What AVT can make of corpora: some findings from the Pavia Corpus of Film Dialogue