Item

Abstract

We introduce the \emph{synonymy graph} as a new angle of looking at spectral retrieval techniques, including latent semantic indexing (LSI) and its many successors. The synonymy graph is defined for each pair of terms in the collection, and our findings suggest that it is at the heart of what makes spectral retrieval work in practice. % We show that LSI and many of its variants can be equivalently viewed as a particular document expansion (not query expansion) process, where each term effects the insertion of some other term if and only if the synonymy graph for that term pair has a certain characteristic shape. We provide a simple, parameterless algorithm for detecting that shape. % We point out inherent problems of every algorithm that bases its expansion decisions merely on individual values of the synonymy graph, as done by almost all existing methods. Our new algorithm overcomes these limitations, and it consistently outperforms previous methods on a number of test collections. % Our synonymy graphs also shed light on the effectiveness of three fundamental types of variations of the basic LSI scheme.