Paper Title
A New Approach to Problem Solving in Education Based on Embedding
Abstract
Most intelligent tutoring systems ITS provid feedback like hints or messages to facilitate problem-solving in
mathematics, but they don’t offer feedback on reasoning strategies. In other words, ITS can’t guide learners towards
appropriate didactic concepts (theorems, definitions, etc.) for problem-solving. Our methodology allows for the generation of
feedback in human learning environments. It predicts the relevant didactic concepts necessary to solve a given exercise
question Q(i) based on existing or previously demonstrated data (Q(i-1), Q(i-2), etc.) using the Math-Bridge ontology. The
approach involves segmenting concept instances in the ontology and exercises into prerequisites and outcomes. The purpose
of this segmentation is to compare the prerequisites and outcomes of concept-exercise pairs (Ci-E) in order to identify the
most suitable concepts for solving a specific exercise, achieved through similarity calculations based on embedding. On the
other hand, our approach utilizes a new strategy for vectorizing mathematical text, enabling the resolution of NLP problems
specific to mathematical texts in order to exploit them in other deep learning issues. For the evaluation phase, we tested our
approach on a set of exercises taken from mathematics education platforms and compared the number of concepts generated
by our approach with those of experts. Our approach shows promising potential for exercise resolution.
Keywords - Ontology, Didactic concept, Semantic similarity, Mathematical problem, Deep learning.