The 2025-2026 Linguistics Colloquium series continues on Monday, October 6, with a talk by Tom McCoy (Yale). The talk will take place in Dwinelle 370 and synchronously via Zoom from 3:10-4:30pm. The title is "Using neural networks to test hypotheses about language acquisition" and the abstract is as follows:
A central challenge in linguistics is understanding how children acquire their first language. A variety of influential hypotheses have been put forth about the learning strategies that children might use (e.g., the syntactic bootstrapping hypothesis: Gleitman 1990). Such hypotheses have received important empirical support in controlled experimental settings (e.g., Naigles 1990), but it is challenging to test whether these hypotheses continue to hold for children’s large-scale, naturalistic input because it would be unethical to perform extensive interventions on a child’s primary linguistic data. In this talk, I will discuss how neural network language models - the type of system powering ChatGPT - can be used to test hypotheses about which strategies are effective for learning from naturalistic child-directed language, providing a source of evidence that complements the controlled experiments that can be run with actual children. I will discuss two case studies that use this paradigm. The first case study focuses on the aforementioned syntactic bootstrapping hypothesis, which postulates that syntax plays a critical role in the acquisition of word meaning, especially for verbs. In support of this hypothesis, we find that syntactic information is central to the ability of neural networks to learn verb meanings. The second case study focuses on the poverty of the stimulus argument as it pertains to English polar questions - the argument that children’s input does not provide strong enough evidence for generic learning algorithms to recognize that English yes/no questions are driven by syntactic tree structure. In accordance with this argument, we find that neural networks trained on child-directed language struggle to learn the syntax-sensitive nature of this phenomenon. Taken together, these results illustrate one way in which neural network models can contribute to linguistic theory. (Work done in collaboration with Xiaomeng Miranda Zhu, Aditya Yedetore, Robert Frank, and Tal Linzen).
