Reranking for Biomedical Named-Entity Recognition

This paper investigates improvement of automatic biomedical named-entity recognition by applying a reranking method to the COLING 2004 JNLPBA shared task of bioentity recognition. Our system has a common reranking architecture that consists of a pipeline of two statistical classifiers which are based on log-linear models. The architecture enables the reranker to take advantage of features which are globally dependent on the label sequences, and features from the labels of other sentences than the target sentence. The experimental results show that our system achieves the labeling accuracies that are comparable to the best performance reported for the same task, thanks to the 1.55 points of F-score improvement by the reranker.