Course AI for research on language learning and use

Course description

Artificial intelligence serves two distinct roles in language research. First, it functions as a tool:

the model performs tasks that would otherwise require substantial human effort, such as

large-scale corpus annotation. In this capacity, performance is the primary metric; a model that

produces accurate and reliable judgments is valuable, regardless of whether its underlying

architecture or training data mirrors human cognition.

Second, artificial intelligence serves as a model of the learner. Here, the objective shifts toward

understanding what a system can learn from specific experiences and identifying the conditions

that facilitate such learning. While a model trained on vast datasets may excel as an annotation

tool, it may prove to be a poor model of human development. Conversely, constraining a learner

or its experience can render the model more informative when addressing developmental

questions.

The course develops these two perspectives. The first day focuses on recent changes in

automatic annotation, especially the move from task-specific supervised models to

general-purpose LLMs used as annotators or judges. The second day asks how artificial

learners can instead be used to study language development.

Lectures will alternate with discussions and practical demonstrations. Language development

and conversational data will provide the main examples. Participants are expected to be able to

read basic Python code, but no experience training language models is assumed.

Day 1: AI as a tool for annotation

Automatic annotation has traditionally followed a supervised machine-learning pipeline.

Researchers manually annotate a sample, use these annotations to train a model for the target

task, and then apply the trained model to a larger dataset. A new annotation problem generally

requires a new labeled dataset and some amount of model development. Recent work on

child–caregiver interaction provides an example: communicative intents were annotated with a

task-specific sequential model, while response contingency required a separately fine-tuned

language model (e.g., Agrawal et al., 2026)

General-purpose LLMs make a different approach possible. The researcher can specify a new

annotation task in natural language, sometimes with only a few examples, rather than train a

new classifier from scratch. This substantially lowers the setup cost and makes it easier to

experiment with new coding schemes or context-sensitive judgments. The same underlying

model can be reused across quite different tasks (Calderon et al., 2025).

Using an LLM as a judge shifts much of the methodological work from model training to task

formulation. Prompt design becomes central: how should a coding criterion be described, how

much context should the model receive, and which examples help it distinguish difficult cases?

We will examine these questions using grammaticality judgments in child-caregiver language in

interaction (Nikolaus et al., 2024). Finally, we also dicuss the important question of model

choice. For instance, for sensitive data, smaller open models can be run locally, instead of

sending data to an external service.

Day 2: AI as a model of learning

The second day starts from a different use of AI. The model is no longer an external tool used to

annotate human behavior. It becomes the learner whose behavior we want to understand.

Artificial learners can then serve as experimental systems: the conditions of learning can be

manipulated and the resulting knowledge measured.

Developmental theories make different claims about those conditions. Some concern the

structure and amount of linguistic input. Others concern information available through

perception or social interaction. These factors typically co-vary in children and can be difficult to

manipulate independently for ethical reasons. Artificial learners allow us to change one aspect

of the learning environment, retrain the model, and observe what changes in the resulting

knowledge.

We will examine several examples of this, including the properties of linguistic input—such as its

distribution and ordering (Feng et al., 2026)—the role of multimodal experience, where a model

learns from paired visual and linguistic data recorded from a child’s point of view (Vong & Lake,

2026), and the effect of social feedback, where input is held fixed while information supplied

through caregiver responses is manipulated (Nikolaus & Fourtassi, 2026).

Aim/Learning outcomes

By the end of the course, participants should:

• Understand the distinction between using AI as an annotation tool and using it as a model of learning.
• Be able to assess the main choices involved in LLM-based annotation, including model selection, prompt design, privacy, and human validation.
• Understand how artificial and human knowledge can be compared.
• Understand how controlled changes to the learner's experience can be used to investigate developmental hypotheses.

Requirements for participation

• Participants are expected to be able to read basic Python code. No experience training language models is assumed.

Target group/Participants

• PhD students within the domains of language sciences, language technologies, developmental psychology, cognitive science and related domains.

Workload

• Course/ teaching hours: The course takes place over two days, 4 modules of 3h each
• Preparation hours: The students have to familiarize themselves with the literature provided.
• Written assignments etc.: There are hands-on assignment during the course, but none outside of the course.

Language 

• English

Lecturers

• Abdellah Fourtassi, Aix-Marseille University

Literature

• Agrawal, A., Favre, B., & Fourtassi, A. (2026). Scaffolding early dialogue: A unified account of response contingency in child–caregiver interaction. Cognitive Science, 50(7), e70245.https://doi.org/10.1111/cogs.70245 Wiley Online Library
• Calderon, N., Reichart, R., & Dror, R. (2025). The alternative annotator test for LLM-as-a-judge: How to statistically justify replacing human annotators with LLMs. In W. Che, J. Nabende, E. Shutova, & M. T. Pilehvar (Eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 16051–16081). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.acl-long.782 ACL Anthology
• Feng, S. Y., Goodman, N. D., & Frank, M. C. (2024). Is child-directed speech effective training data for language models? In Y. Al-Onaizan, M. Bansal, & Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 22055–22071). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.1231 ACL Anthology
• Nikolaus, M., & Fourtassi, A. (2026). Modelling children’s grammar learning via caregiver feedback in natural conversations. Philosophical Transactions of the Royal Society B: Biological Sciences, 381(1943), 20240374. https://doi.org/10.1098/rstb.2024.0374 PubMed
• Nikolaus, M., Agrawal, A., Kaklamanis, P., Warstadt, A., & Fourtassi, A. (2024). Automatic annotation of grammaticality in child-caregiver conversations. In N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, & N. Xue (Eds.), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 1832–1844). ELRA and ICCL. ACL Anthology ACL Anthology
• Vong, W. K., & Lake, B. M. (2026). On the robustness of modeling grounded word learning through a child’s egocentric input. Open Mind, 10, 1253–1278. https://doi.org/10.1162/OPMI.a.377

Venue

• TBA

Course dates:

• 08 April 2027 09:00 - 16:00
• 09 April 2027 09:00 - 16:00