A maximum entropy approach to biomedical named entity recognition
Yi-Feng Lin, Richard Tzong‐Han Tsai, Wen‐Chi Chou, Kuen-Pin Wu, Ting‐Yi Sung, Wen-Lian Hsu · 2004
Machine learning approaches are frequently used to solve name entity (NE) recognition (NER). In this paper we propose a hybrid method that uses maximum entropy (ME) as the underlying machine learning method incorporated with dictionary-based and rule-based methods for post-processing. Simply using ME for NER, inaccurate boundary detection of NEs and misclassification may occur. Some NEs are partially recognized by ME. In the post-processing stage, we use dictionary-based and rule-based methods to extend boundary of partially recognized NEs and to adjust classification. We use GENIA corpus 3.01 to conduct 10-fold crossverification experiments. To evaluate the performance, we consider the longest NE annotations. We evaluate our approach using standard precision (P), recall (R), and F-score, where F-score is defined as 2PR/(P+R). The precision, recall and F-score ([P, R, F]) of our ME module for overall 23 categories is [0.512, 0.538, 0.525], and after the postprocessing the performance becomes [0.729, 0.711, 0.72] for [P, R, F]. For protein, DNA and RNA classes, our method achieves [P, R, F] of [0.77, 0.80, 0.785], [0.653, 0.748, 0.7], and [0.716, 0.788, 0.752], respectively. The post-processing stage significantly improves the performance of our MEbased NER module. 1.