The University of Edinburgh-Uppsala University’s Submission to the WMT 2020 Chat Translation Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 473–478 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 473 The University of Edinburgh-Uppsala University’s Submission to the WMT 2020 Chat Translation Task Nikita Moghe1 Christian Hardmeier2 Rachel Bawden1 1School of Informatics, University of Edinburgh, Scotland 2Department of Linguistics and Philology, Uppsala University {nikita.moghe,rachel.bawden}@ed.ac.uk christian.hardmeier@lingfil.uu.se Abstract This paper describes the joint submission of the University of Edinburgh and Uppsala Uni- versity to the WMT’20 chat translation task for both language directions (English↔German). We use existing state-of-the-art machine trans- lation models trained on news data and fine-tune them on in-domain and pseudo-in- domain web crawled data. We also experiment with (i) adaptation using speaker and domain tags and (ii) using different types and amounts of preceding context. We observe that con- trarily to expectations, exploiting context de- grades the results (and on analysis the data is not highly contextual). However using do- main tags does improve scores according to the automatic evaluation. Our final primary systems use domain tags and are ensembles of 4 models, with noisy channel reranking of out- puts. Our en-de system was ranked second in the shared task while our de-en system outper- formed all the other systems.1 1 Introduction and challenges The task’s aim is to create machine translation (MT) systems to enable task-oriented communication be- tween a service agent and a customer speaking different languages (English and German respec- tively). Like most dialogues, the texts can show strong context sensitivities, as the customer and the agent engage in a common activity and continually react to each other’s utterances (Hardmeier, 2014; Bawden, 2018). However, the dialogues, which re- late to ordering or reserving products and services from a limited set of providers, also follow fairly strong scripts and are anchored in a small discourse universe defined by the products on offer. Their context sensitivity is therefore counterbalanced by domain-specific conventions and expectations. 1http://www.statmt.org/wmt20/ chat-task_results_DA.html Our design choices are informed by an initial manual inspection of the training data and a base- line translation, which revealed that the main chal- lenges relate to idiomaticity: incorrect or poor translation of English idioms, named entities and politeness markers (e.g. formal vs. informal forms of address, or poor translation of English sir) and an incorrect use of domain-specific terminology. Almost always, the problems were the result of an excessively literal translation of the source text, and this literalness also frequently affected the ref- erence translations themselves too. Surprisingly, we found few instances of phenomena explicitly requiring context to be correctly translated (e.g. we did not find pronominal anaphora to be a major problem in the dialogues examined).2 The context- dependent instances we did find were more task- specific (e.g. English Enjoy! should be translated differently depending on whether it is about a pizza (Guten Appetit!) or a film (Viel Spaß!)). We therefore focus on domain adaptation and general context modelling strategies. Our submis- sions are based on existing state-of-the-art MT sys- tems for news translation, which we fine-tune on …

Content truncated to 500 words