train MT systems, particularly NMT ones, is not a trivial task due to the need for correct alignment and, in many cases, human curation. In this context, the au- tomated creation of parallel corpora from freely available resources is extremely important in Natural Language Pro- cessing (NLP). Many parallel corpora are already available, some bilingually aligned (Tiedemann, 2012) and others multilingually aligned, with 3 or more languages (e.g. Eu- roparl from the European Parliament (Koehn, 2005), JRC- Acquis from the European Commission (Steinberger et al., 2006) and OpenSubtitles consisting of film subtitles (Zhang et al., 2014)). The extraction of parallel sentences from patents and sci- entific texts can be a valuable language resource for MT and other NLP tasks. The problem has been researched by several authors to support, for example, translation of biomedical articles (Wu et al., 2011; Neves et al., 2016) and named entity recognition of biomedical concepts (Kors et al., 2015). The most comprehensive dataset released to date for mul- tilingual patent analytics and machine translation is the COPPA parallel corpus developed by the World Intellectual Property Office (WIPO)1. Version 2 of the COPPA corpus contains almost 13 million parallel sentences in 8 language pairs. However, all language pairs are from/to English and highly imbalanced, with the English/French language pair consisting of around 10 million sentences, more than 80% of the entire corpus. In addition, this corpus is copyrighted, being free only for research purposes and requiring an ap- plication process in order to gain access. This work describes the …

Content truncated to 500 words