本研究探討生成式人工智慧(AI)技術在專利申請過程中的應用，特別是其在撰寫專利發明說明書方面的潛力，專利保護對企業的創新至關重要，然而，專利申請書的撰寫耗時且需大量專業知識，隨著需求增加，相關人員的負擔加重，因此，簡化和優化專利申請書的撰寫過程尤為重要。
本研究旨在利用生成式AI技術自動生成專利說明書初稿，提高撰寫效率和品質，減少法務部門的工作負擔，我們開發了一個專利發明說明書自動生成模型，並結合語意相似度模型和使用者問卷調查進行評估，結果顯示，生成文本與現行公開專利的語意相似度平均值達到0.904，表示模型能高度模仿專利文本的語意結構，然而，由於資源限制的挑戰，強化學習從人類反饋中學習(RLHF)技術未能如預期般成功。
根據使用者滿意度問卷指出，模型在生成名稱和摘要方面表現優秀，但在細節描述和創新性內容處理上仍需改進，未來方向包括增強細節生成能力、提升創新處理能力、加入LongLoRA技術以及獲取更多資源並實施RLHF技術。
總結來說，本研究證明了生成式AI技術在專利文本自動生成中的應用價值，為未來技術改進和應用擴展提供了方向。

This study explores the application of generative artificial intelligence (AI) technology in the patent application process, specifically its potential in drafting patent specifications. Patent protection is crucial for corporate innovation; however, writing patent applications is time-consuming and requires extensive expertise. As demand increases, the burden on related personnel intensifies, making it essential to simplify and optimize the patent application drafting process.
The aim of this study is to utilize generative AI technology to automatically generate initial drafts of patent specifications, improving drafting efficiency and quality, and reducing the workload on legal departments. We developed an automated patent specification generation model and evaluated it using a semantic similarity model and user satisfaction surveys. The results show that the generated texts achieved an average semantic similarity score of 0.904 with existing published patents, indicating that the model can highly mimic the semantic structure of patent texts. However, due to resource limitations challenges, the implementation of reinforcement learning from human feedback (RLHF) did not achieve the expected success.
User satisfaction surveys indicate that the model performs excellently in generating names and abstracts, but still needs improvement in detail description and handling innovative content. Future directions include enhancing detail generation capabilities, improving innovation processing abilities, incorporating LongLoRA technology, and obtaining more resources to implement RLHF technology.
In summary, this study demonstrates the application value of generative AI technology in the automatic generation of patent texts, providing a direction for future technological improvements and application expansions.

