The rapid development of microarray and other genomic technologies now enables biologists to monitor the expression of hundreds, even thousands of genes in a single experiment. Interpreting the biological meaning of the expression patterns still relies largely on biologist's domain knowledge, as well as on information collected from the literature and various public databases. Yet individual experts’ domain knowledge is insufficient for large data sets, and collecting and analyzing this information manually from the literature and/or public databases is tedious and time-consuming. Computer-aided functional analysis tools are therefore highly desirable. We describe the architecture of GeneNarrator, a text mining system for functional analysis of microarray data. This system’s primary purpose is to test the feasibility of a more general system architecture based on a two-stage clustering strategy that is explained in detail. Given a list of genes, GeneNarrator collects abstracts about them from PubMed, then clusters the abstracts into functional topics in a first clustering stage. In the second clustering stage, the genes are clustered into groups based on similarities in their distributions of occurrence across topics. This novel two-stage architecture, the primary contribution of this project, has benefits not easily provided by onestage clustering.