01
English pivot dependency
Indirect translation introduced an additional processing step and another opportunity for meaning to be altered.
European multilingual AI infrastructure
NTEU: direct neural translation infrastructure for every official EU language
Neural Translation for the European Union built a large scale farm of specialized neural models for direct translation between European languages, supported by multilingual data collection, corpus preparation, model training and systematic evaluation.
506 directions Direct neural translation between non English EU languages
24 EU languages Infrastructure designed for complete official language coverage
€1.9 million European funding for multilingual digital infrastructure
2019 to 2021 Two year data, training and deployment programme
Pangeanic led Consortium with KantanMT, Tilde and SEDIA
The European language challenge
A multilingual Europe cannot depend on one language as the route between all others
Earlier machine translation infrastructure commonly relied on English as an intermediate language. A translation from Estonian into Portuguese, for example, could be generated first from Estonian into English and then from English into Portuguese.
This pivot model increased the risk of information loss, accumulated errors and reduced performance for languages with fewer digital resources. It also limited the ability of European public administrations to exchange information directly across national language boundaries.
NTEU addressed the problem by building models for direct translation between official EU languages, supported by the collection, cleaning and preparation of large multilingual corpora.
02
Uneven language resources
High resource languages had significantly more parallel and monolingual data available than smaller European languages.
03
Fragmented public data
Valuable multilingual material existed across institutions, repositories and previous projects but required extensive preparation.
04
Cross border services
European administrations needed reliable language infrastructure for digital services, documents and institutional communication.
The NTEU approach
Build each translation direction as a specific data and model problem
NTEU combined large scale data operations with model specialization. Each language direction required suitable corpora, preparation procedures, training decisions and evaluation criteria.
01
Discover multilingual data
Identify parallel and monolingual resources from European institutions, consortium repositories, public sources and previous research programmes.
02
Clean and normalize corpora
Remove noise, duplicates, misaligned segments, encoding problems and unsuitable content before training.
03
Align language pairs
Convert multilingual documents into usable sentence and segment alignments for each required translation direction.
04
Train specialized models
Build neural translation systems adapted to the data conditions and linguistic characteristics of individual pairs.
05
Evaluate performance
Measure output systematically through automated metrics, comparative testing and professional linguistic review.
06
Deploy as infrastructure
Prepare direct language services for integration into European public administration and eTranslation environments.
Pangeanic in NTEU
Consortium leadership, multilingual data operations and neural model development
Pangeanic coordinated the project and connected corpus acquisition, data engineering, neural training, evaluation and European institutional requirements into a single delivery programme.
LEAD
Project coordination
Pangeanic led the consortium, technical planning, delivery coordination and alignment with the objectives of European multilingual digital services.
DATA
Corpus collection and preparation
Large volumes of bilingual and monolingual material were gathered, cleaned, filtered, aligned and prepared for model training.
NMT
Neural model training
Translation systems were trained for direct language directions rather than relying systematically on English as an intermediary.
LOW
Support for lower resource pairs
Data augmentation, multilingual transfer and synthetic data techniques helped address language directions with limited parallel material.
QA
Evaluation and quality control
Automated evaluation and expert linguistic review were used to compare systems and identify models suitable for operational use.
EU
Public sector integration
The resulting engine farm was designed as capacity for secure multilingual services serving European public administrations.
A major Data for AI programme
The engines were the visible output. The multilingual data operation made them possible.
Training hundreds of direct models required far more than running neural architectures. The central challenge was assembling useful, representative and technically consistent training material for language combinations with very different levels of digital availability.
NTEU therefore operated as a large multilingual data programme: sourcing corpora, validating rights and provenance, removing unsuitable material, detecting language, aligning segments, filtering noise and constructing training and evaluation sets.
This work anticipated the current demand for task specific models. Model quality depends on the suitability of the data, the intended domain, the language direction and the evaluation method.
Corpus acquisition
Institutional repositories, translation memories, public corpora and project data.
Data cleaning
Deduplication, language identification, format normalization and noise removal.
Alignment and filtering
Sentence pairing, confidence scoring and rejection of weak or misleading alignments.
Training set construction
Balanced corpora prepared according to the requirements of each language direction.
Evaluation assets
Separate test material for automated benchmarking and professional human assessment.
Project results
Direct translation capacity across the official languages of the European Union
The project complemented language infrastructure that had traditionally concentrated on directions involving English.
506
Direct translation directions
Twenty three non English official EU languages multiplied by the twenty two other possible target languages.
23 × 22
Complete non English matrix
Each language could be translated directly into every other non English official EU language.
0
Required English pivots
Direct models reduced the need to pass European language content through English.
