Taxonomy Learning Using Word Sense Induction

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 82–90, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Taxonomy Learning Using Word Sense Induction Ioannis P. Klapaftis Department of Computer Science The University of York York, UK, YO10 5DD giannis@cs.york.ac.uk Suresh Manandhar Department of Computer Science The University of York York, UK, YO10 5DD suresh@cs.york.ac.uk Abstract Taxonomies are an important resource for a variety of Natural Language Processing (NLP) applications. Despite this, the current state- of-the-art methods in taxonomy learning have disregarded word polysemy, in effect, devel- oping taxonomies that conflate word senses. In this paper, we present an unsupervised method that builds a taxonomy of senses learned automatically from an unlabelled cor- pus. Our evaluation on two WordNet-derived taxonomies shows that the learned taxonomies capture a higher number of correct taxonomic relations compared to those produced by tradi- tional distributional similarity approaches that merge senses by grouping the features of each word into a single vector. 1 Introduction A concept or a sense, s, can be defined as the mean- ing of a word or a multiword expression. A con- cept s can be linguistically realised by more than one word while at the same time a word w can be the lin- guistic realisation of more than one concept. Given a set of concepts S, taxonomy learning is the task of hierarchically classifying the elements in S in an au- tomatic manner. For example, consider a set of con- cepts linguistically realised by the words/multiword expressions LAN, computer network, internet, mesh- work, gauze, snood. Taxonomy learning methods produce taxonomies, such as the ones shown in Fig- ures 1 (a) and 1 (b). By observing Figure 1 (a), we can express IS- A statements, such as Internet IS-A Computer Net- work etc. However, the same does not apply to the Figure 1: A labelled and an unlabelled concept taxonomy taxonomy in Figure 1 (b), since this taxonomy is not fully labelled. Despite this, its hierarchical organ- isation clearly shows that the concepts are divided into groups, which are further subdivided into sub- groups and so forth, until we reach a level where each concept belongs to its own group. Unlabelled taxonomies are typically produced by agglomera- tive hierarchical clustering algorithms (King, 1967; Sneath and Sokal, 1973). The knowledge encoded in taxonomies can be utilised in a range of NLP applications. For in- stance, taxonomies can be used in information re- trieval to expand a user query with semantically re- lated words or to enhance document representation by abstracting from plain words and adding concep- tual information (Cimiano, 2006). WordNet’s (Fell- baum, 1998) taxonomic relations have also been used in Word Sense Disambiguation (WSD) (Nav- igli and Velardi, 2004b). In named entity recog- nition, methods relying on gazetteers could make 82 use of automatically acquired taxonomies (Cimiano, 2006), while question answering systems have also benefited (Moldovan and Novischi, 2002). Despite the wide uses of taxonomies, the majority of methods disregard or …

Content truncated to 500 words