河北农业大学学报2026，Vol.49Issue(2)：95-107,13.DOI:10.13320/j.cnki.jauh.2026.0025
融合对抗训练与胶囊网络的羊疾病命名实体自动标注方法
An automatic annotation method for sheep disease named entities integrating adversarial training and capsule networks
摘要
Abstract
Disease prevention and control are crucial for the healthy development of the sheep industry.However,weak capabilities of the primary diagnosis and over-reliance on expert experience often lead to delay in disease management.Deep learning-based large models offer an innovative solution for assisted diagnosis,for which sheep disease texts serve as core carriers of knowledge in this field.Named Entity Recognition(NER)is a key step towards automatic text annotation.Nevertheless,sheep disease corpora faces challenges such as imbalanced entity distribution,complex semantic relationships,and ambiguous entity boundaries.To address these issues,this study analyzed the characteristics of sheep disease texts,established annotation rules,and constructed a corpus containing 9 categories and 9,635 entities to support NER tasks,proposing a named entity recognition model for sheep disease texts—SD-ATCN(A Named Entity Recognition Model for Sheep Diseases Texts Integrating Adversarial Training and Capsule Networks).First,the RoBERTa model was used to extract contextual embeddings,which were integrated with adversarial samples generated through adversarial training to form the output of the text embedding layer.This enhanced the model's stability in handling noisy data and imbalanced annotation scenarios.Then,capsule networks were incorporated into the global features extracted by BiLSTM to improve the model's ability to capture complex entity relationships.Finally,a CRF layer decoded the entity labels to achieve automatic entity recognition.Additionally,differential learning rates were applied during training to flexibly adjust learning rates across different layers,accelerating convergence and improving performance.Experimental results showed that the model respectively achieved precision,recall,and F,scores as 93.28％,95.57％,and 94.40％,outperforming the baseline RoBERTa-BiLSTM-CRF model by 2.13％,1.20％,and 1.69％.Based on the SD-ATCN model,an automatic annotation system for sheep disease texts was developed,significantly improving annotation efficiency and providing technical support for automated text processing and intelligent diagnosis in sheep disease management.关键词
羊疾病/命名实体识别/RoBERTa/对抗训练/胶囊网络/差分学习率/自动标注Key words
sheep diseases/named entity recognition/RoBERTa/adversarial training/capsule networks/differential learning rates/automatic annotation分类
信息技术与安全科学引用本文复制引用
张泽嘉,孙小华,王超,王斌,袁万哲,王福顺..融合对抗训练与胶囊网络的羊疾病命名实体自动标注方法[J].河北农业大学学报,2026,49(2):95-107,13.基金项目
河北省重点研发计划项目(22327403D) （22327403D）
河北省现代农业产业技术体系羊产业创新团队专项资金项目(HBCT2024250204). （HBCT2024250204）
