🚨 OffensEval NLP

Transformer-Based Offensive Language Detection with Sentiment-Aware Modeling

A production-ready NLP system for offensive language detection, leveraging DistilBERT sentence embeddings, sentiment-aware feature augmentation, and a Logistic Regression classifier, deployed as an interactive Streamlit web application with CI-validated inference and automated testing.

---

🌐 Live Demo

👉 Deployed Application: https://offenseval-nlp-ojkd85r386zbffvjbvrkst.streamlit.app/

The app demonstrates real-time offensive language detection with sentiment transparency and confidence-based predictions.

---

📌 Project Overview

This project implements a complete end-to-end offensive language detection pipeline, covering:

• Text preprocessing & normalization
• Transformer-based semantic embeddings
• Sentiment-aware feature engineering
• Supervised classification
• Model evaluation & validation
• Automated inference testing (CI-ready)
• Interactive web-based deployment

The system is trained and evaluated on the TweetEval – Offensive Language benchmark and designed with interpretability, robustness, and deployability in mind.

---

🎯 Key Features

• 🔍 Transformer-based embeddings (DistilBERT via SentenceTransformers)
• 😊 Sentiment-aware modeling using VADER compound scores
• ⚖️ Interpretable classifier (Logistic Regression)
• 📊 Confidence scores & transparent explanations
• 🧪 Automated inference testing using PyTest
• 🌐 Interactive Streamlit web application
• 🚀 Deployment-ready architecture
• 📁 Clean, modular, industry-standard repository structure

---

🧠 Model Architecture

Input Text
   ↓
Text Cleaning & Normalization
   ↓
DistilBERT Sentence Embeddings (768D)
   ↓
VADER Sentiment Score (+1 feature)
   ↓
Feature Concatenation (769D)
   ↓
Logistic Regression Classifier
   ↓
Offensive / Not Offensive Prediction

---

📊 Dataset

• Dataset: TweetEval – Offensive Language

• Task: Binary classification

  • Offensive
  • Not Offensive

---

📈 Model Performance (Validation)

Metric	Score
Macro F1	~0.72
ROC-AUC	~0.81

Metrics are reported on a held-out validation set following standard NLP evaluation practices.

---

🌐 Web Application Features

The deployed application provides:

• ✔️ Preprocessed text preview (model input transparency)
• ✔️ Predicted label with confidence score
• ✔️ Sentiment analysis (Positive / Neutral / Negative)
• ✔️ Color-coded sentiment bars with percentage distribution
• ✔️ Clear explanation of model reasoning
• ✔️ Explicit disclaimers and limitations

---

🧪 Automated Testing (CI-Ready)

Inference tests validate:

• Model loading and compatibility

• Correct preprocessing output shape

• Robust prediction across:

  • Short, medium, and long inputs
  • Offensive and non-offensive language

• Proper probability normalization

Run locally with:

pytest tests/test_inference.py

---

📁 Repository Structure

OffensEval-NLP/
│
├── app.py                     # Streamlit web application
├── predict.py                 # CLI inference script
├── preprocessing.py           # Centralized preprocessing & features
│
├── model/
│   ├── final_model_tuned_distilbert.joblib
│   └── label_encoder.json
│
├── embeddings/                # Saved embedding artifacts
│
├── tests/
│   └── test_inference.py      # Automated inference tests
│
├── Model_Training_and_Evaluation.ipynb
├── requirements.txt
├── .gitignore
├── LICENSE
└── README.md

---

⚠️ Disclaimer & Limitations

This system predicts offensive language based on patterns learned from social media data.

Limitations:

• Sarcasm and humor may not be interpreted correctly
• Cultural and contextual nuances may be missed
• Predictions are intended to support, not replace, human judgment

Use responsibly.

---

👨‍💻 Developer

Arunava Kumar Chakraborty

Data Analyst | ML Enthusiast

🔗 https://www.linkedin.com/in/arunava-kr-chakraborty

🐙 https://github.com/ArunavaKumar

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

⭐ Acknowledgements

• HuggingFace SentenceTransformers
• VADER Sentiment Analysis
• TweetEval Benchmark
• Streamlit