Knowledge-Enhanced Retrieval: A Scheme for Question Answering

Chinese Knowledge Base Question Answering (CKBQA), as a significant task in natural language processing, has drawn massive attention from both industry and academia. However, previous studies mainly concentrated on multi-hop questions, which may limit the performance of tackling complex natural questions with various forms. To that end, in this paper, we propose a comprehensive technical framework called Knowledge-Enhanced Retrieval Question Answering (KERQA) for tackling complex questions, which could precisely extract the gold answers to these questions from a large-scale knowledge graph. Specifically, our proposed KERQA follows the pipeline with five modules, including the Question Classification module to categorize questions, the Named Entity Recognition module to extract mentions, and the Entity Linking module to match entities in the knowledge graph (KG). Along this line, we further design the Path Generation module to associate the paths in the KG with predefined templates, as well as the Path Ranking module to capture the best path. Extensive validations demonstrate the effectiveness of our KERQA framework, which achieved an F1 score of 78.78% on the final leaderboard of the CCKS 2021 KBQA contest.