End-to-end Plural Coreference Resolution on TV Show Transcripts

This paper introduces the first plural end-to-end coreference resolution model. This coreference system generates spans embeddings, which are optimized to predict the mentions and the coreferent antecedents. This model handles plural mentions and plural speakers. Our approach builds on the higher-order coreference resolution with coarse-to-fine inference by adapting it to the Friends corpus, which has plural speakers as a feature and also has singletons. Additionally, the model predicts plural antecedents as done in previous plural coreference works. These, in combination with the singular antecedents, are used to construct the final clusters, which have a one-to-one correspondence to the entities.