This model is a fine-tuned Multilingual BERT (mBERT) for Urdu sentiment classification. It classifies Urdu text into three categories: Positive, Negative, and Neutral.

Task

Urdu Text Classification for Sentiment Analysis

Model Description

Base Model: bert-base-multilingual-cased
Architecture: Transformer (BERT)
Task: Sentiment Classification
Language: Urdu
Framework: Hugging Face Transformers

This model is optimized for low-resource Urdu NLP using transfer learning from a pretrained multilingual transformer.

Dataset

This model was trained using a publicly available Urdu sentiment dataset from Hugging Face:

https://huggingface.co/datasets/umar178/UrduMultiDomainClassification

Dataset Description

The dataset contains Urdu text samples annotated for sentiment analysis tasks.
It was used to fine-tune the multilingual BERT model for classification into:

Positive
Negative
Neutral

This dataset is suitable for low-resource NLP research in Urdu language understanding.

Training Pipeline

Raw Urdu Text → Tokenization → mBERT Encoder → Classification Head → Sentiment Output

Here is an example of how you can run this model:

python

from transformers import pipeline

model_name = "arifa-batool/urdu-sentiment-analysis-mbert"

classifier = pipeline(
    "text-classification",
    model=model_name,
    tokenizer=model_name
)

text = "یہ فلم بہت اچھی تھی"
result = classifier(text)

print(result)

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "arifa-batool/urdu-sentiment-analysis-mbert"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "یہ بہت بری خبر ہے"

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)

probs = torch.softmax(outputs.logits, dim=1)

pred_id = torch.argmax(probs, dim=1).item()
confidence = torch.max(probs).item()

label = model.config.id2label[pred_id]

print(label, confidence)

Evaluation Results

Accuracy: 0.91
F1 Score: 0.91
Balanced performance across all sentiment classes

Deployment

Available via:

Hugging Face Model Hub
Hugging Face Spaces (Gradio App)
Transformers API

Future Improvements

Multi-domain Urdu dataset expansion
Integration with larger models (XLM-R, DeBERTa)
Social media sentiment optimization

Author

Syeda Arifa Batool | AI/ML Engineer

Live Demo

You can try the model here:

https://huggingface.co/spaces/arifa-batool/urdu-sentiment-classifier