Overview

Project for CSE 256 2026 winter project 1.

This project implements several sentiment classification models, including:

1. Bag-of-Words (BOW) models
2. Deep Averaging Network (DAN) with word-level embeddings
3. Subword-based DAN using Byte Pair Encoding (BPE) All models can be trained and evaluated using a single command.

Requirements

Python 3, PyTorch, scikit-learn, matplotlib

How to Run

1. Train and evaluate BOW models

python main.py --model BOW

2. Train and evaluate word-level DAN

python main.py --model DAN

This runs the Deep Averaging Network (DAN) using: i. GloVe-initialized word embeddings ii. Randomly initialized word embeddings Both settings are trained and evaluated sequentially.

3. Train and evaluate subword-based DAN (BPE)

python main.py --model SUBWORDDAN

Optional arguements for BPE:

python main.py --model SUBWORDDAN --bpe_vocab_size 10000 --subword_emb_dim 50

Arguments: --bpe_vocab_size: vocabulary size for BPE (default: 10000) --subword_emb_dim: embedding dimension for subword embeddings (default: 50)

Notes

1. Word-level DAN uses pre-trained GloVe embeddings when available.

2. Subword-based DAN initializes embeddings randomly and trains them from scratch.

3. All experiments are run for 100 epochs.