The same text. Different languages. Different cognitions at work.
COGNILANG investigates how the language of a text, translational choices, and discursive framing jointly shape what its readers understand, remember, and believe — whatever the text: a news item, a report, a speech, an official document — treating translation as a cognitive process, not a transfer of words, and using AI to objectify interlinguistic transformations at scale.
An international research project directed by Prof. Mathieu Guidère, open to language teams worldwide.
One aligned segment from the pilot corpus — here in English, French, and Arabic, three of the possible languages — with its annotation codes: linguistic LIN, translational TRA, predictability PRD.
What the project does
A text is produced — a news item, a report, a speech, an institutional document. Within hours it is rewritten, adapted, and translated across languages. These versions differ — in framing, in information order, in what is made explicit and what is left out — and those differences are not cosmetic: they change the cognitive work a reader must do. COGNILANG builds comparable corpora of texts in any set of languages (for example English, French, and Arabic in the pilot), annotates them on linguistic, discursive, translational, and predictability variables, uses AI for alignment and divergence detection, and tests the cognitive consequences (comprehension, recall, cognitive load, perceived credibility) with real readers. The protocol is designed to be replicated by other teams in other languages.
Texts differ measurably
Framing, density, and modalisation vary by language; translation leaves observable traces — omission, explicitation, modulation, reorganisation.
Readers feel the difference
Those transformations affect comprehension, recall of key information, and credibility — and originals are not processed like translations.
AI and prediction connect them
AI detects transformation regularities at scale, and linguistic prediction (surprisal) links textual predictability to processing effort.
Built to travel across languages
A fixed scientific core
The constructs, hypotheses, annotation grid, validation rules, and analysis logic are invariant — so results from a Spanish–Chinese team are commensurable with the EN/FR/AR pilot.
A documented adaptation layer
Tokenisation, normalisation, framing-label piloting, and language-model choice are explicitly language-specific, with written rules for how to adapt them and report what you changed.
Shared instruments
An 18-worksheet researcher pack, an AI-assisted analyzer app, code-books with examples, and a six-week workshop plan — everything a new team needs to start.
A growing comparative dataset
Each set of languages adds a panel to the same picture: how mediation across languages reshapes what people read, and what that does to its readers.
Two approaches to translation, by level of analysis
The project's analytical backbone is presented by level of analysis — word, sentence, text, translation — each approach with a pedagogical tool you can try directly.
Cognitive Approach to Translation and Interpretation
Cognitive baggage at word level, cognitive biases at sentence level, cognitive framing of the narrative at text level, and five translation strategies — literal, domesticating, foreignising, explicitating, neutralising.
Explore the levels of analysisNeurolinguistic Approach to Translation
The same four levels read through the bilingual brain: lexical access, real-time sentence processing and prediction, the reader's situation model, and translation strategies as neural routes with real costs.
Explore the levels of analysisLLMs, Cognitive Security & Defence
The project's security and defence component: how large language models can be weaponised for cognitive manipulation — individualised and mass attacks — and how a cognitive immuno-strategy (firewalls, safety barriers) protects readers and institutions.
Read the LLM component