面向"主题—数据—方法"研究路径分析的学术文献实体抽取OACHSSCD

[Purpose/Significance]Knowledge entities in academic literature serve as key carriers of scholarly value.Their automated and high-precision extraction,analysis,and utilization provide an important means for identifying deve-lopment trends in scientific and technological intelligence and revealing disciplinary evolution patterns.[Method/Process]Taking the bibliographic records of 9053 articles published from 2021 to 2025 in 11 Chinese core journals in the field of information science as the research object,this study integrated Chain-of-Thought(CoT)knowledge distillation with QLoRA-based parameter-efficient fine-tuning.A large-parameter teacher model was used to generate reasoning-enhanced CoT data,which were then distilled into lightweight student models to achieve efficient and low-cost knowledge extraction.For the extracted entities,a BERTopic clustering method incorporating BGE semantic representations and LLM-based semantic summarization was employed to identify and summarize three types of clusters,namely research topics,data sources,and research methods(T-D-M),as well as their associative combination paths.[Result/Conclusion]The results show that the proposed training strategy significantly improves the performance of student models,although a performance gap remains compared with the teacher model.Domain-specific LLMs pretrained on scientific corpora outperform general-purpose LLMs with the same architecture.Further clustering and association analyses reveal the major research hotspots and directions in information science in recent years,construct a T-D-M research path map,and uncover differentiated research logics and potential research combinations driven by research topics,data sources,and research methods.