Improving NMT Quality Using Terminology Injection

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4820–4827 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4820 Improving NMT Quality Using Terminology Injection Duane K. Dougal, Deryle W. Lonsdale Brigham Young University Provo, Utah, USA 84602 dkdougal@outlook.com, lonz@byu.edu Abstract Many organizations use domain- or organization-specific words and phrases. This paper explores the use of vetted terminology as an input to neural machine translation (NMT) for improved results: ensuring that the translation of individual terms is consistent with an approved multilingual terminology collection. We discuss, implement, and evaluate a method for injecting terminology and for evaluating terminology injection. Our use of the long short-term memory (LSTM) attention mechanism prevalent in state-of-the-art NMT systems involves attention vectors for correctly identifying semantic entities and aligning the tokens that represent them, both in the source and the target languages. Appropriate terminology is then injected into matching alignments during decoding. We also introduce a new translation metric more sensitive to approved terminological content in MT output. Keywords: neural machine translation, NMT, terminology, term injection, attention vector 1. Introduction Machine translation (MT) is a growing research area within natural language processing, a high-demand area of com- puter science. For many years, statistical MT (SMT) has dominated this research area and much progress has been made in MT using this approach and its variants, such as phrase-based MT. However, many believe, as quoted by a researcher in the field, that “the current approach of statis- tical, phrase-based MT has kind of reached the end of its natural life.”1 (Marking, 2016) With the recent resurgence of neural networks, advances in MT over the last two to three years are due to neural machine translation (NMT). Most organizations use an increasing number of domain- or organization-specific words and phrases. Any associated translation process, whether human or automated, must ac- curately and efficiently use these specific multilingual ter- minology collections. Hasler et al. (2018) make the point that enforcing “terminology...is a requisite, ...for compa- nies wanting to ensure that brand-related information is ren- dered correctly and consistently when translating...and is often more important than translation quality alone.” Cost savings are significant for organizations that manage and apply terminology (SDL, 2017), including in multilingual processes like translation. However, comparatively little has been done to explore the integration of vetted terminology and MT processing, with the goal of improving overall results. In particular, manip- ulating the translation output of NMT systems to adhere to user-provided terminology specifications, despite the im- pressive quality improvements of NMT, remains an open problem (Hasler et al., 2018). The now-common (and perhaps past its prime) SMT and the newer NMT both rely on training data and computed probabilities and weights to achieve their results. However, this does not always guarantee that the output results will contain the approved terminology, as is the case with Dinu et al. (2019). A sentence translated using MT may be gram- matically or linguistically correct, but if it does not reflect …

Content truncated to 500 words