Growing Multi-Domain Glossaries from a Few Seeds using Probabilistic Topic Models

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 170–181, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Growing Multi-Domain Glossaries from a Few Seeds using Probabilistic Topic Models Stefano Faralli and Roberto Navigli Dipartimento di Informatica Sapienza Universit`a di Roma {faralli,navigli}@di.uniroma1.it Abstract In this paper we present a minimally- supervised approach to the multi-domain ac- quisition of wide-coverage glossaries. We start from a small number of hypernymy rela- tion seeds and bootstrap glossaries from the Web for dozens of domains using Probabilis- tic Topic Models. Our experiments show that we are able to extract high-precision glos- saries comprising thousands of terms and def- initions. 1 Introduction Dictionaries, thesauri and glossaries are useful sources of information for students, scholars and ev- eryday readers, who use them to look up words of which they either do not know, or have forgotten, the meaning. With the advent of the Web an increas- ing number of dictionaries and technical glossaries has been made available online, thereby speeding up the definition search process. However, finding definitions is not always immediate, especially if the target term pertains to a specialized domain. Indeed, not even well-known services such as Google Define are able to provide definitions for scientific or tech- nical terms such as taxonomy learning or distant su- pervision in AI or figure-four leglock and suspended surfboard in wrestling. Domain-specific knowledge of a definitional na- ture is not only useful for humans, it is also use- ful for machines (Hovy et al., 2013). Examples include Natural Language Processing tasks such as Question Answering (Cui et al., 2007), Word Sense Disambiguation (Duan and Yates, 2010; Faralli and Navigli, 2012) and ontology learning (Velardi et al., 2013). Unfortunately, most of the Web dictionar- ies and glossaries available online comprise just a few hundred definitions, and they therefore provide only a partial view of a domain. This is also the case with manually compiled glossaries created by means of collaborative efforts, such as Wikipedia.1 The coverage issue is addressed by online aggrega- tion services such as Google Define, which bring to- gether definitions from several online dictionaries. However, these services do not classify textual def- initions by domain: they just present the collected definitions for all the possible meanings of a given term. In order to automatically obtain large domain glossaries, in recent years computational approaches have been developed which extract textual defi- nitions from corpora (Navigli and Velardi, 2010; Reiplinger et al., 2012) or the Web (Velardi et al., 2008; Fujii and Ishikawa, 2000). The methods in- volving corpora start from a given set of terms (pos- sibly automatically extracted from a domain cor- pus) and then harvest textual definitions for these terms from the input corpus using a supervised sys- tem. Web-based methods, instead, extract text snip- pets from Web pages which match pre-defined lex- ical patterns, such as “X is a Y”, along the lines of Hearst (1992). These approaches typically per- form with high precision and low …

Content truncated to 500 words