Revisiting Taxonomy Induction over Wikipedia

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2300–2309, Osaka, Japan, December 11-17 2016. Revisiting Taxonomy Induction over Wikipedia Amit Gupta EPFL Lausanne, Switzerland amit.gupta@epfl.ch Francesco Piccinno University of Pisa Pisa, Italy piccinno@di.unipi.it Mikhail Kozhevnikov Google Inc. Zurich, Switzerland qnan@google.com Marius Pa¸sca Google Inc. Mountain View, California mars@google.com Daniele Pighin Google Inc. Zurich, Switzerland biondo@google.com Abstract Guided by multiple heuristics, a unified taxonomy of entities and categories is distilled from the Wikipedia category network. A comprehensive evaluation, based on the analysis of upward gen- eralization paths, demonstrates that the taxonomy supports generalizations which are more than twice as accurate as the state of the art. The taxonomy is available at http://headstaxonomy.com. 1 Introduction Motivation. As possibly the largest resource of publicly available, semi-structured knowledge, Wikipedia (Remy, 2002) serves as a stepping stone towards the construction of collections of structured data (Remy, 2002; Hoffart et al., 2013; Vrande˘ci´c and Krötzsch, 2014). Data within Wikipedia benefits from new additions and distributed curation by human editors, and has proven beneficial in text analysis tasks ranging from co-reference resolution (Ratinov and Roth, 2012), word sense (Mihalcea, 2007) and entity disambiguation (Ratinov et al., 2011), to information retrieval (Hu et al., 2009) and information extraction (Wu and Weld, 2010; Nastase and Strube, 2013; Hoffart et al., 2013; Dong et al., 2014). Wikipedia links millions of entities (e.g. Barack Obama) to thousands of inter-connected categories of different granularity (e.g. Presidents of the United States, Political office-holders, Politicians) to form what is often referred to as the Wikipedia category network (WCN). However, obtaining a taxonomy of increasingly general categories from WCN is by no means trivial because upward edges in WCN, from entities to categories and also from child to parent categories, are not confined to is-a relations (Ponzetto and Strube, 2007). In fact, consistently discarding not-is-a edges such as Japan↝660 BC or Award win- ners↝Awards, while retaining as many true is-a edges as possible, has been the object of a steady body of research (Ponzetto and Strube, 2007; Hovy et al., 2013; Flati et al., 2014). Recent methods still produce taxonomies with glaring gaps in precision and coverage. More importantly, even if the methods correctly identify individual is-a edges with an accuracy as high as 85% (Flati et al., 2014), it is not uncommon for upward paths to traverse at least some incorrect edges. The resulting taxonomies transitively connect entities (e.g., Natural language processing) to many ancestor categories (e.g., Physical body, Mass)1 that are not true generalizations, thus limiting their utility in practice. Contributions. This paper proposes a novel method for taxonomy induction from WCN. As described in Section 3, the method exploits syntactic evidence in category titles to connect entities (i.e., pages) with increasingly more general categories. A novel, comprehensive framework for taxonomy evaluation is proposed, focusing on the accuracy and granularity of longer generalization paths, as opposed to individual edges. Section 4 describes the evaluation framework and carries out an in-depth comparison of the proposed taxonomy against the …

Content truncated to 500 words