Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PatentMind: A Multi-Aspect Reasoning Graph for Patent Similarity Evaluation

Yongmin Yoo, Qiongkai Xu, Longbing Cao

TL;DR

PatentMind introduces a Multi-Aspect Reasoning Graph (MARG) to address patent similarity by separately evaluating technical features, application domains, and claim scope, then integrating these signals through context-aware dynamic weighting. The framework yields interpretable, dimension-wise judgments and a final similarity score, validated against a human-annotated benchmark (PatentSimBench) with $r=0.938$, outperforming embedding-, prompting-, and even some fine-tuned baselines. Key contributions include the MARG decomposition, the four-stage context-aware weighting process, and the PatentSimBench dataset, which together enable robust, domain-sensitive similarity assessments for tasks such as prior art search and infringement risk evaluation. The approach demonstrates model-agnostic robustness across LLMs and offers a scalable, interpretable foundation for real-world patent analytics and decision-making.

Abstract

Patent similarity evaluation plays a critical role in intellectual property analysis. However, existing methods often overlook the intricate structure of patent documents, which integrate technical specifications, legal boundaries, and application contexts. We introduce PatentMind, a novel framework for patent similarity assessment based on a Multi-Aspect Reasoning Graph (MARG). PatentMind decomposes patents into their three dimensions of technical features, application domains, and claim scopes, then dimension-specific similarity scores are calculated over the MARG. These scores are dynamically weighted through a context-aware reasoning process, which integrates contextual signals to emulate expert-level judgment. To support evaluation, we construct a human-annotated benchmark PatentSimBench, comprising 500 patent pairs. Experimental results demonstrate that the PatentMind-generated scores show a strong correlation ($r=0.938$) with expert annotations, significantly outperforming embedding-based models, patent-specific models, and advanced prompt engineering methods. Beyond computational linguistics, our framework provides a structured and semantically grounded foundation for real-world decision-making, particularly for tasks such as infringement risk assessment, underscoring its broader impact on both patent analytics and evaluation.

PatentMind: A Multi-Aspect Reasoning Graph for Patent Similarity Evaluation

TL;DR

PatentMind introduces a Multi-Aspect Reasoning Graph (MARG) to address patent similarity by separately evaluating technical features, application domains, and claim scope, then integrating these signals through context-aware dynamic weighting. The framework yields interpretable, dimension-wise judgments and a final similarity score, validated against a human-annotated benchmark (PatentSimBench) with , outperforming embedding-, prompting-, and even some fine-tuned baselines. Key contributions include the MARG decomposition, the four-stage context-aware weighting process, and the PatentSimBench dataset, which together enable robust, domain-sensitive similarity assessments for tasks such as prior art search and infringement risk evaluation. The approach demonstrates model-agnostic robustness across LLMs and offers a scalable, interpretable foundation for real-world patent analytics and decision-making.

Abstract

Patent similarity evaluation plays a critical role in intellectual property analysis. However, existing methods often overlook the intricate structure of patent documents, which integrate technical specifications, legal boundaries, and application contexts. We introduce PatentMind, a novel framework for patent similarity assessment based on a Multi-Aspect Reasoning Graph (MARG). PatentMind decomposes patents into their three dimensions of technical features, application domains, and claim scopes, then dimension-specific similarity scores are calculated over the MARG. These scores are dynamically weighted through a context-aware reasoning process, which integrates contextual signals to emulate expert-level judgment. To support evaluation, we construct a human-annotated benchmark PatentSimBench, comprising 500 patent pairs. Experimental results demonstrate that the PatentMind-generated scores show a strong correlation () with expert annotations, significantly outperforming embedding-based models, patent-specific models, and advanced prompt engineering methods. Beyond computational linguistics, our framework provides a structured and semantically grounded foundation for real-world decision-making, particularly for tasks such as infringement risk assessment, underscoring its broader impact on both patent analytics and evaluation.

Paper Structure

This paper contains 66 sections, 3 equations, 4 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Patent Similarity Evaluation
  4. Prompt Engineering
  5. Methodology
  6. Multi-Dimensional Feature Extraction
  7. Dimension-wise Similarity Computation
  8. Context-Aware Dynamic Weight Reasoning
  9. Context Encoding
  10. Multi-perspective Analysis
  11. Weight Derivation and Justification
  12. Final Similarity Calculation
  13. Dataset Construction
  14. Dataset Composition
  15. Annotation Process
  16. ...and 51 more sections

Figures (4)

  • Figure 1: The comparison between multi-aspect reasoning and embedding-based similarity methods.
  • Figure 2: The workflow of PatentMind, structured as a Multi-Aspect Reasoning Graph (MARG).
  • Figure 3: The computational graph of the Context-Aware Dynamic Weight Reasoning framework, with modules annotated by their notations: Domain Relationship Analysis ($f_{rel}, R_{domain}$), Information Distribution Analysis ($f_{dist}, D_{info}$), Dimension Relevance Assessment ($f_{assess}, A_{rel}$), and Cross-validation Reasoning ($f_{valid}, V_{cross}$).
  • Figure 4: The correlation between expert judgments and scores predicted by PatentMind.