UMass Machine Learning and Friends Lunch | Main / Online Ranking Applied To Greedy Coreference Resolution

Coreference resolution is the task of mapping mentions of entities (e.g., noun phrases, person names) into anaphoric sets. This is an important step in the construction of relational databases from unstructured text. In our work we reduce the problem to graph partitioning, which we approximate by a greedy algorithm. We learn a linear scoring function which operates over clusters (rather than pairs) of mentions, and train with an online large-margin method. An error reduction of 30% compared to prior state of the art results is noted.

Presented by Rob Hall