Towards error-based strategies for automatically assessing ESL learners' proficiency

In this paper we propose potential strategies for automatically assessing second language proficiency based on the presence of errors only. We used an open-source grammar and spelling check tool to extract errors from the answers of the written section of an Italian English as a second language (ESL) learners’ corpus annotated with human scores and we automatically generated the respective correct versions. We found a moderate correlation between the presence of errors and the scores assigned by human experts. As such, we believe that error-rate may be particularly suitable for automatic assessment tools. Therefore, we envisage the use of various state-of-the-art machine learning approaches, aiming at developing useful techniques for both ESL learners and teachers.