A hybrid model integrating RoBERTa, TF-IDF, and attention mechanism for medical query intent classification

In medical retrieval scenarios, accurate identification of user query intent is pivotal for enhancing retrieval relevance and knowledge augmentation effectiveness. However, existing approaches exhibit polarized performance: traditional shallow models (e.g., TF-IDF, Bag-of-Words) struggle to capture contextual semantics, while pre-trained language models, despite their strong deep semantic representation capabilities, often overlook keyword weighting and category-specific priors. This leads to suboptimal accuracy for long-tail categories and susceptibility to confusion from emotional or colloquial query formulations. To address these challenges, this study proposes a hybrid model integrating RoBERTa-wwm-ext with sentence-level TF-IDF and category-centroid TF-IDF via a three-branch attention fusion mechanism. The model employs adaptive attention mechanisms to dynamically weight deep-shallow features at the sample level, while incorporating category-centroid vectors as explicit semantic anchors to mitigate class imbalance issues. Experimental evaluations on the KUAKE-QIC dataset demonstrate that our approach achieves an accuracy of 0.824 and a Macro-F1 score of 0.800, outperforming all baseline models by 2.3% points. These results highlight the synergistic benefits of deep-shallow feature interaction, ensuring robust recall for dominant categories while significantly improving precision for minority classes. This work provides a generalizable framework for medical intent classification that balances semantic integrity with keyword-driven priors, laying a more reliable foundation for applications such as clinical question-answering systems, medical search ranking, and intelligent triage systems.