Transcription factor (TF) binding sites or motifs (TFBMs) are functional cis-regulatory DNA sequences that play an essential role in gene transcriptional regulation. Although many experimental and computational methods have been developed, finding TFBMs remains a challenging problem. We propose and develop a novel dictionary based motif finding algorithm, which we call WordSpy. One significant feature of WordSpy is the combination of a word counting method and a statistical model which consists of a dictionary of motifs and a grammar specifying their usage. The algorithm is suitable for genome-wide motif finding; it is capable of discovering hundreds of motifs from a large set of promoters in a single run. We further enhance WordSpy by applying gene expression information to separate true TFBMs from spurious ones, and by incorporating negative sequences to identify discriminative motifs. In addition, we also use randomly selected promoters from the genome to evaluate the significance of the discovered motifs. The output from WordSpy consists of an ordered list of putative motifs and a set of regulatory sequences with motif binding sites highlighted. The web server of WordSpy is available at http://cic.cs.wustl.edu/wordspy.