DOMAIN ADAPTATION FOR AUTOMATED ESSAY SCORING

Abstract

Automated Essay Scoring (AES) is an important task in Natural Language Processing. The research done by various commercial organizations has identi ed the features that correlate well with human scoring. They have built strong AES systems that achieve high agreement with human scoring based on these features. One of these commercial organizations, ETS, even uses their own AES system (erater) as a second rater for their high-stakes exams, GRE and TOEFL. However, most of these AES systems use prompt-speci c features. This means that each time a new prompt is introduced, a large number of essays need to be annotated as training data. This thesis gives an overview of the AES task and shows that domain adaptation can help an AES system to achieve high performance with a small number of annotated essays.