Deep Structure-Aware Approach for QA Over Incomplete Knowledge Bases

Abstract

AbstractThe incompleteness of Knowledge Base (KB) greatly limits the performance of Question Answering (QA) system. Combining documents and incomplete KBs to develop QA system has become a hot spot in the research of Knowledge Base Question Answering (KBQA). Recent work ignores the relevance of KB and documents, and hinder the fusion of structured knowledge and unstructured text. This paper firstly builds a question-related subgraph entity encoder (QRS-Encoder) to gain nodes embedding; secondly, a structure-aware document reader (SAD-Reader) with deep structure information and related question is constructed, which can capture the Meta Dependency Path (MDP) nodes in the dependency graph constructed from question related documents. The MDP nodes are fused into node embedding representation in SAD-Reader, which effectively increases the node’s attention to the non-local dependencies in the document. Empirical results on WebQSP dataset show that our model outperforms state-of-the-art (SOTA) model in terms of both F1 and Hit@1 under incomplete KBs setting, which proves the effectiveness of our model.KeywordsKBQAKnowledge-aware readerMDP nodes