Intelligent Translation Memory Matching and Retrieval with Sentence Encoders

Intelligent Translation Memory Matching and Retrieval with Sentence Encoders Tharindu Ranasinghe♦, Constantin Or˘asan♥ and Ruslan Mitkov♦ ♦Research Group in Computational Linguistics, University of Wolverhampton, UK ♥Centre for Translation Studies, University of Surrey, UK {t.d.ranasinghehettiarachchige, r.mitkov}@wlv.ac.uk c.orasan@surrey.ac.uk Abstract Matching and retrieving previously translated segments from a Translation Memory is the key functionality Translation Memories systems. Howeverin this matching and retrieving process is still limited to algorithms based on edit distance which we have identified as a major drawback in Translation Memories systems. In this paper we introduce sentence encoders to improve the matching and retrieving process in Translation Memories systems - an effective and efficient solution to replace edit distance based algorithms. 1 Introduction Translation Memories (TMs) are “structured archives of past translations“ which store pairs of corresponding text segments1 in source and target languages known as “translation units” (Simard, 2020). TMs are used during the translation process in order to reuse previously translated segments. The original idea of TMs was proposed more than forty years ago when (Arthern, 1979) noticed that the translators working for the European Commission were wasting valuable time by re- translating (parts of) texts that had already been translated before. He proposed the creation of a computerised storage of source and target texts which could easily improve the performance of translators and that could be part of a computer- based terminology system. Based on this idea, c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1Segments are typically sentences, but there are implementations which consider longer or shorter units. many commercial TM systems appeared on the market in the early 1990s. Since then the use of this particular technology has kept growing and recent studies show that it is used on regular basis by a large proportion of translators (Zaretskaya et al., 2018). Translation Memories systems help translators by continuously trying to provide them with so- called matches, which are translation proposals retrieved from its database. These matches are identified by comparing automatically the segment that has to be translated with all the segments stored in the database. There are three kinds of matches: exact, fuzzy and no matches. Exact matches are found if the segment to be translated is identical to one stored in the TM. Fuzzy matches are used in cases where it is possible to identify a segment which is similar enough to the one to be translated, and therefore, it is assumed that the translator will spend less time editing the translation retrieved from the database than translating the segment from scratch. No matches occur in cases where it is not possible to identify a fuzzy match (i.e. there is no segment similar enough to the one to be translated to be worth using its translation). TMs distinguish between fuzzy matches and no matches by calculating the similarity between segments using a similarity measure and comparing it to a threshold. Most of the existing TM systems rely on a variant of …

Content truncated to 500 words