Reading Document and Answering Question via Global Attentional Inference.

Teaching a computer to understand and answer a natural language question pertaining to a given general document is a central, yet unsolved task in natural language processing. The answering of complex questions, which require processing multiple sentences, is much harder than answering those questions which can be correctly answered by merely understanding a single sentence. In this paper, we propose a novel global attentional inference (GAI) neural network architecture, which learns useful cues from structural knowledge via a dynamically terminated multi-hop inference mechanism, to answer cloze-style questions. Here, reinforcement learning is employed to learn an inference gate which can determine whether to keep accumulating cues or to predict an answer. By exploiting structural knowledge, our model can answer complex questions much better than other compared methods.