NLP from first principles

A single notebook covering four small language systems, built on the Stanford Sentiment Treebank and on literary texts.

PyTorch
TF-IDF
NER
n-grams

Context: UCL Natural Language Processing, individual coursework
Date: March 2026
Team: Individual
Stack: Python, NumPy, pandas, scikit-learn, PyTorch, Hugging Face Transformers

What it covers

Sentiment with logistic regression

Tokenisation, hand-built features and word embeddings, compared on the same task.

A neural sentiment classifier

An MLP in PyTorch with hyperparameter tuning, compared against zero-shot and few-shot classification by large language models.

Information retrieval

Preprocessing, named entity recognition, TF-IDF vectors and cosine similarity to retrieve documents from the Sherlock Holmes stories, with a two-dimensional projection of the document vectors.

Text generation

Unigram, bigram and longer n-gram language models, written without libraries.