論文審定書 i
誌謝 ii
摘要 iii
Abstract iv
目錄 vi
圖目錄 ix
第一章 緒論(Introduction) 1
1.1. 研究背景 1
1.2. 研究動機 1
1.3. 研究目的 2
第二章 文獻探討(Background and literature discission) 3
2.1. 大型語言模型(LLM) 3
2.1.1. 數據驅動的學習過程 3
2.1.2. 轉換器(Transformer)架構 3
2.1.3. Mistral介紹 4
2.2. 微調(Fine-tuning) LLM 5
2.2.1. 微調的基礎與實施 5
2.2.2. 挑戰與性能評估 6
2.2.3. MediaTek Research Breeze介紹 6
2.3. 提示工程(Prompt Engineering) 8
2.4. 專利生成文獻探討 9
2.4.1. 生成式AI衝擊專利世界 9
2.4.2. LLM在專利生成的應用 10
2.5. 基於人類回饋的強化學習(RLHF) 11
2.6. 文本相似度(Sentence Similarity) 12
第三章 研究方法(Research methods) 15
3.1. 系統架構 15
3.2. 初始模型比較 16
3.2.1. ChatGPT-3.5 16
3.2.2. Gemini 1.5 Pro 16
3.2.3. MR Breeze-7B 17
3.2.4. 初始模型比較結論 17
3.3. 資料來源 17
3.3.1. 大樣本低品質資料 18
3.3.2. 小樣本高品質資料 18
3.3.3. 大樣本低品質資料與小樣本高品質資料 19
3.3.4. 資料來源結論 19
3.4. 研究流程與進行步驟 20
3.4.1. 資料前處理 20
3.4.2. 模板設計 21
3.4.3. 微調(Fine-tuning) 23
3.4.4. 擬定策略 25
3.4.5. 擬定Prompt 26
3.5. 模型推論 28
3.6. 評分標準-與現行公開專利相似度比對 29
3.7. 評分標準-問卷滿意度 30
3.8. RLHF 31
第四章 實驗結果(Experiment result) 33
4.1. 資料整理與文字處理 33
4.2. 訓練參數與訓練損失 33
4.3. 推理結果 35
4.4. 評分標準-與現行公開專利相似度比對結果 39
4.4.1. 相似度評估方法 39
4.4.2. 評估流程 40
4.4.3. 分析與討論 40
4.4.4. 相似度比對結果 41
4.5. 評分標準-使用者問卷調查滿意度結果 43
4.6. RLHF 對於模型推論的影響 50
4.7. 實驗結果總結 53
4.7.1. 使用者問卷調查結果 54
4.7.2. 語意相似度比對結果 54
第五章 結論(Conclusion) 55
5.1. 研究成果總結 55
5.2. 模型優點 55
5.3. 模型缺點 55
5.4. 未來改進方向 56
5.5. 研究應用價值與貢獻 56
參考文獻(Reference) 58
附錄(Appendix) 61
附錄一 61
附錄二 64
附錄三 66
附錄四 68
附錄五 70
附錄六 74
附錄七 76
附錄八 78
附錄九 80
附錄十 81
附錄十一 83
附錄十二 85
附錄十三 87
附錄十四 89
附錄十五 92
附錄十六 95
附錄十七 97
附錄十八 99

1. Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., ... & Wen, J. R. (2023). A survey of large language models. arXiv preprint arXiv:2303.18223.
2. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
3. Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. D. L., ... & Sayed, W. E. (2023). Mistral 7B. arXiv preprint arXiv:2310.06825.
4. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., ... & Scialom, T. (2023). Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
5. Dodge, J., Ilharco, G., Schwartz, R., Farhadi, A., Hajishirzi, H., & Smith, N. (2020). Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305.
6. Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., ... & Sun, M. (2023). Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3), 220-235.
7. White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., ... & Schmidt, D. C. (2023). A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382.
8. 李昆鴻，生成式 AI 衝擊專利世界－專利工程師的新挑戰與機遇，專利師，第 56 期，1-17 頁，2024年01月.
9. MediaTek Research Breeze-7B: Experience the Latest Highly Efficient Large Language Model . Retrieved May 7, 2024, from https://www.mediatek.com/blog/mediatek-research-breeze-7b
10. Lee, J. S., & Hsiang, J. (2020). Patent claim generation by fine-tuning OpenAI GPT-2. World Patent Information, 62, 101983.
11. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., ... & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35, 27730-27744.
12. Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., & Liu, Z. (2024). Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216.
13. Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2024). Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36.
14. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2021). Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.
15. 智慧芽. (n.d.). 智慧芽官网-全球科技创新和知识产权信息服务商. https://www.zhihuiya.com/
16. Lee, J. S. (2020, October). Controlling patent text generation by structural metadata. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 3241-3244).
17. Chen, Y., Qian, S., Tang, H., Lai, X., Liu, Z., Han, S., & Jia, J. (2023). Longlora: Efficient fine-tuning of long-context large language models. arXiv preprint arXiv:2309.12307.
18. Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving language understanding by generative pre-training.
19. Megahed, F. M., Chen, Y. J., Ferris, J. A., Knoth, S., & Jones-Farmer, L. A. (2024). How generative AI models such as ChatGPT can be (mis) used in SPC practice, education, and research? An exploratory study. Quality Engineering, 36(2), 287-315.
20. Practical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation). (2023, November 19). https://magazine.sebastianraschka.com/p/practical-tips-for-finetuning-llms
21. Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023). QLoRA: efficient finetuning of quantized LLMs (2023). arXiv preprint arXiv:2305.14314, 52, 3982-3992.