Coreference resolution is a classic unsolved problem in natural language processing. We present a novel antecedent ranking model based on hierarchical recurrent neural networks (RNN). The word-level RNN encodes the context into the representation of mention. The mention-level network is trained to learn to exploit these useful representation and few hand-crafted features to detect anaphora and its antecedent by simple attention mechanism. We evaluate our system on CoNLL 2012 shared task and set up a new state-of-the-art.