Using any machine translation source for fuzzy-match repair in a computer-aided translation setting

Using any machine translation source for fuzzy-match repair in a computer-aided translation setting John E. Ortega jeo10@alu.ua.es Felipe S´anchez-Mart´ınez fsanchez@dlsi.ua.es Mikel L. Forcada mlf@dlsi.ua.es Dept. de Llenguatges i Sistemes Inform`atics, Universitat d’Alacant, E-03071, Alacant, Spain Abstract When a computer-assisted translation (CAT) tool does not find an exact match for the source segment to translate in its translation memory (TM), translators must use fuzzy matches that come from translation units in the translation memory that do not completely match the source segment. We explore the use of a fuzzy-match repair technique called patching to repair trans- lation proposals from a TM in a CAT environment using any available machine translation system, or any external bilingual source, regardless of its internals. Patching attempts to aid CAT tool users by repairing fuzzy matches and proposing improved translations. Our results show that patching improves the quality of translation proposals and reduces the amount of edit operations to perform, especially when a specific set of restrictions is applied. 1 Introduction Computer-aided translation (CAT) tools based on translation memories (TM) are one of the most popular technologies among professional translators (Bowker, 2002; Somers, 2003). CAT tools exploit existing, segment-aligned translations to help the translator translate a new docu- ment by recycling as much target-language (TL) text as possible. To do so, CAT tools first split the source-language (SL) document to translate into segments, and for each SL segment s′ they look up the translation memory for segment pairs (s, t) (called translation units) where t is the translation of s and s is similar to s′. These translation units are then shown to the translator in decreasing order of similarity, together with an indication of the words in s that do not match those in s′. Finally, the translator decides which translation unit to use and which parts of its TL segment t have to be edited to produce t′, the desired translation of s′, and performs such editions. The similarity between s and s′ is computed by means of a fuzzy-match score (FMS) func- tion whose output is between 0% (no match at all) and 100% (a perfect match, s = s′). Com- mercial CAT systems implement proprietary versions of FMS, but a reasonable approximation is given by: FMS(s, s′) = � 1 − ED(s, s′) max(|s|, |s′|) � · 100% (1) where ED(s, s′) is the (word-based) edit distance (Wagner and Fischer, 1974) between s and s′ —the minimum number of one-word deletions, insertions and substitutions needed to transform s into s′— and |x| stands for the number of words in segment x. Many times translation tools Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 42 use a fuzzy-match score threshold (FMT), for instance 80%, to reduce the number of translation proposals. When a perfect match is not found in the translation memory, and before making any changes to the TL segment t in the proposed translation unit, the translator has to identify the sub-segments …

Content truncated to 500 words