Online Word Alignment for Online Adaptive Machine Translation

Abstract

A hot task in the Computer Assisted Translation scenario is the integration of Machine Translation (MT) systems that adapt sentence after sentence to the post-edits made by the translators. A main role in the MT online adaptation process is played by the information extracted from source and post-edited sentences, which in turn depends on the quality of the word alignment between them. In fact, this step is particularly crucial when the user corrects the MT output with words for which the system has no prior information. In this paper, we first discuss the application of popular state-of-the-art word aligners to this scenario and reveal their poor performance in aligning unknown words. Then, we propose a fast procedure to refine their outputs and to get more reliable and accurate alignments for unknown words. We evaluate our enhanced word-aligner on three language pairs, namely English-Italian, English-French, and English-Spanish, showing a consistent improvement in aligning unknown words up to 10% absolute F-measure.