Sequence memoizer based model for Biomedical Named Entity Recognition

Biomedical Named Entity Recognition (Bio-NER) plays a fundamental role in biomedical text mining. The improvement of the Bio-NER module's performance always brings notable enhancement to the entire text mining system. Treating the Bio-NER task as a sequence labeling problem, this paper proposes a generative approach to address it. By adopting the spirit of Sequence Memoizer, our approach gives reasonable named entity labels to the elements in the given sequences, based on the better modeling of the context word distributions. Compared with the traditional generative models like Hidden Markov Model, our approach works better on capturing the power-law scaling and the long range dependencies of natural language. Our method is evaluated on the JNLPBA 2004 dataset and compared with the classic generative and discriminative approaches. The experimental results have shown that it outperforms the HMM model and is comparable with the Maxent model.