DKFM: a novel data and knowledge fusion-driven model for difficulty prediction of mathematical exercise

Zhiyi Duan, Hengnian Gu, Di Wang, Dongdai Zhou

Abstract: In online education systems, exercise difficulty prediction holds significant importance for various applications, including personalized exercise recommendation and evaluation of learners’ knowledge proficiency. Traditional methods are mainly based on the manual labeling by experts, which is labor-costing and time-consuming. Accurately predicting the difficulty of exercises is a challenging task, since it is hard to simultaneously mine the rich semantic features existed in the texts and the objective knowledge depth information. To address this issue, we propose a novel data and knowledge fusion-driven model (DKFM) for accurately predicting the difficulty of mathematical exercises. Specifically, we first design a large-scale pre-trained model-based layer to extract semantic features for each exercise. Subsequently, we construct a Mathematical Knowledge Base that enables automated extraction of knowledge depth information associated with the exercises. To consider the varying contributions of different types of features toward the final prediction, we propose an attention-based fusion approach to learn the coefficients for diverse features. Experimental evaluations illustrate that compared with GPT-3, our proposed DKFM achieves an improvement of 22.5% on Algebra dataset, 25.5% on Geometry dataset and 38.8% on Counting dataset in terms of F1-score.

External IDs:dblp:journals/kais/DuanGWZ25