Knowledge Base Question Answering via Structured Query Generation using Question domain

The core problem of Knowledge Base Question Answering(KBQA) is to find queries from user questions to knowledge bases. Specifically, natural language questions need to be transformed into structured queries before associating with knowledge bases, and the answers can be found from the knowledge graph using the structured queries. We found structured queries in similar question domains tend to have repetitive reasoning steps. Also, humans often use cases identical to the question and information from these cases to assist in answering new questions. Hence, we propose a new KBQA framework based on similar question domains. We separately design the inference information retriever module to extract cases with a similar structure to the question and the relation information retriever module to narrow the scope of reasoning relation extraction. Finally, we used the retrieved inference cases and relation candidate sets as auxiliary information and generated an executable Knowledgeoriented Programming Language(KoPL) through the program generation module. Experiments have shown that the model can handle complex question answering and has a strong reasoning ability. Our methodology has resulted in new state-of-the-art performance on WebQSP and CWQ datasets.