Terminology-Constrained Neural Machine Translation at SAP

Terminology-Constrained Neural Machine Translation at SAP Miriam Exel Bianka Buschbeck Lauritz Brandt SAP SE Dietmar-Hopp-Allee 16, 69190 Walldorf Germany firstname.lastname@sap.com Simona Doneva∗ University of Mannheim 68131 Mannheim Germany sdoneva@mail.uni-mannheim.de Abstract This paper examines approaches to bias a neural machine translation model to adhere to terminology constraints in an industrial setup. In particular, we investigate varia- tions of the approach by Dinu et al. (2019), which uses inline annotation of the target terms in the source segment plus source factor embeddings during training and in- ference, and compare them to constrained decoding. We describe the challenges with respect to terminology in our usage sce- nario at SAP and show how far the investi- gated methods can help to overcome them. We extend the original study to a new lan- guage pair and provide an in-depth evalua- tion including an error classification and a human evaluation. 1 Introduction With over one billion words per year, SAP deals with a huge translation volume; covering prod- uct localization and translation of documentation, training materials or support instructions for up to 85 languages. With a wide range of prod- uct lines in different industries, translation settings are diverse. There are over 100 active transla- tion domains for which we maintain translation re- sources such as translation memories and termi- nologies. At SAP we usually train multi-domain neural machine translation (NMT) engines, whose input consists of a multitude of data sources includ- ing the contents of the company-internal transla- tion memories from various domains. The result- ∗Employed as a working student at SAP during this project. © 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. ing NMT system produces high-quality technical translations, but has difficulties generating appro- priate and coherent terminology in specific con- texts. Given the great importance of correct and consistent terminology in technical translation, this is a nuisance for the translators that work in a post- editing scenario as well as for users consuming ma- chine translation (MT) in a self-service scenario. In our translation environment, translators are assigned projects along with the relevant transla- tion domain’s terminology. To achieve term con- sistency, SAP maintains SAPterm1, a large termi- nology database which also specifies viable term translations. Translators can easily select target terms from SAPterm in a computer-assisted trans- lation (CAT) environment, but applying terminol- ogy constraints in NMT is a challenge. As we do not have reliable term recognition or morphologi- cal inflection generation tools for all our productive languages at our disposal, we require an approach that not only enforces the correct terminology but also learns its contextually appropriate inflections. To that end, we investigate the approach pre- sented in Dinu et al. (2019), which combines in- line annotation with source factors (Sennrich and Haddow, 2016), that provide an additional in- put stream with terminology annotation, to show how domain-specific terminology can be enforced in a multi-domain NMT model. The approach should be capable of handling unseen terminology while retaining …

Content truncated to 500 words