Using Curvature and Markov Clustering in Graphs for Lexical Acquisition and Word Sense Discrimination

We introduce two different approaches for clustering semantically similar words. We accommodate ambiguity by allowing a word to belong to several clusters. Both methods use a graph-theoretic representation of words and their paradigmatic relationships. The first approach is based on the concept of curvature and divides the word graph into classes of similar words by removing words of low curvature which connect several dispersed clusters. The second method, instead of clustering the nodes, clusters the links in our graph. These contain more specific contextual information than nodes representing just words. In so doing, we naturally accommodate ambiguity by allowing multiple class membership. Both methods are evaluated on a lexical acquisition task, using clustering to add nouns to the WordNet taxonomy. The most effective method is link clustering.