Pronominal anaphora understanding

We argue that pronominal anaphora understanding must rely on the recovery of argument structure asymmetries in conjunction with principles restricting the set of possible antecedents for pronouns. We provide empirical evidence for the need of deep parsing recovering arguments, both overt and covert, that can be possible antecedents for pronouns. We identify several limits of systems that do not rely on deep parsing for pronominal anaphora resolution. We point to independent evidence sustaining our claim, and we draw more general consequences for Natural Language Understanding.