🧠 Emotion Detector — NLP Text Classification

A machine learning web application that detects human emotions from text using Natural Language Processing. Built with TF-IDF vectorization and Logistic Regression, wrapped in an interactive Streamlit UI.

🔗 Live Demo: emotion-detector-pranjal-pandey.streamlit.app

📌 Table of Contents

• Overview
• Demo
• Emotions Detected
• Project Structure
• NLP Pipeline
• Model Comparison
• Tech Stack
• Installation & Usage
• Dataset
• Results
• Author

---

🔍 Overview

This project builds an end-to-end NLP pipeline that classifies text into one of 6 human emotions. The pipeline includes full text preprocessing, feature extraction using TF-IDF and Bag of Words, and classification using multiple ML models. The best model is served through a Streamlit web app with a clean, interactive UI.

---

🎬 Demo

Run locally with:

python -m streamlit run emotion_app.py

Features of the UI:

• Upload your own training data
• Type any text and get instant emotion prediction
• View confidence scores for all 6 emotions as a bar chart
• See cleaned text after the full preprocessing pipeline
• View training data emotion distribution chart

---

🎭 Emotions Detected

Emotion	Label	Emoji
Sadness	0	😢
Anger	1	😠
Love	2	❤️
Surprise	3	😲
Fear	4	😨
Joy	5	😄

---

📁 Project Structure

emotion-detector-nlp/
│
├── emotion_app.py          # Streamlit web application
├── NLP-emotions.ipynb      # Full model training notebook
├── train.txt               # Training dataset (text;emotion format)
├── requirements.txt        # Python dependencies
└── README.md               # Project documentation

---

🔄 NLP Pipeline

Every input text goes through the following preprocessing steps before prediction:

Raw Text
   │
   ▼
1. Lowercase conversion         → "I AM HAPPY" → "i am happy"
   │
   ▼
2. Remove punctuation           → "hello!!!" → "hello"
   │
   ▼
3. Remove numbers               → "room 404" → "room "
   │
   ▼
4. Remove emojis                → keeps only ASCII characters
   │
   ▼
5. Remove URLs                  → strips http/https/www links
   │
   ▼
6. Remove stopwords (NLTK)      → removes "is", "the", "and" etc. (198 words)
   │
   ▼
Cleaned Text → TF-IDF Vectorizer → Logistic Regression → Predicted Emotion

---

📊 Model Comparison

Three models were trained and evaluated. Here are the results:

Model	Vectorizer	Test Accuracy	Notes
Multinomial Naive Bayes	Bag of Words (CountVectorizer)	~77%	Fast, but lower accuracy
Multinomial Naive Bayes	TF-IDF	~74%	Worse than BoW for this dataset
Logistic Regression	TF-IDF	~86.2%	✅ Best model — selected for deployment

Why Logistic Regression won?

• Handles high-dimensional sparse TF-IDF features well
• Outputs calibrated probabilities for all classes
• max_iter=1000 allowed full convergence of the optimizer
• Significantly outperformed Naive Bayes on this 6-class problem

Why TF-IDF over Bag of Words?

• TF-IDF penalizes very common words that appear across all documents
• Gives higher weight to words that are more unique to specific emotions
• Reduces noise from frequently occurring but uninformative words

---

🛠 Tech Stack

Tool	Purpose
Python 3.8+	Core programming language
Pandas	Data loading and manipulation
NLTK	Stopwords removal and tokenization
Scikit-learn	TF-IDF, CountVectorizer, Logistic Regression, Naive Bayes
Matplotlib	Probability bar charts and distribution plots
Streamlit	Interactive web UI
NumPy	Numerical operations

---

⚙️ Installation & Usage

1. Clone the repository

git clone https://github.com/venom312004/emotion-detector-nlp.git
cd emotion-detector-nlp

2. Install dependencies

pip install -r requirements.txt

3. Run the app

python -m streamlit run emotion_app.py

4. Use the app

• Upload your train.txt file via the sidebar
• Wait a few seconds for the model to train
• Type any text in the input box
• Click Predict Emotion and see the result!

---

📂 Dataset

The dataset is a plain text file (train.txt) with semicolon-separated values:

i feel really happy today;joy
i miss you so much;sadness
this makes me so angry;anger
i love spending time with you;love
i had no idea this would happen;surprise
i am terrified of the dark;fear

• Format: text;emotion
• Classes: 6 (sadness, anger, love, surprise, fear, joy)
• Split: 80% training / 20% testing
• Random state: 42 (reproducible results)

---

📈 Results

Best Model   : Logistic Regression + TF-IDF
Test Accuracy: ~86.2%
Classes      : 6
Max Features : 20,000 (TF-IDF vocabulary)
Max Iter     : 1000
Solver       : lbfgs

The model performs particularly well on joy, sadness, and anger — emotions with strong distinctive vocabulary. Surprise and fear sometimes overlap due to similar language patterns.

---

👤 Author

Pranjal Pandey

• 🎓 B.Tech in Artificial Intelligence
• 💼 Email: pranjalpandey0301@gmail.com
• 🐙 GitHub: @venom312004

---

📄 License

This project is licensed under the MIT License — feel free to use, modify, and distribute.

---

Made with ❤️ using Python & Streamlit