融合 RAG 和大语言模型微调的学术短文本学科分类研究

[Purpose/Significance] Research on disciplinary classification of academic short texts can effectively enhance the bibliometric analysis of scholarly papers.[Method/Process]This paper proposed an academic short-text classification framework that integrates Retrieval-Augmented Generation(RAG)with large language model fine-tuning. By dynamically retrieving discipline-related information and combining parameter-efficient fine-tuning techniques, the framework enhanced the semantic representation of model inputs while achieving deep adaptation to domain-specific tasks. [Result/Conclusion]Experiments demonstrate that, compared to traditional deep learning models and general large language models, the synergistic paradigm integrating LoRA fine-tuning and RAG significantly enhances multi-label classification performance. The classification error rate decreases by 36.8%, while accuracy, coverage, and top-label error rate all achieve optimal levels. The classification framework integrating RAG and large language model fine-tuning exhibits synergistic advantages in multi-label classification of academic short texts. Its modular architecture provides a technical pathway for cross-disciplinary knowledge transfer, holding significant academic value and practical implications.