ReviewsGPT - AI Emotion Detection & Chatbot System

An advanced emotion detection and chatbot system that combines machine learning models with conversational AI to analyze emotions in text and provide intelligent responses.

Table of content

• Introduction
• Installation
• Usage
• Features
• Model Details
• Configuration
• Contributing
• License

Introduction

CustomGPT is a comprehensive AI system that integrates emotion detection with chatbot capabilities. The project features:

• Emotion Detection Pipeline: Uses LightGBM models trained on emotion datasets to detect and classify emotions in text
• Local Chatbot: Powered by Microsoft's Phi-4 model for offline text generation
• OpenAI Integration: Customer service chatbot with emotion-aware responses
• Multi-Model Support: Includes RoBERTa and other transformer models for emotion analysis
• Data Processing: Complete pipeline for emotion dataset preparation and model training

The system is particularly useful for customer service applications, sentiment analysis, and building emotion-aware conversational AI.

Installation

You will need Python 3.8+ installed on your machine to run this project.

Prerequisites

• Python 3.8 or higher
• Git
• Optional: CUDA-capable GPU for faster model inference

Setup

1. Clone the repository:

git clone https://github.com/Luxchar/CustomGPT.git
cd CustomGPT

2. Install the dependencies:

pip install -r requirements.txt

3. For OpenAI integration: Create a .env file in the root directory and add your OpenAI API key:

OPENAI_API_KEY=your_api_key_here

4. For emotion detection: The pre-trained LightGBM models are included in the notebooks/lightgbm_ultimate_saved/ directory.

Usage

1. Local Chatbot (Offline)

Run the local Phi-4 powered chatbot that works entirely offline:

streamlit run src/main.py

2. OpenAI-powered Customer Service Bot

Run the emotion-aware customer service chatbot:

streamlit run src/main_openai.py

3. Emotion Detection Pipeline

Use the emotion detection system programmatically:

from src.emotion_pipeline import EmotionDetectionPipeline

# Initialize the pipeline
pipeline = EmotionDetectionPipeline()
model_path = "./notebooks/lightgbm_ultimate_saved/lightgbm_ultimate_rescued_20250706_060702"
pipeline.load_model(model_path)

# Analyze a single text
result = pipeline.predict_single("I absolutely love this product!")
print(f"Major emotion: {result['major_emotion']} ({result['major_confidence']:.3f})")

# Batch processing
texts = ["Great product!", "Terrible service!", "It's okay, I guess."]
df_results = pipeline.predict_batch(texts, save_csv="predictions.csv")

4. Data Processing and Training

Explore the Jupyter notebooks for data processing and model training:

jupyter notebook notebooks/

Key notebooks:

• data_prep_yann.ipynb: Data preparation and preprocessing
• detecteur_emotion_*.ipynb: Emotion detection model training
• text-generation*.ipynb: Text generation experiments

Features

Emotion Detection

• 30+ Emotion Categories: Detects emotions like joy, anger, sadness, fear, surprise, etc.
• Multi-Model Architecture: LightGBM ensemble with TF-IDF and SVD features
• Batch Processing: Efficient processing of large text datasets
• Confidence Scores: Returns probability scores for all emotions

Chatbot Capabilities

• Local Inference: Phi-4 model runs entirely on your machine
• Streaming Responses: Real-time text generation with streaming UI
• Emotion-Aware Responses: Customer service bot adapts responses based on detected emotions
• Memory: Maintains conversation context across interactions

Data Processing

• Multiple Datasets: Support for GoEmotions and custom emotion datasets
• Feature Engineering: Advanced text preprocessing and feature extraction
• Model Training: Complete pipeline for training custom emotion detection models

Model Details

Emotion Detection Models

• Primary Model: LightGBM Ultimate with TF-IDF and SVD features
• Alternative Models: RoBERTa-based emotion classifiers
• Training Data: GoEmotions dataset and custom review datasets
• Features: Word/character n-grams, semantic embeddings, statistical features

Language Models

• Local Model: Microsoft Phi-4-mini-instruct (CPU optimized)
• Cloud Model: OpenAI GPT-3.5-turbo integration
• Optimization: Quantization and optimization for fast inference

Configuration

Environment Variables

Create a .env file with the following variables:

OPENAI_API_KEY=your_openai_api_key
MODEL_PATH=./notebooks/lightgbm_ultimate_saved/lightgbm_ultimate_rescued_20250706_060702

Model Paths

Update model paths in the configuration files:

• Emotion models: notebooks/lightgbm_ultimate_saved/
• Language models: Downloaded automatically via Hugging Face

Streamlit Configuration

For production deployment, configure Streamlit in .streamlit/config.toml:

[server]
port = 8501
address = "0.0.0.0"

[theme]
primaryColor = "#FF6B6B"
backgroundColor = "#FFFFFF"
secondaryBackgroundColor = "#F0F2F6"

Contributing

If you want to contribute to this project, you can fork this repository and make a pull request with your changes.

Development Guidelines

• Follow PEP 8 style guidelines
• Add docstrings to all functions and classes
• Include unit tests for new features
• Update documentation for any API changes

Areas for Contribution

• Additional emotion categories
• Performance optimizations
• New model architectures
• UI/UX improvements
• Documentation and tutorials

Anyone is welcome to contribute to this project.

License

This project is under the MIT license.

---

Quick Start Example

# Quick emotion detection
from src.emotion_pipeline import quick_predict

result = quick_predict("This product is absolutely amazing!")
print(f"Emotion: {result['major_emotion']} (confidence: {result['major_confidence']:.2f})")

For more examples and detailed documentation, check the notebooks/ directory.