PatentSBERTa_V2
| Property | Value |
|---|---|
| Authors | Bekamiri, H., Hain, D. S., & Jurowetzki, R. |
| Paper | Publication Link |
| Vector Dimension | 768 |
| Architecture | MPNet-based with Mean Pooling |
What is PatentSBERTa_V2?
PatentSBERTa_V2 is a specialized deep learning model developed by Aalborg University Business School's AI Growth-Lab for patent analysis and classification. It's built upon the SBERT architecture and has been specifically optimized for processing patent documents, offering sophisticated semantic understanding of technical and legal patent language.
Implementation Details
The model employs a hybrid architecture based on MPNet with mean pooling strategy, trained using cosine similarity loss. It was trained over 4 epochs with a batch size of 16, using AdamW optimizer with a learning rate of 2e-05 and warmup steps of 664.
- Maximum sequence length of 512 tokens
- Mean pooling implementation for sentence embeddings
- Integrated with both sentence-transformers and HuggingFace frameworks
- Optimized with weight decay of 0.01
Core Capabilities
- Patent distance calculation for similarity analysis
- Dense vector encoding (768-dimensional space)
- Semantic search across patent documents
- Patent classification and categorization
- Cross-patent comparison and analysis
Frequently Asked Questions
Q: What makes this model unique?
PatentSBERTa_V2 is specifically designed for patent analysis, combining SBERT architecture with patent-specific optimizations. Its hybrid approach and specialized training make it particularly effective for technical document processing and classification.
Q: What are the recommended use cases?
The model excels in patent similarity search, classification tasks, technical document comparison, and patent landscape analysis. It's particularly useful for R&D departments, patent offices, and intellectual property researchers.
