Correcting open-answer questionnaires through a Bayesian-network model of peer-based assessment

We have previously shown that, with the help of peer-assessment and of a finite-domain constraint-based model of the student´s decisions, the teacher could have a complete assessment of the answers to open-ended questions, by grading just a subset of the answers (as low as half of the lot) and having the rest of the grading inferred by the supporting system. In this paper we present a probabilistic version of the earlier model, using Bayesian networks instead than constraints. Our aims are both defining the approach and prepare its validation: 1) modeling the peer-assessment activity of a student that evaluates others´ answers, 2) using peer-assessment to help the teacher with a faster/shorter assessment process, 3) inferring the student´s level of competence and ability to judge, from peer-assessment and from (partial) teacher-assessment, 4) learning the conditional probabilistic tables (CPTs) of the model from student data, and 5) comparing the probability distribution of competences in the class at different course phases. The model is under development and test with real data. We are developing a web-based interface to deliver open-answer and peer-assessment questionnaires and to assist the teacher-assessment.