Word Transition Entropy as an Indicator for Expected Machine Translation Quality
Michael H. Carl, Moritz Jonas Schaeffer · 2014
While most machine translation evaluation techniques (BLEU, NIST, TER, METEOR) assess translation quality based on a set of reference translations, we suggest to evaluate the literality of a set of (human or machine generated) translations to infer their potential quality. We provide evidence which suggests that more literal translations are produced more easily, by humans and machine, and are also less error prone. Literal translations may not be appropriate or even possible for all languages, types of texts, and translation purposes. However, in this paper we show that an assessment of the literality of translations allows us to (1) evaluate human and machine translations in a similar fashion and (2) may be instrumental to predict machine translation quality scores,