Question answering over knowledge base with symmetric complementary attention

Knowledge Base Question Answering (KBQA), which aims to answer natural language questions with structured data from a knowledge base is an important Natural Language Processing (NLP) problem. To answer the question, we need to find the fact from the Knowledge Base whose subject and relation best match the question. Most existing methods treat this task as a pipeline of two separate subtasks: subject matching and relation matching. While ignoring the relevance between them. In this paper, we focus on solving this problem through a joint learning method. We present a neural joint model with a shared encoding layer to learn the two subtasks together to improve each other. In particular, we design a Symmetric Bidirectional Complementary Attention module based on the attention mechanism and the gate mechanism to model the relationship between the two subtasks. The experimental results demonstrate that our approach can obtain higher accuracy than the state-of-the-art method.