The Anaphor-Antecedent Match: Issues for Referring Expressions Generation

The design of referring expressions generation algorithms that produce descriptions closer to human performance may benefit from evidence on how human speakers actually interpret these descriptions. To shed light on this issue, in this paper we present an experiment to measure the time required for the interpretation of matching and non-matching anaphoric definite description pairs. Preliminary results suggest that non-matching descriptions, or those that share fewer attributes with the antecedent term, seem to be more difficult to interpret, an insight that may be potentially useful to the design of novel referring expressions generation algorithms.