🧠 Emotion Classification Using BiLSTM + Attention

This project is a complete NLP deep learning pipeline for emotion classification, combined with a Streamlit-based interactive frontend.

It demonstrates:

• Custom Attention mechanism
• Handling of class imbalance
• Modular ML pipeline (embedding → training → inference → deployment)

🚀 Project Overview

This system classifies user-entered text into one of the following emotions:

• 😢 sadness
• 😄 joy
• ❤️ love
• 😡 anger
• 😨 fear
• 😲 surprise

The backend model is a BiLSTM + Attention neural network trained on a large Twitter-based emotion dataset.

The frontend is built with Streamlit for real-time inference and visualization.

📊 Dataset

Source:

• Kaggle Emotion Dataset link - https://www.kaggle.com/datasets/nelgiriyewithana/emotions

Classes:

• 0 → sadness
• 1 → joy
• 2 → love
• 3 → anger
• 4 → fear
• 5 → surprise

Key Properties:

• Large-scale text dataset
• Nearly class-balanced
• No emojis (pure textual patterns)
• Cleaned and deduplicated before training

🏗 Model Architecture

Architecture Flow:

Embedding → BiLSTM → Attention → Dense → Softmax

Details:

• Embedding Dimension: 200
• BiLSTM Hidden Units: 128
• Attention: Custom time-step weighted aggregation
• Dense Layer: 64 units (ReLU)
• Output: 6-class Softmax

Why Attention?

• Helps the model focus on the most emotionally relevant words
• Improves interpretability and robustness over plain LSTM

🔬 Training Details

Loss Function:

• Sparse Categorical Crossentropy

Optimizer:

• Adam (LR = 1e-3)

Batch Size:

• 128

Class Imbalance Handling:

• Dynamic class_weight computation using scikit-learn

Callbacks:

• EarlyStopping
• ReduceLROnPlateau
• ModelCheckpoint (full .keras model)

Final Model Storage:

• Stored as a full model: artifacts/bilstm_attention_emotions.keras

✅ Results & Performance

• Test Accuracy: ~96–97%
• High recall for minority classes after class-weight correction
• Stable training with early convergence (2–4 epochs)

The model performs very strongly on:

• sadness
• joy
• anger
• fear
• surprise

⚠ Known Limitations

• The "love" and "joy" classes show semantic overlap.
• Highly positive affectionate sentences are sometimes classified as joy instead of love.
• This is a dataset + label boundary issue, not an engineering bug.

This limitation is common even in transformer-based emotion models.

🖥 Demo (Streamlit UI Screenshots)

Main Interface:

Prediction Output:

Probability Distribution:

🛠 Tech Stack

• Python
• TensorFlow / Keras
• NumPy
• Pandas
• Scikit-learn
• Streamlit

📁 Project Structure

text-emotion-predictor
| -- streamlit_app.py
|
| -- notebooks/
|    │-- embeddings.ipynb
|    │-- attention.ipynb
|    │-- training.ipynb
|    └-- prediction.ipynb
|
| -- artifacts/
|    │-- tokenizer.pkl
|    └-- bilstm_attention_emotions.keras
|
| -- requirements.txt
| -- README.md
| -- LICENSE
| -- .gitignore

▶ How To Run Locally

1. Install dependencies:

pip install -r requirements.txt

2. Ensure artifacts exist:

artifacts/
├── bilstm_attention_emotions.keras
└── tokenizer.pkl

3. Start the Streamlit app:

streamlit run streamlit_app.py

👤 Author

Developed by: Pradyumn Bisht

Focus Area: Machine Learning, Deep Learning, NLP, AI Deployment