F-coref: Fast Coreference Resolution Model
| Property | Value |
|---|---|
| License | MIT |
| Paper | F-coref Paper |
| Performance | 78.5 F1 on OntoNotes |
| Frameworks | PyTorch, Transformers |
What is f-coref?
F-coref is a groundbreaking coreference resolution model that combines speed and accuracy through innovative distillation techniques. Developed by the BIU-NLP team, it processes 2.8K OntoNotes documents in just 25 seconds on a V100 GPU, significantly faster than existing solutions while maintaining competitive accuracy.
Implementation Details
The model achieves its exceptional speed through two key innovations: model distillation from the LingMess architecture and an efficient batching implementation using a novel "leftover" technique. This results in dramatically reduced processing times compared to traditional models like AllenNLP (12 minutes) and LingMess (6 minutes).
- Efficient batching implementation with leftover technique
- Compact model architecture through distillation
- Optimized memory usage (4.0 GiB)
- Built on PyTorch and Transformers frameworks
Core Capabilities
- Fast processing: 25 seconds for 2.8K documents
- High accuracy: 78.5 F1 score on OntoNotes
- Memory efficient: Only 4.0 GiB usage
- English language support
- Compatible with multi_news and ontonotes datasets
Frequently Asked Questions
Q: What makes this model unique?
F-coref stands out for its exceptional speed-to-accuracy ratio, processing documents up to 28 times faster than traditional models while maintaining strong performance. Its innovative batching technique and distilled architecture make it particularly suitable for large-scale applications.
Q: What are the recommended use cases?
The model is ideal for applications requiring fast, accurate coreference resolution, particularly when processing large document collections. It's especially suitable for production environments where processing speed is crucial while maintaining accuracy.
