ResiAdvNet: A named entity recognition model for potato diseases and pests based on progressive residual structures and adversarial training

Conventional named entity recognition methods based on pretrained models often focus on utilizing the output of the final layer of a pretrained model while ignoring the linguistic features embedded in its internal layers. To utilize pretrained models more fully, this paper proposes a named entity recognition model called ResiAdvNet, which combines progressive residual structures with adversarial training. This proposed model is applied to potato disease and pest identification. MacBERT is utilized as the pretrained model and residual blocks are used to aggregate the outputs of its internal layers, obtaining a final output that emphasizes information from all layers. This output is subsequently fed into a bidirectional long-short term memory network for context modeling and finally passed through a conditional random field to obtain the globally optimal tagging sequence. Additionally, adversarial training is introduced as a means to enhance model robustness. By introducing adversarial examples during training, the model learns more robust feature representations, thereby improving its performance when facing unknown inputs. ResiAdvNet was tested with on a custom dataset called PpdKED over five trials with an average F1 score of approximately 0.9225, significantly outperforming other models. Experimental results demonstrate that the proposed model can efficiently extract entities related to potato diseases and pests, laying a solid foundation for the subsequent tasks of relation extraction and knowledge graph construction.