基于大语言模型微调的学术文献自动分类研究OACHSSCD

[Purpose/Significance]To address the limitations of traditional literature classification methods,such as insufficient accuracy and limited semantic understanding when processing the rapidly increasing volume of academic litera-ture,this study explores an automatic classification approach based on Large Language Models(LLMs).[Method/Process]This study selected several mainstream open-source LLMs,including Llama,Qwen,GLM,Baichuan,and DeepSeek,and combined Low-Rank Adaptation(LoRA)with instruction tuning to optimize these models efficiently.A classification dataset was constructed based on 13 mainstream disciplines from the OpenAlex database,with paper abstracts converted into instruction-format training samples.This study then fine-tuned the selected LLMs on this dataset under unified hard-ware conditions,and evaluated model performance comprehensively using precision,recall,and F1 score.Finally,this study compared the fine-tuned LLMs against traditional deep learning models,specifically SciBERT,RoBERTa,and DeBERTa,as well as against prompt learning methods applied to large-scale models like DeepSeek-V3.2 and Qwen3-Max-Thinking.Additionally,this study analyzed differences in classification performance across various model architec-tures,including base models versus chat-optimized models,across different training data scales to explore the data effi-ciency of LLMs,and across various mainstream disciplines.[Result/Conclusion]The selected LLMs consistently outper-form traditional deep learning models on the academic literature classification task.Among all models,Llama-3.1_Open-Scholar-8B achieves the highest F1 score of 82.82%,surpassing SciBERT by approximately 4 percentage points.Small-scale models,such as Llama-3.2-3B-Instruct,also demonstrate strong classification capabilities,significantly exceeding traditional models while maintaining fast inference speed.The performance difference between chat models optimized for dialogue tasks and standard base models remains marginal,likely because literature classification involves neither multi-turn interaction nor long-text generation.Even under limited training data conditions,LLMs exhibit considerable classifi-cation potential.With only 400 training samples per discipline,Llama-3.1_OpenScholar-8B achieves better performance than RoBERTa and DeBERTa trained on 1 200 samples per discipline,demonstrating superior data efficiency.Perfor-mance varies across disciplines:Physics and Astronomy,Energy,and Neuroscience achieve high F1 scores(above 88%),while Medicine,Computer Science,and Biochemistry,Genetics and Molecular Biology show relatively lower performance due to terminological complexity and interdisciplinary overlap.This study validates the effectiveness and application pro-mise of LLMs for academic literature classification,providing a feasible technical reference for large-scale automatic classifi-cation tasks,especially in resource-constrained scenarios.