Research
What we study
We investigate how language models work and how they can be used reliably. Our research brings linguistic analysis together with knowledge representation and AI security.
The areas below connect related questions across the group. Individual projects often span more than one.
Language & meaning
How do models handle differences between languages?
We use linguistic analysis to understand how models represent meaning and transfer knowledge between languages. This includes formal semantics and typology, as well as evaluation in languages with limited training data.
Our work on language sampling asks which languages a benchmark needs to support a broader claim. Research on information structure examines whether models distinguish what a listener already knows from what a speaker wants to emphasise.
Current questions
- Which linguistic similarities help models transfer knowledge?
- How should Creole and other lower-resourced language tools be evaluated?
- Do models recognise how word order and emphasis change meaning?
Faculty: Johannes Bjerva · Heather Lent · Russa Biswas
Knowledge & reasoning
How should models represent knowledge and support their claims?
We study how to represent and complete knowledge graphs, combining graph structure with text. We also investigate how models can reason over graph paths and connect generated claims to source documents.
These questions extend to scientific communication: what changes when a finding becomes a summary or a public-facing report, and can readers trace each claim to its evidence?
Current questions
- How can graph structure and text improve knowledge-graph completion?
- Can generated claims be traced to relevant passages in long documents?
- How do we measure factuality across languages?
Faculty: Russa Biswas · Dustin Wright · Johannes Bjerva
Security & reliability
Where do language models fail or expose private information?
We test privacy leakage and attacks on multilingual language models. Our work includes recovering text from embeddings, distinguishing memorisation from inference, and identifying models that have been tampered with.
We also examine the reliability of evaluation methods. Security experiments need explicit threat models and responsible disclosure; a useful test must establish what it measures and where its conclusions hold.
Current questions
- Which privacy risks persist across languages and model families?
- Can linguistic or formal methods reveal hidden model failures?
- How can security research reduce harm and support responsible disclosure?
Faculty: Johannes Bjerva · Heather Lent · Xikun Jiang · Russa Biswas · Dustin Wright
Language technology in use
What changes when language tools are used in practice?
Useful NLP depends on the setting. Creole-language research asks which tools communities need. AI:PAGE connects historical sources so that researchers can explore relationships while retaining the context of the documents.
We also study what AI changes beyond an individual task: whose knowledge appears in model outputs, and the environmental consequences of producing and using those systems.
Current questions
- How should community priorities shape the tasks researchers choose?
- Can alignment preserve the diversity of ideas expressed by language models?
- What does sustainable AI require beyond computational efficiency?
Faculty: Heather Lent · Russa Biswas · Dustin Wright · Johannes Bjerva
Research partnerships
Develop these questions with us
We welcome European consortium and MSCA enquiries, especially where our work on security or factuality can contribute to a shared research question.
We choose methods to answer the question, make assumptions explicit, and involve relevant expertise while a project can still change.