Automatically Labeling Semantic Classes

Automatically Labeling Semantic Classes Patrick Pantel and Deepak Ravichandran Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 {pantel,ravichan}@isi.edu Abstract Systems that automatically discover semantic classes have emerged in part to address the limitations of broad-coverage lexical re- sources such as WordNet and Cyc. The cur- rent state of the art discovers many semantic classes but fails to label their concepts. We propose an algorithm labeling semantic classes and for leveraging them to extract is-a relationships using a top-down approach. 1 Introduction The natural language literature is rich in theories of se- mantics (Barwise and Perry 1985; Schank and Abelson 1977). However, WordNet (Miller 1990) and Cyc (Le- nat 1995) aside, the community has had little success in actually building large semantic repositories. Such broad-coverage lexical resources are extremely useful in applications such as word sense disambiguation (Lea- cock, Chodorow and Miller 1998) and question answer- ing (Pasca and Harabagiu 2001). Current manually constructed ontologies such as WordNet and Cyc have important limitations. First, they often contain rare senses. For example, WordNet in- cludes a rare sense of computer that means ‘the person who computes’. Using WordNet to expand queries to an information retrieval system, the expansion of computer will include words like estimator and reckoner. Also, the words dog, computer and company all have a sense that is a hyponym of person. Such rare senses make it difficult for a coreference resolution system to use WordNet to enforce the constraint that personal pro- nouns (e.g. he or she) must refer to a person. The second problem with these lexicons is that they miss many do- main specific senses. For example, WordNet misses the user-interface-object sense of the word dialog (as often used in software manuals). WordNet also contains a very poor coverage of proper nouns. There is a need for (semi-) automatic approaches to building and extending ontologies as well as for validat- ing the structure and content of existing ones. With the advent of the Web, we have access to enormous amounts of text. The future of ontology growing lies in leveraging this data by harvesting it for concepts and semantic relationships. Moreover, once such knowledge is discovered, mechanisms must be in place to enrich current ontologies with this new knowledge. To address some of the coverage and specificity problems in WordNet and Cyc, Pantel and Lin (2002) proposed and algorithm, called CBC, for automatically extracting semantic classes. Their classes consist of clustered instances like the three shown below: (A) multiple sclerosis, diabetes, osteoporosis, cardiovascular disease, Parkinson's, rheumatoid arthritis, heart disease, asthma, cancer, hypertension, lupus, high blood pressure, arthritis, emphysema, epilepsy, cystic fibrosis, leukemia, hemophilia, Alzheimer, myeloma, glaucoma, schizophrenia, ... (B) Mike Richter, Tommy Salo, John Vanbiesbrouck, Curtis Joseph, Chris Osgood, Steve Shields, Tom Barrasso, Guy Hebert, Arturs Irbe, Byron Dafoe, Patrick Roy, Bill Ranford, Ed Belfour, Grant Fuhr, Dominik Hasek, Martin Brodeur, Mike Vernon, Ron Tugnutt, Sean Burke, Zach Thornton, Jocelyn Thibault, Kevin Hartman, Felix Potvin, ... (C) pink, red, turquoise, …

Content truncated to 500 words