Agricultural Text Classification Method Based on ERNIE 2.0 and Multi-Feature Dynamic Fusion

Under the rapid development of agricultural information technology, there has been an explosive growth in agricultural Q&A texts on the Internet, making the accurate classification of these texts a critical research topic in this field. However, the dispersion, complexity, and difficulty in collecting agricultural data pose significant challenges to achieving efficient and accurate text classification. To address this, this paper proposes a text classification model based on multi-feature dynamic fusion—ERNIE 2.0_Bi-GRU_CapsNet_Att. The model uses the ERNIE 2.0 pre-training framework to encode the dataset to learn the dynamic semantic representations of words; it combines Bidirectional Gated Recurrent Units (Bi-GRU) and Capsule Networks to capture high-dimensional sequential and local features of the text; meanwhile, it enriches feature inputs by constructing artificial numerical feature vectors. With the introduction of an attention mechanism, the model can dynamically adjust feature weights to achieve effective multi-feature fusion. We validated the effectiveness of this fusion model on a self-constructed agricultural Q&A dataset. Experimental results, averaged over five runs, show that the proposed model achieved 95.1% precision (±0.1), 94.8% recall (±0.2), and 95.2% F1-score (±0.1), representing a 1.1% improvement in precision, a 0.6% improvement in recall, and a 0.7% improvement in F1-score compared to the best-performing baseline model.These results demonstrate that the proposed model significantly outperforms traditional classification models and achieves robust performance across experiments.