BERT-based Emotion Recognition Pipeline

Overview

This project provides a fully reproducible pipeline for text-based emotion detection using BERT and modern NLP preprocessing. Includes:

• Data Pipeline: Automated merging, advanced preprocessing, and negation-based augmentation.
• BERT Model: Fine-tuned distilbert-base-uncased for multi-class emotion classification.
• Interfaces: CLI interactive mode and a modern Streamlit web application.

---

Features

• Unified data preparation script: prepare_data.py combines, preprocesses, and augments emotion datasets automatically.
• Automated NLTK setup: No manual downloads required; the script handles punkt, stopwords, etc.
• Advanced Preprocessing: Lemmatization, proper stopword handling, and explicit negation marking (e.g., "not happy" -> "NOT_happy").
• Augmentation: Adds robust negated phrase samples for each emotion class to improve "neutral" detection.
• Flexible Inference: Choose between a Command Line Interface (CLI) or a Web UI.

---

Repository Structure

├── data/
│   ├── train.csv
│   ├── test.csv
│   ├── val.csv
│   ├── final_data_aug.csv   # Generated training data
├── prepare_data.py         # Data pipeline: load, preprocess, augment
├── bert_emotion.py         # Model training & CLI inference
├── app.py                  # Streamlit Web Application
├── requirements.txt        # Dependencies
├── README.md               # Documentation
└── bert_emotion_model/     # Saved model artifact (after training)

---

Setup

1. Install requirements

pip install -r requirements.txt

2. NLTK Data

NLTK data is now automatically downloaded when you run the scripts for the first time.

---

Usage

1. Data Preparation

Run the unified script to preprocess and augment data:

python prepare_data.py

• Output: data/final_data_aug.csv

2. Model Training

Train the BERT emotion classifier:

python bert_emotion.py

• Choose Option 1 for training.
• The model will be saved in bert_emotion_model/.

---

Inference

A. Command-line Interface (CLI)

python bert_emotion.py

• Choose Option 2 for interactive detection.
• Type sentences and press Enter for results.

B. Web Application (Streamlit)

For a more visual experience, run the web app:

streamlit run app.py

---

Requirements

The pipeline requires the following Python libraries:

• pandas
• nltk
• torch
• transformers
• scikit-learn
• accelerate
• streamlit (for the web app)

---

Tips

• GPU Acceleration: The scripts automatically detect CUDA-enabled GPUs for faster training and inference.
• Custom Labels: You can tune emotion labels in prepare_data.py.
• Model Tuning: Adjust epochs and batch sizes in bert_emotion.py within TrainingArguments.

---

License

MIT License

---

Developed with automation and ease of use in mind.