Language, meaning & variation
We study linguistic structure, semantics, multilinguality, low-resource languages, and what model representations reveal about transfer and failure.
Explore language & representationNatural Language Processing · Aalborg University
Language is where AI meets the world.
We study how AI represents language and knowledge, connect its outputs to evidence, and test whether systems remain reliable across languages, communities, threats, and real-world constraints.
5faculty with distinct research programmes
18current researchers across career stages
8active funded projects
≈ DKK 38.5mgross current project volume
A shared intellectual framework
Four dimensions of trustworthy language technology
Our faculty pursue distinct programmes across four dimensions: linguistic validity; knowledge representation and grounding; integrity under failure or attack; and real-world fit and consequences for communities, institutions, and the material environment.
Knowledge & grounding
We build and reason over knowledge graphs, connecting model outputs to source documents and other traceable evidence.
Explore knowledge & groundingSecurity, reliability & risk
We study attacks, privacy, uncertainty, robustness, evaluation, and the conditions under which apparently capable systems break.
Explore reliability & riskCommunities & consequences
We examine how community priorities and institutional settings—including environmental and social costs—change what responsible success means.
Explore communities & consequencesFaculty
Distinct programmes, shared foundations
Each faculty member has an independent intellectual centre of gravity. Collaboration grows from genuine overlap across language, knowledge, evidence, reliability, and consequences—not from forcing every project into one agenda.
Professor of Natural Language Processing · Group Leader
Johannes Bjerva
Johannes studies how linguistic structure can make multilingual language technology more capable, secure, private, and trustworthy. He leads AAU NLP and the Copenhagen Section of AAU's Department of Computer Science.
Tenure-Track Assistant Professor · Co-director, AI:PAGE-Lab
Russa Biswas
Russa studies how explicit knowledge graphs and learned representations can strengthen one another, with applications in multilingual factuality and cultural heritage.
Tenure-Track Assistant Professor
Dustin Wright
Dustin studies how AI transforms information, spanning evidence attribution, science communication, factuality and evaluation, epistemic diversity, efficient modelling, and system-level sustainability.
Assistant Professor
Heather Lent
Heather develops community-grounded language technology for Creole and other lower-resourced languages, spanning multilingual evaluation, machine translation, data quality, security, and research ethics.
Assistant Professor · Joint affiliation with Formal Methods for Security & Privacy
Xikun Jiang
Xikun develops trustworthy and responsible AI, with a focus on privacy-preserving methods and AI verification for secure and reliable systems.
Selected programmes
Long-horizon questions, properly resourced
Our projects connect fundamental research with concrete risks and settings, from multilingual model security to educational feedback and cultural heritage.
LM²-SEC
2025–2030
Linguistically Motivated Language Model Security
Uses linguistic analysis to understand and mitigate security and privacy risks in multilingual language models.
Official recordTRUST
2026–2030
Building TRUST in Text
Linguistically Motivated Language Model Detection
Develops linguistically grounded ways to detect poisoned or tampered language models and scalable safeguards across languages and domains.
Official recordMML-RPL
2022–2026
Multilingual Modelling for Resource-Poor Languages
Explores how systematic linguistic similarities can extend useful language technology to languages with limited labelled data and digital resources.
Official recordFS-AIS
2026–2028
Formal Semantic Methods for AI Safety
Combines formal semantic analysis and neural interpretability to identify and explain rare multilingual failures in generative AI.
Official recordHow we work
Setting people up for success
Excellent research depends on individual judgement and a healthy collective system: expertise should be visible, feedback routine, expectations clear, contributions recognised, and collaboration easier.
Our research culture- Rigour Questions and evidence before venue or fashion.
- Feedback Share work while it can still change.
- Responsibility Independence without isolation.
- Trust Open disagreement with respect and support.
Selected work
Research across the programme
A small selection spanning multilingual evaluation, language-model security, structured knowledge, factuality, and reliable information.
TBLS: Transition-Based Language Similarity through Activation Perturbation
Measures language similarity through perturbation-induced cross-lingual transitions, finding stable structure across multilingual models and datasets.
- Multilingual & Lower-Resource NLP
- Language, Meaning & Typology
When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs
Shows that vision-language models respond to discourse context in Hungarian but compress interacting information-structural pressures into narrow patterns.
- Language, Meaning & Typology
- Factuality, Reliability & Evaluation
Few-shot Semantic Recovery Attacks on Image Embeddings
Introduces a few-shot black-box method for recovering captions and sensitive attributes from image embeddings, and tests the utility cost of current defences.
- Security, Privacy & Safety
Do LLMs Really Memorize Personally Identifiable Information? Revisiting PII Leakage with a Cue-Controlled Memorization Framework
Separates genuine memorisation from reconstruction enabled by lexical cues across 32 languages, tightening the evidential standard for claims about PII leakage.
- Security, Privacy & Safety
- Multilingual & Lower-Resource NLP
MultiHal: Multilingual Dataset for Knowledge-Graph Grounded Evaluation of LLM Hallucinations
Builds a multilingual multihop benchmark from 25,900 curated knowledge-graph paths for evaluating and reducing hallucinations.
- Knowledge, Evidence & Reasoning
- Multilingual & Lower-Resource NLP
How Good is Your Wikipedia? Auditing Data Quality for Low-resource and Multilingual NLP
Audits all non-English Wikipedia editions and shows why data quality—not language coverage alone—must be a first-class concern in multilingual NLP.
- Multilingual & Lower-Resource NLP
- Factuality, Reliability & Evaluation
News & milestones
What is happening now
People
Six new researchers join AAU NLP
Since 1 June, Ren Tao, Anahita Baninajjar, Dorielle Lonke, Anna Lackner, Davis Davalos-DeLosh, and Lena Pickartz have joined the group, strengthening our work in linguistics, trustworthy AI, and language-model security.
Publications
Five papers at EMNLP 2026
Four Main Conference papers span multilingual language modelling, vision-language models, embedding security, and epistemic diversity. They include Tao's first paper and the final PhD thesis papers for Marcell and Yiyi. A fifth contribution was accepted to the System Demonstrations track.
Projects
Two new programmes begin
TRUST and Formal Semantic Methods for AI Safety strengthen the group's work on model integrity, formal semantics, multilingual failure analysis, and explainable safeguards.
Work with us
Bring us a hard language problem.
We collaborate across linguistics, semantics, logic, machine learning, AI, security, privacy, HCI, and the domains where language technology has to work in practice.
Start a conversation