Highlighting matched and mismatched segments in translationmemory output through sub-tree alignment
Ventsislav Zhechev · 2010
In recent years, it is becoming more and more clear that the localisation industry does not have the necessary manpower to satisfy the increasing demand for high-quality translation. This has fuelled the search new and existing technologies that would increase translator throughput. As Translation Memory (TM) systems are the most commonly employed tool by translators, a number of enhancements are available to assist them in their job. One such enhancement would be to show the translator which parts of the sentence that needs to be translated match which parts of the fuzzy match suggested by the TM. For this information to be used, however, the translators have to carry it over to the TM translation themselves. In this paper, we present a novel methodology that can automatically detect and highlight the segments that need to be modified in a TM-suggested translation. We base it on state-of-the-art sub-tree align- ment technology (Zhechev,2010) that can produce aligned phrase-based-tree pairs from unannotated data. Our system operates in a three-step process. First, the fuzzy match selected by the TM and its translation are aligned. This lets us know which segments of the source-language sentence correspond to which segments in its translation. In the second step, the fuzzy match is aligned to the input sentence that is currently being translated. This tells us which parts of the input sentence are available in the fuzzy match and which still need to be translated. In the third step, the fuzzy match is used as an intermediary, through which the alignments between the input sentence and the TM translation are established. In this way, we can detect with precision the segments in the suggested translation that the translator needs to edit and highlight them appropriately to set them apart from the segments that are already good translations for parts of the input sentence. Additionally, we can show the alignments—as detected by our system—between the input and the translation, which will make it even easier for the translator to post-edit the TM suggestion. This alignment information can additionally be used to pre- translate the mismatched segments, further reducing the post-editing load.