word2vec

A from-scratch NumPy implementation of Word2Vec (Skip-Gram with Negative Sampling) trained on the NLTK Gutenberg corpus.

Project structure

word2vec/
├── src/
│   ├── __init__.py        # Public API exports
│   ├── vocabulary.py      # Vocabulary class (token <-> index, subsampling, neg-sampling probas)
│   ├── model.py           # Word2Vec class (SGNS training, embeddings, nearest neighbours)
│   ├── data.py            # Gutenberg corpus loader
│   └── logger.py          # Singleton logger
├── main.py                # Entry point
└── requirements.txt

Requirements

• Python 3.10+
• numpy
• nltk

Install dependencies:

pip install -r requirements.txt

Usage

python main.py

To limit the number of sentences loaded (useful for quick experiments):

python main.py --n-samples 5000

This will:

1. Download the NLTK Gutenberg corpus (first run only).
2. Build a vocabulary (words with min_freq >= 5).
3. Train a Skip-Gram model with Negative Sampling for 5 epochs.
4. Print the nearest neighbours for a set of query words.