Detecting Multiple Facets of an Event using Graph-Based Unsupervised Methods

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 609–616 Manchester, August 2008 Detecting multiple facets of an event using graph-based unsupervised methods Pradeep Muthukrishnan Dept of EECS University of Michigan mpradeep@umich.edu Joshua Gerrish School of Information University of Michigan jgerrish@umich.edu Dragomir R. Radev Dept of EECS & School of Information, University of Michigan radev@umich.edu Abstract We propose a new unsupervised method for topic detection that automatically iden- tifies the different facets of an event. We use pointwise Kullback-Leibler divergence along with the Jaccard coefficient to build a topic graph which represents the com- munity structure of the different facets. The problem is formulated as a weighted set cover problem with dynamically vary- ing weights. The algorithm is domain- independent and generates a representa- tive set of informative and discriminative phrases that cover the entire event. We evaluate this algorithm on a large collec- tion of blog postings about different news events and report promising results. 1 Introduction Finding a list of topics that a collection of docu- ments cover is an important problem in informa- tion retrieval. Topics can be used to describe or summarize the collection, or they can be used to cluster the collection. Topics provide a short and informative description of the documents that can be used for quickly browsing and finding related documents. Inside a given corpus, there may be multiple top- ics. Individual documents can also contain multi- ple topics. Traditionally, information retrieval systems re- turn a ranked list of query results based on the similarity between the user’s query and the docu- ments. Unfortunately, the results returned will of- ten be redundant. Users may need to reformulate c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. their search to find the specific topic they are in- terested in. This active searching process leads to inefficiencies, especially in cases where queries or information needs are ambiguous.For example, a user wants to get an overview of the Virginia tech shootings, then the first query he/she might try is “Virginia tech shooting”. Most of the results re- turned would be posts just mentioning the shoot- ings and the death toll. But the user might want a more detailed overview of the shootings. Thus this leads to continuously reformulating the search query to discover all the facets of the event. 2 Related Work Topic detection and tracking was studied exten- sively on newswire and broadcast collections by the NIST TDT research program (Allan et al., ). The large number of people blogging on the web provides a new source of information for topic de- tection and tracking. The TDT task defines topics as “an event or ac- tivity, along with all directly related events and ac- tivities.” In this paper we will stay with this defini- tion of topic. Zhai et al. proposed several methods for dealing with a related task, which they called subtopic re- trieval (Zhai et al., 2003). This is an information retrieval task …

Content truncated to 500 words