A subgraph-representation-based method for answering complex questions over knowledge bases.

Abstract

Knowledge-based question answering has attracted a lot of attention in the research communities of natural language processing and information retrieval. However, existing studies do not adequately address the problem of answering complex questions which involve multiple entities and require extraction of facts from multiple relations. To address this issue, we propose a novel approach which learns the distributional representations of questions and candidate answers in a unified deep-learning framework based on directed-acyclic-graph-structured long short-term memory and memory networks. Specifically, the questions are encoded to match candidate directed acyclic subgraphs of the knowledge base, which are able to include information related to multiple entities and relations in the complex questions. The experimental results show that the proposed approach outperforms other methods on the widely used dataset SPADES, especially when dealing with complex questions with multiple entities.