CS918 Natural Language Processing - Sentiment Analysis Project

This project explores machine learning and deep learning approaches for Twitter sentiment analysis, classifying tweets as positive, negative, or neutral.

📋 Overview

We implement and compare traditional ML models (Naive Bayes, Logistic Regression, SVM) and deep learning models (Bi-LSTM, Bi-LSTM with Attention, BERT). The workflow covers data preprocessing, EDA, model training, and evaluation.

🗂️ Structure

CS918-natural-language-processing/
├── README.md
├── requirements.txt
├── solution.ipynb
├── data/
│   ├── glove.6B.100d.txt
│   ├── twitter-training-data.txt
│   ├── twitter-dev-data.txt
│   ├── twitter-test1.txt
│   ├── twitter-test2.txt
│   └── twitter-test3.txt
├── models_weights/
│   ├── naive_bayes_model.joblib
│   ├── logistic_regression_model.joblib
│   ├── svm_model.joblib
│   ├── lstm_model.pt
│   ├── lstm_with_attention_model.pt
│   ├── bert_raw_tweets.pt
│   └── bert_cleaned_tweets.pt
└── scripts/
    ├── __init__.py
    ├── data_loading_utils.py
    ├── model_training_utils.py
    ├── models.py
    ├── plotting_utilities.py
    ├── text_preprocessing_utils.py
    └── tweet_data_set.py

📊 Dataset

• Training: 45,101 tweets (46% neutral, 35% positive, 18% negative)
• Development: 2,000 tweets (similar distribution)

🔧 Preprocessing

Pipeline includes normalization, mention/URL removal, hashtag and emoji handling, slang/contraction expansion, tokenization (NLTK TweetTokenizer), and cleaning.

🤖 Models

• Naive Bayes: TF-IDF, best CV score 0.615
• Logistic Regression: TF-IDF/Count n-grams, accuracy 0.63
• SVM: TF-IDF, accuracy 0.63
• Bi-LSTM: GloVe embeddings, 2 layers, bidirectional
• Bi-LSTM + Attention: Adds self-attention
• BERT: bert-base-uncased, tested on raw and cleaned tweets

📈 Results (Validation)

Model	Precision	Recall	F1	Acc
Naive Bayes	0.62	0.56	0.58	0.62
Logistic Regression	0.63	0.63	0.62	0.63
SVM	0.64	0.63	0.61	0.63

• Class imbalance: Negative class recall is lowest.
• Neutral class: Highest performance due to size.
• Traditional ML and DL: Comparable results.

📊 EDA

• Tweet length by sentiment
• N-gram and word cloud analysis
• UMAP visualization

🚀 Getting Started

1. Clone repo & install dependencies:

   pip install -r requirements.txt

2. Download NLTK data:

   import nltk; nltk.download('stopwords')

3. Run solution.ipynb in Jupyter.

📦 Key Dependencies

• pandas, numpy, nltk, ekphrasis, emoji, contractions
• scikit-learn, joblib
• PyTorch, torchtext, transformers
• matplotlib, seaborn, wordcloud, umap-learn

See requirements.txt for details.

📝 Modules

• models.py: LSTM, LSTM+Attention, BERT architectures
• text_preprocessing_utils.py: Cleaning, normalization
• data_loading_utils.py: Data and embedding loaders
• model_training_utils.py: Training, metrics, plots
• tweet_data_set.py: PyTorch datasets

---

Note: This project is part of an academic assignment for CS918 at the University of Warwick.