AI-Powered Patent Examination Support System

Overview

The Korean Intellectual Property Office (KIPO) currently ranks fourth globally in patent application volume. With applications in key 4th Industrial Revolution sectors growing at an average annual rate of 13.5%, the increasing complexity of examinations necessitates an improved workflow environment. To overcome the high costs of infrastructure and initial investment required for Large AI development, KIPO collaborated with domestic companies leading the Generative AI sector to build a specialized support system.

Project description

Due to the highly specialized nature of intellectual property, a consortium was formed with Saltlux and GenIP. Saltlux managed data collection, pre-processing, model training, and UI development, while GenIP provided legal counsel on IP law, ensured regulatory compliance of refined data, and validated the model's accuracy. The collaboration resulted in four core functions: Chatbot-based Examination Q&A, Automated Search Query Generation, Similar Trial/Judicial Precedent Search, and Written Opinion Summarization.

The Examination Q&A function uses RAG to provide conversational answers within a defined scope of patent laws, examination guidelines, and KIPO publications. It prevents hallucinations by providing the original source text alongside the answer. The Search Query Generation function analyzes existing examination patterns to extract keywords and expand them through LLM-driven synonym/antonym mapping. This standardizes the "prior art search" process, which previously relied on individual expertise and apprenticeship-style learning. For the Similar Precedent Search, a dual structure of keyword and vector search (utilizing cosine similarity and Nearest Neighbor algorithms) allows examiners to find relevant trial records accurately. Finally, the Written Opinion Summarization tool allows examiners to upload lengthy submissions from applicants to receive concise summaries and engage in a follow-up Q&A to ensure no critical points are overlooked.

Key technological features

• Embedding Search: Development of Context-aware Chunking capable of understanding and segmenting context for domain-specific indexing optimization
• RAG (Retrieval-Augmented Generation): LLM training to mitigate hallucinations
  • Pre-training: Additional learning of domain-specific expressions and vocabulary to understand specialized technical knowledge
  • Instruction-Tuning: Fine-tuning based on instructions expected by professional domain users
  • Alignment-Tuning: Optimization based on human preferences using RLHF (Reinforcement Learning from Human Feedback) and DPO (Direct Preference Optimization)

Expected outcomes

As application volumes and technical complexity rise, there is increasing social pressure for higher examination quality. Current dissatisfaction among applicants often stems from inconsistent standards regarding clarity or inventive steps across different examiners. The tools developed in this project will enhance the efficiency and quality of work for examiners and trial judges by providing fast, accurate information. Specifically, the search query generation tool will serve as a standardized guideline for new examiners, reducing the learning curve.

Furthermore, AI support will minimize subjective errors in judgment, thereby increasing the consistency and reliability of the entire examination and trial process. This is expected to lead to higher satisfaction among the public and patent applicants. Through continuous system refinement and advancement, KIPO aims to build a more efficient patent examination environment, strengthening national intellectual property competitiveness and solidifying South Korea’s status as a global leader in the IP sector.

Launch year: 2024

Attribution

Case study courtesy of the National Information Society Agency, Republic of Korea, Apolitical's Content Partner. Sourced from Use Cases of Public AI Service, Vol.2, https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do?cbIdx=31975&bcIdx=29182&parentSeq=29223