Wanting to be Understood: Modeling Interaction in Early Language Learning
eScholarship (California Digital Library) · 18.06.2025
DOI: 10.48448/4gcf-5s19 ↗Kurz gefasst
Fachartikel in der Zeitschrift eScholarship (California Digital Library), erschienen am 18.06.2025. Verfasst von Cognitive Science Society 2025 und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Human language acquisition involves diverse learning resources, including self-supervised learning (sequence prediction) and communicative interactions (talking to caregivers). While recent advancements in language models highlight the power of self-supervised learning, the role of communicative interaction remains unclear. This study uses Reinforcement Learning (RL) and parent-child agent simulations to model interactions and investigate their role in language acquisition, as well as whether RL-like mechanisms may function in children. We pretrained a small transformer model as a child agent, which then interacted with Google’s Gemini, acting as a parent agent, to learn language with the goal of being understood. Model evaluations show that the interactive training enhances intelligibility of model’s communication and increases behavioral similarity to real child speech. However, minimal pertaining alone provides noticeable syntactic and semantic competence, with RL yielding no consistent gains. These findings imply that interaction may play a more critical role in pragmatic aspects of language learning than in the development of linguistic structures, and that learning through interaction is a mechanism used by children.
Schlagwörter
Zitieren
APA
2025, C. S. S., Meares, M., Ralston, Robert, Sloutsky, V., & Xu, Q. (2025). Wanting to be Understood: Modeling Interaction in Early Language Learning. eScholarship (California Digital Library). https://doi.org/10.48448/4gcf-5s19
Vancouver
2025 CSS, Meares M, Ralston, Robert, Sloutsky V, Xu Q. Wanting to be Understood: Modeling Interaction in Early Language Learning. eScholarship (California Digital Library). 2025. doi:10.48448/4gcf-5s19
BibTeX
@article{2025,
author = {2025, Cognitive Science Society and Meares, Madison and Ralston and Robert and Sloutsky, Vladimir and Xu, Qihui},
title = {Wanting to be Understood: Modeling Interaction in Early Language Learning},
journal = {eScholarship (California Digital Library)},
year = {2025},
doi = {10.48448/4gcf-5s19}
}
EUVIMED