Duluth : Word Sense Induction Applied to Web Page Clustering

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 202–206, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics Duluth : Word Sense Induction Applied to Web Page Clustering Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812 USA tpederse@d.umn.edu http://senseclusters.sourceforge.net Abstract The Duluth systems that participated in task 11 of SemEval–2013 carried out word sense induction (WSI) in order to cluster Web search results. They relied on an approach that repre- sented Web snippets using second–order co– occurrences. These systems were all imple- mented using SenseClusters, a freely available open source software package. 1 Introduction The goal of task 11 of SemEval–2013 was to clus- ter Web search results (Navigli and Vannella, 2013). The test data consisted of the top 64 Google results for each of 100 potentially ambiguous queries, for a total of 6,400 test instances. The Web snippets re- turned for each query were clustered and evaluated separately, with an overall evaluation score provided for each system. The problem of Web page clustering is one of the use cases envisioned for SenseClusters (Peder- sen and Kulkarni, 2007; Pedersen, 2010a), a freely available open source software package developed at the University of Minnesota, Duluth starting in 2002. It supports first and second–order clustering of contexts using both co–occurrence matrices (Pu- randare and Pedersen, 2004; Kulkarni and Pedersen, 2005) and Latent Semantic Analysis (Landauer and Dumais, 1997). SenseClusters has participated in various forms at different SenseEval and SemEval shared tasks, in- cluding SemEval-2007 (Pedersen, 2007), SemEval- 2010 (Pedersen, 2010b) and also in an i2b2 clinical medicine task (Pedersen, 2006). 2 Duluth System While we refer to three Duluth systems (sys1, sys7, and sys9), in reality these are all variations of the same overall system. All three are based on second– order context clustering as provided in SenseClus- ters. The query terms are treated exactly like any other word in the snippets, which is called headless clustering in SenseClusters. 2.1 Common aspects to all systems The input to sys1, sys7, and sys9 consists of 64 Web search snippets, each approximately 25 words in length. All text was converted to upper case prior to processing. The goal was to group the 64 snip- pets for each query into k distinct clusters, where k was automatically determined by the PK2 method of SenseClusters (Pedersen and Kulkarni, 2006a; Ped- ersen and Kulkarni, 2006b). Each discovered clus- ter represents a different underlying meaning of the given query term that resulted in those snippets be- ing returned. Word sense induction was carried out separately on the Web snippets associated with each query term, meaning that the algorithm was run 100 times and clustered 64 Web page snippets each time. In second–order context clustering, the words in a context (i.e., Web snippet) to be clustered are re- placed by vectors that are derived from some cor- pus of text. The corpora used are among the main differences in the Duluth systems. …

Content truncated to 500 words