NewsNER-DistilBERT

A fine-tuned DistilBERT-based Named Entity Recognition (NER) model for extracting people, organizations, locations, and miscellaneous entities from English news text.

The model was fine-tuned on the CoNLL-2003 dataset and evaluated using entity-level precision, recall, and F1-score.

Model Overview

Property	Details
Model	DistilBERT
Task	Named Entity Recognition
Domain	English News Articles
Base Model	distilbert-base-uncased
Dataset	CoNLL-2003
Entity Types	PER, ORG, LOC, MISC
Evaluation	Entity-level
Overall F1	88.79%

Entity Labels

The model recognizes four main entity categories:

• PER — Person
• ORG — Organization
• LOC — Location
• MISC — Miscellaneous entities

Training

The model was fine-tuned using the CoNLL-2003 Named Entity Recognition dataset.

The training pipeline includes:

1. Tokenization using the DistilBERT tokenizer
2. Alignment of word-level NER labels with tokenized inputs
3. Fine-tuning DistilBERT for token classification
4. Evaluation using entity-level precision, recall, and F1-score
5. Comparison against traditional and spaCy-based baselines

The resulting model is designed to provide a lightweight Transformer-based alternative for practical NER inference.

Dataset

The model was trained and evaluated on the CoNLL-2003 dataset.

Dataset splits:

Split	Samples
Train	14,041
Validation	3,250
Test	3,453

The dataset contains news articles annotated with named entities including persons, organizations, locations, and miscellaneous entities.

Evaluation Results

The official entity-level evaluation on the test set achieved:

Metric	Score
Precision	87.92%
Recall	89.68%
F1	88.79%

Per-Entity Performance

Entity	Precision	Recall	F1
PER	95.02%	94.50%	94.76%
ORG	84.28%	87.78%	85.99%
LOC	91.84%	91.07%	91.45%
MISC	72.82%	79.77%	76.14%

The strongest performance was achieved on PER and LOC entities, while MISC remains the most challenging category.

Baseline Comparison

The fine-tuned DistilBERT model was evaluated against several baselines:

Approach	F1
Rule-Based	10.15%
spaCy Small	56.22%
spaCy Medium	57.19%
Fine-tuned DistilBERT	88.79%

This comparison demonstrates the improvement achieved by fine-tuning a Transformer-based model for the target NER task.

Example

Input

Apple CEO Tim Cook announced a new investment in India during a meeting in New Delhi. Microsoft also plans to expand its operations in London.

Extracted Entities

Apple        → ORG
Tim Cook     → PER
India        → LOC
New Delhi    → LOC
Microsoft    → ORG
London       → LOC

The deployed application also provides confidence scores for individual predictions.

How to Use

Install the required libraries:

pip install transformers torch

Load the model using the Hugging Face Transformers pipeline:

from transformers import pipeline

ner_pipeline = pipeline(
    "token-classification",
    model="AbdelrahmanAkl/NewsNER-DistilBERT",
    aggregation_strategy="simple"
)

text = """
Apple CEO Tim Cook announced a new investment in India.
"""

results = ner_pipeline(text)

for entity in results:
    print(
        entity["word"],
        entity["entity_group"],
        entity["score"]
    )

Intended Use

This model is intended for experimentation, research, education, and practical NLP applications involving English news text.

Potential applications include:

• News article analysis
• Information extraction
• Document processing
• Entity-based search
• Knowledge graph construction
• News analytics
• NLP preprocessing pipelines

Limitations

The model was trained on the CoNLL-2003 news domain and may not generalize equally well to other domains such as:

• Medical documents
• Legal documents
• Financial reports
• Social media
• Informal conversations
• Highly technical documents

Performance may also decrease on text containing entities, terminology, or writing styles that differ significantly from the training distribution.

The MISC category has lower performance than PER, ORG, and LOC and should therefore be treated with additional caution.

Ethical Considerations

Named Entity Recognition systems can be used to automatically extract information about people, organizations, and locations.

Users should consider:

• Privacy when processing sensitive documents
• Potential downstream errors
• Biases present in the training data
• The consequences of incorrect entity extraction

Predictions should be validated before being used in high-impact automated decision-making systems.

Project

This model is part of the NewsNER-AI project, a complete Named Entity Recognition application built around a fine-tuned Transformer model.

The project includes:

• Model fine-tuning
• Entity-level evaluation
• Baseline benchmarking
• Streamlit inference application
• Hugging Face model deployment
• Production-oriented documentation

Resources

GitHub Repository

https://github.com/AbdelrhmanAkl/News-NER-Entity-Recognition

Live Demo

https://news-ner-entity-recognition.streamlit.app/

Author

Abdelrahman Akl

AI Engineer | NLP | LLMs | RAG | Agentic AI

GitHub: https://github.com/AbdelrhmanAkl

---

If you use this model in a project, please consider referencing the original CoNLL-2003 dataset and the underlying DistilBERT architecture.