Topic Detection and Tracking Based on Event Ontology

In recent years, Topic Detection and Tracking (TDT) has served as a core technology for searching, organizing and structuring news oriented textual materials from a variety of internet news and social media. The biggest challenges of TDT are the sparsity and complexity of data, organization of topic granularity, unexpectedness of emergency topics, and the unpredictability of topics evolution. This paper proposes a new TDT method based on event ontology for hierarchical topic detection and tracking topic evolution, named TDTEO. As domain-oriented event knowledge base, the event ontology provides event classes hierarchy based on domain common sense, as well a set of scenario models that describe the occurrence and evolution of different types of emergency events. The proposed method solves the problem that new emerging events are easy to be missed in the process of topic detection, and solves the problem that the topic model is easy to result in semantics drift due to the dynamic evolution of topic, and it can effectively improve the accuracy of topic detection and tracking. Experiments show our models achieve satisfactory performance of topic detection with a maximum macro-F1 value of 85.25%, and the (C
_Det
)Norm of topic tracking in the datasets is as low as 0.1028. Experimental results show that hierachical topic model can effectively detect the topics and tracking model based on event scenario can reflect the trend of topic evolution.