{"pk":21677,"title":"A systematic investigation of learnability from single child linguistic input","subtitle":null,"abstract":"Language models (LMs) have demonstrated remarkable profi-\nciency in generating linguistically coherent text, sparking dis-\ncussions about their relevance to understanding human lan-\nguage learnability. However, a significant gap exists between\nthe training data for these models and the linguistic input a\nchild receives. LMs are typically trained on data that is or-\nders of magnitude larger and fundamentally different from\nchild-directed speech (Warstadt & Bowman, 2022; Warstadt\net al., 2023; Frank, 2023a). Addressing this discrepancy,\nour research focuses on training LMs on subsets of a sin-\ngle child's linguistic input. Previously, Wang, Vong, Kim,\nand Lake (2023) found that LMs trained in this setting can\nform syntactic and semantic word clusters and develop sen-\nsitivity to certain linguistic phenomena, but they only consid-\nered LSTMs and simpler neural networks trained from just one\nsingle-child dataset. Here, to examine the robustness of learn-\nability from single-child input, we systematically train six dif-\nferent model architectures on five datasets (3 single-child and\n2 baselines). We find that the models trained on single-child\ndatasets showed consistent results that matched with previous\nwork, underscoring the robustness of forming meaningful syn-\ntactic and semantic representations from a subset of a child's\nlinguistic input.\nKeywords: learnability; single-child; distributional learning;\nrobustness; language models","language":"eng","license":{"name":"","short_name":"","text":null,"url":""},"keywords":[{"word":"Artificial Intelligence; Linguistics; Psychology; Concepts and categories; Language development; Language learning; Natural Language Processing; Computational Modeling"}],"section":"Papers with Poster Presentation","is_remote":true,"remote_url":"https://escholarship.org/uc/item/9986685c","frozenauthors":[{"first_name":"Yulu","middle_name":"","last_name":"Qin","name_suffix":"","institution":"New York University","department":""},{"first_name":"Wentao","middle_name":"","last_name":"Wang","name_suffix":"","institution":"New York University","department":""},{"first_name":"Brenden","middle_name":"","last_name":"Lake","name_suffix":"","institution":"NYU","department":""}],"date_submitted":null,"date_accepted":null,"date_published":"2024-01-01T19:00:00+01:00","render_galley":null,"galleys":[{"label":"PDF","type":"pdf","path":"https://journalpub.escholarship.org/cognitivesciencesociety/article/21677/galley/11276/download/"},{"label":"PDF","type":"pdf","path":"https://journalpub.escholarship.org/cognitivesciencesociety/article/21677/galley/22070/download/"}]}