Solutions · Language
Language, translation and localisation
Adapting language while preserving meaning, terminology and intent.
On one page
Systems that analyse, generate, translate and localise content while maintaining meaning, terminology, register and intent.
- 01Adapting a piece of content without losing what must remain. Translating and localising consists of adapting language, variety, register, references and conventions while preserving the meaning, the terminology and the relationships that hold the content together.
- 02Contrastive research and invariants. Localisation adapts language, variety, terminology, register, references, conventions, formats and audience expectations.
- 03Varieties, register and audience. The same information may need different expressions for specialists, management, users, students, clients or automated systems.
Adapting a piece of content without losing what must remain.
Translating and localising consists of adapting language, variety, register, references and conventions while preserving the meaning, the terminology and the relationships that hold the content together.
A piece of content contains intent, terminology, register, relationships, references, cultural assumptions and domain constraints, as well as its words.
We build systems that research, generate and evaluate controlled adaptations across languages, language varieties, audiences and formats.
This line draws on Excorpora, our research and evaluation work for locating real differences between varieties, domains and contexts without introducing unnecessary changes.
Contrastive research and invariants
Localisation adapts language, variety, terminology, register, references, conventions, formats and audience expectations.
The system declares invariants: meaning, figures, names, relationships, conditions, obligations, units and protected terminology.
Contrastive research and adversarial review help to detect surface changes that look natural yet shift the sense. The evaluation contrasts the result with sources, invariants and purpose.
Varieties, register and audience
The same information may need different expressions for specialists, management, users, students, clients or automated systems. The adaptation modifies selection, explanation, examples, density and vocabulary.
Excorpora develops this line with particular attention to national varieties and semantic fidelity.
NLG composition and multimodality
Natural language generation — NLG — turns informational structures into language through rules, templates, calculations and generative models. The deterministic part protects what must remain exact; the generative part adapts selection, explanation and style.
When the content reaches voice, subtitles or video, the adaptation additionally takes in pronunciation, pacing, duration, synchronisation, legibility and correspondence between voice and image.
Adversarial auditing
We evaluate fidelity, terminology, naturalness, register, necessary change, omission, addition, coherence and fitness for the audience separately. Fluency can coexist with a shift of meaning, and that is precisely the condition adversarial review looks for.
This page goes deeper into one application of informational intelligence; the corresponding technical capability lives in Linguistic engineering and knowledge-based generation.
What we build
- 01Linguistic researchLocating evidence about differences of meaning, usage, register and context.
- 02TerminologyKeeping glossaries, concepts and decisions coherent throughout the project.
- 03ContextTaking in audience, format, domain, intent and communicative situation.
- 04GenerationProposing translations and adaptations appropriate to the defined conditions.
- 05EvaluationExamining semantic fidelity, unnecessary changes, omissions, register and consistency.
- 06ReviewPresenting doubts and alternatives in a way that is useful to the person responsible for the content.
What can be adapted
- 01languages
- 02national varieties
- 03registers
- 04specialised domains
- 05channels
- 06formats
- 07levels of knowledge
The system from the inside
Maturity · Evaluation- How it is organised
- We separate what must remain invariant — figures, names, terminology, legal obligations — from what may be reformulated. A glossary and a translation memory fix the former; the language model works on the latter, with automatic checks of terminology and of non-translatable elements. Deviations are flagged for review instead of being corrected silently.
- What can be examined
- We can show a multilingual alignment demonstrator over public corpora, the glossary and the invariance rules, and real examples of a change of meaning detected during review.
- How it is checked
- Besides the usual automatic metrics, we use human review to evaluate the change of meaning that is linguistically natural, which is precisely where the metrics fail. We measure terminological compliance and the preservation of the invariant elements separately.
Conditions and limits
We are evaluating the system by specific language pairs and domains; a good result in one does not transfer automatically to another. Register and cultural conventions still require human judgement. Detecting subtle shifts of meaning remains an open question for us.