Given a set of scores between pairs of mentions, form clusters to optimize a coreference metrics.
Luo et al. (2004) apply beam-search over Bell trees which they define as search trees obtained by going from an one-mention root node to full-clusters leave nodes by either starting a new cluster or adding a mention to an existing one. They didn't compare their algorithm to anything else and reported scores on true mentions only. Nicolae and Nicolae (2006) later found no improvement of their method compared to a greedy baseline.
Nicolae and Nicolae (2006) propose BestCut, a slightly modified version of the well-known MinCut algorithm. They show that BestCut works better than beam search and greedy decoding.
- Luo, X., Ittycheriah, A., Jing, H., Kambhatla, N., & Roukos, S. (2004). A mention-synchronous coreference resolution algorithm based on the Bell tree. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics.
- Nicolae, C., & Nicolae, G. (2006). BestCut: A Graph Algorithm for Coreference Resolution. EMNLP 2006, 275–283.