Incorporating Topic and Property For Knowledge Base Synchronization

Abstract

Abstract Open domain knowledge bases have been widely used in many applications, and it is critical to maintain their freshness. Most existing studies update an open knowledge base by predicting the change frequencies of the entities and then updating those unstable ones. In this work, we propose a novel topic-aware entity stability prediction framework which incorporates property and topic features of the entities to facilitate the prediction on their stability with graph structures, so that a knowledge base could be updated accordingly with favorable time and computation efficiency. Specifically, we first build an entity property graph for each entity, with its property names as edges and the property values as nodes, which models the various properties of the entities. Then, with the constructed entity property graph, we develop a Topic Classifier to label the topic information for the entity via unsupervised clustering. Finally, we treat the prediction task as a binary classification problem and solve it with an Entity Stability Predictor, which is designed to comprise two layers of Graph Convolution Networks, one pooling layer and one fully connected layer.The Entity Stability Predictor then predicts the stability of an entity based on its revision history acquired from the source encyclopedia webpage, wherein the topic information serves as strong supervision. Extensive experiments on collections of real-world entities have demonstrated the superior performance of our proposed method, and also well shown the benefits of each new module in our framework.