曾元顯Tseng, Yuan-Hsien郭珮涵Kuo, Pei-Han2024-12-179999-12-312024https://etds.lib.ntnu.edu.tw/thesis/detail/6176b55d3629434dad9974c7ff980a24/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/122423隨著科學論文發表數量的快速增長，引用來源的多樣性和格式差異增加了參考文獻剖析的難度。本研究旨在探討如何自動化擷取科學論文中的參考文獻，並利用人工智慧工具進行剖析，藉以簡化工作流程，降低人力和時間成本，並提升圖書館的知識傳播效能。本文提出了從中文期刊文章檔案中自動化擷取參考文獻的方法，並評估使用人工智慧工具剖析參考文獻的可行性。本研究實驗分為三個部分，第一部分設計程式，擷取期刊文章中的參考文獻章節；第二部分評估不同人工智慧工具在參考文獻剖析任務中的效能；第三部分根據第二部分的實驗結果修正實驗方法，並評估和比較修正後的成果。實驗結果如下：1. 在參考文獻擷取實驗中，基於規則方法的程式能夠自動擷取文章中的參考文獻內容，用於建立資料集作為後續研究基礎。2. 在參考文獻剖析實驗中，本研究比較了spaCy和ChatGPT兩種基於Transformer架構的人工智慧工具的效能。實驗結果顯示，ChatGPT在各欄位的F1-score表現優於spaCy，具有較高的準確性和穩定性。3. 在第三部分實驗中，選擇了第二部分中效能較佳的ChatGPT進行提示修正。實驗結果顯示，經過提示調整後，ChatGPT在各欄位的F1-score表現均有所提升。本研究結果顯示了使用人工智慧工具自動化剖析參考文獻的可行性，並展現了大型語言模型在這一任務中的潛力和優勢。未來研究可以進一步嘗試結合多種人工智慧工具，探討利用不同模型優勢提升參考文獻剖析的準確性，同時探討減低剖析成本的可能性。With the rapid growth in the number of scientific publications, the diversity of citation styles has increased the difficulty of reference parsing. This thesis aims to discuss how to automate the extraction and parsing of references from scientific papers using artificial intelligence tools, thereby simplifying workflows, reducing time costs, and enhancing the efficiency of knowledge dissemination in libraries. This paper proposes a method for extracting references from Chinese journal articles and evaluates the feasibility of parsing these references using AI tools.The study is divided into three parts. The first part involves extracting reference sections from journal articles. The second part assesses the performance of different AI tools in the task of reference parsing. The third part modifies the experimental methods based on the results of the second part and evaluates and compares the outcomes after these adjustments. The experimental results are as follows:1. In the first experiment, the rule-based program successfully extracted reference content from the articles in their entirety.2. The second experiment compared the performance of two AI tools, spaCy and ChatGPT, both based on the Transformer architecture, in reference parsing. Results showed that ChatGPT outperformed spaCy in terms of F1-score, indicating higher accuracy and stability.3. In the third experiment, ChatGPT, which demonstrated better performance in the second part, was selected for model adjustments. We optimized the prompt, and the results indicated that after adjustments, ChatGPT's F1-score performance improved across all fields. In summary, the results of this study demonstrate the feasibility of parsing references using AI tools and reveal the potential of large language models in this task. Future research could explore further integration of various artificial intelligence tools to enhance the accuracy of this task, as well as possibilities for reducing the costs.人工智慧自然語言處理命名實體識別大型語言模型參考文獻剖析Artificial IntelligenceNatural Language ProcessingNamed Entity RecognitionLarge Language ModelsBibliographic Reference Parsing學術論文