BM-TOK
Balancing multilingual token vocabularies
How to build the vocabulary of a multilingual language model so that less-resourced languages and non-Latin scripts are represented fairly.
Languages, dialects and multilinguality
How language models deal with dialects, scripts and less-resourced languages, and how to make them fairer across languages.
Language models work best for a handful of large languages. We study why, starting from the text itself: how words are split into pieces (tokenization), how scripts and dialects are represented inside models, and how much information text carries in different languages.
The goal is language technology that serves speakers of all languages and dialects more equally. Our work on tokenization bias in dialects received a Senior Area Chair award at EMNLP 2025.
Projects
LPHI-mRAG
Language preference and hallucination in multilingual retrieval-augmented generation
Do language models prefer some languages over others when answering from documents? We test this for chatbots that answer questions from a collection of documents in several languages.
M2P2
Medical, multilingual and privacy-preserving NLP in the clinical domain
Makes modern language technology usable in Swiss hospitals: across the country's languages and without exposing patient data.
ELDI
Dialect identification in short texts
Developed methods to recognise the dialect of short texts such as social media posts, focusing on varieties of Arabic.
INCdid
Dialect identification in noisy text
Developed methods to identify dialects in short, noisy texts, especially between closely related language varieties.
StageAI
Conversational recommendation of performers
With Stagend, improved a platform that matches performers to events, using multilingual question answering to fill in missing information.
People
Publications
- Are Language Models Script-Aware?
- Script Choice in LLMs: Evidence for Late-Layer Commitment
- Are the LLMs Capable of Maintaining at Least the Language Genus?
- Presenting LLMs' Collective Intelligence Approach for Multilingual Hallucination Detection
- Cross-lingual Transfer-learning Approach to Negation Scope Resolution