The naive Bayes and maximum entropy approaches to text classification are typically discussed as completely unrelated techniques. In this paper, however, we show that both approaches are simply two different ways of doing parameter estimation for a common log-linear model of class posteriors. In particular, we show how to map the solution given by maximum entropy into an optimal solution for naive Bayes according to the conditional maximum likelihood criterion. Naive Bayes Model We denote the

more »

... ass variable by c = 1, . . . , C, the word variable by d = 1, . . . , D, and a document of length L by d L 1 = d 1 d 2 · · · d L . The joint probability of occurrence of c, ⋆