Enhancing Coreference Classifiers Using a Ranking-Aware Feature
Khai Nguyen, Ryutaro Ichise · 2017
A coreference refers to different instances of the same real-world entity. Coreference classification is an important problem in knowledge and data management. The basic idea is to predict whether two instances are matched or non-matched, based on their similarity vector. Previous efforts on coreference classification share a common weakness. It is the unawareness of the ambiguity variation among different clusters of instance-pairs. We discuss the importance of clusterwise instance-pairs local ranking, which is an effective solution for the ambiguity variation. Since then, we study the inclusive possibility of the ranking factor in a classifier globally trained from all clusters. Finally, we propose to include an extra element in the original similarity vector of the instancepairs. Such extra element is a ranking-aware feature, which represents the preference of an instance-pair against its cluster. The experiment results confirm that the proposed feature significantly boosts the performance of many classifiers.