Classifying Taxonomic Relations between Pairs of Wikipedia Articles

International Joint Conference on Natural Language Processing, pages 788–794, Nagoya, Japan, 14-18 October 2013. Classifying Taxonomic Relations between Pairs of Wikipedia Articles Or Biran Columbia University Department of Computer Science orb@cs.columbia.edu Kathleen McKeown Columbia University Department of Computer Science kathy@cs.columbia.edu Abstract Natural language generation systems rely on taxonomic thesauri for tasks such as lexical choice and aggregation. WordNet is one such taxonomy, but it is limited in size. Motivated by the needs of a genera- tion system in the scientific literature do- main, we present a method for building a taxonomic thesaurus from Wikipedia arti- cles, where each article represents a poten- tial concept in the taxonomy. We propose framing the problem of creating a taxon- omy as a classification task of the potential relations between individual Wikipedia ar- ticle pairs, and show that a supervised al- gorithm can achieve high precision in this task with very little training data. 1 Introduction Thesauri are useful resources for many NLP appli- cations. In particular, taxonomic thesauri which contain synonymy and hypernymy relations are important for natural language generation (NLG) systems which must make decisions regarding lex- ical choice and aggregation. WordNet (Fellbaum, 1998) is one such thesaurus which has many uses in generation (Jing, 1998), but its set of concepts (called synsets) is quite limited. It does not contain many domain-specific concepts, nor does it con- tain technical concepts that emerged very recently. This work is motivated by the needs of a NLG sys- tem in the scientific literature domain, where these missing concepts are absolutely necessary for any practical application. Our goal is to generate a the- saurus containing synonomy and hypernomy rela- tions between scientific terms which a generation system can use to select the most appropriate term given a context. The English Wikipedia has over 4 million arti- cles, and over 8.6 million titles if redirects, which are alternative titles for the articles, are included. These titles are essentially lexical terms referring to concepts. Crucially, it contains articles describ- ing many domain-specific concepts, and, in partic- ular, scientific and technological concepts. For ex- ample, Wikipedia contains articles with titles such as Supersymmetric String Theory, Gorilla Glass and Sentiment Analysis, all of which are missing from WordNet. While there have been attempts to build ontologies from Wikipedia, these tended to focus (in their optimization and evaluation) on entities such as people, places and events. There is still a need for a WordNet-like taxonomy which would contain accurate synonymy and hypernomy relations for highly specialized terms from vari- ous scientific domains (for our purposes) and other specialized domains. Unlike previous approaches, which tend to rely on WordNet’s hierarchy and/or on Wikipedia’s pseudo-hierarchy of categories, we frame the problem as a binary classification task for a pair of Wikipedia article titles - deciding whether the term representing the concept in the first article is a hypernym of the term representing the second or not. This enables us to handle specialized con- cepts which are far from the established concepts in the …

Content truncated to 500 words