Refining an Almost Clean Translation Memory Helps Machine Translation

Refining an Almost Clean Translation Memory Helps Machine Translation Shivendra Bhardwaj shivendra.bhardwaj@umontreal.ca David Alfonso-Hermelo david.alfonso.hermelo@gmail.com Philippe Langlais felipe@iro.umontreal.ca RALI/DIRO, Universit´e de Montr´eal, Montr´eal, QC H3C 3J7, Canada Gabriel Bernier-Colborne Gabriel.Bernier-Colborne@nrc-cnrc.gc.ca Cyril Goutte Cyril.Goutte@nrc-cnrc.gc.ca Michel Simard Michel.Simard@nrc-cnrc.gc.ca Multilingual Text Processing, National Research Council, Ottawa, ON K1A 0R6, Canada Abstract While recent studies have been dedicated to cleaning very noisy parallel corpora to improve Machine Translation training, in this work we focus on filtering a large and mostly clean Trans- lation Memory. This problem of practical interest has not received much consideration from the community, in contrast with, for example, filtering large web-mined parallel corpora. We experiment with an extensive, multi-domain proprietary Translation Memory and compare five approaches involving deep-, feature-, and heuristic-based solutions. We propose two ways of evaluating this task, manual annotation and resulting Machine Translation quality. We report significant gains over a state-of-the-art, off-the-shelf cleaning system, using two MT engines. 1 Introduction Major language service providers that handle huge numbers of translation requests in various domains typically call upon a large pool of translators (internal and freelance) whose transla- tions are fed into one or several internal Translation Memories (TM). Translators access these TMs through a dedicated interface to match requested translations with previously completed ones, speeding up the translation process. Given their provenance and purpose, it is normally assumed that TMs are mostly exempt of incorrect translations (henceforth “noise”). Unfortu- nately, because of the high number of translators involved, with varying levels of expertise, tight deadlines, and other technological issues, noise inevitably accumulates over the years. TM noise falls under two broad categories. First, mechanical noise, which occurs be- cause of the pipeline used to populate the TM, whereby text is extracted from source and target documents, segmented into sentences and then aligned — all three processes producing occa- sional errors. Second, human-induced noise, which can arise for a variety of reasons, including spelling or morphosyntax that do not meet established norms, missing translation units on the target side, as well as typical translation errors such as calques, the use of false friends, etc. These errors reduce the usefulness of a TM, and motivate our investigation into whether Paral- lel Corpus Filtering methods may be useful to increase the quality of a large TM. This contrasts with the typical use of these methods, which are more commonly applied to large, very noisy bilingual corpora automatically extracted from the Web (or other uncontrolled sources). In the following section (Sec. 2), we discuss related work in more details. We then describe Proceedings of the 15th Biennial Conference of the Association for Machine Translation in the Americas Orlando, USA, September 12-16, 2022. Volume 1: Research Track 215 the large TM on which our experiments are based in Section 3, and the corpus filtering methods we implemented in Section 4. We evaluate the performance of these methods by measuring their impact on Neural Machine Translation (NMT) in Section 5, and report our results and analysis in Section 6. We show that, surprisingly, …

Content truncated to 500 words