Evaluating MT with translations or translators: what is the difference?
Martin Volk, Søren Harder · Zurich Open Repository and Archive (University of Zurich) · 2007
This paper describes a project on building a Machine Translation system for television and film subtitles. We report on the specific properties of the text genre, the language pair Swedish-Danish, and the large training corpus. We focus on the evaluation of the system output against independent and post-edited translations. We show that evaluation results against post-edited translations are higher by a margin of up to 19 points BLEU score.