Researchers at UMass Amherst and Eedi are placing exercise questions and dialogue sub-tasks on the same difficulty scale, to better track the progress of pupils’ skills.
Models that estimate a student's level are almost always based on their answers to questions. However, with the generalization of tutoring, human or AI, the exchanges between tutor and student contain much finer clues about what the student masters or not. Andrew Lan's team proposes to take advantage of both sources at the same time.
Dialogues transformed into exercises
The method begins by identifying, in the tutoring dialogues, the mathematical subtasks asked by the tutor, rewriting them into autonomous questions and determining whether the student has answered them correctly. A model derived from the theory of response to the item then places questions and subtasks on the same scale: it estimates their difficulty from the text, follows the evolution of the student's abilities from one interaction to the next and exploits the tutor's choices between two successive subtasks to better calibrate their difficulty.
Evaluated on two real-world math teaching datasets, including dialogues led by human tutors and AIs, the framework goes beyond solid reference methods, both for predicting the difficulty of questions and for anticipating whether an answer will be correct.
Read the full article on the original source
Go to the source →