The coef_ and intercept_ on MultinomialNB don't behave like other linear models because they contain probabilities even in the binary case. They should contain log-odds ratios.

The ones in BernoulliNB (not documented) are even worse: they're not the coef and intercept used to make the actual prediction (see BernoulliNB._joint_log_likelihood) so they can't be used for model inspection. This is a hairy issue in terms of backward compat.