Performance on test2000 of maximum entropy classifier for ranked classifications. The maximum entropy classifier assigns for each code a probability of its relevance to the unclassified document. We ranked each code by its probability for the documents in test2000 and have calculated accuracy for each rank (light gray bars). Some of the documents in test2000 have multiple correct classifications. The articles with more than a single correct classification were removed and accuracy was recalculated for each rank (dark gray bars). Although the accuracy of the highest-rank prediction is only slightly reduced from 72.8% to 72.0%, the accuracy of the second- and third-ranked classes is somewhat more reduced from 17.7% to 13.7% and 6.2% to 4.2%, respectively.
