Leveraging Contextual Interaction and Semantic Knowledge for Multi-Hop Question Answering

Knowledge Graph Question Answering (KGQA) is a fundamental task in Knowledge Graph (KG). Existing works of KGQA deduce path information from the neighborhood of the subgraph and obtain the answer to the question. However, the global semantic information may not be obtained for sparse KG. In addition, the interaction between question and the knowledge base is ignored. In this paper, we propose a model called Contex-tual Interaction and Semantic Knowledge (CISK) for multi-hop question answering. We leverage the attention mechanism to fuse contextual information to enhance the question. Additionally, we use knowledge graph embedding (KGE) to ensure the integrity of the KG. Compared to the baseline (EmbedKGQA), our model achieves 0.7% and 2.0 % accuracy improvements on 2-hop questions for MetaQA_full and MetaQA_half datasets, 0.3% and 1.6 % on 3-hop questions, respectively, and 3.8 % accuracy improvements on WebQSP _full and 0.9% on WebQSP _half, respectively.