Sentiment Analysis with NLP Models

Project Overview

This project implements and compares multiple machine learning models for sentiment analysis on the Stanford Sentiment Treebank dataset. The goal is to classify sentences from movie reviews as either positive or negative based on their sentiment.

I've implemented and evaluated four different models:

1. Log-Linear Model with One-Hot Encoding
2. Log-Linear Model with Word2Vec Embeddings
3. LSTM Model with Word2Vec Embeddings
4. Transformer Model

To do so I used the following libraries; torch, numpy, matplotlib, transformers.

Results & Comparisons

• Full results can be seen in the report file.

Key Insights

1. One-Hot Encoding Performs Poorly - Since one-hot encoding doesn’t capture word relationships, it fails on unseen words and negations.
2. Word2Vec Improves Performance - Pre-trained embeddings allow the model to generalize better by capturing word meanings.
3. LSTM Handles Negation & Sequential Context Better - Captures dependencies across words, leading to better negation handling.
4. Transformers Achieve the Best Accuracy - Handles long-range dependencies and understands context more effectively than RNN-based models.

Installation and Setup

1. Open Google Drive and create a new folder named: NLP_Project
2. Place the following files inside the NLP_Project folder: stanfordSentimentTreebank, data_loader.py
3. Open Google Colab Notebooks: For Log-Linear and LSTM Models open project_models.ipynb and for the Transformer Model open transformer.ipynb
4. Once setup is complete, you can run the notebook cells to train and evaluate the models.
5. Alternatively, if you prefer running the models from Python I have provided the respective files.