Graph Based Unsupervised Word Sense Disambiguation with Semantic Network and Graph Centrality Algorithms

Abstract

This dissertation describes an approach to solve the problem of Word Sense Disambiguation through semantic network of word senses constructed from Word Net as a reference dictionary.Word Sense Disambiguation is the task of assigning a meaning to word, among many available from a dictionary, which is most appropriate dependingupon the context of a word. Starting from the single sense vertex graphs, graphs are expanded by adding semantically related senses from Word Net until a graph containing senses of all words in a sentence is constructed. In a graph thus obtained containing senses of all words, graph-based centrality algorithms are applied in order to figure out the most important sense vertex for each word. The method assumes the same word indifferent sentences may have different meanings. Disambiguation is done by sentencewise, taking sentence as a context for disambiguation. The approach is completely unsupervised and do not require any manually sense tagged resources, but only require a sense repository that provides explicit relation between senses. Weights on edge between senses are assigned using ETF-INF. The method is evaluated by disambiguating texts from standard sense tagged corpora from SemCor and Senseval. The method is suitable for all words domain independent sense disambiguation.