Konferensartikel

One of the issues in automatically evaluating learner input in the context of Intelligent Tutoring Systems is learners’ use of incorrect forms and non-standard language. Grammatical Error Correction (GEC) systems have emerged as a way of automatically correcting grammar and spelling mistakes, often by approaching the task as machine translation of individual sentences from non-standard to standard language. However, due to the inherent lack of context awareness, GEC systems often do not produce a contextually appropriate correction. In this paper, we investigate how current neural GEC systems can be optimized for educationally relevant tasks such as Short Answer Assessment. We build on a recent GEC system and train a reranker based on context (e.g. similarity to prompt), task (e.g. type and format) and answer-level (e.g. language modeling) features on a Short Answer Assessment dataset augmented with crowd worker corrections. Results show that our approach successfully gives preference to corrections that are closer to the reference.

Grammatical Error Correction, Reranking, Short Answer Assessment, Intelligent Tutoring Systems