Taking a Closed-Book Examination: Decoupling KB-Based Inference by Virtual Hypothesis for Answering Real-World Questions

Complex question answering in real world is a comprehensive and challenging task due to its demand for deeper question understanding and deeper inference. Information retrieval is a common solution and easy to implement, but it cannot answer questions which need long-distance dependencies across multiple documents. Knowledge base (KB) organizes information as a graph, and KB-based inference can employ logic formulas or knowledge embeddings to capture such long-distance semantic associations. However, KB-based inference has not been applied to real-world question answering well, because there are gaps among natural language, complex semantic structure, and appropriate hypothesis for inference. We propose decoupling KB-based inference by transforming a question into a high-level triplet in the KB, which makes it possible to apply KB-based inference methods to answer complex questions. In addition, we create a specialized question answering dataset only for inference, and our method is proved to be effective by conducting experiments on both AI2 Science Questions dataset and ours.