A system chooses which of two candidate nouns correctly fills a sentence based on contextual knowledge. The dataset construction uses filtering intended to reduce exploitable annotation patterns. Accuracy reflects performance on this particular coreference-style task.