Predicting Machine Translation Adequacy

Abstract

As Machine Translation (MT) becomes more popular among end-users, an increasingly relevant issue is that of estimating the quality of automatic translations for a particular task. The main application for such quality estimates has been selecting good enough translations for human post-editing. The endusers, in this case, are fluent speakers of both source and target languages and the quality estimates reflect post-editing effort, for example, the time to post-edit a sentence. This paper focuses on quality estimation to address the challenging problem of making MT more reliable to a different type of end-user: those who cannot read the source language. We propose a number of indicators contrasting the source and translation texts to predict the adequacy of such translations at the sentence-level. Experiments with Arabic-English MT show that these indicators can yield improvements over previous work using general quality indicators based on source complexity and target fluency. 1