BBA Agricultural Text Classification Model with Confrontation Training

To address the issues of poor consistency in data augmentation and potential shortcomings in generalization ability of existing text warning models when facing disturbances, this paper proposes an innovative agricultural text warning model. The new model combines contrastive learning with adversarial training methods. By integrating these two techniques, it enhances the capability to identify potential risks in agricultural texts. In agricultural intelligent knowledge services, this approach can achieve real-time monitoring and analysis of agricultural text data, providing timely risk warnings and decision support for farmers and practitioners. Additionally, the model can be used to analyze and predict information such as agricultural policies and market trends, offering more targeted advice and guidance for agricultural practitioners, thereby improving decision-making abilities and market competitiveness. Experimental results demonstrate that, compared to other models such as the BERT-BiLSTM-Attention model (BBA), the BERT-BiLSTM model with integrated adversarial training (AT-BERT-BiLSTM), and the BERT-BiLSTM model, the BBA model with integrated adversarial training performs the best among all models. It achieves an average precision of 87.97%, an average recall of 88.55%, and a harmonic mean of 88.26%. Compared to the existing agricultural text classification model, MacBERT-DPCNN with Attention Fusion (MDA), the BBA agricultural text classification model with integrated adversarial training demonstrates greater stability and consistency across different categories.