This system had the highest coreference resolution performance on the shared task, and this paper is a write up describing the system.

There were two interesting things about this paper. The first is that it is a state-of-the-art system that is purely deterministic and rule-based. Thus there is no training of the system required, and prediction is probably very fast. The second is the nature of their rule-based system. The idea is that they apply each of their rules one at a time, going from highest precision to progressively lower precision and higher recall, hoping to pick up coreferent mentions that earlier rules (or “sieves”) missed. One of their sieves uses information from Wikipedia or Freebase, which I thought was interesting, but it actually hurt performance a little bit in the end and was disabled for their final submission.