What is a corpus?

Corpus – a collection of language

The word corpus comes from the Latin corpus, body. A corpus, or text corpus, is a collection of texts or recorded speech that has been selected and organized so it can be studied.

What sets a corpus apart from an ordinary pile of documents is that it has been compiled with a purpose. There is often information about each text: source, date, genre or speaker.

From linguistics …

In linguistics, corpora have long been used to understand how language is actually used. Which words are common? How does language change over time? How does spoken language differ from written language?

Across Europe, national language banks and libraries – such as Språkbanken in Sweden – collect corpora and language resources and make them available for research.

… to AI

The same idea lies behind today's language models. A model learns language by processing a very large corpus. The better the corpus reflects a language, the better the model gets at that language.

A Nordic example is GPT-SW3, which AI Sweden built on a Nordic corpus. But smaller European languages make up a small share of global training data. OpenEuroLLM, for example, has had to generate synthetic data because there isn't enough authentic text.

Norway shows why it matters. There, the National Library found that licensed books and newspapers improved Norwegian language models. So the quality and origin of the corpus affect the result, not just its size.

What makes a corpus useful?

Whether it is used in research or for AI training, the same properties usually determine its value:

• Representativeness – do the texts reflect how the language is actually used, in different contexts and genres?
• Size – larger corpora give more examples to learn from, especially for pre-training.
• Metadata – details about source, date and type of text make the corpus easier to filter and use.
• Quality – clean, coherent language is more valuable than duplicates, templates and automated messages.
• Rights – for AI training it must be clear who may grant access to the material, and on what terms.

Read more about how training data is used in pre-training, fine-tuning and for AI agents on our page about AI training data.

Why we're called CorpusMarket

European companies are sitting on corpora without calling them that: email archives, documentation, support tickets, published texts, podcasts and video archives. It is authentic language, often in Nordic and other local European languages, that rarely exists on the open web.

That is the data CorpusMarket helps companies license. We broker licenses between companies that have data and AI companies that need it as training data. You keep ownership, the licenses are limited and you approve every deal. To find out more about us, read about us. If you have published content, text, image and audio is often the simplest way in.