针对临床医疗记录中的复杂语义实体和长短距离依赖关系识别准确率低的难题,文章提出了一种双向语义与残差注意力网络的医疗文本命名实体识别方法。利用BERT-wwm预训练模型捕捉语义特征,结合双向门控循环单元BiGRU用于处理复杂长程语义关...针对临床医疗记录中的复杂语义实体和长短距离依赖关系识别准确率低的难题,文章提出了一种双向语义与残差注意力网络的医疗文本命名实体识别方法。利用BERT-wwm预训练模型捕捉语义特征,结合双向门控循环单元BiGRU用于处理复杂长程语义关联;增加残差连接的注意力Attention结构,保障专注于关键信息的同时,不会丢失捕捉到的整体序列特征;条件随机场CRF负责最后的序列标注预测,对前序多层神经网络抽取的特征序列进行最优路径解码。实验结果表明,通过本方法能够有效提升命名实体识别的准确率。Aiming at the challenge of low recognition accuracy for complex semantic entities and long- and short-range dependencies in clinical medical records, this paper proposes a medical text named entity recognition method that integrates bidirectional semantics with a residual attention network. The method leverages the BERT-wwm pre-trained model to capture semantic features and combines it with a Bidirectional Gated Recurrent Unit (BiGRU) to handle complex long-range semantic associations. An Attention mechanism with residual connections is added to ensure focus on key information while preserving the overall sequence characteristics captured. A Conditional Random Field (CRF) is responsible for the final sequence labeling prediction, performing optimal path decoding on the feature sequences extracted by the preceding multi-layer neural networks. Experimental results demonstrate that this approach can effectively improve the accuracy of named entity recognition.展开更多
文摘针对临床医疗记录中的复杂语义实体和长短距离依赖关系识别准确率低的难题,文章提出了一种双向语义与残差注意力网络的医疗文本命名实体识别方法。利用BERT-wwm预训练模型捕捉语义特征,结合双向门控循环单元BiGRU用于处理复杂长程语义关联;增加残差连接的注意力Attention结构,保障专注于关键信息的同时,不会丢失捕捉到的整体序列特征;条件随机场CRF负责最后的序列标注预测,对前序多层神经网络抽取的特征序列进行最优路径解码。实验结果表明,通过本方法能够有效提升命名实体识别的准确率。Aiming at the challenge of low recognition accuracy for complex semantic entities and long- and short-range dependencies in clinical medical records, this paper proposes a medical text named entity recognition method that integrates bidirectional semantics with a residual attention network. The method leverages the BERT-wwm pre-trained model to capture semantic features and combines it with a Bidirectional Gated Recurrent Unit (BiGRU) to handle complex long-range semantic associations. An Attention mechanism with residual connections is added to ensure focus on key information while preserving the overall sequence characteristics captured. A Conditional Random Field (CRF) is responsible for the final sequence labeling prediction, performing optimal path decoding on the feature sequences extracted by the preceding multi-layer neural networks. Experimental results demonstrate that this approach can effectively improve the accuracy of named entity recognition.
