基于改进图注意力网络的医疗领域命名实体识别

Named Entity Recognition in Medical Field Based on Improved Graph Attention Network

面向医疗领域的命名实体识别是实现医疗数字化的重要任务之一,领域相关资源文本实体数量庞大且词汇构成复杂多样,存在着严重的实体嵌套及词边界模糊问题。以原协作图网络为基础,提出一种基于改进图注意力网络的命名实体识别方法。该方法融合多种特征交互逻辑,完成整体交互关系图的构造,进而结合图注意力网络实现多特征的自适应聚合,降低了计算开销,提升了多特征传导交互效率。此外,模型还使用一种具备方向感知的TENER模型建模文本深层次依赖,强化方向、位置敏感型特征的捕获能力,进一步提升了模型在限定领域实体识别任务中的性能表现。实验结果表明,所提方法相比于原模型在CCKS-2019和Resume数据集上的F 1值分别取得了0.0071和0.0025的提升,证明了改进的有效性。

Named entity recognition for medical field is one of the important tasks to realize medical digitization.The number of text entities of domain related resources is large and the lexical composition is complex and diverse,and there are serious problems of entity nesting and word boundary ambiguity.Based on the original cooperative graph network,this paper proposes a named entity recognition method based on the improved graph attention network.This method integrates multiple feature interaction logic to realize the construction of an overall interactive graph,and then combines the graph attention network to realize the adaptive aggregation of multi-features,which reduces the computing cost and improves the efficiency of multi-feature transmission interaction.In addition,the model also uses a directional sensing TENER model to model the deep dependency of text,strengthening the ability to capture directional and location-sensitive features,and further improving the performance of the model in the limited domain entity recognition task.Experiments show that compared with the original model,the F 1 values of CCKS-2019 and Resume data sets are improved by 0.0071 and 0.0025 respectively,which proves the effectiveness of the improved method.