Knowledge Graph Based Absolute Difficulty Prediction of Multiple Choice Question

The difficulty of a multiple choice question (MCQ) is related not only to the semantic distribution of the options but also to the knowledge points involved in the question. Currently, models for the difficulty prediction of MCQ mostly rely only on the features of the question, neglecting the relationship between a question and the involved knowledge points and the connections among these knowledge points. Aiming at this shortcoming, a Knowledge Graph enhanced Neural Network for Absolute Difficulty Prediction (KGNN-ADP) is proposed in this paper. KGNN-ADP utilizes Bi-LSTM to extract textual representations, and a knowledge graph is used to get knowledge points related to a question. KGNN-ADP predicts the absolute difficulty of a MCQ based on both knowledge difficulty and option difficulty. Experiments on actual datasets show the effectiveness of the KGNN-ADP.