Abstract: The existing discipline withdrawal and researcher redistribution rely on manual decision-making, which made it difficult to effectively coordinate discipline associations. In this context, large language models with strong knowledge analysis capabilities were utilized to provide new ideas for discipline optimization based on discipline withdrawal and researcher redistribution. However, university research data represented by scientific research information faced challenges such as difficult to understand professional terms and obvious long-tail distribution. In response to the above challenges, a multi-stage coupled decision framework MCRF (Multistage Coupled Redeployment Framework) for the redistribution of researchers in revoked disciplines was proposed. MCRF was composed of four stages: recall, semantic enhancement, pairing, and rearrangement, which effectively decomposed difficult problems into multiple relatively simple sub-problems. Firstly, a subject research word cloud association dataset was constructed to alleviate the problem where general models had difficulty understanding specialized academic terms; secondly, an association recall algorithm was designed to quickly recall Top-K related subjects of scientific research information, thereby reducing the overall decision-making time overhead; additionally, an implicit optimization module was introduced to generate diverse representations of scientific research information, ensuring that tail subject research information could be fully associated with researchers' research directions, and accurate semantic matching was achieved through a fine-grained scientific research project ranking model. The experimental results showed that in the recall stage, the recall rate of the proposed framework on the dataset reached 92%, and the accuracy rate in the rearrangement stage was 95%. These results effectively verify the effectiveness of MCRF in the task of subject structure optimization.

Key words: Large Language Model (LLM), scientific research word cloud, discipline structure optimization, research information；semantic matching

摘要： 现有学科撤销与科研人员重分配依赖人工决策，难以有效统筹学科关联。在此背景下，拥有强大知识分析能力的大语言模型给基于学科撤销与科研人员重分配的学科优化提供了新的思路，但在以科研信息为代表的高校科研数据上也面临着专业术语难理解、长尾分布明显等挑战。针对上述挑战，提出一种面向撤销学科科研人员重分配的多阶段耦合决策框架MCRF（Multistage Coupled Redeployment Framework）。MCRF包含召回、语义增强、配对和重排4个阶段，能有效将困难问题分解为多个相对简单的子问题。首先，构建学科科研词云关联数据集，以缓解通用模型难以理解专用学术名词的问题；其次，设计了关联召回算法以快速召回科研信息的Top-K关联学科，进而降低整体决策时间开销；此外，引入了隐式优化模块，从而生成多样化的科研信息表述，确保尾部学科科研信息能与科研人员研究方向全面关联，并通过细粒度科研项目排序模型实现精准语义匹配。实验结果表明，在召回阶段，所提框架在数据集上的召回率达到了92%，在重排阶段准确率为95%，这些结果有效验证了MCRF在学科结构优化任务中的有效性。

关键词: 大语言模型, 科研词云, 学科结构优化, 科研信息, 语义匹配