TLDR: This research paper introduces a hybrid AI framework for strategically pruning patent portfolios to identify high-value assets for technology transfer. It combines a Learning-to-Rank (LTR) model for quantitative evaluation with a unique “Need-Seed” agent system that uses NLP and LLMs to match patent capabilities with documented market demands. A Human-in-the-Loop (HITL) protocol ensures expert validation. The framework generates a “Core Ontology Framework” providing actionable insights for divestment decisions, as demonstrated in a case study involving SanDisk and Samsung.
Managing vast collections of patents, known as patent portfolios, is a complex and often inefficient task for companies and research institutions. Traditional methods for valuing these patents often look backward, relying on historical data, or involve time-consuming manual analysis. This can lead to underutilized assets and missed opportunities for technology transfer – the process of moving technology from one organization to another for commercialization.
A new research paper introduces an innovative, multi-stage hybrid artificial intelligence (AI) framework designed to automate and enhance this process. The framework aims to strategically prune patent portfolios, identifying high-value assets that are ripe for technology transfer by integrating advanced AI techniques with a deep understanding of market needs.
A Hybrid Approach to Patent Valuation
The core of this framework lies in its hybrid intelligence approach, combining quantitative analysis with qualitative market insights. It moves beyond static indicators to offer a dynamic, forward-looking analysis grounded in current market realities. The system is built upon three main contributions:
First, it uses a sophisticated Learning-to-Rank (LTR) model. This model conducts a multi-faceted quantitative analysis, evaluating patents against more than 30 legal, technical, and commercial parameters. Unlike simpler models, LTR can learn complex, non-linear relationships between these factors, providing a more nuanced ranking of patents based on their potential for commercial transfer.
Second, an innovative “Need-Seed” agent architecture is introduced. This dual-agent system uses advanced Natural Language Processing (NLP) and fine-tuned Large Language Models (LLMs) to bridge the gap between a patent’s technological capabilities (the “Seed”) and documented market demands (the “Need”).
Finally, the framework incorporates a robust Human-in-the-Loop (HITL) validation protocol. This ensures that the AI-driven insights are credible, defensible, and align with the accountability structures of high-stakes business decisions. Human experts review and approve critical stages of the analysis, and their feedback is used to continuously refine the AI models.
The Need-Seed Nexus: Connecting Technology to Market Demand
The “Need Agent” acts as a market intelligence engine. It continuously mines vast amounts of unstructured market and industry data – including market research reports, financial news, corporate disclosures, and scientific literature – using NLP techniques like Named Entity Recognition, Sentiment Analysis, and Relation Extraction. This agent builds a dynamic Knowledge Graph of explicit industry needs, strategic challenges, and technological “wants.” This graph helps identify problems that need solutions.
Complementing this is the “Seed Agent.” This agent uses LLMs, specifically fine-tuned on patent documents, to deeply understand the technological solutions offered by patents. It performs semantic claim analysis, identifying core components, assessing claim breadth, mapping problem-solution connections, and even providing preliminary assessments of how difficult it would be for competitors to design around the patent. This agent generates a structured “Seed Profile” for each patent, detailing its core technological solution and legal protection.
The culmination of this analysis is the Core Ontology Framework. This framework matches the “Seed Profile” of high-potential patents with the documented “Needs” in the Knowledge Graph. When a strong match is found, the system generates a comprehensive report. This report includes details about the patent, the specific market need it addresses, quantitative and qualitative scores, estimated opportunity size, risk profile, and actionable strategic recommendations, such as initiating targeted licensing discussions.
Real-World Application: The SanDisk Case Study
The paper illustrates the framework’s power with a use case involving SanDisk’s patent portfolio. Tasked with monetizing non-core assets, SanDisk’s team used the framework to analyze nearly 19,000 patents. By selecting a “Quick Monetization/Non-Core Assets” strategic profile, the LTR model prioritized patents with high forward citation velocity and strong litigation history, identifying a cluster of 116 high-value patents.
The Need Agent then analyzed public data related to Samsung’s memory division, identifying a strategic crisis due to competitive lag in High-Bandwidth Memory (HBM) and technical hurdles. Crucially, it found that some of SanDisk’s patents were direct obstacles to Samsung’s R&D. The Seed Agent then deconstructed the technological solutions within SanDisk’s patents, identifying core pillars like in-memory computing and advanced memory architecture.
The framework successfully matched SanDisk’s technological “Seed” with Samsung’s acute market “Need,” generating a Core Ontology Framework that provided a complete, evidence-backed business case for a targeted divestment. This demonstrated the system’s ability to not just value patents, but to identify precisely who would find them most valuable and why, leading to actionable strategic recommendations. You can read more about this innovative framework in the full research paper available here.
Also Read:
- AI-Powered Scenario Generation for Complex Service Environments
- Streamlining E-commerce Product Matching with Multi-Agent AI
Future Directions
The researchers plan to further refine the training data for the AI models, enhance the Seed Agent’s capabilities for assessing claim breadth and design-around difficulty, and explore multimodal analysis by integrating computer vision for patent drawings. Expanding to cross-jurisdictional analysis and incorporating causal inference models for the Need Agent are also key future objectives, along with adapting the framework for other forms of intellectual property like trademarks and trade secrets.
