结合CRF的边界组合生物医学命名实体识别

许多的生物医学命名实体识别(Bio-NER)工作都集中于提取扁平化的实体,而忽略了嵌套实体和不连续实体.此外,大多数生物医学命名实体都未遵循统一的命名法,具有许多典型的领域特征,但其使用效率较低.为此提出一种结合CRF的边界组合命名实体识别方法,有效地利用了生物医学实体特征.该方法包括边界检测、边界组合和实体筛选三个步骤.首先使用神经网络模型和基于特征的CRF模型识别实体开始和结束边界,然后经过边界组合产生候选实体,最后使用多输入的卷积神经网络模型对候选实体进行筛选并分类.实验表明,该方法能够有效地识别生物医学文献中的嵌套和不连续实体,在GENIA数据集上达到81.89％的F值.

CRF-combined boundary assembly method for biomedical named entity recognition

Many biomedical named entity recognition(Bio-NER) works focus on extracting flat entities, ignore nested entities and discontinuous entities. In addition most biomedical named entities do not follow a unified nomenclature and have many typical field features, but there is only lower efficiency. To this end, this paper proposed a CRF-combined boundary assembly method, it effectively utilized the features of biomedical entities. This method consisted of three steps: boundary detec ting, boundary assembling and entity discriminating. Firstly, it used a neural network model and a feature-based CRF model to identify the start and end boundaries of entities, and then generated candidate entities through boundary combination. Finally it screened and classified the candidate entities using a multi-input convolutional neural network model. Experiments show that this method can effectively recognize nested and discontinuous entities in biomedical literature, and achieve an F-score of 81 89% on the GENIAdata set.