| 潘雪峰,陈洁,王超.基于大语言模型多模态融合驱动的图书分类优化方法研究[J].情报工程,2025,11(3):015-026 |
| 基于大语言模型多模态融合驱动的图书分类优化方法研究 |
| Research on Book Classification Optimization Method Driven by Multimodal Fusion of Large Language Models |
| |
| DOI： |
| 中文关键词: 单模态；多模态；大语言模型；图书分类 |
| 英文关键词: Unimodal; Multimodal; Large Language Model; Book Classification |
| 基金项目:2023 年度教育部人文社会科学研究规划基金“‘人工智能驱动的科学研究’对科研数据基础设施建设的影响机制研究”（23YJA870002）。 |
| 作者 | 单位 | | 潘雪峰 | 辽宁工业大学图书馆 锦州 121000 | | 陈洁 | 桂林理工大学图书馆 桂林 541000 | | 王超 | 辽宁工业大学图书馆 锦州 121000 |
|
| 摘要点击次数: 59 |
| 全文下载次数: 39 |
| 中文摘要: |
| [目的/意义]探索基于多模态数据的大语言模型图书分类方法，解决传统图书分类方法存在的瓶颈。[方法/过程]构建单模态数据集和多模态数据集，通过小样本学习策略构建提示语，对比测试GPT-4 和DeepSeek 模型图书分类结果。建立人工编目团队参照系，进行4 次实验，每次实验通过5 组对照量化模型效能。[ 结果/ 结论] 单模态分类中，DeepSeek 平均准确率（80.69%）显著高于GPT-4（64.20%），提示语信息增加使DeepSeek 准确率提升，GPT-4 因信息冗余出现性能波动。多模态场景下，DeepSeek在精确率（94.5%）、召回率（84.0%）及F1 值（88.9%）上全面优于GPT-4，资深编目员与DeepSeek 协同准确率达99.0% 以上。大语言模型通过多模态协同表征与动态知识蒸馏机制，显著优化图书分类效能。 |
| 英文摘要: |
| [Objective/Significance] To explore a large language model book classification method based on multimodal data and solve the bottleneck of traditional book classification methods. [Methods/Processes] A unimodal dataset and a multimodal dataset were constructed. Prompt language was constructed using a small sample learning strategy, and the book classification results of GPT-4 and DeepSeek models were compared and tested. Establishing a reference frame for the manual cataloging team, conduct 4 experiments, and quantify the effectiveness of the model through 5 control groups in each experiment. [Results/Conclusions]In unimodal classification, the average accuracy of DeepSeek (80.69%) was significantly higher than that of GPT-4 (64.20%).The increase in prompt information improved the accuracy of DeepSeek, while GPT-4 experienced performance fluctuations due to information redundancy. In multimodal scenarios, DeepSeek outperforms GPT-4 in accuracy (94.5%), recall (84.0%), and F1 score (88.9%), with a collaborative accuracy rate of over 99.0% achieved by experienced catalogers and DeepSeek. The large language model significantly optimizes the efficiency of book classification through multimodal collaborative representation and dynamic knowledge distillation mechanism. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
