Quality Estimation of MT-Engine Output Using Language Models for Post-editing and Their Comparative Study

Abstract

Machine Translation (MT) systems do not have real-world knowledge or contextual awareness. MT errors are possible at any level: lexical, grammatical, syntactic, etc., MT systems give 10–70 % accurate output, so human post-editing(HPE) is required for final output. But HPE is very expensive and slow, if we can filter out good translations out of all translations, those can make correct via miner edits then our HPE would be fast and less expensive. We can estimate good quality of a sentence using language model (LM). There are different LMs available. We showed in our experiment that Kneser-Ney smoothing LM is the right choice for measuring MT-Engine-output’s quality for the post-editing.