RAG-Enhanced Decision Modeling for Patent Office Actions
AI system helping improve patent office action submission and prior-art matching using domain-specific retrieval models.
Project Overview
Mentored by professors at Tsinghua University AI Laboratories. Led a three-member research team to design a retrieval-augmented generation (RAG) based system for automating patent office action drafting and prior-art matching. Collected, standardized, and preprocessed 100,000+ multilingual patent records from USPTO and CNIPA sources, building robust text-cleaning, tokenization, and entity-extraction pipelines to ensure consistency. Developed domain-specific embeddings and semantic similarity models using transformer-based architectures (e.g., BERT, SentenceTransformers) to link patent claims with supporting evidence, improving retrieval precision by 20% over baseline. Designed a multi-stage retrieval–ranking–re-ranking framework that incorporated statistical weighting and cosine similarity thresholds, enhancing both result interpretability and legal transparency for expert review.
Technologies Used
Impact
Improved retrieval precision by 20% and enhanced interpretability in legal AI workflows