Learning Selectional Preferences for Use in Resolving Associative Anaphora

Anaphora resolution is well recognised as one of the more important -- and most difficult -- tasks in natural language processing. It has applications in a wide variety of areas machine translation, text summarisation, information extraction, and information retrieval. Most work in the area to date has focussed on cases involving pronouns (and certain full noun phrases) that are coreferent with other noun phrases in a text. This paper presents some preliminary results from experiments aimed at extending the coverage of an anaphora resolution system to deal with certain cases of associative anaphora -- where the anaphor and the expression that allows it to be interpreted (the antecedent) do not refer to the same thing. The technique that has been used involves automatically acquiring semantic relationships from a parsed corpus, using the WordNet thesaurus as a resource to overcome the problem of data sparseness.