ComQA: Question Answering Over Knowledge Base via Semantic Matching

Question answering over knowledge base (KBQA) is a powerful tool to extract answers from graph-like knowledge bases. Here, we present ComQA-a three-phase KBQA framework by which end-users can ask complex questions and get answers in a natural way. In ComQA, a complex question is decomposed into several triple patterns. Then, ComQA retrieves candidate subgraphs matching the triple patterns from the knowledge base and evaluates the semantic similarity between the subgraphs and the triple patterns to find the answer. It is a long-standing problem to evaluate the semantic similarity between the question and the heterogeneous subgraph containing the answer. To handle this problem, first, a semantic-based extension method is proposed to identify entities and relations in the question while considering the underlying knowledge base. The precision of identifying entities and relations determines the correctness of successive steps. Second, by exploiting the syntactic pattern in the question, ComQA constructs the query graphs for natural language questions so that it can filter out topology-mismatch subgraphs and narrow down the search space in knowledge bases. Finally, by incorporating the information from the underlying knowledge base, we fine-tune general word vectors, making them more specific to ranking possible answers in KBQA task. Extensive experiments over a series of QALD challenges confirm that the performance of ComQA is comparable to those state-of-the-art approaches with respect to precision, recall, and F1-score.