Matching utterances to rich knowledge structures to acquire a model of the speaker's goal