Lisa Bylinina (U. of Utrecht) will teach a two-part workshop on AI and language models in linguistic research. You will find an abstract of the workshop below.
Workshop schedule:
Tuesday 20th October 2026, 10.00-12.00 and 13.00-15.00 (lunch 12-13), room U06, Nedre Lysthus
Thursday 22th October 2026, 10.00-12.00 and 13.00-15.00 (lunch12-13), room 2.072, ILP building.
The workshop will also be streamed online via Teams.
We will offer welcome coffee and lunch on both days.
If you would like to take part (in person or via Teams), please fill in the following registration form by October 1:
Abstract:
Language models are appearing more and more often in experimental and theoretical linguistics, as a source of estimates for psycholinguistic measures, and increasingly as objects of study in their own right. The first part of this workshop covers what these models actually are: how transformers are built, what happens during training and what the training objective optimises for, and how scale, architecture and training data shape the behaviour we observe. No background in machine learning is assumed. Alongside the large models that get most of the attention, we will look at small ones trained on developmentally plausible amounts of text, and at multilingual models -- often the more useful instruments for linguistic research, and small enough to run and intervene on.
The second part turns to research practice. We will go through the measures that can be extracted from a model, how to obtain them and how they can be used, working through concrete cases from recent research with an emphasis on methods. We will view language models as learners with controllable linguistic and extralinguistic experience. This allows us to design targeted manipulations of the learning environment -- remove a construction from the input, change word order, add or withhold visual grounding -- and observe the consequences. This raises the question of observation: how to measure, in a faithful and interpretable way, what a model has learned about language as a result of training -- by probing its internal representations, and by modelling its behaviour as a function of its experience. Examining those representations also gives us a way to quantify language similarity on different levels.
The practical sessions run in the browser, with no local installation, and are set up to be as compatible as possible with already existing workflows; no prior experience with Python is required. Participants are welcome to bring their own questions and data.
Lisa Bylinina is an Assistant Professor in Computational Linguistics at Utrecht University, in the Language and Communication group at the Institute for Language Sciences. She works at the intersection of theoretical linguistics and natural language processing. Her recent research uses language models to study how linguistic knowledge can be acquired from linguistic input, and what kinds of evidence in that input support generalisation beyond it. She is a co-author of 'Semantics and Deep Learning' (Cambridge University Press, 2025). Before Utrecht she worked at the University of Groningen, as an NLP engineer in industry, and at Leiden University. She teaches a master's course on transformers and their applications in Utrecht's Applied Data Science programme.