Anagha will be defending her Master's thesis next week! ----------------------- COMPUTER SCIENCE COLLOQUIUM
Unsupervised Context Discrimination and Automatic Cluster Stopping ANAGHA KULKARNI Computer Science Graduate Student Wednesday, July 5 1:00 p.m. HELLER HALL 306 ABSTRACT Context discrimination is the task of separating a set of written contexts into different groups based on their mutual similarity or dissimilarity. More specific instances of this task are problems such as name discrimination, word sense discrimination, and email clustering. Name discrimination seeks to distinguish between different entities that share the same name. For example, George Bush - the 41st President of the USA, versus George Bush - the 43rd. In word sense discrimination the task is to disambiguate the occurrences of a word with multiple meanings. For example, the word tape can refer to a measuring device, a fastening piece of cloth or paper, etc. The goal of email clustering is to group together messages that discuss similar topics. In this thesis we approach each of these problems and more generally the context clustering problem by adapting and extending the unsupervised word sense discrimination techniques proposed by Purandare and Pedersen [2004]. One of our main contributions is the development of three cluster-stopping measures: PK2, PK3 and the Adapted Gap Statistic. Given a collection of contexts it is a non-trivial task for a user to determine how many different clusters might exist. Hence, we automate the process of predicting the optimal number of clusters for any given dataset by considering the properties of the raw data. We find that our measures can be successfully applied to each of the above problems with equal ease. Of the three cluster-stopping measures, PK3 shows the highest agreement when compared to manually determined clusters. ------------------------ Yahoo! Groups Sponsor --------------------~--> Check out the new improvements in Yahoo! Groups email. http://us.click.yahoo.com/6pRQfA/fOaOAA/yQLSAA/x3XolB/TM --------------------------------------------------------------------~-> Yahoo! Groups Links <*> To visit your group on the web, go to: http://groups.yahoo.com/group/nlpatumd/ <*> To unsubscribe from this group, send an email to: [EMAIL PROTECTED] <*> Your use of Yahoo! Groups is subject to: http://docs.yahoo.com/info/terms/
