Product Feedback Emotion Analysis API

This project provides a containerized solution for analyzing product feedback using emotion classification. The implementation uses FastAPI with uvicorn workers for efficient async processing.

Model Choice

For this demonstration, we're using the bhadresh-savani/distilbert-base-uncased-emotion model, which is a DistilBERT model fine-tuned for emotion classification. This model was chosen because:

1. It can detect 6 different emotions:

   • Sadness 😭 - Disappointment or unhappiness
   • Joy 😀 - Happiness and delight
   • Love ❤️ - Affection and strong positive feelings
   • Anger 🤬 - Strong negative emotions
   • Fear 😨 - Anxiety or concern
   • Surprise 😲 - Unexpected or astonished reactions

2. It's well-suited for analyzing customer feedback

3. It provides probability scores for each emotion

4. It has good performance while maintaining reasonable resource requirements

Setup

1. Build the Docker container:

docker build -t emotion-analysis .

2. Run the container:

docker run -p 8000:8000 emotion-analysis

The API will be available at http://localhost:8000

API Endpoints

POST /analyze

Accepts a list of feedback texts and returns emotion analysis results.

Request body:

{
    "texts": [
        "I'm absolutely delighted with this purchase!",
        "I'm really disappointed with this product.",
        "I'm worried about the safety of this product."
    ]
}

Response:

{
    "predictions": [
        {
            "text": "I'm absolutely delighted with this purchase!",
            "emotions": {
                "sadness": 0.01,
                "joy": 0.85,
                "love": 0.10,
                "anger": 0.01,
                "fear": 0.01,
                "surprise": 0.02
            },
            "dominant_emotion": "joy"
        },
        {
            "text": "I'm really disappointed with this product.",
            "emotions": {
                "sadness": 0.75,
                "joy": 0.01,
                "love": 0.01,
                "anger": 0.15,
                "fear": 0.02,
                "surprise": 0.06
            },
            "dominant_emotion": "sadness"
        },
        {
            "text": "I'm worried about the safety of this product.",
            "emotions": {
                "sadness": 0.05,
                "joy": 0.01,
                "love": 0.01,
                "anger": 0.05,
                "fear": 0.85,
                "surprise": 0.03
            },
            "dominant_emotion": "fear"
        }
    ]
}

GET /health

Health check endpoint that returns the status of the service.

Demo Script

The included demo.ipynb script demonstrates how to make parallel requests to the API. To run it:

1. Make sure the Docker container is running
2. Open demo.ipynb file and run the cells.

The scripts inside the notebook shows:

• How to make parallel requests using ThreadPoolExecutor
• Performance metrics for parallel processing
• Detailed emotion analysis results for each feedback
• Examples of all 6 emotions in product feedback

Architecture

• FastAPI: Modern, fast web framework for building APIs
• Uvicorn: ASGI server with multiple worker support
• Transformers: HuggingFace's library for state-of-the-art NLP
• Docker: Containerization for easy deployment

The container is configured to run 4 uvicorn workers by default, which can be adjusted based on your CPU cores and requirements.