Abstract

Knowledge base question answering (KBQA) aims to answer user questions innatural language using rich human knowledge stored in large KBs. As currentKBQA methods struggle with unseen knowledge base elements at test time,weintroduce SG-KBQA: a novel model that injects schema contexts into entityretrieval and logical form generation to tackle this issue. It uses the richersemantics and awareness of the knowledge base structure provided by schemacontexts to enhance generalizability. We show that SG-KBQA achieves stronggeneralizability, outperforming state-of-the-art models on two commonly usedbenchmark datasets across a variety of test settings. Our source code isavailable at https://github.com/gaosx2000/SG_KBQA.

 

Quick Read (beta)