Scholarlink

This study explores the application of generative artificial intelligence (AI) in automating the assignment of classification numbers to books acquired by libraries, focusing on the Nippon Decimal Classification (NDC) system. Using a dataset from the National Diet Library (NDL) consisting of 1,000 to 1.2 million samples, the research evaluated the impact of training data volume on classification accuracy. Fine-tuning was conducted using OpenAI’s GPT-3.5 Turbo model with default parameter settings. Results indicated that even with 1,000 training samples, the model could correctly assign NDC division (two-digit) numbers to about half of the books, achieving over 70% accuracy for NDC class (one-digit) numbers. As the number of training samples increased, accuracy improved logarithmically, reaching 70.5% with 800,000 samples. However, when the sample size reached 1.2 million, there was only a marginal improvement observed to 70.7%. This trend was consistent across NDC class, division, and section (three-digit) levels, with more pronounced improvements observed in the more detailed classification levels. This study not only highlights the potential of generative AI for library classification tasks, but also suggests that more training data may not necessarily lead to a substantial improvement in accuracy. The findings indicate that while generative AI holds potential for automating library classification, the effectiveness of this approach depends on the quality and volume of the data used for training.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Using Generative AI to Improve Library Book Classification Accuracy: The Role of Increased Training Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Takashi Harada,
  • Sho Sato,
  • Minako Nishiura

摘要

This study explores the application of generative artificial intelligence (AI) in automating the assignment of classification numbers to books acquired by libraries, focusing on the Nippon Decimal Classification (NDC) system. Using a dataset from the National Diet Library (NDL) consisting of 1,000 to 1.2 million samples, the research evaluated the impact of training data volume on classification accuracy. Fine-tuning was conducted using OpenAI’s GPT-3.5 Turbo model with default parameter settings. Results indicated that even with 1,000 training samples, the model could correctly assign NDC division (two-digit) numbers to about half of the books, achieving over 70% accuracy for NDC class (one-digit) numbers. As the number of training samples increased, accuracy improved logarithmically, reaching 70.5% with 800,000 samples. However, when the sample size reached 1.2 million, there was only a marginal improvement observed to 70.7%. This trend was consistent across NDC class, division, and section (three-digit) levels, with more pronounced improvements observed in the more detailed classification levels. This study not only highlights the potential of generative AI for library classification tasks, but also suggests that more training data may not necessarily lead to a substantial improvement in accuracy. The findings indicate that while generative AI holds potential for automating library classification, the effectiveness of this approach depends on the quality and volume of the data used for training.