On the Performance of Latent Semantic Indexing-based Information Retrieval

Abstract

Conventional vector based Information Retrieval (IR) models, Vector Space Model (VSM) and Generalized Vector Space Model (GVSM), represents documents and queries as vectors in a multidimensional space. This high dimensional data places great demands for computing resources. To overcome these problems, Latent Semantic Indexing (LSI): a variant of VSM, projects the documents into a lower dimensional space, computed via Singular Value Decomposition. It is stated in IR literature that LSI model is 30% more effective than classical VSM models. However statistical significance tests are required to evaluate the reliability of such comparisons. But to the best of our knowledge significance of performance of LSI model is not analyzed so far. Focus of this paper is to address this issue. We discuss the tradeoffs of VSM, GVSM and LSI and empirically evaluate the difference in performance on four testing document collections. Then we analyze the statistical significance of these performance differences.