🛡️ Toxic Comment Classification

Multi-label toxicity detection system with 3 ML models deployed on AWS — from classical ML to multilingual Transformers.

---

Overview

This project builds a production-ready multi-label toxic comment classifier that detects 6 categories of harmful content in text. Three models — ranging from a fast classical ML baseline to a 560M-parameter multilingual Transformer — are each deployed as independent AWS Lambda microservices behind a unified API Gateway.

The system includes a real-time Wikipedia edit analyzer, monitoring toxicity across 6 language editions (EN, FR, DE, ES, IT, AR) — directly relevant to content moderation at scale, a core challenge in AI Safety.

🔗 Live Application : toxic-classifier-frontend

Page	Description
Analyzer	Real-time comment analysis with model selection
Dashboard	Side-by-side comparison of all 3 models
Wikipedia Live	Live toxicity monitoring of Wikipedia edits

---

Key Results

Model	F1-Score	Latency	Languages	Size
XGBoost + TF-IDF	0.76	~50ms	English	—
RoBERTa	0.80	~300ms	English	125M params
XLM-RoBERTa	~0.78	~400ms	100+	560M params

Best accuracy : RoBERTa (F1 = 0.80) · Best multilingual coverage : XLM-RoBERTa (100+ languages)

---

AI Safety Relevance

Toxic content detection is a foundational problem in AI Safety and content moderation :

• Scalable oversight : automated flagging of harmful content reduces human reviewer burden
• Multilingual coverage : XLM-RoBERTa enables safety monitoring beyond English-centric systems
• Model comparison : the dashboard exposes tradeoffs between speed, accuracy, and coverage — critical for deployment decisions
• Real-world grounding : Wikipedia live monitoring tests the system on genuine adversarial content

---

Dataset

• Source : Kaggle Jigsaw Toxic Comment Classification Challenge
• Size : 159,571 annotated Wikipedia comments
• Task : Multi-label classification (a comment can belong to multiple categories)

Label	Description
toxic	General toxicity
severe_toxic	Extreme toxicity
obscene	Obscene language
threat	Threats of violence
insult	Personal insults
identity_hate	Hate speech targeting identity

---

Architecture

┌─────────────────────────────────────┐
│         AWS S3 (Frontend)           │
│   React App · Dashboard · Wikipedia │
└───────────────┬─────────────────────┘
                │
                ▼
┌─────────────────────────────────────┐
│         AWS API Gateway             │
│  /xgboost  /roberta  /multilingual  │
└──────────┬──────────┬───────────────┘
           │          │          │
           ▼          ▼          ▼
  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
  │Lambda XGBoost│ │Lambda RoBERTa│ │Lambda XLM-R  │
  │  (Docker)    │ │  (Docker)    │ │  (Docker)    │
  │   512 MB     │ │   2048 MB    │ │   3008 MB    │
  └──────────────┘ └──────────────┘ └──────────────┘

Each model runs in an isolated Docker container on AWS Lambda, enabling independent scaling and versioning.

---

API Reference

Base URL : https://0hik6heuhc.execute-api.us-east-1.amazonaws.com/prod

Endpoint	Method	Model
/xgboost/predict	POST	XGBoost + TF-IDF
/roberta/predict	POST	RoBERTa
/multilingual/predict	POST	XLM-RoBERTa
/*/health	GET	Health check

Request

curl -X POST https://0hik6heuhc.execute-api.us-east-1.amazonaws.com/prod/multilingual/predict \
  -H "Content-Type: application/json" \
  -d '{"text": "You are absolutely worthless."}'

Response

{
  "is_toxic": true,
  "toxic_probability": 0.89,
  "language_detected": "en",
  "confidence": "high",
  "model": "XLM-RoBERTa Multilingual"
}

---

Project Structure

toxic-comment-classification/
├── deployment/
│   ├── lambda-xgboost/          # XGBoost microservice
│   │   ├── app.py               # FastAPI + Mangum handler
│   │   ├── Dockerfile
│   │   └── requirements.txt
│   ├── lambda-roberta/          # RoBERTa microservice
│   ├── lambda-multilingual/     # XLM-RoBERTa microservice
│   ├── frontend/                # React application
│   └── dashboard/               # Static comparison dashboard
├── documentation/
│   ├── XGBOOST_MODEL.md
│   ├── ROBERTA_MODEL.md
│   └── MULTILINGUAL_MODEL.md
└── README.md

---

Quick Start

git clone https://github.com/Bassongo/toxic-comment-classification.git
cd toxic-comment-classification

# Deploy XGBoost Lambda
cd deployment/lambda-xgboost
docker build -t toxic-xgboost .
aws ecr get-login-password | docker login --username AWS --password-stdin <account>.dkr.ecr.us-east-1.amazonaws.com
docker tag toxic-xgboost:latest <account>.dkr.ecr.us-east-1.amazonaws.com/toxic-xgboost:latest
docker push <account>.dkr.ecr.us-east-1.amazonaws.com/toxic-xgboost:latest

Prerequisites : AWS CLI · Docker Desktop · Node.js 18+

---

Tech Stack

Layer	Technologies
ML / NLP	XGBoost · PyTorch · HuggingFace Transformers
Backend	Python 3.11 · FastAPI · Mangum
Cloud	AWS Lambda · API Gateway · S3 · ECR
Frontend	React · Chart.js · Axios
DevOps	Docker

---

Limitations & Future Work

• F1 ceiling : XGBoost (0.76) struggles on minority labels (severe_toxic, threat) — upsampling or cost-sensitive learning could help
• African languages : XLM-RoBERTa covers Swahili but not Wolof, Bambara, or Mooré — a natural extension given the African NLP gap
• RLHF integration : the classifier could serve as a reward signal in a Constitutional AI-style training loop
• Evaluation depth : adding AUC-ROC per label and calibration curves would better expose model reliability

---

Author

Marc MARE — Statistics & ML Engineer
ENSAE Dakar | MSc SEP, University of Reims (2026)

---

License

MIT License — see LICENSE for details.