Abstract:

[Purpose/Significance] The rapid evolution of artificial intelligence (AI) has led to a surge in patent applications, placing immense pressure on traditional patent examination and management systems. While automated classification has gained attention, existing methods often suffer from "semantic drift," hierarchical conflicts, and a lack of interpretability, primarily because they treat classification as a flat probabilistic task rather than a structured logical inference. This study aims to develop a hierarchical automatic patent classification framework that is not only efficient but also hierarchy-consistent and deeply aligned with the professional logic of patent examiners. By shifting the paradigm from black-box probabilistic guessing to knowledge-driven steady-state inference, this study provides a scalable and reliable pathway for intelligent patent classification in high-density technical domains. [Method/Process] The proposed framework was built upon a three-stage mechanism: technical content extraction, technical theme condensation, and hierarchical mapping, utilizing DeepSeek-V3 as the core semantic engine. First, the study constructed an IPC classification standard library and a patent classification knowledge base. A key innovation here is the "Hierarchical Fusion Strategy," which explicitly encodes the examination logic by embedding parent-level technical definitions into child-level descriptions to provide a complete semantic boundary. This ensures that the model perceives the nested structure of the IPC system rather than treating categories as independent labels. Second, the framework performs a semantic-anchored extraction of technical information. Unlike traditional methods that rely on raw text, this process utilizes the IPC standard library as a reference to filter and condense patent claims and descriptions into structured "technical themes". This intermediate representation mitigates the risks of semantic hallucination and handles data sparsity by compressing the semantic space into a more consistent and discriminative form. Third, a "bottom-up" hierarchical mapping strategy was implemented. The system prioritizes matching at the most granular level (the IPC subgroup) and then derives higher-level categories through the established hierarchical chain. To ensure robustness, a dual-path verification mechanism - parallel comparison between independent matching and hierarchical mapping - was introduced. When results conflicted, the system employed a logic of confidence priority to perform local error correction, ensuring that the final output was both fine-grainedly accurate and hierarchically consistent. [Results/Conclusions] Experimental validation conducted on a dataset of Chinese AI invention patents from 2021 to 2025 demonstrates the superior "architecture stability" of the framework. The optimal fusion strategy achieved accuracy rates of 100%, 97.32%, 91.84%, 86.48%, and 71.25% at the IPC section, class, subclass, main group, and subgroup levels, respectively, significantly outperforming the PatentBERT baseline and direct large language model (LLM) classification. Ablation studies confirmed that the integration of IPC knowledge guidance, the condensation of technical themes, and the bottom-up mapping strategy are all critical contributors to performance gains. The results demonstrate that by encoding examination logic into the model, the inherent randomness of LLMs can be effectively constrained within a structured logical track. This framework essentially functions as a "classification skill" for AI Agents, capable of being integrated into intelligent examination systems via an API for constant and automated category updates. Despite limitations in domain coverage, the model-agnostic nature of the architecture suggests high potential for migration to other complex technical fields, providing a foundational methodology for the unified representation and analysis of multi-source innovation data.

Key words: patent classification, examination logic, classification knowledge base, hierarchical classification, large language model, artificial intelligence