| Property | Value |
|---|---|
| Parameter Count | 135M |
| License | AFL-3.0 |
| Supported Languages | 10 (ar, de, en, es, fr, it, lv, nl, pt, zh) |
| Task Type | Named Entity Recognition |
What is distilbert-base-multilingual-cased-ner-hrl?
This is a specialized Named Entity Recognition (NER) model built on the DistilBERT architecture, designed to identify three types of entities (Person, Organization, Location) across 10 high-resource languages. The model leverages knowledge distillation techniques to maintain performance while reducing computational requirements compared to BERT.
Implementation Details
The model is implemented using the Transformers library and can be easily deployed using the HuggingFace pipeline API. It utilizes a distilled version of multilingual BERT, trained on carefully curated datasets from various sources including CoNLL 2002/2003, ANERcorp, and other language-specific corpora.
- Token Classification Architecture with B-I-O tagging scheme
- Supports real-time inference with F32 precision
- Trained on NVIDIA V100 GPU with optimized hyperparameters
Core Capabilities
- Multilingual NER across 10 languages with single model
- Detection of Person (PER), Organization (ORG), and Location (LOC) entities
- Distinguishes between beginning (B) and inside (I) tokens for consecutive entities
- 37K+ downloads indicating strong community adoption
Frequently Asked Questions
Q: What makes this model unique?
The model's key strength lies in its ability to perform NER across 10 different languages while maintaining a relatively compact size through distillation. It's particularly valuable for multilingual applications where deploying separate models for each language would be impractical.
Q: What are the recommended use cases?
This model is ideal for processing news articles, documents, and general text analysis requiring entity extraction in any of the supported languages. It's particularly well-suited for applications in news analysis, content categorization, and information extraction systems.