Tracking events via automatically generated profiles

In Topic Detection and Tracking (TDT), topics are at different levels of granularity. Some topics concentrate on a specific event, while others consist of a series of related events. In this research, we focus on tracking the topics that originate and evolve from a specific event such as an earthquake. Intuitively, a few of key elements of a target event, such as the date, the location, and the persons involved, would be sufficient for making a decision on whether a test story is on-topic or not. We, therefore, focus on how to automatically generate event profiles for tracking relevant news stories. Several techniques including named entity recognition, feature selection, and text summarization are combined to help obtain human readable event profiles from one or more given on-topic stories, and such derived profiles are expected to contain most, if not all, of the key semantic elements of a target event. Experimental results on TDT2 mandarin corpus demonstrate the effectiveness of this profile-based tracking method for topics centered on a specific event. Moreover, with the proposed method, human intervention becomes possible and the tracking system could absorb new and precise knowledge in due course.