Discourse-Related Language Contrasts in English-Croatian Human and Machine Translation

Proceedings of the Third Conference on Machine Translation (WMT), Volume 1: Research Papers, pages 36–48 Belgium, Brussels, October 31 - Novermber 1, 2018. c⃝2018 Association for Computational Linguistics https://doi.org/10.18653/v1/W18-64005 Discourse-Related Language Contrasts in English-Croatian Human and Machine Translation Margita Šoštari´c†∗ †University of Zagreb Dept. of English margita.sostaric1@ffzg.hr Christian Hardmeier∗ Sara Stymne∗ ∗Uppsala University Dept. of Linguistics and Philology first.last@lingfil.uu.se Abstract We present an analysis of a number of coref- erence phenomena in English-Croatian human and machine translations. The aim is to shed light on the differences in the way these struc- turally different languages make use of dis- course information and provide insights for discourse-aware machine translation system development. The phenomena are automati- cally identified in parallel data using annota- tion produced by parsers and word alignment tools, enabling us to pinpoint patterns of inter- est in both languages. We make the analysis more fine-grained by including three corpora pertaining to three different registers. In a sec- ond step, we create a test set with the challeng- ing linguistic constructions and use it to evalu- ate the performance of three MT systems. We show that both SMT and NMT systems strug- gle with handling these discourse phenomena, even though NMT tends to perform somewhat better than SMT. By providing an overview of patterns frequently occurring in actual lan- guage use, as well as by pointing out the weak- nesses of current MT systems that commonly mistranslate them, we hope to contribute to the effort of resolving the issue of discourse phe- nomena in MT applications. 1 Introduction Every natural language has means of marking ele- ments belonging to the same coreference chain in order to achieve cohesion and coherence in run- ning text. These discourse phenomena are crucial for understanding texts and their misrepresenta- tion harms text intelligibility. Despite their sig- nificance, machine translation (MT) systems have been known to struggle with adequately transfer- ring coreference relations from the source to the target language, which is partly due to the great differences in the way languages express these re- lations. In our approach, we extend the framework introduced by Lapshinova-Koltunski and Hard- meier (2017), who identify discourse discrepan- cies in parallel data for English and German by predefining and automatically extracting discourse patterns of interest, and then utilize word align- ment information to determine which of the ex- tracted phenomena lack the equivalent counterpart in the other language. We use the same procedure to automatically extract phenomena, but extend the methodology to include cases where the phe- nomenon does have an equivalent construction in the other language, despite the alignment data sug- gesting that it is more frequently left unaligned. In this research, we perform an in-depth study of the way in which diverse discourse phenomena are handled in translation from English to Croa- tian. We investigate both human translation and the output of different types of MT systems. In the first step, we use the extended methodology of Lapshinova-Koltunski and Hardmeier (2017) to extract interesting diverging discourse patterns that …

Content truncated to 500 words