Investigating automatic and manual filtering methods to produce MT-ready glossaries from existing ones

Investigating automatic and manual filtering methods to produce MT-ready glossaries from existing ones Maria Afara AI R&D team, Acolad mafara@acolad.com Randy Scansani AI R&D team, Acolad rscansani@acolad.com Loïc Dugast AI R&D team, Acolad ldugast@acolad.com Abstract Commercial Machine Translation (MT) providers offer functionalities that allow users to leverage bilingual glossaries. This poses the question of how to turn glos- saries that were intended to be used by a human translator into MT-ready ones, re- moving entries that could harm the MT output. We present two automatic filtering approaches – one based on rules and the second one relying on a translation mem- ory – and a manual filtering procedure car- ried out by a linguist. The resulting glos- saries are added to an MT model. The outputs are compared against a baseline where no glossary is used and an output produced using the original glossary. The present work aims at investigating if any of these filtering methods can bring a higher terminology accuracy without negative ef- fects on the overall quality. Results are measured with terminology accuracy and Translation Edit Rate. We test our filters on two language pairs, En–Fr and De–En. Re- sults show that some of the automatically filtered glossaries may help reach a bet- ter balance between accuracy and overall quality, replacing the costly manual pro- cess. 1 Introduction The ability to correctly and consistently translate domain-specific or customer-specific terminology is key in the field of translation. To accommodate © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. for this need, Machine Translation (MT) providers have started to offer terminology features that en- force glossary entries at runtime.1 The availabil- ity of such features can be particularly advanta- geous for Language Service Providers (LSPs), giv- ing them an opportunity to offer a terminology ac- curate MT output in a scenario in which training a model from scratch is not an option. However, such glossaries were created to be used by human translators, relying on their ability to, e.g., disambiguate terms before inserting them in the target text. Also, glossaries are often created by customers without the help of terminologists. As a result, they might not be ready to be used by MT, since they might contain entries that harm the output quality (Bergmanis et al., 2021; Guerrero, 2020; Scansani and Dugast, 2021). The creation of a pipeline to clean glossaries can help MT users leverage their terminology data base. The pipeline can be based on a manual in- tervention, which can be time-consuming, or rely on an automatic procedure. Either way, two oper- ations may be involved, i.e. removing entries that are not helpful and/or editing them. Automatically editing entries is not a trivial task. For example, automatically editing term entries where several term alternatives are separated by slashes poses the question of which alternative(s) to keep. Also, in some cases the slash is used to separate parts of a compound or of a multi-word term (e.g. …

Content truncated to 500 words