Using Names and Topics for New Event Detection

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 121–128, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Using Names and Topics for New Event Detection Giridhar Kumaran and James Allan Center for Intelligent Information Retrieval Department of Computer Science University of Massachusetts Amherst Amherst, MA 01003, USA {giridhar,allan}@cs.umass.edu Abstract New Event Detection (NED) involves monitoring chronologically-ordered news streams to automatically detect the stories that report on new events. We compare two stories by finding three cosine simi- larities based on names, topics and the full text. These additional comparisons sug- gest treating the NED problem as a bi- nary classification problem with the com- parison scores serving as features. The classifier models we learned show statis- tically significant improvement over the baseline vector space model system on all the collections we tested, including the lat- est TDT5 collection. The presence of automatic speech recog- nizer (ASR) output of broadcast news in news streams can reduce performance and render our named entity recognition based approaches ineffective. We provide a so- lution to this problem achieving statisti- cally significant improvements. 1 Introduction The instant and automatic detection of new events is very useful in situations where novel informa- tion needs to be detected from a real-time stream of rapidly growing data. These real-life situations occur in scenarios like financial markets, news anal- yses, and intelligence gathering. In this paper we focus on creating a system to immediately identify stories reporting new events in a stream of news - a daunting task for a human analyst given the enormous volume of data coming in from various sources. The Topic Detection and Tracking (TDT) pro- gram, a DARPA funded initiative, seeks to develop technologies that search, organize and structure mul- tilingual news-oriented textual materials from a va- riety of broadcast news media. One of the tasks in this program, New Event Detection (NED), involves constant monitoring of streams of news stories to identify the first story reporting topics of interest. A topic is defined as “a seminal event or activity, along with directly related events and activities” (Al- lan, 2002). An earthquake at a particular place is an example of a topic. The first story on this topic is the story that first carries the report on the earth- quake’s occurrence. The other stories that make up the topic are those discussing the death toll, the res- cue efforts, the reactions from different parts of the world, scientific discussions, the commercial impact and so on. A good NED system would be one that correctly identifies the article that reports the earth- quake’s occurrence as the first story. NED is a hard problem. For example, to dis- tinguish stories about earthquakes in two different places, a vector space model system would rely on a tf-idf weighting scheme that will bring out the dif- ference by weighting the locations higher. More often then not, this doesn’t happen as the differ- ences are buried in the mass of …

Content truncated to 500 words