This paper proposes two ensemble approaches to learning-based coreference resolution. Experimental results on the ACE data sets demonstrate that both approaches outperform a non-ensemble-based coreference system significantly.