Answer Extraction with Graph Attention Network for Knowledge Graph Question Answering

Abstract

In the knowledge graph question answering, the graph neural network can be used to encode the subgraph nodes related to the question entity to select the correct answer node. However, existing researches mainly focus on the modalities for the node encoding with graph neural network, ignoring that different types of subgraphs have different requirements for encoding information. To overcome the problem, this paper divides the subgraph into two types: the searching graph and the extending graph. Then we propose an answer extraction method with graph attention network for the searching graph, which can weight the information of neighbor nodes with different attention instead of the average. The hierarchical attention is also introduced to integrate question information into the subgraph node embedding to obtain the node presentation with question dependency. The accuracy of 48.2% is achieved on the CommonsenseQA dataset, which is much higher than the random guess (20%). In addition, the accuracy of the simplified model with no hierarchical attention decreases by 3.5%, which indicates the hierarchical attention mechanism can improve the predictive performance of the proposed model.