Issues in Anaphora Annotation: The Case of Encyclopaedic Animal Descriptions

Anaphora resolution has been a widely studied problem in natural language processing almost since work in the field began (see [Hirst 1981, Mitkov 1999] for summaries). However, most work focuses on the resolution of pronouns, or, at best, on a wider range of anaphoric expressions which co-refer with their antecedents. Real natural language texts exhibit a much richer range of anaphoric phenomena. In this paper, we describe an experiment in annotating a corpus for the variety of anaphors found, and indicate some of the problematic cases that robust natural language processing techniques have to deal with. As a result of this analysis we have identified several subtypes of associative anaphoric relationships that occur frequently in the corpus, and determined that it may be possible to identify the most common relationships using some fairly simple heuristics.