Extracting and Evaluating Ontologies of Human Activities from Linked Open Data and Social Media

Many of the problems we face today in artificial intelligence, e.g., real-world activity navigation, can only be solved if we have adequate knowledge bases of human behavior. Although some of the knowledge is available as Linked Open Data (LOD), the amount of data that is available is not enough to realize truly useful navigation applications. The purpose of this paper is to propose and evaluate a method that can automatically extract an ontology of human activities. This ontology is composed of a graph based on ConceptNet5 whose nodes represent activities and whose edges represent semantics between activities. The challenges to developing a truly useful activity navigation system are the following: (i) improving the currently low coverage rate of extracted activities, and (ii) allowing and predicting multiple semantics between activities. The proposed method has two steps: (1) extract mutually related activities from social media, and (2) predict the semantic label between activities via supervised learning by leveraging the semantics on ConceptNet5 as labeled data. Comparisons with ConceptNet5 show that the proposed method can discover more activities than ConceptNet5 and predict semantic relations between activities more accurately, about 10% more accurately, than the baseline method.