Neural Fuzzy Repair: Integrating Fuzzy Matches into Neural Machine Translation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1800–1809 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1800 Neural Fuzzy Repair: Integrating Fuzzy Matches into Neural Machine Translation Bram Bult´e Centre for Computational Linguistics (CCL) KU Leuven bram.bulte@ccl.kuleuven.be Arda Tezcan Language and Translation Technology Team (LT3) Ghent University arda.tezcan@ugent.be Abstract We present a simple yet powerful data aug- mentation method for boosting Neural Ma- chine Translation (NMT) performance by leveraging information retrieved from a Trans- lation Memory (TM). We propose and test two methods for augmenting NMT training data with fuzzy TM matches. Tests on the DGT- TM data set for two language pairs show con- sistent and substantial improvements over a range of baseline systems. The results suggest that this method is promising for any transla- tion environment in which a sizeable TM is available and a certain amount of repetition across translations is to be expected, especially considering its ease of implementation. 1 Introduction Even though Machine Translation (MT) quality may have increased considerably over the past years, most notably with advances in the field of Neural Machine Translation (NMT), Translation Memories (TMs) still offer some advantages over MT systems. They are not only able to trans- late previously seen sentences ‘perfectly’ but they also offer ‘near perfect’ translation quality when highly similar source sentences are retrieved from the TM. As a result, in Computer-Assisted Trans- lation (CAT) workflows, the MT system is often used as a backoff mechanism when the TM fails to retrieve high fuzzy matches above a certain thresh- old (Rossi and Chevrot, 2019; Federico et al., 2012), even though it has been shown that this ba- sic integration method is not always the most op- timal TM-MT combination strategy (Simard and Isabelle, 2009). Our aim in this paper is to integrate the advan- tages of TMs into NMT systems in order to im- prove MT quality by utilizing existing translations for highly similar source sentences in a given TM. We propose a simple method for TM-NMT in- tegration that is based on augmenting the source data with retrieved fuzzy TM targets by means of concatenation. We train both dedicated Neu- ral Fuzzy Repair (NFR) systems that deal specif- ically with query sentences for which a (suffi- ciently high-scoring) match is found in the TM as well as unified systems capable of translat- ing any query sentence. Several configurations are tested on the DGT-TM data set (Steinberger et al., 2013) for the language directions English into Dutch (EN→NL) and English into Hungarian (EN→HU). In the next section, we provide an overview of previous research on TM-MT integration. Sec- tion 3 details the approach proposed in this paper. The experimental setup is presented in section 4, and the results in section 5. This is followed by the discussion (section 6) and conclusion (section 7). 2 Research background The idea to combine the advantages of TM and MT is certainly not new. Early TM-MT integra- tion approaches …

Content truncated to 500 words