This paper concerns supervised classification of text. Rocchio, the method we choose for its efficiency and extensibility, is tested on three reference corpora “20NewsGroups”, “OHSUMED” and “Reuters”, using several similarity measures. Analyzing statistical results, many limitations are identified and discussed. In order to overcome these limitations, this paper presents two main solutions: first constituting Rocchio-based classifier committees, and then using semantic resources (ontologies) in order to take meaning into consideration during text classification. These two approaches can be combined in a Rocchio-based semantic classifier committee.