15M
Target sentence scale
For many language pairs, the programme targeted at least fifteen million high quality training sentences where data availability permitted.
Direct language architecture
European languages communicating without an obligatory intermediary
The objective was a network of specialized directions rather than a system in which every path passed through the same central language.
Estonian → Portuguese
Maltese → Greek
Croatian → Italian
Latvian → Spanish
Finnish → Bulgarian
Slovenian → Danish
Romanian → Lithuanian
Slovak → Dutch
These examples illustrate the architecture. NTEU addressed the full matrix of direct directions between the 23 official EU languages other than English.
European consortium
Language technology companies and Spanish public sector coordination
The consortium combined multilingual data resources, neural machine translation platforms, experience with lower resource languages and institutional coordination.
Coordinator
Public sector partner
Pangeanic
Project leadership, multilingual data operations, neural model development, evaluation and European delivery.
Technology partnerKantanMT
Neural machine translation technology, training infrastructure and model development.
Technology partnerTilde
Multilingual data, neural models and expertise in Baltic and lower resource European languages.
SEDIA
Spanish public sector participation and alignment with national language technology strategy.
NTEU in 2026
An early demonstration of task specific multilingual AI
NTEU was designed before small and specialized models became a central enterprise AI discussion. Its architecture already reflected a principle that remains relevant: different tasks, languages and operating contexts benefit from models trained and evaluated for their specific purpose.
A general model may offer broad linguistic coverage. Operational quality, however, still depends on suitable domain data, representative evaluation sets, controlled deployment and measurable performance in the intended workflow.
NTEU also reinforced a capability that continues through Pangeanic’s current work: transforming multilingual documents and corpora into governed data assets for training, adapting, aligning and evaluating AI systems.
SLM
Task specific models
Model architecture and data can be adapted to a defined language direction, domain and service requirement.
DATA
Multilingual training assets
Clean parallel corpora remain valuable for machine translation, language models, evaluation and model alignment.
EVAL
Language specific evaluation
Aggregate benchmarks cannot replace evaluation across individual languages, domains and operational conditions.
SOV
European control
Public language infrastructure can be deployed under European governance, security and data protection requirements.
From research to production
What NTEU demonstrates for enterprises and public institutions
The project provides a practical blueprint for developing multilingual AI systems where data quality, specialization, evaluation and deployment control matter.
01
For public administrations
Build secure multilingual services for documents, portals, cross border procedures and institutional communication.
02
For enterprise model teams
Prepare multilingual training and evaluation data for task specific language models and controlled applications.
03
For language technology providers
Source, clean, align and evaluate parallel corpora for adaptation, fine tuning and quality benchmarking.
04
For sovereign AI programmes
Develop models and datasets under defined infrastructure, provenance, security and data governance conditions.
Public evidence
Project documentation, academic publication and European references
These sources document the project objective, consortium, engine matrix, dates and technical approach.
Academic publication
Neural Translation for the European Union Project
EAMT paper describing the direct language architecture, consortium and project period.
Read publication →European project profile
NTEU project facts
Public project profile documenting the consortium, coordination, funding and principal objectives.
View project profile →Technical presentation
The 506 engine architecture
Presentation explaining the 23 by 22 matrix and the objective of eliminating English pivoting.
View presentation →Pangeanic project update
NTEU at META Forum
Contemporary project information covering the engine farm, language combinations and public administration use.
Read project update →European Commission
European AI translation services
Current European Commission information on secure multilingual AI tools and official language coverage.
Explore EU language tools →Pangeanic project portfolio
European AI and language technology projects
NTEU within Pangeanic’s wider record in multilingual data, translation, privacy and European AI infrastructure.
Explore all projects →Continue exploring
From multilingual neural infrastructure to cultural and audiovisual AI
NTEU forms part of a longer European trajectory connecting language data, specialized models, human validation and operational AI systems.
European project portfolio
Explore Pangeanic’s European AI and language technology projects
Review projects covering multilingual datasets, machine translation, anonymization, cultural heritage, audiovisual AI and sovereign infrastructure.
View all projects → Recent related projectMOSAIC Media
Discover how multilingual translation evolved into multimodal services for transcription, subtitles, synthetic voice, metadata and automatic dubbing.
Explore MOSAIC Media → Related data projectAI4Culture
Explore reusable AI tools, datasets, metadata translation and human validation for European cultural heritage collections.
Explore AI4Culture →Multilingual data and models
Build language AI around your languages, data and operational requirements
Pangeanic supports enterprises, public institutions and model developers with multilingual data collection, corpus alignment, model adaptation, evaluation, human validation and sovereign deployment.