基于大型语言模型的学术海报摘要自动摘录方法研究

The description of digital resources is a key link in the digital transformation of university libraries, which can promote the discovery, sharing and personalized knowledge services of digital resources. Among them, the academic poster resources are displayed in the form of pictures, and involve many disciplines, so the description work is difficult to carry out, especially the efficiency of abstracts is low. The rapid development of large language models provides a new way to solve these problems. Therefore, combining optical character recognition technology and large language model, this paper proposes a set of automatic extraction methods suitable for the abstract of academic posters in the form of pictures, which involves four stages： format conversion and alignment of academic lecture posters in the data set preprocessing stage；In the stage of lecture information recognition, text detection and optical character recognition technology are used to locate and extract poster content；In the summary generation stage, prompts are used to guide large language models to generate summaries；In the post-processing evaluation stage, the quality of the summary is evaluated from both quantitative and qualitative perspectives. The results show that the abstracts generated by the proposed method are superior to those written by librarians in terms of conciseness, accuracy and objectivity, and the scores of Rouge-L and novelty are increased by 8.95% and 13.34% respectively, indicating that large language models can help librarians improve the quality and efficiency of writing academic lecture poster abstracts.