A Study of Sense-Disambiguated Networks Induced from Folksonomies

A Study of Sense-disambiguated Networks Induced from Folksonomies Hans-Peter Zorn and Iryna Gurevych Ubiquitous Knowledge Processing Lab Technische Universit¨at Darmstadt Hochschulstraße 10 64289 Darmstadt, Germany Abstract. Lexical-semantic resources are fundamental building blocks in natural language processing (NLP). Frequently, they fail to cover the informal vocabulary of web users as represented in user-generated content. This paper aims at exploring folksonomies as a novel source of lexical-semantic information. It analyzes two prototypical examples of folk- sonomies, namely BibSonomy and Delicious, and utilizes NLP and word sense induction techniques to turn the folksonomies into word sense–disambiguated networks representing the vocabulary and the word senses found in folksonomies. The main contribution of the paper is an in-depth analysis of the resulting resources, which can be combined with conven- tional wordnets to achieve broad coverage of user-generated content. 1 Introduction Lexical-semantic resources are fundamental building blocks of natural language processing sys- tems. For a long time, WordNet has been widely deployed in a great variety of tasks. A few years ago, the attention of researchers has turned to the so-called collaboratively created lexical- semantic resources such as Wikipedia1 and Wiktionary2 (Gurevych and Wolf, 2010). They have been found to perform well as sources of background knowledge in multiple NLP tasks. While collaboratively created resources represent an excellent addition to conventional lexical- semantic resources, their coverage is insufficient in terms of the represented domains. Also, there is often lack of contextual information about individual sense descriptions. This is why, in this pa- per, we turn to folksonomies as a promising source of lexical-semantic information. Folksonomy is a term used first by Vander Wal (2004) and defined as “tagging is ontology that works”, in ref- erence to social bookmarking systems like Delicious. Such applications allow users to assign tags to web sites or other resources. By doing so, they create a structure consisting of the three entities, “tag, the object being tagged and [the user’s] identity” which is referred to as a folksonomy. Our goal is to derive a resource that can effectively complement conventional resources, e.g., WordNet, as folksonomies are known to contain special vocabulary reflecting users’ interests, current trends, or neologisms typically lacking in conventional resources. To achieve this, we investigate two widely deployed folksonomies, Delicious and BibSonomy. We constructed three types of tag-based graphs utilizing the tag co-occurrence with resources, users, and other tags. Then, we apply graph clustering techniques to perform word sense induction and obtain a word sense–disambiguated lexical-semantic network. Based on this graph, we present a detailed analy- sis of the resulting networks in terms of their graph-theoretic properties and compare them with the properties of conventional resources. We also compare their coverage and find that a folksonomy- derived resource includes web-specific vocabulary that is lacking in other resources. Furthermore, Copyright 2011 by Hans-Peter Zorn and Iryna Gurevych 1 http://wikipedia.org 2 http://wiktionary.org 25th Pacific Asia Conference on Language, Information and Computation, pages 323–332 323 we investigate the nature of the resulting word sense distributions for individual lexemes and relate them to …

Content truncated to 500 words