Question Answering over Knowledge Graphs for Thai Retail Banking Products

Abstract

Question Answering over Knowledge Graphs (KGQA) extracts the answer entity directly from the graph, given a natural language question, offering scalability to applications that need to readily provide information to the end users, such as chatbots. Nevertheless, KGQA specifically designed for Knowledge Graphs in Thai has not yet been well investigated. In this paper, we adapt multi-hop KGQA using Graph Embedding to handle Thai dataset while being able to extract answer entities that do not have explicit relation to the head node. We also construct a Thai Knowledge Graph with the ontology based on retail banking products. The model achieves a HITS @ 1 score of 80.8 on our annotated dataset. The results confirm that, aside from reaching multi-hop answers, using Graph Embedding in KGQA helps improve the overall score, especially in sparse Knowledge Graphs. Moreover, augmenting the training questions to include more entities in the graph can further help increase the performance.