Assisting teachers and annotators in their tasks: semi-automatic ensemble-based correction of errors

Anonymous

Abstract: We present in this paper an experiment to support the work of correcting errors in learner texts. Our system suggests possible corrections for spans of text that have been manually marked as erroneous, such as when teachers correct learner essays or when linguists annotate errors in a learner corpus. Our system uses a two-step ensemble approach relying first on several components capable of suggesting candidate corrections, most of them being specifically trained machine translation (MT) systems, and then on a random forest machine learning (ML) classifier that categorizes candidate corrections as correct or incorrect. The results we obtained on German data show that our approach has tangible potential, but also that a number of challenges remain to be addressed in order to truly exploit it in practice.

Paper Type: long