What is natural language processing?

Natural language processing is the field concerned with making computers work with human language: classifying it, extracting information from it, translating it, summarising it and generating it. Since a computer only manipulates numbers, every NLP system begins by converting text into numerical representations, and most of the field's difficulty lies in doing that conversion without losing meaning.

What is tokenisation?

Tokenisation is cutting text into the units a model will process. Modern systems use subword tokenisation, which keeps common words whole and splits rare ones into fragments, so a fixed vocabulary of perhaps fifty thousand tokens can represent any word including ones never seen before. It also explains why models count text in tokens rather than words, and why some languages consume more tokens for the same content.

What is a word embedding?

An embedding is a vector of numbers representing a word, learned rather than assigned, positioned so that words used in similar contexts sit near each other. This turns meaning into geometry: similarity becomes a dot product, and relationships between words appear as directions in the space.

Why did contextual embeddings replace fixed ones?

A fixed embedding gives a word one vector regardless of use, so bank in river bank and bank account receives the same representation, which is plainly wrong. Contextual embeddings, produced by transformer models, compute a different vector for a word depending on the sentence it appears in, which resolves ambiguity and is a large part of why modern NLP works so much better.

Which NLP tasks actually work reliably?

Classification, sentiment analysis, named entity recognition, machine translation between well-resourced languages, speech transcription and extractive summarisation are all reliable enough for production with human review. What remains unreliable is anything requiring factual accuracy without verification, genuine reasoning over long documents, or robust performance in languages with little training data.

Do I need to build NLP models from scratch?

No, and you should not. Pre-trained language models are freely available and fine-tuning one on a few hundred to a few thousand labelled examples outperforms anything you would train yourself. The valuable skills are choosing an appropriate model, preparing data well, and evaluating honestly.

Natural Language Processing: how machines work with human language

Language is the thing humans do best and computers found hardest. This course explains why it was so hard, what changed, and where the field genuinely stands in 2026 — including what still does not work.

What this course sets out to do: make the pipeline legible. Text becomes tokens, tokens become vectors, vectors become predictions, and every one of those steps loses something you should know about.

What this course does not do: have you build an NLP system. The premium catalogue covers that with notebooks and datasets.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	Why language is hard	Ambiguity, context and everything left unsaid	8 min
2	Tokenisation	How text becomes units, and what that costs	8 min
3	Embeddings: meaning as geometry	From word counts to vectors that capture sense	9 min
4	The tasks	What NLP is used for, and which tasks are solved	9 min
5	Where it still fails	Facts, bias, languages, and evaluation that lies	9 min
6	Recap and FAQ	Synthesis, a task guide, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You want to process text — reviews, tickets, contracts, documents — and want to know what is realistic.
You use language models and want to understand what happens before the model sees your text.
You keep meeting "token", "embedding" and "fine-tune" and want them precise.
You are evaluating an NLP vendor and want to ask questions that reveal something.

Recommended first: Deep Learning, particularly the lesson on transformers.

What you will be able to do at the end

Explain why language resists computation, beyond "it is complicated".
Describe tokenisation and why models bill in tokens rather than words.
Explain what an embedding is and why similarity becomes arithmetic.
Say why contextual embeddings were a step change over fixed ones.
Name the tasks NLP performs and judge which are production-ready.
Recognise the failure modes: fabricated facts, inherited bias, and languages the field neglects.

Estimated time

Around 45 to 55 minutes of reading.

Prerequisites and next steps

Prerequisites: Introduction to AI and Deep Learning.

Natural continuation:

Large Language Models — where NLP arrives today
Generative AI — producing text and images
Ethics of AI — bias in language data, in detail

Frequent questions, answered in one line

Is NLP the same thing as an LLM?

No. NLP is the field; a large language model is currently its most prominent tool. Plenty of production NLP does not involve an LLM at all: a fine-tuned classifier with a few hundred million parameters routinely beats a giant model on a specific classification task, runs far cheaper, and gives more predictable outputs.

Can NLP work on languages other than English?

Yes, and unevenly. English, Chinese, Spanish, French, German and a dozen others are well served. Thousands of languages have too little digitised text to train on, and models for them are markedly worse — a disparity that mirrors and reinforces existing inequalities in who technology serves well.

Does a model understand what it reads?

It captures statistical structure well enough to translate, classify and summarise, and it holds no meaning, intention or model of the world. The practical consequence is that it cannot notice when a question is absurd or when its own answer is impossible, which is why review matters wherever the output is acted upon.

What is the difference between NLP and text mining?

Largely emphasis. Text mining traditionally means extracting structured information and statistics from document collections. NLP covers that and also understanding and generation. In practice the terms overlap heavily and the distinction rarely matters.

Want to build NLP systems rather than read about them?

The premium catalogue covers fine-tuning, RAG and production text pipelines, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​