a negative mi sounds like a bug. as far as i understand, mi can be thought of as the kullback-liebler divergence between an independent and a dependent model. kl diverengence is always postive. (also check Manning + Schutze p 72) Miles