GoSt-ParC-Sign: Gold Standard Parallel Corpus of Sign and spoken language

GoSt-ParC-Sign Gold Standard Parallel Corpus of Sign and spoken language Mirella De Sisto∗, Vincent Vandeghinste†, Lien Soetemans‡, Caro Brosens§, Dimitar Shterionov∗ ∗Tilburg University, †Instituut voor de Nederlandse Taal, ‡KU Leuven, §Vlaams Gebarentaalcentrum m.desisto@tilburguniversity.edu, vincent@ccl.kuleuven.be, lien.soetemans@kuleuven.be, caro.brosens@vgtc.be, d.shterionov@tilburguniversity.edu 1 Introduction In the last decade, there has been an increasing in- terest in extending MT from only focusing on Spo- ken Languages (SpLs) to also targeting Sign Lan- guages (SLs); nevertheless, the advances of this field are still limited, and this is due to a number of reasons (e.g. challenges related to data avail- ability, lack of notation conventions, etc.). Besides the technological gap between SpLMT and SLMT, a severe difference lies in the avail- ability of high-quality (training) data. SpLMT can count on open and free datasets, such as Europarl (Koehn, 2005) and OPUS (Tiedemann and Ny- gaard, 2004), and on several MT platforms which allow training on specific datasets.1 The availabil- ity of sufficient amounts of high-quality (training) data drives the MT performance up. Furthermore, well-designed test sets allow to adequately assess quality and fairly compare MT systems. For SLs, instead, training data is scarce and scattered. Parallel datasets, with one side in a SL and the other in a SpL, are extremely lim- ited. In addition, most of the available datasets consist in broadcasts with subtitles/autocues as a written form of a SpL as the source and interpreta- tion into a SL as the target (Camgoz et al., 2018); this leads to various concerns related to their qual- ity: SL as the result of interpretation or transla- tion is heavily influenced by the source language2 as well as by the interpreting process; in addition, © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. 1See, for instance, Nematus com/EdinburghNLP/nematus),(https://github. OpenNMT (https://opennmt.net/), MarianMT (https: //marian-nmt.github.io/), 2This phenomenon is referred to as translationese (Graham et al., 2020) even though in some cases hearing interpreters are CODA’s (children of deaf adults), most often the interpretation is made by a hearing interpreter for whom the SL is the L2. In some cases, corpora with SL as source are available, such as the Corpus Vlaamse Gebarentaal3 (VGT) (Van Herreweghe et al., 2015) (Corpus of Flemish Sign Language); nevertheless, as annotation of the data is ongoing, the transla- tions available are too insufficient for quality (au- tomatic) SL translation (SLT). Additionally, as the data contain videos of the signer’s faces, strict GDPR rules apply, and signed informed consent forms are required from each of the signers. The SignON project4 aims to build SLT engines and hence gathers available SL data; throughout this process, we faced a number of issues,5 which led us to identify the need for a gold standard par- allel corpus of SL - SpL. The collection, organi- sation and (public) release of such a corpus, will provide a common ground for advancing the field of SLT. 2 Gost-Parc-Sign The goal of this project is to create a gold stan- dard parallel corpus …

Content truncated to 500 words