A computer-implemented method for processing data includes receiving an initial set of records including terms describing respective items in specified categories. Based on the initial set of records, respective term weights are calculated for at least some of the terms with respect to at least some of the categories. Each term weight indicates, for a given term and a given category, a likelihood that a record containing the given term belongs to the given category. Upon receiving a new record, not included in the initial set, respective assignment metrics are computed for two or more of the categories using the respective term weights of the particular terms in the new record with respect to the two or more of the categories. The new record is classified in one of the two or more of the categories responsively to the assignment metrics.