Scholarlink

<p>Patent Landscape Studies (PLS) provide companies with a comprehensive view of the technological status within specific fields, playing a crucial role in patent analysis. The primary task of PLS is to categorize patents into user-defined categories. However, since the classification in PLS is determined by users based on their analytical objectives, it necessitates extensive additional data annotation for deep learning models, and the classification accuracy is influenced by the annotation strategy and the meaning of the labels. Addressing the limitations of existing research, this study proposes a novel method for automating PLS in low-resource scenarios using large language models (LLMs)—GPT-PLS. We designed a prompt template specifically for PLS classification tasks and employed an in-context learning strategy, providing demonstrations to aid the model in understanding the relationship between patent context and classification labels. We conducted experiments on three public datasets: InjVal, Rito, and Atz. The experimental results confirm the capability of LLMs to support PLS in low-resource scenarios, outperforming traditional machine learning and deep learning models, and offering greater interpretability than conventional models.</p>

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

Pioneering exploration in patent landscape studies: leveraging large language models and in-context learning for deeper insights

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Jianliang Yang,
  • Mengchun Zhao,
  • Minghui Qian,
  • Zhian Ying,
  • Chi Wang,
  • Shulin Guo

摘要

Patent Landscape Studies (PLS) provide companies with a comprehensive view of the technological status within specific fields, playing a crucial role in patent analysis. The primary task of PLS is to categorize patents into user-defined categories. However, since the classification in PLS is determined by users based on their analytical objectives, it necessitates extensive additional data annotation for deep learning models, and the classification accuracy is influenced by the annotation strategy and the meaning of the labels. Addressing the limitations of existing research, this study proposes a novel method for automating PLS in low-resource scenarios using large language models (LLMs)—GPT-PLS. We designed a prompt template specifically for PLS classification tasks and employed an in-context learning strategy, providing demonstrations to aid the model in understanding the relationship between patent context and classification labels. We conducted experiments on three public datasets: InjVal, Rito, and Atz. The experimental results confirm the capability of LLMs to support PLS in low-resource scenarios, outperforming traditional machine learning and deep learning models, and offering greater interpretability than conventional models.