融合技术特征提取与引用关系重构的专利权利要求书智能修订方法OACHSSCD

[Purpose/Significance]The automated revision of patent claims is a highly specialized task that requires the precise alignment of technical accuracy with rigorous legal logic.Traditional automated drafting tools often suffer from logical collapses or incomplete feature coverage,failing to meet the professional standards required for intellectual property protection.This study proposes an intelligent patent revision framework that integrates structured constraints with implicit reasoning.By transforming non-structured patent specifications and implicit claim dependencies into machine-understandable structural priors,this study aims to significantly enhance the efficiency of computer-aided patent drafting while ensuring the legal validity and quality of the generated claims.[Method/Process]This paper introduced a Patent Feature-extraction and Citation-linkage reconstruction framework(Pat-FCL)consisting of three core modules.First,the Technical Feature Extraction module decoupled core technical solutions from redundant background descriptions in the pa-tent specification through an LLM.This process transformed semantic fragments into a structured Technical Feature List(TFL)comprising structural components,functional limitations,and constraints,which alleviated the computational pre-ssure of long contexts and ensured comprehensive feature coverage.Second,the Citation Relationship Reconstruction mo-dule explicated the hierarchical dependencies between claims.By formalizing each claim as a four-tuple(including index,type,text,and direct citation sets)and utilizing graph traversal algorithms,the system constructed a directed acyclic graph to map both direct and indirect citations.This module then converted the structural prior into a natural language logi-cal description,effectively preventing logical collapse during long-text generation.Finally,the Claim Revision module integrated the TFL and the explicit citation relationship into a joint input.Under a Zero-Shot strategy,the Pat-FCL frame-work executed end-to-end revisions,where clear content and structural boundaries suppressed the stochasticity of free gene-ration.[Result/Conclusion]Experimental evaluations in the fields of"Electrical Engineering"and"Chemistry",using LLaMA-3-8B-Instruct,Flan-T5-Large,Mistral-7B-Instruct,and Qwen-2-7B-Instruct as benchmarking baselines,demonstrate the effectiveness of the proposed approach.The results show that the Pat-FCL method achieves a substantial performance leap across all primary metrics.In the Electrical Engineering domain,the proposed method reaches a SARI score of 35.44%,representing an increase of 7.59 percentage points(a 27.3%relative improvement)over the Flan-T5-Large baseline.Additionally,the Pat-FCL method achieves a peak ROUGE-L score of 71.04%,representing a maximum increase of 14.45 percentage points(a relative improvement of 25.5%)over the baseline model.Although Flan-T5-Large exhibits higher BLEU scores due to its conservative"copy-and-paste"strategy,the significantly higher SARI values of the Pat-FCL method indicate more substantive and valuable revisions that align with human expert standards.Furthermore,leading results in BERTScore and LLM-Eval(proxy-scored by GLM-4-flash)confirm that the method ensures high seman-tic consistency and professional logical rigor.The framework effectively mitigates the issues of incomplete technical feature coverage and insufficient legal defense gradients between dependent and independent claims,offering a robust and innova-tive technical pathway for intelligent patent drafting and quality control.