Perception to Cognition: Patent Value Assessment Based on Large Language Model Semantic Enhancement and Ensemble Learning

Abstract Purpose To address the limitations of traditional patent metrics in capturing technical substance and the high cost of expert review, this study proposes a hybrid evaluation framework integrating Large Language Models (LLMs) with machine learning to achieve automated, highly accurate identification of high-value patents. Design/methodology/approach Adopting a “Virtual Assessor” paradigm, we constructed a dataset based on the China Patent Gold Awards. The study integrated semantic scores from three diverse LLMs (DeepSeek, Qwen, GLM) under zero-shot and few-shot prompt strategies into a Stacking ensemble learning model (combining XGBoost, Random Forest, and SVM) to predict patent value across nine comparative experimental setups. Findings Direct LLM evaluation revealed a “Knowledge Injection Paradox,” where explicit expert prior knowledge caused negative transfer and reduced accuracy due to over-conditioning. However, the Stacking model successfully rectified these biases, transforming subjective LLM evaluations into robust predictive features. The hybrid model achieved over 97 % accuracy in identifying high-value patents, demonstrating strong robustness even in high-noise environments. Research limitations The study relies on a binary classification of extreme samples (Gold Award vs. non-awarded), potentially oversimplifying the continuous distribution of patent value. Furthermore, the interpretability of the “black box” feature fusion mechanism requires further exploration. Practical implications The proposed framework offers IP managers and policymakers a scalable, cost-effective tool for automated patent screening, effectively bridging the gap between qualitative expert intuition and quantitative data precision. Originality/value This research introduces a “Semantic Enhancement + Algorithmic Rectification” paradigm. It empirically demonstrates how machine learning can correct LLM hallucinations and biases, marking a significant shift from data-driven perception to AI-driven cognitive decision-making in patent valuation.