Contrastive learning enhanced retrieval-augmented few-shot framework for multi-label patent classification.

Area of Science:

• Intellectual Property Analysis
• Machine Learning
• Computer Science

Background:

• Patent databases are rapidly expanding, creating challenges for multi-label patent classification.
• Conventional methods struggle with high annotation costs, scalability, and capturing semantic patent structures.

Purpose of the Study:

• To develop a scalable multi-label patent classification framework using retrieval-enhanced few-shot learning.
• To address limitations of existing methods by incorporating patent-specific contrastive pre-training and semantic retrieval.

Main Methods:

• A retrieval-enhanced few-shot learning framework combining contrastive pre-training and semantic retrieval.
• Domain-adapted embeddings capturing multi-label co-occurrence patterns.
• Retrieval-augmented few-shot learning with structured reasoning for reduced annotation dependency.

Main Results:

• Achieved Macro-F1 of 0.847 and Micro-F1 of 0.892 on a drone patent dataset, outperforming few-shot baselines by 30% and 23%.
• Contrastive pre-training improved performance on underrepresented categories by up to 16% compared to transformer-based methods.
• Demonstrated significant improvements in multi-label classification accuracy and efficiency.

Conclusions:

• The proposed framework offers an effective and resource-efficient solution for multi-label patent classification.
• Enhances scalability and accessibility of intellectual property analysis.
• Highlights the benefits of combining contrastive learning and retrieval for complex patent data.