DiHuTra: a Parallel Corpus to Analyse Differences between Human Translations

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1751–1760 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1751 DiHuTra: a Parallel Corpus to Analyse Differences between Human Translations Ekaterina Lapshinova-Koltunski1, Maja Popovi´c2, Maarit Koponen3 1Language Science and Technology, 2ADAPT Centre, 3Foreign Languages and Translation Studies 1Saarland University, 2Dublin City University, 3University of Eastern Finland 1e.lapshinova@mx.uni-saarland.de, 2maja.popovic@adaptcentre.ie, 3maarit.koponen@uef.fi Abstract This paper describes a new corpus of human translations which contains both professional and students translations. The data consists of English sources – texts from news and reviews – and their translations into Russian and Croatian, as well as of the subcorpus containing translations of the review texts into Finnish. All target languages represent mid-resourced and less or mid-investigated ones. The corpus will be valuable for studying variation in translation as it allows a direct comparison between human translations of the same source texts. The corpus will also be a valuable resource for evaluating machine translation systems. We believe that this resource will facilitate understanding and improvement of the quality issues in both human and machine translation. In the paper, we describe how the data was collected, provide information on translator groups and summarise the differences between the human translations at hand based on our preliminary results with shallow features. Keywords: translation, human translation, parallel corpus, multilingual corpus, multilinguality, Russian, Croatian, Finnish, translation variation, machine translation, news translation, review translation 1. Introduction In this paper, we describe a new parallel corpus con- taining professional and student translations of two dif- ferent text registers – news and reviews. This cor- pus has been compiled to address an important prob- lem affecting machine translation (MT), namely vari- ation in human translation. Besides being an essen- tial resource for MT evaluation, this corpus will also allow us to study important differences between pro- fessional and student translations across different lan- guages (Russian, Croatian and Finnish) and text reg- isters (news and reviews). We believe this kind of work is timely because there are studies showing ev- idence of variation between texts generated by differ- ent translators in terms of language patterns which may have an impact on MT evaluation too (Rubino et al., 2016; Popovi´c, 2020). The corpus is available from the CLARIN repository1, so it can be found though the Vir- tual Language Observatory2. Additionally, a GitHub repository3 contains the data and some additional in- formation. In the MT community, we observe a growing aware- ness of the phenomena related to human translations, i.e. features of translated texts that make them sta- tistically different from non-translated texts (Geller- stam, 1986; Baker, 1996; Toury, 1995). These fea- tures not only reflect the variation between text produc- tion types (translations and non-translations), but also help to detect the level of competence (professional vs. 1http://hdl.handle.net/21.11119/ 0000-000A-1BA9-A 2https://vlo.clarin.eu 3https://github.com/katjakaterina/ dihutra novice) in translations (Kunilovskaya and Lapshinova- Koltunski, 2020). At the same time, we still observe a widespread lack of deep understanding of these phenomena, especially in relation to …

Content truncated to 500 words