How infants acquire language
Is the human capacity to learn a first language built in, or does it emerge from general learning and social interaction?
Summary
Why it is open
The main argument for innate structure is the poverty of the stimulus. Robert Berwick, Paul Pietroski, Beracah Yankama and Noam Chomsky put it this way: experience "warrants selection of the grammars acquired only if humans assume, a priori, that selectable grammars respect substantive constraints" [9]. Their main candidates include the structure dependence of grammatical rules, in particular constraints on question formation [9]. In Chomsky's example, the question formed from "The dog that is in the corner is hungry" is "Is the dog that is in the corner hungry?", not "Is the dog that in the corner is hungry?", which fronting the first auxiliary verb would give [4]. Usage-based accounts answer that children learn concrete expressions from what they hear and then, "using their general cognitive and social-cognitive skills", categorise, schematise and combine them, as Michael Tomasello put it [3].
Part of the disagreement is about which built-in principles are needed. Steven Piantadosi, who argues against innate grammar, grants that "there are no doubt some principles required for language"; the question, he writes, is whether they are language-specific, innate and of the form generative theorists propose [13]. Whether the mind in general is built from innate, specialised systems is a separate problem, the modularity of mind.
The problem is also several questions: how infants find sounds and words, how they reach syntax, why the age of exposure matters and what social interaction adds. No body sets a bar for a resolution. The sources point to two kinds of test. One counts input: whether the data that separate the right rule from a wrong one are absent from children's experience, the claim that Geoffrey Pullum and Barbara Scholz examined in four cases [4]. The other builds learners: Roni Katzir argues that a language model can serve as a theory of human linguistic cognition only if, as "a necessary condition of adequacy", it behaves like humans "after training on the same amount of linguistic data" [12].
What is known
What infants extract from the input (1996–2003). Jenny Saffran, Richard Aslin and Elissa Newport found that 8-month-old infants segmented words from fluent speech after 2 minutes of exposure, "based solely on the statistical relationships between neighboring speech sounds" [1]. Infants can at first tell apart the phonetic units of all languages, but between 6 and 12 months their discrimination of foreign-language units declines sharply [7]. Patricia Kuhl, Feng-Ming Tsao and Huei-Mei Liu found that 12 laboratory sessions with native Mandarin speakers reversed this decline in 9-month-old American infants, while the same material from audiovisual or audio-only recordings "had no effect" [7].
Structure without a model (1998–2004). Susan Goldin-Meadow and Carolyn Mylander found that deaf children of hearing parents, with little usable linguistic input, themselves introduce language-like structure into their gestures at the level of sentence and word; the systems of children in American and Chinese families shared structural similarities [2]. In Nicaragua, deaf people created a new sign language in the 25 years before 2004 [8]. Ann Senghas, Sotaro Kita and Aslı Özyürek found that the children who first created it analysed complex events into basic elements and combined them into hierarchically structured expressions, by principles not seen in the gestures of the surrounding hearing speakers, and that later cohorts extended the system [8]. They conclude that "children naturally possess learning abilities capable of giving language its fundamental structure" [8]; the conclusion does not say whether those abilities are specific to language.
Age of exposure (2002–2018). Rachel Mayberry, Elizabeth Lock and Hena Kazmi found that deaf and hearing people exposed to a language in infancy later learned a new language comparably well, whether that early language was signed or spoken, while deaf people with little early language experience performed poorly [6]. From 669,498 native and non-native English speakers, Joshua Hartshorne, Joshua Tenenbaum and Steven Pinker estimated that the ability to learn the grammar of a second language is preserved almost to 17.4 years of age and then declines steadily, which they read as a sharply defined critical period that ends later than previously thought [10].
Richness of the input (2000–2011). Tomasello reviewed evidence that most of young children's early language is organised around concrete, item-based schemas, from which more abstract constructions develop only gradually [3]. Pullum and Scholz examined four published poverty-of-the-stimulus arguments and found that none held up [4]. For auxiliary fronting they estimated, from preliminary counts, that the relevant questions are at least 1% of the questions in a typical corpus, so that even a child who hears only 10 million words by the age of three would hear about 7,500 [4]. Julie Anne Legate and Charles Yang replied that such evidence exists but is insufficient: among the questions addressed to two children in the CHILDES database they found it in about 0.068% (Nina) and 0.045% (Adam), against about 1.2%, which they estimate a child needs to settle a comparable binary choice by around the third birthday [5]. In 2011 Berwick and colleagues held that these arguments "remain an important source of support" for a priori structure-dependent constraints [9].
Recent claims: language models (2023–2024). In 2023 Piantadosi circulated a manuscript, published in 2024 as a book chapter, arguing that large language models "undermine strong claims for the innateness of language" [13]. Built without the constraints that generative linguists held necessary, the models, he argues, "essentially lay" the poverty-of-the-stimulus issue "to rest"; he grants that they are trained on more data than children receive and argues that most of it is probably needed for meaning rather than syntax [13]. Katzir replied in Biolinguistics in December 2023 that the models fail his test: in an April 2023 exchange GPT-4 preferred a sentence that breaks a constraint on gaps in coordinated clauses, which he says every English-speaking child grasps, and four models tested by his group failed on it [12]. Julie Kallini and colleagues tested the claim, which they attribute to Chomsky and others, that language models learn possible and impossible languages equally well; GPT-2 small models trained on altered versions of the BabyLM training corpus "struggle to learn impossible languages" compared with English [15].
Main approaches
Innate, language-specific constraints. Children's grammars respect constraints that their input does not fix, so the constraints must be part of the human endowment [9]. Strongest evidence: the rarity of crucial evidence for structure-dependent question formation in child-directed speech [5], and children who build structured systems from little or no linguistic input [2, 8]. Main difficulty: how much evidence is enough is itself disputed [4, 5], and the authors of the Nicaraguan study describe what children bring as learning abilities [8].
