Improving complex knowledge base question answering via structural information learning

Abstract

Responding to complex questions is one of the most difficult and valuable goals of KBQA. Current efforts mainly follow two approaches to extract the in-depth semantics of questions. Information retrieval-based methods tend to encode questions directly and ignore the explicit analysis of the question structure. Besides, although retaining the analysis ability of question structure, semantic parsing-based methods rely on the expensive query graph labels and suffer from sparse reward due to wrong explorations. To benefit from both sides, this paper proposes a novel semantic parsing model, structural information restraint (SIR) for KBQA. SIR applies structural information of questions for reinforcement-based path reasoning for the first time. Specifically, SIR synthesizes the dependency tree, constituency tree, and the first token to build a composited structural attention (StrucAtt) and realizes reasoning without expensive query graphs labels. Such an attention mechanism improves the efficiency of path reasoning by distinguishing different knowledge paths, based on the relevance between path features and question structure. In addition, we also design a type-assisted reward based on answer concepts (person, location, etc.) instead of simple variable types (string, number, etc.), which alleviates the sparse reward problem effectively. Besides, the experiment results clearly show that our model achieves SOTA on CWQ, CQ, and WQSP datasets.