“Hi, how can I help you?” Improving Machine Translation of Conversational Content in a Business Context

“Hi, how can I help you?” Improving Machine Translation of Conversational Content in a Business Context Bianka Buschbeck∗ Jennifer Mell∗ Miriam Exel Matthias Huck SAP SE Dietmar-Hopp-Allee 16, 69190 Walldorf, Germany firstname.lastname@sap.com Abstract This paper addresses the automatic trans- lation of conversational content in a busi- ness context, for example support chat dia- logues. While such use cases share charac- teristics with other informal machine trans- lation scenarios, translation requirements with respect to technical and business- related expressions are high. To succeed in such scenarios, we experimented with curating dedicated training and test data, injecting noise to improve robustness, and applying sentence weighting schemes to carefully manage the influence of the dif- ferent corpora. We show that our approach improves the performance of our models on conversational content for all 18 in- vestigated language pairs while preserv- ing translation quality on other domains – an indispensable requirement to integrate these developments into our MT engines at SAP. 1 Introduction At SAP we build machine translation systems to cope with a huge translation volume, cover- ing product localization and translation of docu- mentation, training materials or support instruc- tions for up to 85 languages. We usually train mixed-domain neural machine translation (MT) engines, whose training input consists of a mul- titude of data sources including the contents of the company-internal translation memories from vari- ous domains. The resulting MT systems produce ∗Equal contribution. © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. high-quality technical translations but have diffi- culties generating appropriate output for conversa- tional content, required for multilingual chatbots or product support chats. For an enhanced user ex- perience, it becomes more and more important that our customers can communicate in the language of their choice. Therefore, we started a project to im- prove translation quality for business-related con- versations. This includes better translations, but also more robustness towards typos and poor input quality. We focus on methods that allow for im- provements in one area without degrading quality in others, since we aim to deliver a single model per language pair. The remainder of this paper is structured as fol- lows. Section 2 introduces the particular chal- lenges to be addressed when training an MT sys- tem for corporate conversational content. The methods we explored are presented in Section 3 and experimental results on selected language pairs can be found in Section 4. Section 5 con- tains results on 18 language pairs of major interest at SAP using the final experimental configuration. We point out open research questions in Section 6. Related problems and approaches are highlighted in Section 7 before we summarize our findings in Section 8. 2 Challenges Our baseline MT system is built on large amounts of technical documents and UI-style training data with the consequence that it performs unsatisfac- torily in business conversations, where the content is technical, but style is conversational. An exam- ple conversation is given in Table 1 (column Orig- inal …

Content truncated to 500 words