UNIPI Participation in the Evalita 2011 Anaphora Resolution Task

摘要

The Anaphora Resolution task of Evalita 2011 was intended to measure the ability of participating systems to recognize mentions of the same real-world entity within a given document. The UNIPI system is based on the analysis of dependency parse trees and on similarity clustering. Mention detection relies on parse trees obtained by re-parsing texts with DeSR, and on ad-hoc heuristics to deal with specific cases, when mentions boundaries do not correspond to sub-trees. A binary classifier, based on Maximum Entropy, is used to decide whether there is a coreference relationship between each pair of mentions detected in the previous phase. Clustering of entities is performed by a greedy clustering algorithm.