Unit 1: Tokenisation and embeddings
This unit covers tokenisation and embeddings, two fundamental concepts of modern NLP. Tokenisers split text into smaller units such as words, subwords, or characters. Embeddings are fixed-size vector representations of tokens that can be learned from data and optimised for different tasks.
Lectures
The lectures start with traditional word-based tokenisation and then present the Byte Pair Encoding (BPE) algorithm, which is used by most current language models. In the second half of the lectures, you will learn about embeddings and different methods for how they can be learned.
Section | Title | Video | Slides | Quiz |
---|---|---|---|---|
1.1 | Introduction to tokenisation | video | slides | quiz |
1.2 | The Byte Pair Encoding algorithm | video | slides | quiz |
1.3 | Introduction to embeddings | video | slides | quiz |
1.4 | Word embeddings | video | slides | quiz |
1.5 | Learning word embeddings: Matrix decomposition | video | slides | quiz |
1.6 | Learning word embeddings: The skip-gram model | video | slides | quiz |
You must complete the quizzes no later than 2025-09-16.
Online meeting
In the online meeting for this unit, we will discuss the issue of bias in word embeddings. We will examine how biases arise, how to detect and measure them, and what mitigation strategies exist – along with their trade-offs and limits. These issues have broad consequences for real-world uses of NLP systems.
The meeting will take place on 2025-09-17 between 18:00–20:00. A Zoom link will be sent out via the course mailing list.
Lab
In lab 1, you will build an understanding of how text can be transformed into representations that computers can process and learn from. Specifically, you will code and analyse a tokeniser based on the Byte Pair Encoding (BPE) algorithm, and then explore embeddings in the context of a simple text classifier architecture.
If you want a written review of this lab, you must submit it (via Lisam) no later than 2025-10-31.