Racist-Detection-App

This is a production-ready, end-to-end system developed to detect and classify racist tweets using advanced Natural Language Processing (NLP) techniques. Built on top of BERTweet (vinai/bertweet-base) and fine-tuned with a robust, k-fold cross-validation training pipeline, powered by streamlit UI!

📝 Racist/Sexist Tweet Detection App

1. Overview

This project implements a robust racist/sexist tweet detection system using the BERTweet (vinai/bertweet-base) model, fine-tuned on domain-specific labeled data. It includes a training pipeline, preprocessing, inference module, and a Streamlit web application for real-time predictions.

2. Features

• Fine-tuning with 5-fold cross-validation and best-model selection
• Advanced tweet preprocessing pipeline (including spell correction)
• Offline-friendly reusable inference module
• User-friendly Streamlit web application

3. Model: vinai/bertweet-base

• A transformer-based language model pre-trained on 850M English tweets.
• Fine-tuned the model on a binary classification task:
  • non_racist
  • racist

4. Workflow

4.1 Training (bertweet_train.py)

• Loads cleaned and oversampled dataset.
• Performs 5-fold stratified cross-validation.
• Applies oversampling to balance the minority class.
• Preprocesses tweets using text_preprocessor.py.
• Fine-tunes vinai/bertweet-base using HuggingFace Trainer.
• Saves metrics and confusion matrix per fold.
• Selects best fold based on F1-score.

4.1 Best Model Consolidation

• Automatically finds the fold with the highest F1-score by reading eval_results.json in each fold directory.

4.3 Prediction (predict.py)

• Loads the best fold model automatically.
• Runs predictions on the test set or new data.
• Prints classification report and saves predictions to CSV.

4.4 Spell Correction (symspellpy_local.py)

• Prepares and saves the SymSpell dictionary locally for use during inference.

4.5 Inference (inference.py)

• Loads the best fold model.
• Preprocesses text (including spell correction, emojis, hashtags).
• Performs inference on single tweet input.
• Returns label_id, label_name, and confidence score.

4.6 Streamlit App (app.py)

• Real-time web interface.
• Server-side preprocessing and model inference.
• Displays prediction with confidence.

5. Evaluation

• Validation metrics per fold:
• Accuracy, Precision, Recall, F1-score
• Classification report
• Confusion matrix

6. Preprocessor: text_preprocessor.py

• Preprocessing steps:
  • Converts to UTF-8
  • Removes irrelevant characters and stopwords
  • Corrects spelling with SymSpell
  • Extracts meaningful elements: emojis, hashtags, mentions
  • Ensures consistency with training preprocessing

7. Web App: 🌐 Streamlit

• Run the web app for interactive predictions:
• bash streamlit run app.py
• Features:
  • Single tweet input
  • Clean and preprocess
  • Predict and display label_name & confidence

8. Project Structure

racist-detection/ ├── app.py # Streamlit UI ├── bertweet_train.py # Training with 5-fold CV ├── predict.py # Batch predictions & evaluation ├── inference.py # Single-tweet inference logic ├── symspellpy_local.py # Save SymSpell dictionary locally ├── text_preprocessor.py # Text preprocessing logic ├── bertweet_train_dir2/ # Fine-tuned models & metrics │ └── vinai-bertweet-base-foldX/ ├── requirements.txt ├── README.md

9. Installation

• Install dependencies:
  • bash pip install -r requirements.txt

10. Example Inference Output

• Sample prediction: (1, 'racist', 0.84)

11. Future Work

• Batch predictions with CSV upload
• More granular sentiment categories
• Explainable AI integration (e.g., SHAP)
• Deploy on HuggingFace Spaces or Docker

12. Acknowledgments

• vinai/bertweet-base by VinAI Research
• 🤗 HuggingFace Transformers & Datasets
• Streamlit