🎭 CustEmotiTrack

---

📚 Overview

CustEmotiTrack is an AI-powered Customer Emotion Analysis System designed to analyze and categorize customer feedback efficiently. It leverages Machine Learning (ML), NLP, and FAISS-based similarity retrieval to extract emotions, detect hierarchical topics, and compute sentiment scores.

---

🚀 Key Features

✅ Emotion Detection: Extracts user emotions using DistilBERT embeddings and FAISS-based similarity search.
✅ Topic & Subtopic Recognition: Hierarchical topic modeling with MiniLM L6 Sentence Transformer.
✅ Adorescore Computation: Weighs sentiment intensities and computes an engagement score between -100 to 100.
✅ Multilingual Support: Google Translate API enhances cross-language compatibility.
✅ Streamlit UI: Real-time feedback analysis with an interactive web interface.

---

🧠 System Architecture Overview

The architecture consists of several core components:

1.1 Core Components

• 🎯 Data Preprocessing Module: Cleans, normalizes, and balances customer feedback.
• 🎭 Emotion Detection Module: DistilBERT embeddings + FAISS HNSW similarity search.
• 🕵️ Topic & Subtopic Recognition: MiniLM embeddings with cosine similarity mapping.
• 🌐 Multilingual Processing: Google Translate API for cross-language support.
• 📊 Adorescore Computation: Sentiment intensity with topic-weighted scoring.
• 🎨 Streamlit-based UI: Provides real-time analysis and data visualization.

1.2 Model Selection

• Emotion Classification: DistilBERT (base-uncased-go-emotions) + FAISS HNSW retrieval.
• Topic Modeling: MiniLM L6 Transformer embeddings with cosine similarity-based clustering.
• Sentiment Quantification: Intensity scaling with topic-relevant weight adjustments.

---

📊 Data Flow Diagram (DFD)

2.1 Level 0 (Context Diagram)

• Input: Raw customer feedback.
• Processing:
  • Emotion extraction
  • Topic detection
  • Sentiment computation
• Output: JSON response with detected emotions, topics, and computed Adorescore.

2.2 Level 1 (Detailed Data Flow)

• User Input: Feedback received via Streamlit UI or API.
• Preprocessing: Text cleaning, tokenization, and normalization.
• Feature Extraction:
  • DistilBERT embeddings for emotion detection
  • MiniLM embeddings for topic recognition
• Classification:
  • Emotion: FAISS HNSW similarity search.
  • Topics: Cosine similarity clustering.
• Sentiment Scoring: Intensity scaling weighted by topic impact.
• Output Generation: JSON response for UI or downstream applications.

---

Project Structure

• 📂 CustEmotiTrack
  • 📂 config/
  • 📂 datasets/ (Raw & Processed Data)
  • 📂 evaluation/ (scripts)
    • evaluate_emotions.py
    • evaluate_topic.py
  • 📂models/ (Trained ML Models)
    • emotion_model.pt
    • topic_model.pkl
    • tokenizer/
    • load_model.py (emotion module)
    • load_topic_model.py (topic module)
    • save_model.py
    • save_topic_model.py
  • 📂streamlit_app/ (Frontend UI)
    • app.py
  • gdrive_auth.py
  • README.md
  • requirements.txt (Dependencies)
  • LICENSE

---

🔥 API Specifications

4.1 Emotion Detection API

• Endpoint: /predict-emotion
• Input:

{
  "feedback": "The product quality is amazing!"
}

• Processing:
  • DistilBERT feature extraction
  • FAISS HNSW similarity classification
• Output:

{
  "emotions": [{"joy": 0.92}, {"admiration": 0.85}],
  "confidence": 0.95
}

---

4.2 Topic Recognition API

• Endpoint: /predict-topic
• Input:

{
  "feedback": "The delivery was very fast and free!"
}

• Processing:
  • MiniLM embeddings
  • Cosine similarity for topic and subtopic mapping
• Output:

{
  "topic": "Delivery",
  "subtopic": "Fast Shipping",
  "relevance": 0.88
}

---

4.3 Sentiment Scoring API

• Endpoint: /compute-adorescore
• Input:

{
  "emotions": [{"joy": 0.92}, {"admiration": 0.85}],
  "topics": [{"Delivery": 0.88}]
}

• Processing: Weighted sentiment scoring.
• Output:

{
  "Adorescore": 75.4
}

---

🧪 Test Cases & Evaluation

Test Case	Scenario	Expected Outcome
Emotion Detection	"Service was terrible!"	{ "anger": 0.91 }
Topic Recognition	"Fast and free shipping"	{ "topic": "Delivery", "subtopic": "Fast Shipping" }
Multilingual Support	"El servicio fue excelente!"	{ "joy": 0.89 }
Adorescore Calculation	Valid emotion-topic input	Computed Adorescore with relevance score

---

🧠 Findings & Recommendations

✅ Findings

• FAISS-based retrieval enhances classification speed over deep neural models.
• Hybrid feature extraction embeddings boost classification accuracy.
• Google Translate API provides fast multilingual support but lacks contextual depth.

🚀 Recommendations

1. Hybrid FAISS + Transformer Classification: Combine retrieval search with neural models for improved results.
2. Advanced Topic Modeling: Expand topic granularity using hierarchical clustering.
3. Contextual Sentiment Weighting: Enhance Adorescore calculations by dynamically adjusting sentiment intensities.

---

💻 Hardware & System Requirements

7.1 Hardware

• CPU: Intel i5/i7 or AMD Ryzen 5/7 for local inference.
• GPU: NVIDIA RTX 2060 or higher for model training.
• RAM: Minimum 16GB for processing.
• Storage: SSD with at least 50GB free space.

7.2 Software & Frameworks

• Python: 3.8+
• Libraries: PyTorch, FAISS, Hugging Face Transformers, Scikit-Learn, NLTK.
• Cloud Services:
  • Google Firebase (Data Storage)
  • Google Translate API (Multilingual Support)
• Version Control: GitHub for code and dataset versioning.

---

⚡ Working Instructions

8.1 Setup and Run

# Navigate to the project directory
cd CustEmotiTrack

# Install dependencies
pip install -r requirements.txt

# Save emotion and topic models
python models/save_model.py
python models/save_topic_model.py

# Load trained models
python models/load_model.py
python models/load_topic_model.py

# Run Streamlit app
streamlit run streamlit_app/app.py

---

🎨 UI and Visuals

Checkout the following images in project folder 'output' for reference:

• img1.png, img2.png, img3.png, img4.png, img5.png, test.png