資料載入中.....
|
請使用永久網址來引用或連結此文件:
https://nccur.lib.nccu.edu.tw/handle/140.119/159101
|
| 題名: | 探討負責任 AI 與大型語言模型
Explore the Responsible AI and LLM |
| 作者: | 温存正
Wen, Cun-Zheng |
| 貢獻者: | 蔡瑞煌
林怡伶
Tsaih, Rua-Huan
Lin, Yi-Ling
温存正
Wen, Cun-Zheng |
| 關鍵詞: | 負責任 AI
大型語言模型營運
透明性
Responsible AI (RAI)
LLMOps
Transparency |
| 日期: | 2025 |
| 上傳時間: | 2025-09-01 15:06:29 (UTC+8) |
| 摘要: | 隨著生成式人工智慧技術的快速發展，大型語言模型（LLMs）已廣泛應用於醫療、金融與教育等多個領域。儘管這些模型大幅提升了系統的智能化與生產力，但同時也引發了倫理、隱私、偏誤與幻覺（hallucination）等風險。為因應此類挑戰，近年興起兩大關鍵框架：負責任 AI（Responsible AI）與大型語言模型營運管理（LLMOps）。然而，這兩者目前多以孤立方式運作，缺乏一套整合倫理治理與營運實務的統一機制。本研究提出一個整合式框架，將核心的 Responsible AI 原則（如公平性、透明性、隱私保護、非傷害性、包容性、穩健性與問責性）系統性地嵌入 LLMOps 各階段，包括資料處理、模型開發、部署、生成內容管理與系統監控等，以實現兼具倫理遵循與營運可靠性的 LLM 系統。本研究以金融服務領域為應用場域，著重提升模型輸出過程的透明性與可解釋性。我們導入思路鏈（Chain of Thought, CoT）推理策略與少量提示（Few-shot Prompting）技巧，以增強模型回應的邏輯展現與資訊說明能力。為驗證框架成效，本研究設計一項結構化問卷實驗，邀請具有金融背景的領域專家，針對兩種提示設計情境（如 PO 與 PDE）所產生的模型回應進行比較與評估，並回饋其對資訊完整性、系統可信度與可解釋性的主觀感受。實驗結果顯示，本研究所提整合框架有助於提升模型回應的解釋性與用戶信任感，亦具備實務應用潛力。
With the rapid advancement of Generative AI, large language models (LLMs) have found applications in domains such as healthcare, finance, and education. While enhancing productivity, these models raise serious concerns around ethics, privacy, bias, and hallucinations. To address such challenges, two complementary paradigms have emerged: Responsible AI (RAI) and Large Language Model Operations (LLMOps). However, these frameworks remain largely siloed, lacking an integrated mechanism that bridges ethical oversight and operational practice.
To bridge this gap, this study proposes a unified framework that systematically embeds core RAI principles—such as fairness, transparency, and accountability—across the LLMOps lifecycle, spanning stages from data preprocessing to model monitoring. Within the financial services domain, we particularly focus on enhancing the transparency of LLM-generated outputs. To this end, we employ Chain-of-Thought reasoning and Few-shot Prompting techniques to improve the explainability of system responses.
An empirical evaluation was conducted via a structured questionnaire involving financial domain experts. Participants assessed the outputs under two prompting scenarios (e.g., PO vs. PDE) and provided feedback on informational adequacy, trust, and perceived interpretability. Results demonstrate the efficacy of the proposed framework in aligning model behavior with ethical expectations while enhancing user trust. |
| 參考文獻: | Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language Models are Few-Shot Learners. https://commoncrawl.org/the-data/
Gebru, T., Morgenstern, J., Vecchione, B., Wortman Vaughan, J., Wallach, H., Daumé Iii, H., & Crawford, K. (2021). review articles Datasheets for Datasets Documentation to facilitate communication between dataset creators and consumers. COMMUNICATIONS OF THE ACM, 64(12). https://doi.org/10.1145/3458723
Huang, K., Wang, W., & Manral, V. (2024a). From LLMOps to DevSecOps for GenAI. 241–269. https://doi.org/10.1007/978-3-031-54252-7_8
Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., & Zhu, C. (2023). G-EVAL: NLG Evaluation using GPT-4 with Better Human Alignment. https://github.com/nlpyang/geval
Lundberg, S. M., Allen, P. G., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. https://github.com/slundberg/shap
Mittal, S., Thakral, K., Singh, R., Vatsa, M., Glaser, T., Ferrer, C. C., & Hassner, T. (2023). On Responsible Machine Learning Datasets with Fairness, Privacy, and Regulatory Norms.
