Experiments in Mutual Exclusion Bootstrapping

Experiments in Mutual Exclusion Bootstrapping Tara Murphy and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {tm, james}@it.usyd.edu.au Abstract Mutual Exclusion Bootstrapping (MEB) was designed to overcome the problem of se- mantic drift suffered by iterative bootstrap- ping, where the meaning of extracted terms quickly drifts from the original seed terms (Curran et al., 2007). MEB works by ex- tracting mutually exclusive classes in paral- lel which constrain each other. In this paper we explore the strengths and limitations of MEB by applying it to two novel lexical-semantic extraction tasks: ex- tracting bigram named entities and WordNet lexical file classes (Fellbaum, 1998) from the Google Web 1T 5-grams. 1 Introduction Extracting lexical semantic resources from text with minimal supervision is critical to overcoming the knowledge bottleneck in Natural Language Process- ing (NLP) tasks ranging from Word Sense Disam- biguation to Question Answering. Template-based extraction is attractive because it is reasonably efficient, works on small and large datasets, and requires minimal linguistic pre- processing, making it fairly language independent. Hearst (1992) proposed template-based extraction for identifying hyponyms using templates like X, Y, and/or other Z where X and Y are hyponyms of Z. Riloff and Shepherd (1997) proposed iterative bootstrapping where frequent neighbours to terms from a given semantic class are extracted in multi- ple bootstrap iterations. Roark and Charniak (1998) improved its accuracy by optimising the bootstrap- ping parameters. In mutual bootstrapping (Riloff and Jones, 1999) the terms, and the contexts they oc- cur in, are extracted. Similar approaches have been used in Information Extraction (IE) for identifying company headquarters (Agichtein et al., 2000) and acronym expansions (Sundaresan and Yi, 2000). In Mutual Exclusion Bootstrapping (MEB), we as- sume the semantic classes partition terms into dis- joint sets, that is, the classes are mutually exclusive (Curran et al., 2007). Each class is extracted in par- allel using separate bootstrapping loops that each race to collect terms and contexts. Although this as- sumption is clearly false, it significantly reduces the extraction errors of existing approaches. This paper presents two applications of MEB that allow some insight into MEB’s strengths and limi- tations. First, we extend MEB to extracting bigram BBN named entity types (Weischedel and Brunstein, 2005). We discover that both unigram and bigram MEB are very sensitive to the context window sur- rounding the extracted terms. Surprisingly, MEB is insensitive to the order the semantic classes are pre- sented and the noise in the terms themselves. Second, we extract common nouns using 25 se- mantic classes defined by the WordNet lexical files (Fellbaum, 1998). We use a closed vocabulary of WordNet unigram nouns, so the evaluation can be performed automatically against WordNet. We find that MEB performs well on classes with narrow def- initions and thus more coherent contexts, such as animal, but performs poorly on classes like cogni- tion. We also find that increasing the number of seed terms increases the accuracy significantly. Proceedings of the Australasian Language Technology Workshop 2007, pages 66-74 66 …

Content truncated to 500 words