基于大模型多智能体协作的高价值专利识别方法研究OACHSSCD

[Purpose/Significance]International research on patent value assessment has progressively developed quantitative indicators(e.g.,citation counts,family size,claim metrics)and machine-learning classifiers built on engi-neered features.However,three interrelated limitations remain.First,most methods rely on surface-level indicators or feature aggregation and therefore fail to capture deep semantic relations,causal signals,and argumentative structure embedded in patent texts.Second,prevalent end-to-end discriminative models operate as"black boxes",offering little human-readable justification and limiting trust for decision-makers and policymakers.Third,many approaches show weak transferability:they perform poorly in low-resource settings and across technological domains because they lack prin-cipled,multi-dimensional evaluation frameworks.To address this gap,the paper proposes an interpretable,generative framework for high-value patent identification that leverages multi-agent collaboration built on large language models(LLMs).[Method/Process]The study first defined a patent evaluation framework covering three complementary dimen-sions:technical,economic,and legal.Using an LLM's contextual and chain-of-thought capabilities,the study synthe-sized explicit evaluation criteria for each dimension and iteratively refined them into a structured guideline.Empirically,the study used IncoPat's global patent database.Based on the"Classification of Strategic Emerging Industries and Corre-spondence Table with International Patent Classification",the study constructed retrieval queries and obtained 375591 granted invention patents in the artificial intelligence field.The study adopted IncoPat's"He Xiang value score"as the value indicator,treating patents with a score of 10 as high-value(positive class)and the remainder as non-high-value(negative class).From this corpus,the study randomly sampled 1000 high-value patents and 1000 non-high-value pa-tents to ensure class balance,and partitioned them into training,validation,and test sets in a 6∶2∶2 ratio.The core analy-sis used the DeepSeek-R1 to implement three evaluative agents(technical,economic,legal)and a collaborative reasoning protocol that aggregates agent outputs and generates human-readable rationales.[Result/Conclusion]On the AI patent test set,the proposed multi-agent LLMs framework achieves an accuracy of 0.81 and an F1 score of 0.80.It consistently pro-duces structured,logic-grounded rationales,overcoming the interpretability gap that persists in most state-of-the-art international models.Moreover,in low-resource and cross-domain experiments,the method demonstrates stronger gene-ralization than mainstream indicator-driven or end-to-end discriminative approaches.By integrating multi-dimensional patent value criteria with generative reasoning mechanisms,this work provides a scalable and transparent alternative to black-box evaluation pipelines that currently dominate patent analytics.The proposed paradigm aligns with the global shift toward trustworthy AI and explainable intellectual property analytics,offering actionable insights for technology evaluation,strategic decision-making,and innovation governance in both industrial and policy contexts.