近年大型語言模型（Large Language Model, LLM）廣泛用於學術寫作，協助學者改善文章的語法與結構，提升寫作效率與文本可讀性。然而在享受其便利性的同時，也衍生值得關注的潛在風險，包括學術誠信與作者責任的界定、原創性與抄襲爭議，以及內容正確性與可驗證性等隱憂。有鑑於此，本研究旨在建構兼具可解釋性與泛化能力的統計分析，判斷文章是否由人工智慧生成。除了以分類效能指標（如偽陽性率與偽陰性率）為評估依據，本研究亦嘗試從文本特徵萃取、辨識寫作風格，提升模型解釋能力與實務應用價值。
方法上，本研究從語意壓縮與資訊覆蓋兩個觀點出發，建構具可解釋性的摘要寫作風格文字特徵，並以提升模型泛化能力及評估特徵選取穩定性為原則進行設計。建構之特徵涵蓋詞彙多樣性、資訊量指標（如相異字比例、熵）、句長與結構統計量，以及反映用詞語境與句法結構的語言特徵。資料方面選擇民國107 至109 年經濟學門碩士論文，並蒐集GPT-5、Gemini 2.5 Pro 與Claude Opus 4 根據人類摘要生成的修改文本，比對兩者的差異。研究結果顯示，常見詞彙的使用為最主要之辨識訊號，語境與句法特徵則可作為輔助資訊。以最佳特徵組合搭配羅吉斯迴歸，即可達到與深度學習相當之辨識效能，且兼具較高之可解釋性與運算效率。然而，無論是本文方法或深度學習模型，在不同生成模型之辨識準確率皆明顯下降，顯示現有文本辨識方法之泛化能力仍有其限制，際應用時須審慎評估其適用範圍與侷限性。

Large language models (LLMs) have recently been widely adopted in academic writing, helping scholars refine the grammar and structure of their texts and improving both writing efficiency and readability. Alongside these conveniences, however, their use raises concerns that warrant attention. These include defining the boundaries of academic integrity and authorial responsibility, disputes over originality and plagiarism, and doubts regarding the correctness and verifiability of content. In light of these issues, this study aims to construct a statistical framework that is both interpretable and generalizable for determining whether a text is generated by artificial intelligence. Beyond relying on classification performance metrics (such as the false positive rate and the false negative rate), the study further seeks to identify writing style through text-feature extraction, thereby enhancing the model’s explanatory power and practical value.
Methodologically, this study proceeds from two perspectives—semantic
compression and information coverage—to construct interpretable stylistic
features of abstract writing, designed with the goals of improving model generalization and assessing the stability of feature selection. The constructed
features encompass lexical diversity, information-content measures (e.g., the
proportion of unique words and entropy), sentence-length and structural statistics, and linguistic features reflecting lexical context and syntactic structure. Regarding the dataset, master’s thesis abstracts in the field of economics from 2018 to 2020 were selected as human texts, and revised texts generated by GPT-5, Gemini 2.5 Pro, and Claude Opus 4 based on the human abstracts were collected for comparison. The results show that the use of common words is the most important discriminative signal, whereas contextual and syntactic features serve as auxiliary information. Pairing the best feature combination with logistic regression attains recognition performance comparable to that of deep learning while offering higher interpretability and computational efficiency. Nevertheless, for both the proposed method and the deep learning model, recognition accuracy drops markedly across different generative models, indicating that the generalization ability of existing text detection methods remains limited; practical applications therefore require a careful assessment of their scope and limitations.

一、中文文獻
中央研究院詞庫小組（CKIP Lab） （2021） 。CKIP Transformers： 繁體中文自然語言處理工具[電腦軟體]。https://github.com/ckiplab/ckip-transformers
朱君輝、王夢焰、楊爾弘、聶錦燃、王譽傑、岳岩、楊麟兒（2023）。人工智能生成語言與人類語言對比研究——以ChatGPT 為例。第二十二屆中國計算語言學大會論文集（CCL 2023），523–534。
余清祥（1998）。統計在紅樓夢的應用。政大學報，76，303–327。
余清祥、葉昱廷（2020）。以文字探勘技術分析臺灣四大報文字風格。數位典藏與數位人文，6，67–94。
余清祥、張祐瑜（2026）。以統計方法辨識ChatGPT 生成之論文摘要。中國統計學報，64，148–167。
李賢平（1987）。《紅樓夢》成書新說。復旦學報（社會科學版），5，3–16。
杜協昌（2012）。利用文本採礦探討《紅樓夢》的後四十回作者爭議。第三屆數位典藏與數位人文國際研討會（DADH 2012）。
張祐瑜（2025）。以統計方法辨識碩士及ChatGPT 生成之經濟學類論文摘要。國立政治大學統計學系學位論文，2025，1–58。
劉貞莉（2024）。臺灣碩博士論文之文字分析——以商業及管理學門摘要為例。國立政治大學統計學系學位論文。2024，1–88。
二、英文文獻
Argamon, S., Koppel, M., Fine, J., & Shimoni, A. R. (2003). Gender, genre, and writing style in formal written texts. Text & Talk, 23(3), 321–346.
Berriche, L., & Larabi-Marie-Sainte, S. (2024). Unveiling ChatGPT text using writing style. Heliyon, 10(12), e32976.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., …Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). Association for Computing Machinery.
Desaire, H., Chua, A. E., Isom, M., Jarosova, R., & Hua, D. (2023). Distinguishing academic science writing from humans or ChatGPT with over 99% accuracy using off the shelf machine learning tools. Cell Reports Physical Science, 4(6), 101426.
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In J. Burstein, C. Doran, & T. Solorio (Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (pp. 4171–4186).Association for Computational Linguistics.
Georgiou, G. P. (2025). Differentiating between human-written and AI-generated texts using automatically extracted linguistic features. Information, 16(11), 979.
Holmes, D. I. (1998). The evolution of stylometry in humanities scholarship. Literary and Linguistic Computing, 13(3), 111–117.
Muñoz-Ortiz, A., Gómez-Rodríguez, C., & Vilares, D. (2024). Contrasting linguistic patterns in human and LLM-generated news text. Scientific Reports, 14(1), 22446.
OpenAI. (2022, November 30). Introducing ChatGPT. https://openai.com/index/chatgpt/
Reinhart, A., Markey, B., Laudenbach, M., Pantusen, K., Yurko, R., Weinberg, G., & Brown, D. W. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. Proceedings of the National Academy of Sciences, 122(8), e2422455122.
Stamatatos, E., Fakotakis, N., & Kokkinakis, G. (2000). Automatic text categorization in terms of genre and author. Computational Linguistics, 26(4), 471–495.
Thamdee, N., & Man-Khongdi, P. (2025). A comparative linguistic perspective on AI language use: Examining Gemini and Claude beyond technical performance. Journal of Arts, Humanities and Social Science, 8(2).
Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1), 267–288.
Tukey, J. W. (1977). Exploratory data analysis. Addison-Wesley.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998–6008.
Yue, C. J., Huang, P., Hsu, S.-S., & Fan, C.-Y. (2026). A statistical approach of distinguishing patent abstracts written by human from those generated by ChatGPT. World Patent Information, 85, 102457.