Terminology in Neural Machine Translation: A Case Study of the Canadian Hansard

Terminology in Neural Machine Translation: A Case Study of the Canadian Hansard Rebecca Knowles Samuel Larkin Marc Tessier Michel Simard National Research Council Canada 1200 Montreal Road, Ottawa, Ontario, CANADA K1A 0R6 FirstName.LastName@nrc.gc.ca Abstract Incorporating terminology into a neural machine translation (NMT) system is a feature of interest for many users of ma- chine translation. In this case study of English–French Canadian Parliamentary text, we examine the performance of stan- dard NMT systems at handling terminol- ogy and consider the tradeoffs between po- tential performance improvements and the efforts required to maintain terminological resources specifically for NMT. 1 Introduction Incorporating data from a specialized or particu- lar lexicon is a commonly-desired property of neu- ral machine translation (NMT) systems used in computer-aided translation settings. A number of approaches have been proposed for this, includ- ing modifications to decoding, training systems for special behavior, and training with external lexi- cons. Results vary, highlighting the fact that they navigate a difficult compromise between impos- ing specific lexical choices on the decoder, and interfering as little as possible with its behavior (Yvon and Rauf, 2020). Parallel to this, lexical resources developed by terminologists and trans- lators are not necessarily designed and formatted with NMT requirements in mind, and not all terms they contain naturally lend themselves to incorpo- ration: for example, it may be difficult to process terms with many morphological variants or terms whose translation depends on the context. Extract- ing these resources’ content for NMT and main- © 2023 His Majesty the King in Right of Canada. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC-BY-ND. taining the two resources in sync may pose practi- cal challenges. In light of this, it is reasonable to ask when, how, and whether it is worth implement- ing these methods in a real-life, practical setting. Here, we use the scenario of Canadian Parlia- mentary translation as a case study to examine questions about terminology and machine trans- lation performance. The data we use consists of transcriptions and translations of speech in the Canadian House of Commons (the proceedings, or Hansard), with most speech originally in English (then translated to French), a much smaller part spoken in French (then translated to English), and a very small fraction in other languages. Parlia- mentary translators have access to a document that provides guidance on terminology, from which we have manually extracted word and phrase pairs. We are interested in the following questions: • In this specific case, should we attempt to ex- plicitly handle terminology in our NMT sys- tems? If so, how? • More generally, in which scenarios does it make sense to incorporate terminology into an NMT system? What tradeoffs might re- searchers and users want to consider? With this data, we begin by examining just how “usable” the terminology actually is for NMT in- corporation, and how consistently it is used in hu- man translations. We then compare how an NMT system (without any special terminology handling) performs …

Content truncated to 500 words