Spotting topics with the singular value decomposition

Abstract

The singular value decomposition, or SVD, has been studied in the past as a tool for detecting and understanding patterns in a collection of documents. We show how the matrices produced by the SVD calculation can be interpreted, allowing us to spot patterns of characters that indicate particular topics in a corpus. A test collection, consisting of two days of AP newswire traffic, is used as a running example.