Automatic Grading of Short Medical Answers : exploring Diverse Approaches

Details

Supervisors

Fairon, Cédrick

Faculty

Faculté de philosophie, arts et lettres

Degree label

Master [120] en linguistique, à finalité spécialisée: traitement automatique du langage

Abstract

The research surrounding automatic evaluation of student answers has received substantial interest in recent years. Automatically grading student answers would not only lead to a reduction of time dedicated by teachers to the task of evaluating answer, but could also enhance the objectivity and consistency of the grading process. In the present Master's thesis, attempts to automatically evaluate short answers connected to the medical domain are presented. The studied question prompts students to summarize the current problem of a patient presenting himself to a doctor's consultation. While not aiming at the development of a functioning evaluation system for summaries answering to this type of question, the present work's objective is to build a base for future work. Both a concept matching approach based on word overlap, as well as two cosine-similarity-based approaches, leveraging either LSA models or sentence transformers, were tested and satisfying results could be achieved. Important insights surrounding each approach's fitness for the task of grading medical short answers could be gained, and major challenges, as well as possible improvements are discussed.