M3TRA: integrating TM and MT for professional translators

M3TRA: integrating TM and MT for professional translators Bram Bult´e CCL – KU Leuven Tom Vanallemeersch CCL – KU Leuven firstname.lastname@ccl.kuleuven.be Vincent Vandeghinste CCL – KU Leuven Abstract Translation memories (TM) and machine translation (MT) both are potentially use- ful resources for professional translators, but they are often still used independently in translation workflows. As translators tend to have a higher confidence in fuzzy matches than in MT, we investigate how to combine the benefits of TM retrieval with those of MT, by integrating the results of both. We develop a flexible TM-MT in- tegration approach based on various tech- niques combining the use of TM and MT, such as fuzzy repair, span pretranslation and exploiting multiple matches. Results for ten language pairs using the DGT-TM dataset indicate almost consistently better BLEU, METEOR and TER scores com- pared to the MT, TM and NMT baselines. 1 Introduction While software for professional translators has in- cluded translation memories (TMs) since several decades, especially in the context of specialized documents, the use of machine translation (MT) in such software is more recent. Even though certain commercial translation tools now offer function- alities such as automatic fuzzy match repair, TM and MT technologies are often still used indepen- dently, i.e. either a match for a query sentence or an MT output is provided. This is not ideal, as translators tend to have a higher confidence in ‘hu- man’ TM than in MT. It has to be kept in mind, however, that only exact matches provide a trans- c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. lation of the query sentence; ‘fuzzy’ matches of- fer a translation of a similar sentence. In contrast, MT systems provide a translation for any sentence, but they have problems with a number of, often linguistic, issues, such as complex morphologi- cal phenomena, long distance dependencies and word order (Bisazza and Federico, 2016; Sudoh et al., 2010). We investigate how to combine the confidence in fuzzy match retrieval with full sen- tence translation by integrating TM and MT out- put. We develop M3TRA,1 a method which per- forms a TM match preprocessing step before run- ning a standard phrase-based statistical MT (PB- SMT) system trained on the TM. M3TRA com- bines different approaches, and is flexible in sev- eral respects: it applies various fuzzy match score thresholds, allows for more than one match to be used per query sentence, and can use several fuzzy metrics. It comprises two main components: (a) fuzzy repair, automatically editing high-scoring fuzzy matches, and (b) span pretranslation, con- straining MT output by including certain consis- tently aligned spans of one or more TM matches. We perform tests on ten language pairs which involve multiple language families, using the DGT-TM dataset (Steinberger et al., 2013). We ap- ply PBSMT without span pretranslation as a base- line, as well as ‘pure’ TM and a standard NMT system, and evaluate the translations using several metrics. …

Content truncated to 500 words