Using LSI to evaluate the quality of hypertext links

Useful hypertext is constrained by the need for users to be able to nd documents about similar topics without extensive navigation. We show how examining the properties of a graph built by a document's hypertext links can be used to evaluate the usefulness of the document. To formally measure the quality of hypertext linking in a corpus, we compare the semantic similarity of pairs of documents with the minimum number of links between their corresponding nodes in an analogous hypertext graph. We use the measure of document-todocument similarity computed using latent semantic indexing as our measure of semantic similarity. Our method has been applied to a corpus composed of Usenet messages.