🇵🇰 Urdu AI Sentiment System

Fine-tuning mBERT for Low-Resource Urdu NLP + Real-Time Deployment

Project Overview

This project presents an end-to-end Urdu sentiment analysis system built by fine-tuning a multilingual transformer model on an Urdu dataset. It classifies Urdu text into Positive, Negative, and Neutral sentiments and is deployed as a real-time web application using Gradio and Hugging Face Spaces.

It demonstrates a complete NLP pipeline from

data preprocessing → model training → evaluation → deployment.

Problem Statement

Urdu is a low-resource language in Natural Language Processing (NLP), with limited labeled datasets and pretrained models.

This creates challenges for building accurate sentiment analysis systems in real-world Urdu applications.

Objectives:

• Build a robust Urdu sentiment classification system
• Fine-tune a transformer-based model on Urdu dataset
• Develop an end-to-end deployable AI pipeline
• Enable real-time sentiment prediction through a web interface

Key Project Highlights

• 🇵🇰 Focus on Urdu language NLP
• Fine-tuned mBERT transformer model
• Real-time inference with Gradio
• Deployed on Hugging Face Spaces
• Full NLP pipeline training + evaluation + deployment

Model Used

Use a pretrained multilingual transformer:

👉 mBERT

Why mBERT?

• Supports Urdu language
• Strong performance on low-resource NLP tasks
• Pretrained on multilingual corpora
• Easy fine-tuning for classification tasks

Dataset

• Source: HuggingFace Dataset Hub
• Dataset: Urdu Multi-Domain Classification

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

Data Preprocessing

The dataset was cleaned and prepared using the following steps:

• Removed irrelevant columns: topic, intent, binary
• Filtered dataset to keep only 3 sentiment classes
• Removed low-frequency class request
• Ensured balanced class distribution

Text Normalization

Urdu text was normalized using LughaatNLP.

✔ Benefits:

• Removes noise and inconsistencies
• Standardizes Urdu spelling variations
• Improves tokenization quality
• Enhances model generalization

🔁 Project Pipeline

Raw Urdu Text
      ↓
Text Normalization (LughaatNLP)
      ↓
Tokenization (mBERT Tokenizer)
      ↓
Fine-tuned Transformer Model
      ↓
Sentiment Prediction (Positive / Negative / Neutral)
      ↓
Gradio Web Interface

Application UI

Tech Stack

• Python
• PyTorch
• Hugging Face Transformers
• Datasets Library
• Gradio UI
• Scikit-learn

Model Training

• Model: mBERT multilingual BERT
• Task: Sentiment Classification
• Loss Function: CrossEntropyLoss
• Optimizer: AdamW
• Evaluation: Accuracy, Precision, Recall, F1-score
• Best model saved using Hugging Face Trainer

Evaluation Results

The model achieves strong performance on Urdu sentiment classification:

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

Deployment

The model is deployed using:

• Gradio interface
• Hugging Face Spaces

It supports real-time prediction of Urdu text sentiment.

🌐 Live Demo

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

Key Features

✔ End-to-end NLP pipeline

✔ Low-resource language support Urdu

✔ Transformer-based fine-tuning

✔ Real-time web deployment

✔ Clean and interactive UI

Conclusion

This project demonstrates a complete end-to-end NLP system for Urdu sentiment analysis, covering:

• Fine-tuning transformer models for Urdu NLP
• Effective preprocessing for low-resource languages
• Evaluation using standard ML metrics
• Deployment using Gradio and Hugging Face Spaces

Run Locally

git https://github.com/batoolarifa/urdu-ai-sentiment-system.git

pip install -r requirements.txt
streamlit run app.py

👤 Author

Syeda Arifa Batool
SE @ Karachi University | AI/ML Engineer | Applying technology to create real-world value 📈

🔗 Connect with Me

• LinkedIn: Syeda Arifa Batool
• Kaggle: Syeda Arifa Batool
• Email: thearifabatool@gmail.com

⭐ If you find this project useful, feel free to star the repository!