A Graph Model for Unsupervised Lexical Acquisition

A Graph Model for Unsupervised Lexical Acquisition Dominic Widdows and Beate Dorow Center for the Study of Language and Information 210 Panama Street Stanford University Stanford CA 94305-4115 {dwiddows,beate}@csli.stanford.edu Abstract This paper presents an unsupervised method for assembling semantic knowledge from a part-of- speech tagged corpus using graph algorithms. The graph model is built by linking pairs of words which participate in particular syntactic relationships. We focus on the symmetric rela- tionship between pairs of nouns which occur to- gether in lists. An incremental cluster-building algorithm using this part of the graph achieves 82% accuracy at a lexical acquisition task, eval- uated against WordNet classes. The model nat- urally realises domain and corpus specific am- biguities as distinct components in the graph surrounding an ambiguous word. 1 Introduction Semantic knowledge for particular domains is increasingly important in NLP. Many applica- tions such as Word-Sense Disambiguation, In- formation Extraction and Speech Recognition all require lexicons. The coverage of hand- built lexical resources such as WordNet (Fell- baum, 1998) has increased dramatically in re- cent years, but leaves several problems and challenges. Coverage is poor in many criti- cal, rapidly changing domains such as current affairs, medicine and technology, where much time is still spent by human experts employed to recognise and classify new terms. Most languages remain poorly covered in compari- son with English. Hand-built lexical resources which cannot be automatically updated can of- ten be simply misleading. For example, using WordNet to recognise that the word apple refers to a fruit or a tree is a grave error in the many situations where this word refers to a computer manufacturer, a sense which WordNet does not cover. For NLP to reach a wider class of appli- cations in practice, the ability to assemble and update appropriate semantic knowledge auto- matically will be vital. This paper describes a method for arranging semantic information into a graph (Bollob´as, 1998), where the nodes are words and the edges (also called links) represent relationships be- tween words. The paper is arranged as follows. Section 2 reviews previous work on semantic similarity and lexical acquisition. Section 3 de- scribes how the graph model was built from the PoS-tagged British National Corpus. Section 4 describes a new incremental algorithm used to build categories of words step by step from the graph model. Section 5 demonstrates this algo- rithm in action and evaluates the results against WordNet classes, obtaining state-of-the-art re- sults. Section 6 describes how the graph model can be used to recognise when words are poly- semous and to obtain groups of words represen- tative of the different senses. 2 Previous Work Most work on automatic lexical acquisition has been based at some point on the notion of semantic similarity. The underlying claim is that words which are semantically similar occur with similar distributions and in similar con- texts (Miller and Charles, 1991). The main results to date in the field of au- tomatic lexical acquisition are concerned with extracting lists of words reckoned to …

Content truncated to 500 words