DiHuTra: a Parallel Corpus to Analyse Differences between Human Translations

DiHuTra: a Parallel Corpus to Analyse Differences between Human Translations Ekaterina Lapshinova-Koltunski1, Maja Popovi´c2, Maarit Koponen3 Language Science and Technology1, ADAPT Centre2, Foreign Languages and Translation Studies3 Saarland University1, Dublin City University2, University of Eastern Finland3 e.lapshinova@mx.uni-saarland.de1, maja.popovic@adaptcentre.ie2, maarit.koponen@uef.fi3 Abstract The DiHuTra project aimed to design a corpus of parallel human translations of the same source texts by professionals and students. The resulting corpus consists of English news and reviews source texts, their translations into Russian and Croa- tian, and translations of the reviews into Finnish. The corpus will be valuable for both studying variation in translation and evaluating machine translation (MT) sys- tems. 1 Description Many studies have demonstrated that translated texts have different textual features than texts orig- inally written in the given language (originals). Furthermore, some studies have shown evidence of variation between human translations generated by different translators (Rubino et al., 2016; Popovi´c, 2020; Kunilovskaya and Lapshinova-Koltunski, 2020). Nevertheless, the number of such studies is still very small and limited to comparable cor- pora where different translators translated differ- ent source texts. Therefore, exact comparisons be- tween human translations are not possible. The DiHuTra project, formed by Saarland Uni- versity, ADAPT Centre and University of Eastern Finland in 2021–2022 has aimed to design a paral- lel corpus to address these issues. Each source text originally written in English has been translated into three target languages: Croatian, Russian and Finnish, by two groups of translators: profession- als and students. These parallel human translations c⃝ 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. will enable a better comparison of various text fea- tures as well as impact of automatic MT evaluation when used as references. 2 Data sets The source texts consist of two sub-sets of publicly available data sets from two distinct domains: Amazon product reviews1 contain unique product reviews from Amazon written in English with overall ratings from 1 to 5, 1 and 2 referring to negative, 3 to neutral and 4 and 5 to positive. We selected a balanced set of reviews from 14 cate- gories (e.g., “Sports and Outdoors”, “Books”, etc.) with an equal number of positive and negative re- views (14 from each of the 14 topics). In total, we included 196 reviews, containing 5.4 sentences and 93.2 words on average. News texts were imported from the WMT (2019 and 2020) shared task2 News test corpus. The top- ics vary between politics, sports, crime, health, etc. The news are longer than reviews, with 9.9 sen- tences and 221.7 words on average. The WMT shared tasks also contain a set of human transla- tions of the English source texts into several lan- guages including Russian, however, neither Croat- ian nor Finnish. We selected only texts which were originally written in English and had professional translations into Russian. In total, we included 68 news articles from different sources. 3 Translation process Each English review was translated into the three target languages, Croatian, Russian and …

Content truncated to 500 words