The Fantastic For: Review Moderation Demo

This project is an AI-powered review moderation system inspired by real-world challenges on platforms like Google map, where review content must be quickly and fairly evaluated for policy violations.

---

🌟 Shining Points

• Modern UI:
  A clean, interactive web interface built with HTML, CSS, and JavaScript, allowing users to submit reviews, see detailed moderation results, and provide feedback on the system’s accuracy.

• Multi-Aspect AI Moderation:
  The backend uses a combination of zero-shot classification, sentiment analysis, NSFW image detection, and heuristic rules to evaluate reviews for:

  • Advertisement/promotional content
  • Relevance to the business category
  • Visit authenticity (bot/spam detection)
  • Review quality
  • Offensive/insulting language
  • Image policy violations (including NSFW and stock photo detection)

• Zero-Shot & Sentiment Models:
  Utilizes Hugging Face’s distilbert-base-uncased-mnli for semantic relevance and insult detection, and distilbert-base-uncased-finetuned-sst-2-english for sentiment analysis—no custom training required.

• Image Moderation:
  Detects NSFW content and promotional/stock images using lightweight vision models.

• Feedback Loop:
  Users can provide feedback on moderation accuracy, which is logged for future benchmarking and model improvement.

• Robust Error Handling:
  Handles missing or malformed data gracefully, both in the backend and the UI.

---

🚀 How It Works

1. User submits a review (text, star rating, optional image, and business info) via the web UI.
2. Backend (Flask) analyzes the review using a combination of:
   • Zero-shot classification to check if the review is relevant to the business category/description.
   • Sentiment analysis to ensure the review’s tone matches the star rating.
   • Heuristics to catch generic, spammy, or bot-like reviews.
   • Image analysis for NSFW or promotional content.
3. Results are displayed in the UI with detailed breakdowns and a recommended action (APPROVE, FLAG, REMOVE).
4. User can give feedback on the moderation result, which is stored for benchmarking.

---

🛠️ Techniques Used

• Zero-shot classification (distilbert-base-uncased-mnli) for semantic relevance and insult detection.
• Sentiment analysis (distilbert-base-uncased-finetuned-sst-2-english) for review tone.
• NSFW image detection (e.g., Falconsai/nsfw_image_detection).
• Regex and keyword heuristics for spam, promo, and bot detection.
• Dynamic UI with feedback buttons and real-time result display.
• CSV logging for user feedback and moderation benchmarking.

---

🎬 TikTok Context

This project demonstrates how AI and modern web tech can be combined to build a robust, interactive moderation tool—perfect for platforms like TikTok, where content safety and user experience are critical.
The UI is designed for clarity and engagement, making it easy for both users and moderators to understand and trust the moderation process.

---

🖥️ Demo Screenshot

The UI shows moderation results, detailed policy checks, and feedback buttons.

---

🏁 Getting Started

1. Install dependencies:

   pip install -r requirements.txt
   

2. Run the app:

   python app.py
   

3. Open your browser:
   Go to http://localhost:5000

4. It is possible to use through API: The API is documented in the file ./api_documentation.md

---

📂 File Structure

• app.py — Flask backend and moderation logic
• policy.py — AI and heuristic moderation functions
• templates/index.html — Web UI
• feedback_log.csv — Feedback log (auto-generated)
• requirements.txt — Python dependencies
• api_documentation.md - API documents for endpoints

---

🙌 Feedback & Contributions

Feedback is logged and used to improve the moderation system.
Pull requests and suggestions are welcome!