Factor semantics for document retrieval

Principal component analysis is a useful technique for reducing the dimensionality of datasets, a paramount need in high dimensional term spaces. We study three neural networks with Hebbian-like learning that approximately produce the principal components of a document database. The explained variance of the solutions shows how much information the reduced space retains. In this database the first factor is strongly described by words that are usually included in a stop-word list, other factors show high loadings in content specific terms; this indicates that the networks are learning the semantic space. The retrieval performance on the reduced space is compared with other methods