A Novel Framework for Nominal Entity Recognition

The re-ranking algorithm is a common method to use the results of subsequent stages, such as coreference resolution, to improve entity recognition. The nature of re-ranking is to select the most possible candidate from the entire candidate set. But if all of the candidates are incorrect, this method still can not give a right result. We propose a two-layer model to utilize the results of subsequent stages. This novel framework is able to overcome the disadvantage of re-ranking method, and correct the errors introduced by the first tagging. The experiments on the ACE2004 Chinese corpus show that the proposed framework are more effective than re-ranking method.