Tweet Sentiment Analysis using Deep Learning & Transfer Learning

Course Project | NLP | Loyalist College

This project focuses on classifying tweets into positive, negative, or neutral sentiments using multiple deep learning architectures and traditional machine learning models. It leverages pre-trained embeddings (GloVe, word2vec), performs empirical tuning, and uses explainability tools like SHAP and LIME to interpret predictions.

Overview

This project was developed as part of an academic assignment in NLP. We explored a multi-model pipeline where five different architectures were built, trained, and tuned using a custom preprocessed tweet dataset.

Pipeline Structure

1. Preprocessing: Cleaned tweets (punctuation, emoji, case-folding)
2. Feature Engineering: tweet_length, lexicon_score
3. Embedding: GloVe, word2vec, TF-IDF
4. Modeling:
   • Unidirectional RNN with TF-IDF
   • Unidirectional LSTM with word2vec
   • Bidirectional GRU with learned embedding
   • Bidirectional LSTM with learned embedding
   • SVM with GloVe and TF-IDF
5. Evaluation: Accuracy, F1-score, AUC, Confusion Matrix
6. Explainability: SHAP, LIME

Models Implemented

• RNN + TF-IDF
• LSTM + word2vec (64 to 300 vector sizes, CBOW/Skipgram)
• BiGRU (learned embedding)
• BiLSTM (learned embedding)
• SVM + GloVe/TF-IDF

Results Summary

Model	Accuracy	F1 (Positive)	Macro AUC
SVM (TF-IDF)	56%	0.59	0.64
LSTM (word2vec)	52%	0.58	0.69
BiGRU (learned embedding)	53%	0.60	0.66
BiLSTM (learned embedding)	54%	0.61	0.68
RNN (TF-IDF)	51%	0.56	0.62

Tech Stack

• Languages: Python
• Libraries: TensorFlow, Keras, scikit-learn, Gensim, SHAP, LIME
• Embeddings: GloVe, word2vec, TF-IDF
• Tools: Google Colab, GitHub, Matplotlib, Seaborn

Project Structure

├── data/
│   ├── processed/ (train/test/val CSVs)
│   └── raw/ (Sentiment_Data.csv)
├── model-tuning/
│   └── model/ (saved .h5 and .pkl files, encoders, tokenizers)
├── notebooks/
│   └── Assignment-2.ipynb

Interpretability

To make model predictions transparent, we used:

• LIME: Local explanations for individual predictions
• SHAP: Global and local feature importance visualizations

These tools helped debug model behavior and identify key drivers behind sentiment classification.

Future Improvements

• Integrate BERT-based models (e.g., DistilBERT, RoBERTa)
• Add multilingual support for tweets
• Fine-tune embeddings for domain-specific sentiment

Authors & Acknowledgements

Developed by Group of students as part of the NLP course assignment at Loyalist College.

Feel free to fork, improve, or explore!