Shapes of Mind ML Classifier

A sophisticated 3-model hybrid pipeline for sarcasm-aware emotion classification. This system combines dual sarcasm detection with advanced emotion classification to provide contextually accurate emotional analysis.

🧠 Architecture

3-Model Hybrid Pipeline:

1. 🎭 Sarcasm Detection Model #1 (Primary): bharatiyabytes/flan-t5-sarcasm
2. 🎭 Sarcasm Detection Model #2 (Ensemble): AventIQ-AI/Sarcasmdetection
3. 😊 Emotion Classification: SamLowe/roberta-base-go_emotions

🔄 Processing Logic:

graph TD
    A[Input Text] --> B[Sarcasm Model 1]
    A --> C[Sarcasm Model 2]
    B --> D{Both Agree?}
    C --> D
    D -->|Yes| E[Use Consensus]
    D -->|No| F[Trust Model 1]
    E --> G[Emotion Classification]
    F --> G
    G --> H{Is Sarcastic?}
    H -->|No| I[Return Raw Emotion]
    H -->|Yes| J[Apply Sarcasm Correction]
    J --> K[Return Corrected Emotion]

Loading

✨ Features

• 🎭 Dual Sarcasm Detection: Ensemble approach for robust sarcasm identification
• 😊 27+ Emotion Classification: Based on Google's GoEmotions dataset
• 🔄 Sarcasm-Aware Correction: Intelligent emotion reinterpretation for sarcastic content
• 📊 Batch Processing: Efficient analysis of multiple texts
• 🔍 Detailed Analysis: Comprehensive JSON output with confidence scores
• 💬 Interactive Testing: Real-time text analysis interface

🚀 Setup

1. Create Virtual Environment

# Create virtual environment using pyenv
pyenv virtualenv 3.11 shapes-of-mind
pyenv activate shapes-of-mind

2. Install Dependencies

pip install -r requirements.txt

📖 Usage

Basic Analysis

from shapes_of_mind_classifier import ShapesOfMindClassifier

# Initialize the classifier
classifier = ShapesOfMindClassifier()

# Analyze text
result = classifier.analyze("Oh great, another Monday morning!")
print(result['final_emotion'])  # 'anger' (corrected from 'joy')

Complete Analysis

# Get detailed analysis
result = classifier.analyze("I just LOVE waiting in traffic!", verbose=True)

# JSON output structure:
{
  "text": "I just LOVE waiting in traffic!",
  "is_sarcastic": true,
  "sarcasm_confidence": "high",
  "raw_emotions": [
    {"label": "joy", "score": 0.85},
    {"label": "optimism", "score": 0.12},
    ...
  ],
  "final_emotion": "anger",  # Corrected due to sarcasm
  "correction_applied": true
}

Batch Processing

texts = [
    "I love this new policy!",        # Could be sarcastic
    "This weather is perfect.",       # Could be sarcastic  
    "I'm genuinely excited!"          # Likely genuine
]

results = classifier.batch_analyze(texts, verbose=True)

🧪 Testing

Run Test Suite

python test_shapes_classifier.py

Interactive Testing

python test_shapes_classifier.py
# Choose 'y' for interactive mode

Demo

python shapes_of_mind_classifier.py

🎯 Model Information

Sarcasm Detection Models

• Primary: bharatiyabytes/flan-t5-sarcasm (Text-to-text generation)
• Secondary: AventIQ-AI/Sarcasmdetection (Classification)

Emotion Classification Model

• Model: SamLowe/roberta-base-go_emotions
• Base: RoBERTa-base
• Emotions: 27+ emotions from Google's GoEmotions dataset
• Training: 58k carefully curated comments

Sarcasm Correction Rules

When sarcasm is detected, specific emotions are reinterpreted:

Original Emotion	Corrected Emotion	Reasoning
joy, optimism, gratitude	anger	Positive emotions often inverted in sarcasm
neutral, approval	annoyance	Mild emotions become irritation
admiration, excitement	anger	Enthusiastic emotions become frustration

📁 Files

• shapes_of_mind_classifier.py - Main hybrid classifier implementation
• test_shapes_classifier.py - Comprehensive test suite
• requirements.txt - Python dependencies
• README.md - This documentation

🔧 Requirements

• Python: 3.8+
• PyTorch: 2.0+
• Transformers: 4.21+
• NumPy: 1.21+
• SciPy: 1.9.0+
• Scikit-learn: 1.1.0+

📊 Performance Notes

• First Run: Downloads ~800MB of models (cached locally)
• GPU Support: Automatic CUDA detection and utilization
• Processing Speed: ~200-500ms per text (CPU), ~50-100ms (GPU)
• Memory Usage: ~2-3GB RAM for all three models

🎯 Use Cases

• Social Media Analysis: Detecting sarcasm in tweets, comments
• Customer Feedback: Understanding true sentiment behind reviews
• Content Moderation: Identifying disguised negative sentiment
• Market Research: Analyzing authentic vs. sarcastic responses
• Chatbot Training: Improving response appropriateness

🔬 Example Outputs

# Sarcastic text
"Oh wonderful, another software update!" 
→ Sarcasm: True | Final: anger (was joy)

# Genuine text  
"I'm genuinely excited about this project!"
→ Sarcasm: False | Final: joy (unchanged)

# Ambiguous text
"This is totally fine."
→ Sarcasm: depends on models | Final: varies

📄 License

This implementation is provided as-is. Please refer to the original models' licenses on Hugging Face.

🙏 Acknowledgments

• bharatiyabytes for the FLAN-T5 sarcasm detection model
• AventIQ-AI for the ensemble sarcasm detection model
• SamLowe for the GoEmotions RoBERTa implementation
• Google Research for the original GoEmotions dataset