Real-Time Hate Speech Detection

Overview

This project implements a real-time hate speech detection system using DistilBERT, a lightweight transformer model. Designed to classify text into Hate and NoHate. The model is trained on a balanced dataset and deployed with the aim of creating a complete end-to-end application.

Deployment

This project is deployed and accessible online:

Frontend (Netlify): https://hate-speech-detection-007.netlify.app/ enter text and instantly see whether it is classified as hate speech or not, along with the model's confidence score.

Backend API (Hugging Face Spaces): https://architrawat25-real-time-hate-speech-detection.hf.space FastAPI backend serves predictions from a transformer-based model and exposes REST endpoints for both single and batch predictions.

Objective

• Prymary objective: to build a hate speech detection model using DistilBERT.
• Balance the dataset using Back Translation to improve model performance.
• Develop an API for real-time inference.
• Deploy the API using Docker and integrate it with cloud services for real-time inference

Dataset

The dataset is sourced from https://huggingface.co/datasets/odegiber/hate_speech18. This original dataset was highly imbalanced with approximately 9:1 ratio of non-hate to hate speech samples. To address this, Back Translation was applied using the facebook/m2m100_418M model, creating a more balanced dataset for training. https://huggingface.co/architrawat25

Approach

1. Data Preprocessing:

   • Clean text data by removing special characters and lowercasing.
   • Convert text files to a structured CSV format.

2. Dataset Balancing:

   • Applied Back Translation to augment minority class samples.
   • Generated balanced dataset with an equal number of Hate and NoHate samples.

3. Model Training:

   • Fine-tuned DistilBERT on the balanced dataset.
   • Achieved high performance with 94% accuracy on the test set.

4. Model Saving & Hosting:

   • Saved the fine-tuned model.
   • Uploaded the model to Hugging Face for easy access and deployment. https://huggingface.co/architrawat25

5. API Development (Upcoming):

   • Develop an API using Flask or FastAPI.
   • Provide endpoints for text classification.

6. Cloud Deployment (Upcoming):

   • Deploy the API on cloud platforms like Heroku or AWS.

Results

The model achieved the following performance on the test dataset:

Metric	Value
Accuracy	94%
Precision	94%
Recall	94%
F1 Score	94%

The model demonstrated consistent performance across both hate and non-hate classes.

Dependencies

• Python
• Transformers
• PyTorch
• Hugging Face
• NLTK
• Pandas
• Scikit-learn

Next Steps

• Build API for real-time text classification.
• Dockerise and then deploy the API to the cloud.
• Create a interface for easier user interaction.

Author

Archit Rawat

License

This project is licensed under the MIT License.