Like any learner, the naive Bayesian classifier manipulates descriptions of individuals. The classical naive Bayesian classifier uses an attribute-value language, a representation formalism that is commonly used in machine learning. Logically speaking an attribute-value language can be mapped to unary predicate logic, where hypotheses use a single universally quantified variable to express generalisation over all individuals, and examples are variable-free conjunctions concerning single individuals. Whereas this representation is, strictly speaking, not propositional, a learning system keeping track of the distinction between examples and hypotheses can actually drop the syntactic distinction and express both in a variable-free, essentially propositional formalism (the single representation trick). In Section 2.1 we recall the propositional naive Bayesian classifier. In Section 2.2 we discuss the general problem of upgrading it to deal with non-propositional representations.