RI-SSGE: a framework with rule inference and sentence schema graph embedding for knowledge base query construction

As knowledge graph becomes popular in recent years, more and more attention has been paid to Knowledge Base Question-Answer (KBQA) systems. For KBQA systems, Question Understanding, as the first stage, aims to convert factual question into the interpretable form to machine just like Λ-DCS. And some latest works used query subgraph to change the Question Understanding task into the Question to Subgraph(Question2Subgraph) task with which the subgraph can be simply and directly mapped to Λ-DCS. In this paper, we focus on factual question to subgraph task (Qƒ, G) and prove that more complex questions can be easily solved based on it. Then, we propose a novel framework with Rule Inference and Sentence Schema Graph Embedding (RI-SSGE) to solve (Qƒ, G) task. Inspired by isomeride structures in Chemistry, we concentrate RI-SSGE on structure detection of questions to avoid the problem of poor generalization in other models, which are based on templates on various specific domain knowledge graphs. To address the problem of error propagation, RI-SSGE creatively combines the traditional rule inference method and the graph representation method together, and thus guarantees the performance of the whole framework. Having observed that human can exploit the hidden relations by joining the question and the knowledge graph structure together, we raise a novel Sentence-Schema-Graph (SSG) in the last network representation learning stage of RI-SSGE, which is designed to imitate human's way of thinking. We experimented on Geoquery-880 and AceQG[11] datasets which has 133,143 (Factual Question, Subgraph) pairs on an open academic knowledge graph and results demonstrate the advantages of RI-SSGE over other baselines.