Ong, J. C. L., Chang, S. Y.-H., William, W., Butte, A. J., Shah, N. H., Chew, L. S. T., Liu, N., Doshi-Velez, F., Lu, W., Savulescu, J., & Ting, D. S. W. (2024). Medical Ethics of Large Language Models in Medicine. NEJM AI, 1(7). https://doi.org/10.1056/AIRA2400038
Peters, D., Vold, K., Robinson, D., & Calvo, R. A. (2020). Responsible AI—Two Frameworks for Ethical Design Practice. IEEE Transactions on Technology and Society, 1(1), 34–47. https://doi.org/10.1109/TTS.2020.2974991
Pornphol, P., & Chittayasothorn, S. (2024). Using LLM Artificial Intelligence Systems as Complex SQL Programming Assistants. 2024 12th International Conference on Information and Education Technology, ICIET 2024, 187–191. https://doi.org/10.1109/ICIET60671.2024.10542806
Prakash, K., Rao, S., Hamza, R., Lukich, J., Chaudhari, V., & Nandi, A. (2024). Integrating LLMs into Database Systems Education. ACM International Conference Proceeding Series, 33–39. https://doi.org/10.1145/3663649.3664371
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?” Explaining the Predictions of Any Classifier. 97–101. https://github.
Shan, R., & Shan, T. (2024). Enterprise LLMOps: Advancing Large Language Models Operations Practice. Proceeding - 2024 IEEE Cloud Summit, Cloud Summit 2024, 143–148. https://doi.org/10.1109/CLOUD-SUMMIT61220.2024.00030
Shankar Zamfrescu-Pereira Björn Hartmann, S. J., Parameswaran Ian Arawjo, A. G., Shankar, S., Zamfrescu-Pereira, J., Hartmann, B., Parameswaran, A. G., & Arawjo, I. (2024). Who Validates the Validators? Align-ing LLM-Assisted Evaluation of LLM Outputs with Human Preferences. https://doi.org/10.1145/3654777.3676450
Sheng, Y., Cao, S., Li, D., Zhu, B., Li, Z., Zhuo, D., Gonzalez, J. E., & Stoica, I. (2024). Fairness in Serving Large Language Models. https://github.com/Ying1123/VTC-artifact.
Sinha, M., Menon, S., & Sagar, R. (2024). LLMOps: Definitions, Framework and Best Practices. International Conference on Electrical, Computer, and Energy Technologies, ICECET 2024. https://doi.org/10.1109/ICECET61485.2024.10698359
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi Quoc, E. H., Le, V., & Zhou, D. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models Chain-of-Thought Prompting.
Yu, T., Zhang, R., Yang, K., Yasunaga, M., Wang, D., Li, Z., Ma, J., Li, I., Yao, Q., Roman, S., Zhang, Z., & Radev, D. R. (2019). Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task. 3911–3921. http://www.databaseanswers.org/
Zhang, Y., Wang, X., Wu, L., & Wang, J. (2024). Pattern-Aware Chain-of-Thought Prompting in Large Language Models.
Zhao, H., Yang, F., Liu, N., Chen, H., Deng, H., Cai, H., Wang, S., Yin, D., & Du, M. (2024b). Explainability for Large Language Models: A Survey. ACM Trans. Intell. Syst. Technol, 15, 38. https://doi.org/10.1145/3639372
McKinsey & Company. (2023). 2023年人工智能发展现状：生成式AI的突破之年 [The State of AI in 2023: The Breakthrough Year of Generative AI]. https://www.mckinsey.com.cn/wp-content/uploads/2023/09/%E7%94%9F%E6%88%90%E5%BC%8FAI%E7%9A%84%E7%AA%81%E7%A0%B4%E4%B9%8B%E5%B9%B4_2023-%E9%87%91%E8%9E%8D%E5%AD%A3%E5%88%8AGenAI0908.pdf
Google Cloud Architecture Center. (2020). MLOps: Continuous delivery and automation
pipelines in machine learning.
Staab, R., Vero, M., Balunović, M., & Vechev, M. (2023). Beyond memorization: Violating privacy via inference with large language models (arXiv:2310.07298). arXiv. https://doi.org/10.48550/arXiv.2310.07298 |
| 描述: | 碩士
國立政治大學
資訊管理學系
112356047 |
| 資料來源: | http://thesis.lib.nccu.edu.tw/record/#G0112356047 |
| 資料類型: | thesis |
| 顯示於類別: | [資訊管理學系] 學位論文
|
文件中的檔案:
| 檔案 |
描述 |
大小 | 格式 | 瀏覽次數 |
|---|
| 604701.pdf | | 4778Kb | Adobe PDF | 0 | 檢視/開啟 |
|
在政大典藏中所有的資料項目都受到原著作權保護.
|
