自動化專利答辯：大語言模型的應用

Abstract

在美國專利申請過程（U.S. Patent Prosecution）中，對於審查階段所收到的官方審查意見通知書（Office Action, OA）進行及時且有效的答辯（OA-Response）是成功取得專利的關鍵。然而，由於答辯資料取得不易，使得過去在自動化和 AI 領域難以進行深入研究。為了補足該差距，我們引入了審查意見通知書答辯智能系統（Patent Office Action Response Intelligence System, PARIS）及其進階版本——利用大型語言模型強化的 PARIS 系統（Large Language Model （LLM）Enhanced PARIS, LE-PARIS）和 PARIS 與代理式檢索增強生成（PARIS with Agentic RAG Optimization, PARIS-PRO），以提高 OA 答辯的效率。系統的關鍵功能包括建立 OA 主題資料庫（OA Topics Database）、開發答辯模板（Response Templates）和推薦系統（Recommender Systems）以及基於 LLM 的答辯生成（LLM-based Response Generation）。為了證明系統的有效性，我們對 OA 資料和基於過去使用者與系統互動的縱向數據進行研究分析。透過七項分析，我們檢驗了 OA 主題的分類（分析 1 和 2）；為 OA 答辯所設計的推薦系統有效性（分析 3）；生成答辯的質量（分析 4）；以及透過使用者研究探討系統在現實世界場景中的實際價值（分析 5）；PRO 及其結果比較（分析 6）；三個系統質性研究（分析 7）。結果表明，系統在關鍵性能指標上對提升使用者的效率，產生了正面影響。本研究所提出的系統架構，亦可作為醫療、法律等其他高度專業領域應用 AI 與人機協作時的參考。

In U.S.patent prosecution, a timely and effective response to Office Actions (OAs) is essential for securing patent rights. Despite the critical nature of this process, previous research on automation and artificial intelligence (AI) has largely overlooked this domain. To bridge this gap, we introduce the Patent Office Action Response Intelligence System (PARIS) and its advanced iterations: the Large Language Model (LLM) Enhanced PARIS (LE-PARIS) and PARIS with Agentic RAG Optimization (PARIS-PRO). These systems are designed to augment the efficiency of patent attorneys by leveraging AI-assisted collaboration. Key components include the development of an OA Topics Database, the creation of response templates, the integration of recommender systems with LLM-based response generation, and the utilization of Agentic AI framework. To evaluate the effectiveness of our approach, we conducted a comprehensive analysis using the USPTO Office Action database, supplemented by six years of longitudinal data from attorney interactions with our systems. Our investigation comprised seven studies: the first two assessed the constructiveness of OA topics using topic modeling and a Delphi process; the third evaluated the performance of our hybrid LLM-based recommender system specifically tailored for OA responses; the fourth examined the quality of the generated responses; the fifth explored the practical utility of the systems in real-world scenarios through user studies; the sixth examined the performance outcomes of PARIS-PRO；and the seventh qualitative research on three systems. The results demonstrate a significant positive impact on key performance indicators, underscoring the potential of AI-driven solutions in enhancing the patent prosecution process.