Question Difficulty Prediction with External Knowledge

Abstract

The difficulty of test questions is an important indicator for educational examination and recommendation of personalized learning resources. Its evaluation mainly depends on the experience of experts, which is subjective. In recent years, question difficulty prediction (QDP) using neural networks has attracted more and more attention. Although these methods improve the QDP efficiency, it works ill for questions involving abstract concepts, such as numerical calculation, date, and questions whose answers require background knowledge. Therefore, we propose a difficulty prediction model based on rich knowledge fusion (RKF+), which solves the problem that the difficulty prediction models cannot obtain conceptual knowledge and background knowledge. The key is to introduce the attentional mechanism with a sentry vector, which can dynamically obtain the text representation and external knowledge representation of test questions. To further fusion the acquired external knowledge, our model added a bi-interaction layer. Finally, the validity of this model is verified on three different datasets. Besides, the importance of attentional mechanism and external knowledge representation is further analyzed by ablation experiment. In addition, based on a real English reading comprehension test dataset, we explore the influence of two kinds of external knowledge on the question difficulty prediction model.