A Method for Intent Classification of Paddy Field Pest and Disease Questions Based on RoBHBB-Att

Accurate intent classification of paddy field pest- and disease-related questions is crucial for farmers to obtain targeted pest control knowledge and for agricultural question-answering systems to provide effective guidance. To address the lack of publicly annotated datasets in this field and the poor adaptability of general models (due to short question texts and diverse intents), a specialized paddy pest question dataset was constructed, and the RoBHBB-Att intent classification method was proposed in this study, integrating attention with the RoBHBB hybrid architecture. First, RoBERTa efficiently encodes high-dimensional word vectors of paddy pest texts for subsequent feature extraction; after text vectorization, the data is fed into HAN to identify multi-level core semantic elements in questions. Then, Bahdanau BiLSTM mines global context to capture the semantic coherence of questions, which is followed by an attention mechanism that dynamically weights its output features to focus on key semantics and enhance text representation. Finally, the weighted features are subjected to dimensionality reduction via a fully connected layer, and the softmax classifier outputs the final intents. Experiments verify RoBHBB-Att’s excellent classification accuracy and practical value on the dataset.