Séminaire de Recherche en Linguistique

Structured Paradigms for Linguistic Rule Learning and Explanation in Language Models

Language models are often tested on isolated sentences or minimal pairs, but such settings provide only limited evidence about how linguistic knowledge is represented and used. 

In this talk, I present a line of work that studies language models through structured paradigms: small sets of systematically related sentences organised as pattern-completion tasks. 

I discuss three stages: a case study on English verb alternations (the causative-inchoative class), a broader framework that transforms minimal pairs into structured paradigms across many grammatical phenomena in English, and a multilingual extension to Chinese, Russian, and Turkish. 

Across these studies, the main result is that structured contexts provide richer evidence than isolated judgments: they can help small models learn with limited data, and they make the behaviour of large language models easier to interpret by revealing not only whether a model succeeds, but also how it fails.