Assessing linguistically aware fuzzy matching in translation memories

Assessing linguistically aware fuzzy matching in translation memories Tom Vanallemeersch, Vincent Vandeghinste Centre for Computational Linguistics, University of Leuven Blijde Inkomststraat 13 B-3000 Leuven, Belgium {tom,vincent}@ccl.kuleuven.be Abstract The concept of fuzzy matching in trans- lation memories can take place using lin- guistically aware or unaware methods, or a combination of both. We designed a flexible and time-efficient framework which applies and combines linguistically unaware or aware metrics in the source and target language. We measure the correlation of fuzzy matching metric scores with the evaluation score of the suggested translation to find out how well the usefulness of a sugges- tion can be predicted, and we measure the difference in recall between fuzzy match- ing metrics by looking at the improve- ments in mean TER as the match score de- creases. We found that combinations of fuzzy matching metrics outperform single metrics and that the best-scoring combina- tion is a non-linear combination of the dif- ferent metrics we have tested. 1 Introduction Computer-aided translation (CAT) has become an essential aspect of translators’ working environ- ments. CAT tools speed up translation work, create more consistent translations, and reduce repetitive- ness of the translation work. One of the core com- ponents of a CAT tool is the translation memory system (TMS). It contains a database of already translated fragments, called the translation mem- ory (TM), which consists of translation units: seg- ments of a text together with their translation. c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. Given a sentence to be translated, the traditional TMS looks for source language sentences in a TM which are identical (exact matches) or highly sim- ilar (fuzzy matches), and, upon success, suggests the translation of the matching sentence to the translator (Sikes, 2007). Formally, a TM consists of a set of source sen- tences S1, ..., Sn and target sentences T1, ..., Tn, where (Si, Ti) form a translation unit. Let us call the sentence that we want to translate Q (the query sentence). The TMS checks whether Q already occurs in the TM, i.e. whether ∃Si ∈ S1, ..., Sn : Q = Si. If this is the case, Q needs no new translation and the translation Ti can be retrieved and used as the translation of Q. This is an exact match. If the TMS cannot find a perfect match, fuzzy matching is applied using some function Sim, which calcu- lates the best match Sb of Q in the TM, i.e. the most similar match, as in (1): Sb = max Si Sim(Q, Si) (1) If Sim(Q, Sb) >= θ (a predefined minimal threshold, which is typically 0.7 in CAT tools)1), Tb, the translation of Sb, is retrieved from the TM and provided as a suggestion for translating Q. If the threshold is not reached, the TMS assumes that Tb is of no value for the translator and does not provide it as a translation suggestion. Similarity calculation can be done …

Content truncated to 500 words