Computing discourse anaphora from grammatical representation