Using word alignments to assist computer-aided translation users by marking which target-side words to change or keep unedited

Using word alignments to assist computer-aided translation users by marking which target-side words to change or keep unedited Miquel Espl`a and Felipe S´anchez-Mart´ınez and Mikel L. Forcada Departament de Llenguatges i Sistemes Inform`atics Universitat d’Alacant, E-03071 Alacant, Spain {mespla,fsanchez,mlf}@dlsi.ua.es Abstract This paper explores a new method to im- prove computer-aided translation (CAT) systems based on translation memory (TM) by using pre-computed word alignments be- tween the source and target segments in the translation units (TUs) of the user’s TM. When a new segment is to be translated by the CAT user, our approach uses the word alignments in the matching TUs to mark the words that should be changed or kept unedited to transform the proposed trans- lation into an adequate translation. In this paper, we evaluate different sets of align- ments obtained by using GIZA++. Experi- ments conducted in the translation of Span- ish texts into English show that this ap- proach is able to predict which target words have to be changed or kept unedited with an accuracy above 94% for fuzzy-match scores greater or equal to 60%. In an ap- pendix we evaluate our approach when new TUs (not seen during the computation of the word-alignment models) are used. 1 Introduction Computer-aided translation (CAT) systems based on translation memory (TM) (Bowker, 2002; Somers, 2003) and, optionally, additional tools such as terminology databases (Bowker, 2003), are the translation technology of choice for most profes- sional translators, especially when translation tasks are very repetitive and effective recycling of previ- ous translations is feasible. When using a TM-based CAT system to trans- late a source segment s′, the system provides the c⃝ 2011 European Association for Machine Translation. set of translation units (TUs) {(si, ti)}N i=1 whose fuzzy-match score is above a given threshold Θ, and marks which words in each source-language (SL) segment si differ from those in s′. It is how- ever up to the translator to identify which target words in the corresponding target-language (TL) segments ti should be changed to convert ti into t′, an adequate translation of s′. The method we propose and evaluate in this pa- per is aimed at recommending the CAT user which words of ti should be changed by the translator or kept unedited to transform ti into t′. To do so, we pre-process the user’s TM to compute the word alignments between the source and target segments in each TU. Then, when a new segment s′ is to be translated, the TUs with a fuzzy-match score above the threshold Θ are obtained and the alignment be- tween the words in si and ti are used to mark which words in ti should be changed or kept unedited. Related work. In the literature one can find dif- ferent approaches that use word or phrase align- ments to improve existing TM-based CAT systems; although, to our knowledge, none of them use word alignments for the purpose we study in this pa- per. Simard (2003) focuses on the creation of TM- based CAT …

Content truncated to 500 words