本文旨在探討 ChatGPT 自動生成 MARC21 書目紀錄的可行性研究 ，採取正確率 、錯誤率 、新增率 、缺漏率與相似度等多元指標 ，以評量 ChatGPT 自動生成書目紀錄的效益。本文自國家圖書館全國圖書書目資訊網（National Bibliographic Information Network，簡稱NBINet） 書目資料庫選取 55 筆預行編目資料為研究樣本 ，以包含預行編目資料的提示語要求 ChatGPT-4o 生成對應的 MARC21 書目格式紀錄後 ，再以 NBINet MARC21 紀錄為基準 ，評估 ChatGPT-4o之自動化後設資料生成的能力。結果顯示 ChatGPT 生成書目紀錄之正確率為 98.4%、錯誤率為 1.6%，而相較於 NBINet 書目紀錄 ，ChatGPT 生成書目紀錄之新增率為 102% 與缺漏率為 27%。在相似率方面 ，委由 ChatGPT 以五等級李克特表評判 ChatGPT 生成紀錄與預行編目資料間之平均相似度可達 3.39。除此之外 ，本文也發現ChatGPT 具有 MARC21 書目資料之自動化格式對照與轉換 、新增 、修正與補充等能力外 ，也提出錯誤的類型（如資料之錯誤與重複）。

This study aims to investigate the feasibility of using ChatGPT to generate MARC21 bibliographic records in terms of accuracy, error, addition and missing rate. A total of 55 cataloging in publication records from the bibliographic database of “National Bibliographic Information Network” (NBINet) provided by National Central Library in Taiwan were selected as the study sample and were employed as part of prompt to request ChatGPT-4o to generate MARC21 records. Records from the NBINet have served as the benchmark for evaluation of MARC21 records generated by ChatGPT-4o. The findings indicate that ChatGPT-4o achieved an accuracy rate of 98.4%, an error rate of 1.6%, an addition rate of 102%, and a missing rate of 27%. In terms of similarity, the average similarity between the ChatGPT-generated records and the preliminary cataloging data was found to be 3.39, as evaluated by ChatGPT using a five-point Likert scale. Moreover, this study identifies ChatGPT’s capability for automatic mapping and conversion from one format to the other, addition, correction and supplementation handling in MARC21 bibliographic records, as well as categorizes the types of errors observed (e.g., error and duplication).

人工智慧、ChatGPT、資訊組織、自動化後設資料生成、自動化編目

Artificial intelligence, Information organization, Automatic metadata generation, Automatic cataloging