Abstract:

[Background] Agricultural science data serve as a key foundation for supporting agricultural technological innovation and modern agricultural development. These datasets are continuously expanding, feature highly specialized textual descriptions, and have hierarchical subject labels, which pose significant challenges for classification. [Objective] To address the complexity of hierarchical labels in agricultural science data and the low efficiency caused by manual annotation, this study develops a label-semantic-aware hierarchical multi-label classification model, Bert-BiGRU-HiGCN, aiming to achieve efficient automated classification and improve data management and retrieval. [Methods] An experimental dataset is constructed based on metadata from the National Agricultural Science Data Center. In the model, a Graph Convolutional Network was used to capture the hierarchical structure of agricultural subject labels and extract label features containing hierarchical information. A cross-attention mechanism was then applied to fuse textual and label features, obtaining text representations enriched with label semantics. The improved model was evaluated on the constructed dataset through comparative experiments. [Results] Experimental results indicate that the proposed Bert-BiGRU-HiGCN model achieves a precision of 78.1%, recall of 76.3%, and F1-score of 77.2% in the hierarchical multi-label classification task of agricultural science data, outperforming baseline models and effectively improving classification accuracy and efficiency.

Key words: agricultural science data, text classification, hierarchical multi-label classification, deep learning