Question-aware Graph Convolutional Network for Educational Knowledge Base Question Answering

• Main Author: LIN Qika, ZHANG Lingling, LIU Jun, ZHAO Tianzhe
• Format: Article
• Language: zho
• Published: Journal of Computer Engineering and Applications Beijing Co., Ltd., Science Press 2021-10-01
• Series: Jisuanji kexue yu tansuo
• Subjects: graph convolutional network (gcn); attention; educational knowledge base; knowledge base question answering (kbqa); knowledge graph
• Online Access: http://fcst.ceaj.org/CN/abstract/abstract2911.shtml

In recent years, with the continuous informatization of education and the accumulation of massive education resources and teaching data, some educational knowledge bases have been proposed, which provides a good deve-lopment condition for data-driven intelligent education. The question answering method based on educational know-ledge base can provide learners with instant tutoring, and then effectively improve their learning interest and efficiency. However, there are few studies on educational knowledge base question answering (KBQA), and most of the open domain KBQA methods independently model question sentences and candidate answer entities, so the effect of modeling is limited. Based on this, this paper proposes a question answering method of educational knowledge base based on question-aware graph convolutional network (GCN). Firstly, for a specific question, the description information and query entity set of the question are extracted. And they are processed respectively by Transformer and pre-trained embeddings of the knowledge base. Secondly, the subgraph of candidate answer set is extracted from the knowledge base according to the query entity set, and the node information is updated by the GCN with two attention mechanisms. The scores of attention are expressed by the question description and the query entity set respectively. In this way, the question-aware GCN is realized. Finally, the query description information, query entity set and candidate entity representation are fused to calculate the score and predict the answer. Experiments are carried out on the real data set MOOC Q&A, and the prediction accuracy and mean reciprocal rank are used to evaluate. The experimental results show that the proposed method is superior to the benchmark models.