Towards Precise Lexicon Integration in Neural Machine Translation

Proceedings of Recent Advances in Natural Language Processing, pages 1084–1095 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_122 1084 Towards Precise Lexicon Integration in Neural Machine Translation Ogün Öz Cobrainer GmbH Munich, Germany ogun.oz@cobrainer.com Maria Sukhareva Data Analytics Lab, Siemens AG Nuremberg, Germany maria.sukhareva@siemens.com Abstract Terminological consistency is an essential re- quirement for industrial translation. High- quality, hand-crafted terminologies contain en- tries in their nominal forms. Integrating such a terminology into machine translation is not a trivial task. The MT system must be able to disambiguate homographs on the source side and choose the correct wordform on the target side. In this work, we propose a simple but effective method for homograph disambigua- tion and a method of wordform selection by in- troducing multi-choice lexical constraints. We also propose a metric to measure the termino- logical consistency of the translation. Our re- sults have a significant improvement over the current SOTA in terms of terminological con- sistency without any loss of the BLEU score. All the code used in this work will be published as open-source. 1 Motivation The importance of consistent terminology has long been discussed by translation experts (Dagan and Church, 1994; Merkel, 1998; Itagaki et al., 2007; Saraireh, 2001; Byrne, 2006). Terminological stan- dardisation is a critical task for technical and non- technical industrial translation. Patents, technical manuals, and medical instructions rely on consis- tent usage of technical terminology. But also non- technical news releases, marketing texts, promotion materials, legal and financial documents need to adhere to the same terminology. Byrne (2006) cor- rectly points out that many large companies have their own terminologies that should be used in all texts. Such terminologies prescribe the correct us- age of terms and provide not only a list of words that are to be used but also a list of their synonyms that should not be used by writers and translators (so-called negative terms). Sukhareva et al. (2020) describe such terminology for an automotive com- pany and its usage in detail. Not adhering to these rules can be not only confusing for a reader but can also lead to serious legal and financial conse- quences if it is proven that damage was caused by the ambiguity of the instructions. Morphologically rich languages also pose a very practical problem for terminology integration: ter- minological entries are provided in their nominative singular form (Susanto et al., 2020). The SOTA approaches rely on the assumption that the termino- logical entry can be found as is in the translated text. This is not the case for Slavic languages (e.g. Rus- sian), for which finding the correct wordform on the target side is a key challenge for the terminology integration. Morphologically poor languages (e.g. English), on the contrary, pose a very different challenge. Ho- mographs appear in such languages not only due to polysemy and homonymy but also due to poor derivational morphology (e.g. a report vs. to re- port), thus, becoming a very common phenomenon. Liu et al. (2018) show that SOTA neural machine translation (NMT) fails …

Content truncated to 500 words