Emotion Detection from Text using NLP & Machine Learning

A machine learning project that classifies the emotion expressed in a piece of text (e.g., joy, sadness, anger, fear, love, surprise) using classical NLP preprocessing and ML classifiers. The final model is deployed as an interactive Streamlit web app.

🔗 Live Demo: emotiondetector-bipan.streamlit.app

---

📌 Project Overview

This project takes raw text data, cleans and preprocesses it using standard NLP techniques, converts it into numerical features using Bag of Words and TF-IDF, and trains multiple machine learning models to predict the emotion behind the text. The best-performing model is then saved and used to power a simple web UI where users can type a sentence and get an instant emotion prediction.

---

🗂️ Dataset

• File: train.txt
• Format: Semicolon-separated (;) text file with two columns — text and emotion
• Each row contains a sentence and its corresponding emotion label.

---

🧹 Text Preprocessing Pipeline

The raw text was cleaned using the following steps, in order:

1. Lowercasing – Converts all text to lowercase for consistency.
2. Removing Punctuation – Strips punctuation marks using Python's string module.
3. Removing Numbers – Removes digit characters from the text.
4. Removing Emojis & Special Characters – Keeps only ASCII characters.
5. Removing Stopwords – Removes common English stopwords (e.g., "the", "is", "and") using NLTK's stopword list.

After cleaning, the text was split into training (80%) and testing (20%) sets using train_test_split with random_state=42.

---

🔢 Feature Extraction

Two vectorization techniques were used to convert text into numerical form:

Technique	Description
Bag of Words (CountVectorizer)	Represents text as raw word frequency counts.
TF-IDF (TfidfVectorizer)	Weighs words by importance, reducing the influence of frequently occurring but less informative words.

---

🤖 Models Trained & Results

Three model + feature combinations were trained and evaluated on accuracy:

Model	Feature Extraction	Accuracy
Multinomial Naive Bayes	Bag of Words (CountVectorizer)	76.81%
Multinomial Naive Bayes	TF-IDF	66.09%
Logistic Regression	TF-IDF	🏆 86.28%

🏆 Best Model: Logistic Regression + TF-IDF

Logistic Regression combined with TF-IDF features achieved the highest accuracy of 86.28%, outperforming both Naive Bayes variants. This is the model used in the deployed Streamlit application.

Why did Logistic Regression + TF-IDF perform best?

• TF-IDF down-weights common, less-informative words while emphasizing distinctive ones.
• Logistic Regression can learn more nuanced decision boundaries between classes compared to Naive Bayes' independence assumption, which tends to oversimplify relationships between words.

---

📁 Project Structure

├── app.py                    # Streamlit web app for live predictions
├── train.txt                 # Dataset (text;emotion pairs)
├── requirements.txt           # Python dependencies
├── README.md                   # Project documentation
├── models/
│   ├── model_LR.pkl            # Saved trained Logistic Regression model
│   ├── tfidf_vectorizer.pkl     # Saved TF-IDF vectorizer (fitted on training data)
│   └── emotion_numbers.pkl      # Saved label mapping (emotion name <-> number)
└── notebooks/
    └── NLP-with-ML.ipynb        # Main notebook: preprocessing, training, evaluation

---

⚙️ Installation & Setup

1. Clone the repository

   git clone https://github.com/Bipan101/Emotion-Detector.git
   cd Emotion-Detector

2. Install dependencies

   pip install -r requirements.txt

3. Run the Streamlit app

   streamlit run app.py

4. Open the local URL shown in your terminal (usually http://localhost:8501) in your browser.

---

🖥️ How the App Works

1. User types a sentence into the text box.
2. The input goes through the same preprocessing pipeline used during training (lowercasing, punctuation/number/emoji removal, stopword removal).
3. The cleaned text is vectorized using the saved TF-IDF vectorizer.
4. The saved Logistic Regression model predicts the emotion class.
5. The predicted label is decoded back to its original emotion name and displayed, along with a confidence score bar chart.

---

🛠️ Tech Stack

• Python 3
• Pandas / NumPy – Data handling
• NLTK – Stopword removal, tokenization
• Scikit-learn – Vectorization (CountVectorizer, TfidfVectorizer), models (MultinomialNB, LogisticRegression), evaluation
• Streamlit – Web app interface
• Matplotlib / Seaborn – Visualization (EDA)

---

📈 Future Improvements

• Experiment with more advanced models (SVM, Random Forest, XGBoost).
• Try word embeddings (Word2Vec, GloVe) or transformer-based models (BERT) for potentially higher accuracy (out of scope here, as this project focuses on classical NLP + ML rather than deep learning).
• Add cross-validation and hyperparameter tuning (e.g., GridSearchCV) for more robust model selection.
• Handle class imbalance if present in the emotion labels.

---

👤 Author

Bipan Neupane

• GitHub: @Bipan101
• Portfolio: bipanneupane.com.np

---

📄 License

This project is open source and available under the MIT License.