Improving Patent Matching with Self-Knowledge Extraction
Listen to the summary
Uses a voice available on your device
Audio options
On this page 4 sections
Related concepts 2 concepts
Key Takeaways
- The framework achieves an 81.3% accuracy rate using the GLM-4-Chat-9B model as the primary backbone.
- It outperforms traditional RAG and Chain-of-Thought approaches by using self-mined entities to expand search queries.
- The method solves limitations in existing patent matching, such as the high cost of manual labeling and catastrophic forgetting in models.
- The approach demonstrated consistent performance improvements across both English and Chinese patent datasets.
Summary & Methodology Analysis
The proposed framework addresses the inherent complexity of patent matching, where specialized terminology and technical structures often lead to poor retrieval accuracy. Instead of relying on static keyword matching, the system implements a self-knowledge mining strategy. This strategy parses incoming queries to extract key technical entities and map them into hierarchical ontologies. By concatenating this extracted knowledge with the original patent abstract, the system effectively performs query expansion, which creates a more robust input for the retrieval phase. The integration of FAISS for K-nearest neighbor search using bge-large embeddings ensures that the system handles large-scale vector sets efficiently before the final generative step. Once similar patents are retrieved, the system constructs a structured instruction set containing the query, the retrieved matches, and the self-mined ontology. This instruction set is then passed to an LLM to perform final generative matching, allowing the model to analyze patent relationships more autonomously than standard RAG configurations. The framework was evaluated against multiple backbones, including Qwen2-Instruct-7B, Qwen2.5-Instruct-14B, MoZi-7B, PatentGPT-1.5B, and PatentGPT-1.0-Dense-70B. While the GLM-4-Chat-9B backbone delivered the highest accuracy at 81.3%, the research highlights clear paths for future improvement. Current limitations include a lack of support for multi-modal data integration and the need for more advanced, dynamic ontology generation techniques to keep pace with evolving technical fields.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the primary goal of this patent matching framework?
The goal is to improve the accuracy of matching patents by better analyzing technical structures and relationships using LLMs.
Q2. Why do traditional patent matching methods struggle?
They struggle due to the complexity of technical structures, specialized terminology, and multi-modal information.
Q3. What is the key performance result of the paper?
The method achieved an accuracy of 81.3% when using the GLM-4-Chat-9B model.
Q4. How does the self-knowledge mining strategy function?
It extracts key technical entities from the patent query and constructs a hierarchical ontology to provide more context to the LLM.
Q5. What role does FAISS play in this architecture?
FAISS is used for K-nearest neighbor search to retrieve relevant patents from large-scale vector sets.
Q6. Which specific models were used as backbones in the experiments?
The study tested PatentMatch, Qwen2-Instruct-7B, GLM-4-Chat-9B, Qwen2.5-Instruct-14B, MoZi-7B, PatentGPT-1.5B, and PatentGPT-1.0-Dense-70B.
Q7. How does this method compare to standard RAG?
The method outperforms standard RAG and Chain-of-Thought baselines by leveraging self-mined entity-ontology structures for generative matching.
Q8. What are the current limitations of the system?
The system currently lacks support for dynamic ontology generation and multi-modal data integration.
Q9. Does the paper specify the inference cost or latency?
The paper does not specify these metrics.