Clustering by committee

• Authors: Dekang Lin;Patrick Andre Pantel
• Venue: Clustering by committee
• Year: 2003

Text contains a wealth of knowledge about who we are, what we know, how we think, and how we communicate. We are just beginning to tap into the information that is available in the tales we read to our children, the narratives that capture our thoughts, and the stories that shape our world. In this work, we present some recent advances in automatically acquiring knowledge from text. We propose a general-purpose clustering algorithm called CBC (Clustering By Committee) from which we will organize documents according to topics as well as discover concepts and word senses. We will explore the value of these systems by experimenting with two novel evaluation methodologies that attempt to define what a word sense is and define the quality of a particular clustering. CBC addresses the general goal of clustering, which is to group data elements such that the intra-group similarities are high and the inter-group similarities are low. Using sets of representative elements called committees, CBC attempts to discover cluster centroids that unambiguously describe the members of a possible class. CBC will be shown to outperform several common clustering algorithms in document clustering and concept discovery tasks. Document clustering is practical in many information retrieval tasks such as document browsing and the organization and viewing of retrieval results. Broad-coverage lexical resources such as WordNet are extremely useful but are mostly hand generated. They often include many rare senses while missing domain-specific senses. Automatically generating them is useful for many applications such as word sense disambiguation, question answering and ontology construction. Sample concepts discovered by CBC include baking ingredients, symptoms, academic departments, Impressionists, Canadian provinces, musical instruments, and emotions. We present two novel evaluation methodologies. The first is based on the editing distance between output clusters and a manually constructed answer key. It defines how much work is necessary in order to convert from one to the other. For the word sense discovery system, we present an evaluation methodology for measuring the precision and recall of discovered senses. Using WordNet, we formulate what is a correct sense of a word.