bohrariyanshi/pii-ner-extraction

Multilingual NER Model for PII Detection

This model is a fine-tuned version of [bert-base-multilingual-cased](https://huggingface.co/bert-base-multilingual-cased) on the WikiANN dataset for Named Entity Recognition (NER).

Model Description

• Developed by: bohrariyanshi
• Model type: Token Classification (NER)
• Language(s): Multilingual (primarily English)
• Base model: bert-base-multilingual-cased

---

Intended Uses & Limitations

Intended Uses

• Named Entity Recognition for Person (PER), Organization (ORG), and Location (LOC)
• Text analysis and information extraction
• PII (Personally Identifiable Information) detection

Limitations

• Trained on WikiANN (multilingual) but evaluated primarily on English subsets
• May have lower performance on non-English text
• Limited to PER, ORG, LOC entity types

Training Data

The model was fine-tuned on the WikiANN dataset:

• Training examples: 20,000
• Validation examples: 10,000
• Test examples: 10,000
• Entity types: PER (Person), ORG (Organization), LOC (Location)

Training Procedure

Training Hyperparameters

• Learning rate: 2e-5
• Training epochs: 3
• Batch size: 16
• Max sequence length: 256
• Optimizer: AdamW
• Weight decay: 0.01

Performance

The model achieves high confidence predictions on standard NER tasks:

• High confidence predictions (>90%): 19/21 entities in test cases
• Average inference time: ~264ms per sentence
• Entity types detected: PER, ORG, LOC with high accuracy

Usage

from transformers import pipeline

# Load the model
ner = pipeline("ner", model="bohrariyanshi/pii-ner-extraction", aggregation_strategy="simple")

# Example usage
text = "Barack Obama was born in Hawaii."
entities = ner(text)
print(entities)
# Output: [{'entity_group': 'PER', 'score': 0.968, 'word': 'Barack Obama', 'start': 0, 'end': 12}, ...]

Model Architecture

• Base: BERT-base-multilingual-cased
• Parameters: 177M
• Architecture: Transformer with token classification head
• Task: Named Entity Recognition (NER)

Evaluation Results

The model demonstrates superior performance compared to base BERT:

• Confident predictions: 19 high-confidence entities vs 0 for base BERT
• Precision: High accuracy in entity detection
• Speed: ~264ms per sentence (acceptable for production use)

Environmental Impact

Training was performed on a Google Colab T4 GPU for a short duration (fine-tuning only).
The overall environmental impact is minimal compared to large-scale pretraining runs.

Citation

If you use this model, please cite:

@model{bohrariyanshi-pii-ner-extraction,
  author = {bohrariyanshi},
  title = {Multilingual NER Model for PII Detection},
  year = {2025},
  url = {https://huggingface.co/bohrariyanshi/pii-ner-extraction}
}