Publication at Faculty of Mathematics and Physics |
2013
Abstract
In this paper, we discuss the possibility of making the automatic evaluation of machine translation (MT) more accurate by targeted paraphrasing. The most common automatic metric for MT evaluation BLEU disregards synonymous phrases and word order variants.
One way how to alleviate this drawback is to transform the reference translation into a one that is closer to the MT output and keeps its original meaning. We start with a basic algorithm for lexical substitution and discuss extending the process with word order changes and grammar checking.