Anomalies in the WordNet Verb Hierarchy

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 729–736 Manchester, August 2008 Anomalies in the WordNet Verb Hierarchy Tom Richens Aston University Aston Triangle Birmingham B4 7ET, England richentm@aston.ac.uk Abstract The WordNet verb hierarchy is tested, with a view to improving the perform- ance of its applications, revealing topo- logical anomalies and casting doubt on its semantic categories. Encoded tro- ponyms frequently misrepresent other kinds of entailment. Approaches are pro- posed for correcting these anomalies in- cluding a new top ontology. 1 Introduction WordNet is a lexical database widely used for NLP tasks. Any application of WordNet which measures semantic distance employs WordNet relations to do so. As part of a wider project to build an improved and enriched lexical database using existing re- sources, in the hope of improving on the per- formance of WordNet, this study investigates the correctness of the hypernymy/troponymy rela- tions between verbs. The broader project would ideally model these hierarchical relations as trees but for the phe- nomenon of multiple inheritance, which is in- vestigated here to see if it is semantically justi- fied. Moreover it seems intuitively likely that anomalies will be concentrated where the rela- tional structure is more complex. 1.1 Definitions The only document found to specify the Word- Net verbal relations is Fellbaum (1998), who de- fines troponymy (the verb equivalent of hy- ponymy) as a special case of entailment, where © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. the entailing and entailed verb are temporally co- extensive, distinguishing it from causation, tem- poral inclusion, and backward presupposition, of which the last two are not encoded as distinct relations in WordNet, but are combined under the general heading of entailment. The expres- sion sister terms is used to designate a pair of synsets which share a hypernym. This study distinguishes between transitive causation (causing another to do something) and reflexive causation (causing oneself to do something). The term quale (plural: qualia) is borrowed from Pustejovsky (1991). As applied to verbs within the context of this study, the formal quale means what is physically done, while the telic quale means the purpose or intended result of the action. 1.2 Application of WordNet Relations Banerjee & Pedersen (2003) have employed WordNet relations in an extension to the Lesk (1986) algorithm for word sense disambiguation. In order to establish the relatedness of two words, the glosses of their WordNet relatives are compared. Their results are noticeably inferior for verbs than for nouns. Moreover, while the most useful relations for disambiguating nouns were hyponymy and meronymy, in the case of verbs, the example sentences proved more useful than any relations. Their best results were ob- tained by using all relations between verbs indis- criminately. This finding reflects on the distinc- tions between kinds of verb relations and sug- gests that at least where verbs are concerned, the limited success achieved by algorithms relying on WordNet relations arises from the probability …

Content truncated to 500 words