REBEL-Large
| Property | Value |
|---|---|
| Parameter Count | 406M |
| License | CC-BY-NC-SA 4.0 |
| Architecture | BART-based Seq2Seq |
| Task | Relation Extraction |
What is rebel-large?
REBEL (Relation Extraction By End-to-end Language generation) is a state-of-the-art model that revolutionizes relation extraction by treating it as a sequence-to-sequence task. Developed by Babelscape, this model achieves impressive performance on multiple benchmarks, including 93.4 F1 score on NYT dataset.
Implementation Details
Built on the BART architecture, REBEL-large employs a novel linearization approach that enables end-to-end relation extraction for over 200 different relation types. The model processes raw text and generates structured triplets (subject, relation, object) through autoregressive sequence generation.
- Uses special tokens for triplet structure (<triplet>, <subj>, <obj>)
- Supports batch processing with customizable generation parameters
- Implements beam search with configurable parameters
Core Capabilities
- End-to-end relation extraction without pipeline requirements
- Handles multiple relation types simultaneously
- Achieves SOTA performance on major benchmarks
- Flexible integration with Hugging Face Transformers library
Frequently Asked Questions
Q: What makes this model unique?
REBEL's uniqueness lies in its end-to-end approach to relation extraction, eliminating the need for complex pipelines while supporting a vast number of relation types. Its seq2seq architecture enables more natural and flexible extraction compared to traditional methods.
Q: What are the recommended use cases?
The model excels in knowledge base population, fact-checking, and information extraction tasks. It's particularly suitable for applications requiring structured relationship extraction from unstructured text, such as document analysis and automated knowledge graph construction.
