SciPar: A Collection of Parallel Corpora from Scientific Abstracts

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2652–2657 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2652 SciPar: A Collection of Parallel Corpora from Scientific Abstracts Dimitris Roussis, Vassilis Papavassiliou, Prokopis Prokopidis, Stelios Piperidis, Vassilis Katsouros Institute for Language and Speech Processing (ILSP) Athena Research and Innovation Center, Athens, Greece {dimitris.roussis, vpapa, prokopis, spip, vsk}@ athenarc.gr Abstract This paper presents SciPar, a new collection of parallel corpora created from openly available metadata of bachelor theses, master theses and doctoral dissertations hosted in institutional repositories, digital libraries of universities and national archives. We describe first how we harvested and processed metadata from 86, mainly European, repositories to extract bilingual titles and abstracts, and then how we mined high quality sentence pairs in a wide range of scientific areas and sub-disciplines. In total, the resource includes 9.17 million segment alignments in 31 language pairs and is publicly available via the ELRC-SHARE repository. The bilingual corpora in this collection could prove valuable in various applications, such as cross-lingual plagiarism detection or adapting Machine Translation systems for the translation of scientific texts and academic writing in general, especially for language pairs which include English. Keywords: Parallel Corpus, Machine Translation, Scientific Research, European Languages 1. Introduction The significant improvements in Machine Translation (MT) that have been achieved with the use of SMT (Statistical Machine Translation) and especially NMT (Neural Machine Translation) systems in recent years, have been mainly driven by the compilation of large-scale parallel corpora. Many of these corpora cover general language use, while domain-specific ones able to capture and model different word/term/phrase translations and styles of writing are rather scarce. Chu and Wang (2018) highlighted that there are few or no parallel corpora for the majority of language pairs and domains, although there has been significant work in data acquisition over the last years. In fact, the legal and biomedical domains are among the few cases research has focused on to propose new methods for training domain- adapted MT systems. The broad domain of scientific research and academic texts is one of those remaining relatively under-resourced, despite its importance. This importance is becoming increasingly apparent by recent empirical evidence that the research performance of a university declines as the linguistic distance of its local language from English increases, because English is the dominant language of scientific publications (Cao et al., 2021). Institutional repositories of universities, digital libraries and national archives contain a wealth of content and metadata from bachelor and master theses, doctoral dissertations and other scientific publications. The openly available abstracts and titles are quite often translated by their respective authors in at least 2 languages: English and their native language. To the best of our knowledge, these rich sources of parallel data have not been exploited systematically to create parallel corpora related to scientific research. The few notable exceptions concern high- resource language pairs (e.g. English-Spanish, English- Portuguese, English-French), as mentioned in Section 2. 1 https://opus.nlpl.eu/CAPES.php In this paper, we …

Content truncated to 500 words