The field of molecular biology is growing at an astounding rate and research findings are being deposited into public databases, such as Swiss-Prot. Many of the over 200,000 protein entries in Swiss-Prot 49.1 lack annotations such as subcellular localization or function, but the vast majority have references to journal abstracts describing related research. These abstracts represent a huge amount of information that could be used to generate annotations for proteins automatically. Training classifiers to perform text categorization on abstracts is one way to accomplish this task. We present a method for improving text classification for biological journal abstracts by generating additional text features using the knowledge represented in a biological concept hierarchy (the Gene Ontology). The structure of the ontology, as well as the synonyms recorded in it, are leveraged by our simple technique to significantly improve the F-measure of subcellular localization text classifiers by as much as 0.077 and we achieve F-measures as high as 0.935.