Answer Graph-based Interactive Attention Network for Question Answering over Knowledge Base

Question answering over knowledge base is one of the promising tasks to access knowledge in knowledge bases. Existing information retrieval based methods mainly map candidate answer into a vector space by aggregating different answer aspects (i.e., entity types, relation paths and context). However, these answer aspects are simply embedded uniformly, while neglecting both question-related and candidate-related answer aspects are dominant to find the optimal answer. To address the above issue, we propose a novel Answer Graph-based Interactive Attention Network, which explicitly constructs an answer graph for each candidate answer. The answer graph consists of most of the possible answer aspects, and we selectively find partial answer aspects that are most relevant to the question using Gated Graph Neural Networks. With the guidance of both question-related and candidate-related answer aspects, the optimal candidate answer can greatly approximate the question, and hence can answer the question more accurately. We conduct extensive experiments on the Web Questions dataset. Results demonstrate that our approach outperforms the previous state-of-the-art methods.