Dynamic Adaptation of Neural Machine-Translation Systems Through Translation Exemplars

Dynamic Adaptation of Neural Machine-Translation Systems Through Translation Exemplars Arda Tezcan LT3, Language and Translation Technology Team Ghent University Belgium Arda.Tezcan@ugent.be Abstract This project aims to study the impact of adapting neural machine translation (NMT) systems through similar transla- tions retrieved from translation memories, determine the optimal metric(s) for mea- suring similarity, and, verify the usefulness of this approach for domain adaptation of NMT systems. 1 Introduction Translation exemplars, i.e. previously observed or generated translations, whose source is similar to new input, play an important role in the translation process by providing explicit information on con- text, exceptions, and irregularities, which are diffi- cult to generalise. In the context of machine trans- lation (MT), the information provided by similar translations can also be considered complemen- tary to the neural models, which are good at mak- ing generalisations. “Dynamic Adaptation of Neu- ral Machine-Translation Systems Through Trans- lation Exemplars” is a three-year, post-doctoral re- search project (October 2020 – September 2023), funded by the Research Foundation – Flanders (FWO) with the following scientific objectives: (i) study the impact of adapting NMT systems through similar translations retrieved from TMs, (ii) determine the optimal metrics for measuring translation similarity, and (iii) verify the useful- ness of this approach for domain adaptation. All research activities are carried out at the Depart- ment of Translation, Interpreting and Communica- tion, Ghent University, Belgium. © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. The project has been built upon the method- ology used in the translation memory (TM) — NMT integration approach, neural fuzzy repair (NFR) (Bult´e and Tezcan, 2019).1 In NFR, each source sentence in the TM is augmented (con- catenated), with the translation of the most sim- ilar fuzzy match (FM), retrieved from the same TM using edit distance, when at least one FM is found above the minimum similarity threshold of λ. The augmented TM is merged with the origi- nal TM to train an NMT model. During inference, the same technique is applied to augment source sentences prior to obtaining translations from the trained model. If no FMs above the minimum sim- ilarity threshold can be found, the translations are obtained for the original source sentence, using the same model. Tests on multiple language pairs showed that this method results in substantial gains in trans- lation quality compared to (i) baseline MT sys- tems, (ii) the FMs themselves when used as fi- nal output, even when they correspond to near- perfect translations in high similarity ranges, and (iii) a ‘fuzzy match repair’ approach, which relies on editing highly similar FMs to arrive at the fi- nal translation (Bult´e et al., 2018). These exper- iments also demonstrated that using the NFR sys- tem starts being advantageous with a minimum FM score of λ = 0.5, and augmenting source sen- tences with FMs of higher similarity scores leads to higher translation quality. While this study showed the usefulness of adapting NMT systems through similar …

Content truncated to 500 words