MULTILINGUAL HSTE SPEECH WALL

• 

  COMMENT GUIDELINES

• 

  A DISCIPLINED COMMENT

• 

  THE COMMENT HAS BEEN POSTED AS IT IS NOT TOXIC

• 

  ABUSIVE COMMENT THAT CAN'T BE POSTED

• 

  AS THE ABUSIVE COMMENTS ARE NOT POSTED IT IS NOT STORED IN THE DATABASE

Inspiration

Online platforms in India face a serious challenge with toxic and abusive comments written in multiple regional languages. Most moderation systems focus only on English and fail to understand multilingual and code-mixed content. During our research, we noticed how harmful comments often bypass moderation due to language gaps. This inspired us to build a real-time multilingual toxicity detection system that works for Indian languages.

What the Project Does

Multilingual Comment Security Wall (HateWall) is a real-time content moderation system that detects and blocks toxic comments across 12 Indian languages. It analyzes user comments instantly and classifies them as safe or toxic before they are posted publicly. This helps platforms create safer digital spaces without relying on manual moderation.

🛠 How We Built It

The system uses a multilingual deep learning model trained on an 800K+ dataset covering multiple Indian languages. The backend is built using Python and FastAPI, with REST APIs handling comment submissions. For toxicity detection, we used a lightweight optimized model and integrated Google Perspective API as an additional confidence layer. Data retrieval and similarity checks are handled using FAISS for efficient vector search. The application is deployed on Render with optimizations to ensure low latency.

Challenges I Faced

One major challenge was deploying heavy machine learning models on limited cloud infrastructure. Platforms like Render have memory and compute constraints, which forced us to optimize the model using ONNX Runtime. Another challenge was handling code-mixed and low-resource languages where labeled data is scarce. Ensuring real-time performance while maintaining accuracy was also difficult.

What I Learned

Through this project, we gained hands-on experience in multilingual NLP, model optimization, and real-world deployment constraints. We learned how to balance accuracy with performance and how retrieval-based systems can complement deep learning models. Most importantly, we learned how to design AI systems responsibly for real-world social impact.

Built With

• css
• flask
• html
• javascript
• keras
• perspective
• python
• tensorflow