Assessment Modeling: Fundamental Pre-training Tasks for Interactive Educational Systems.

Like many other domains in Artificial Intelligence (AI), there are specifictasks in the field of AI in Education (AIEd) for which labels are scarce andexpensive, such as predicting exam score or review correctness. A common way ofcircumventing label-scarce problems is pre-training a model to learnrepresentations of the contents of learning items. However, such methods failto utilize the full range of student interaction data available and do notmodel student learning behavior. To this end, we propose Assessment Modeling, aclass of fundamental pre-training tasks for general interactive educationalsystems. An assessment is a feature of student-system interactions which canserve as a pedagogical evaluation. Examples include the correctness andtimeliness of a student's answer. Assessment Modeling is the prediction ofassessments conditioned on the surrounding context of interactions. Although itis natural to pre-train on interactive features available in large amounts,limiting the prediction targets to assessments focuses the tasks' relevance tothe label-scarce educational problems and reduces less-relevant noise. Whilethe effectiveness of different combinations of assessments is open forexploration, we suggest Assessment Modeling as a first-order guiding principlefor selecting proper pre-training tasks for label-scarce educational problems.