A machine learning-powered sentiment analysis tool that classifies text as positive or negative sentiment. This project won Second Place for accuracy at the 2020 Ignition Hacks Sigma Division.

The Sentiment Analyzer uses a logistic regression model with TF-IDF vectorization to predict sentiment polarity (positive = 1, negative = 0) in text data. The model achieved an F1 score of approximately 0.82.

Project Link: Devpost Submission

• Features
• Prerequisites
• Installation
• Project Structure
• Usage
• Data Cleaning
• Model Optimization
• Performance
• Authors
• License
• Acknowledgments

• ✅ Binary sentiment classification (positive/negative)
• ✅ Preprocessing pipeline with punctuation removal
• ✅ TF-IDF vectorization with bigrams and unigrams
• ✅ Optimized logistic regression model
• ✅ GridSearchCV for hyperparameter tuning
• ✅ F1 score of ~0.82

To run this project on your local machine, you'll need:

• Python 3.6+
• Required Python packages:
  • pandas - Data manipulation and analysis
  • numpy - Numerical computing
  • scikit-learn (sklearn) - Machine learning library
  • nltk - Natural Language Toolkit
  • joblib - Model serialization

1. Clone the repository:

   git clone https://github.com/georgehtliu/sentiment-analyzer.git
   cd sentiment-analyzer

2. Install required packages:

   pip install pandas numpy scikit-learn nltk joblib

3. Download NLTK data (if needed):

   import nltk
   nltk.download('punkt')
   nltk.download('stopwords')

Alternatively, you may run this project on a cloud-hosted development environment such as Google Colab.

Sentiment-Analyzer/
├── README.md                      # This file
├── LICENSE.md                     # MIT License
├── training_data.csv              # Training dataset
├── contestant_judgment.csv        # Judgement/test dataset
├── predicted_labels.csv           # Output predictions
├── submission_training.ipynb      # Model training notebook
├── submission_createcsv.ipynb     # Prediction generation notebook
└── submission_extras.ipynb        # Additional experiments and analysis

Note: The trained model files (SentimentNewton_Log.pkl and Vectorizer.pkl) are generated when running the training notebook.

Run the training notebook to train and save the sentiment analysis model:

jupyter notebook submission_training.ipynb

This notebook will:

• Load and preprocess the training data
• Train a logistic regression model with optimized hyperparameters
• Save the trained model as SentimentNewton_Log.pkl
• Save the TF-IDF vectorizer as Vectorizer.pkl

After training, use the prediction notebook to classify new text:

jupyter notebook submission_createcsv.ipynb

This notebook will:

• Load the trained model and vectorizer
• Process the judgement dataset (contestant_judgment.csv)
• Generate predictions and save them to predicted_labels.csv

Note: Make sure contestant_judgment.csv is in the same directory, or update the file path in the notebook.

The preprocessing pipeline includes:

• Punctuation removal using regex patterns
• TF-IDF vectorization with the following features:
  • Removal of stopwords
  • Extraction of both unigrams and bigrams
  • Tested lemmatization (results documented in submission_extras.ipynb)

The team experimented with various cleaning techniques, and found that punctuation removal alone provided the best accuracy for this dataset.

The model was developed through extensive experimentation:

• Tested multiple classifiers including:
  • Neural Networks (MLPClassifier)
  • Support Vector Machines (SVC)
  • Logistic Regression (selected)
  • SGD Classifier
• Logistic Regression was chosen for optimal performance

• Used GridSearchCV to optimize logistic regression parameters
• Optimized TF-IDF vectorizer settings:
  • Tuned n-gram range (unigrams and bigrams)
  • Experimented with stopword removal
  • Tested different lemmatization approaches

Detailed exploration of alternative approaches and results can be found in submission_extras.ipynb.

• F1 Score: ~0.82
• Task: Binary sentiment classification
• Model: Optimized Logistic Regression with TF-IDF features

• David Chen
• George Liu
• David Wang
• Michael Yang

This project is licensed under the MIT License - see the LICENSE.md file for details.

• scikit-learn - Extensive use of machine learning tools and algorithms
• NLTK - Natural language processing capabilities
• Ignition Hacks 2020 - Competition organizers and judges