Weighted Mutual Exclusion Bootstrapping for Domain Independent Lexicon and Template Acquisition

Weighted Mutual Exclusion Bootstrapping for Domain Independent Lexicon and Template Acquisition Tara McIntosh and James R. Curran School of IT University of Sydney NSW 2006, Australia {tara,james}@it.usyd.edu.au Abstract We present the Weighted Mutual Exclu- sion Bootstrapping (WMEB) algorithm for simultaneously extracting precise seman- tic lexicons and templates for multiple cat- egories. WMEB is capable of extracting larger lexicons with higher precision than previous techniques, successfully reduc- ing semantic drift by incorporating new weighting functions and a cumulative tem- plate pool while still enforcing mutual ex- clusion between the categories. We compare WMEB and two state-of-the- art approaches on the Web 1T corpus and two large biomedical literature collections. WMEB is more efficient and scalable, and we demonstrate that it significantly outper- forms the other approaches on the noisy web corpus and biomedical text. 1 Introduction Automatically acquiring semantic lexicons and templates from raw text is essential for overcom- ing the knowledge bottleneck in many natural lan- guage processing tasks, e.g. question answer- ing (Ravichandran and Hovy, 2002). These tasks typically involve identifying named entity (NE) classes which are not found in annotated cor- pora and thus supervised NE recognition mod- els are not always available. This issue be- comes even more evident in new domains, such as biomedicine, where new semantic categories are often poorly represented in linguistic resources, if at all (Hersh et al., 2007). There are two common approaches to extract semantic lexicons: distributional similarity and template-based bootstrapping . In template-based bootstrapping algorithms, templates that express a particular semantic type are used to recognise new terms, and in turn these new terms help iden- tify new templates iteratively (Riloff and Jones, 1999). These algorithms are attractive as they are domain and language independent, require min- imal linguistic preprocessing, are relatively effi- cient, and can be applied to raw text. Unfortunately, semantic drift often occurs when ambiguous or erroneous terms or patterns are introduced into the lexicon or set of templates. Curran et al. (2007) developed Mutual Exclusion Bootstrapping (MEB) to reduce semantic drift by forcing semantic classes to be mutually exclusive. We introduce a new algorithm, Weighted Mu- tual Exclusion Bootstrapping (WMEB), that auto- matically acquires multiple semantic lexicons and their templates simultaneously. It extends on the Curran et al. (2007) assumption of mutual exclu- sion between categories by incorporating a novel cumulative template pool and new term and tem- plate weighting functions. We compare WMEB against two state-of-the- art mutual bootstrapping algorithms, MEB (Cur- ran et al., 2007) and BASILISK (Thelen and Riloff, 2002). We have evaluated the terms and templates these algorithms extract under a range of condi- tions from three raw text collections: noisy web text, biomedical abstracts, and full-text articles. We demonstrate that WMEB outperforms these existing algorithms in extracting precise lexicons and templates from all three datasets. WMEB is significantly less susceptible to semantic drift and so can produce large lexicons accurately and effi- ciently across multiple domains. 2 Background Hearst (1992) pioneered the use of templates for information extraction, focussing on acquiring …

Content truncated to 500 words