Usage-based learning. Children learn concrete constructions and generalise from them with general cognitive and social skills [3]. Strongest evidence: the item-based character of early speech [3], and phonetic learning in infancy that occurs with live speakers but not with recordings [7]. Main difficulty: the generalisations for which crucial input is rare, the cases Legate and Yang and Katzir cite [5, 12].
Statistical learning. General mechanisms track the statistics of the input. Strongest evidence: word segmentation from the statistics of neighbouring sounds after 2 minutes [1], and a generic network that learned many word–referent mappings from 61 hours of one child's head-camera recordings, made between 6 and 25 months [14]. Main difficulty: both results concern words and meanings, while the poverty-of-the-stimulus arguments concern syntax [1, 9, 14].
Language models as tests of learnability. Train models on child-scale input and see what they acquire [11]. Strongest evidence: the BabyLM Challenge caps the language data for training at 100 million words because children acquire language from less than that, while large models typically use three or four orders of magnitude more; in 2023 the winning submissions outperformed models trained on trillions of words on the challenge's evaluations [11]. Main difficulty: models differ from children in input and architecture, and Katzir argues that what they learn from a corpus is "entirely different" from what humans learn from a comparable amount [12].
Open sub-questions
- Crucial questions at scale. Across the full CHILDES database, what fraction of questions to children distinguishes structure-dependent from first-auxiliary fronting, and does it approach the 1.2% that Legate and Yang treat as sufficient [4, 5]?
- A child-sized learner. Does a model trained on fewer than 100 million words of child-directed input acquire the constraint on gaps in coordinated clauses on which four language models tested by Katzir's group failed [11, 12]?
- Impossible languages. Do architectures other than GPT-2, trained on the same child-scale corpus, also learn possible languages more easily than impossible ones [15]?
- Social input and words. Does the advantage of live over recorded speech for infant phonetic learning [7] also hold for the word–referent mappings that Vong et al.'s model learned from one child's recordings [14]?
- Late first language. In deaf people first exposed to a sign language late, who learn later languages poorly [6], is command of structure-dependent rules [9] impaired more than vocabulary?
Work on this in Papex
If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.
Key references
- Saffran, J. R., Aslin, R. N. & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science 274(5294), 1926–1928. 10.1126/science.274.5294.1926 · Checked 15 Oct 2026
- Goldin-Meadow, S. & Mylander, C. (1998). Spontaneous sign systems created by deaf children in two cultures. Nature 391(6664), 279–281. 10.1038/34646 · Checked 15 Oct 2026
- Tomasello, M. (2000). The item-based nature of children's early syntactic development. Trends in Cognitive Sciences 4(4), 156–163. 10.1016/S1364-6613(00)01462-5 · Checked 15 Oct 2026
- Pullum, G. K. & Scholz, B. C. (2002). Empirical assessment of stimulus poverty arguments. The Linguistic Review 19(1–2), 9–50 (Crossref and OpenAlex list the volume as 18; the article's own header and the DOI give 19). 10.1515/tlir.19.1-2.9 · Checked 15 Oct 2026
- Legate, J. A. & Yang, C. D. (2002). Empirical re-assessment of stimulus poverty arguments. The Linguistic Review 19(1–2), 151–162 (Crossref lists the volume as 18). 10.1515/tlir.19.1-2.151 · Checked 15 Oct 2026
- Mayberry, R. I., Lock, E. & Kazmi, H. (2002). Linguistic ability and early language exposure. Nature 417(6884), 38. 10.1038/417038a · Checked 15 Oct 2026
- Kuhl, P. K., Tsao, F.-M. & Liu, H.-M. (2003). Foreign-language experience in infancy: effects of short-term exposure and social interaction on phonetic learning. Proceedings of the National Academy of Sciences 100(15), 9096–9101. 10.1073/pnas.1532872100 · Checked 15 Oct 2026
- Senghas, A., Kita, S. & Özyürek, A. (2004). Children creating core properties of language: evidence from an emerging sign language in Nicaragua. Science 305(5691), 1779–1782. 10.1126/science.1100199 · Checked 15 Oct 2026
- Berwick, R. C., Pietroski, P., Yankama, B. & Chomsky, N. (2011). Poverty of the stimulus revisited. Cognitive Science 35(7), 1207–1242. 10.1111/j.1551-6709.2011.01189.x · Checked 15 Oct 2026
- Hartshorne, J. K., Tenenbaum, J. B. & Pinker, S. (2018). A critical period for second language acquisition: evidence from 2/3 million English speakers. Cognition 177, 263–277. 10.1016/j.cognition.2018.04.007 · Checked 15 Oct 2026
- Warstadt, A., Mueller, A., Choshen, L., Wilcox, E., Zhuang, C., Ciro, J. et al. (2023). Findings of the BabyLM Challenge: sample-efficient pretraining on developmentally plausible corpora. In Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning, 1–6 (Crossref page range; the paper is longer). 10.18653/v1/2023.conll-babylm.1 · Checked 15 Oct 2026
- Katzir, R. (2023). Why large language models are poor theories of human linguistic cognition: a reply to Piantadosi. Biolinguistics 17, e13153. 10.5964/bioling.13153 · Checked 15 Oct 2026
- Piantadosi, S. T. (2024). Modern language models refute Chomsky's approach to language. In E. Gibson & M. Poliak (eds.), From fieldwork to linguistic theory: A tribute to Dan Everett, Language Science Press, 353–414; DOI 10.5281/zenodo.12665933 (DataCite). First circulated as lingbuzz/007180 (version 1, March 2023). Source · Checked 15 Oct 2026
- Vong, W. K., Wang, W., Orhan, A. E. & Lake, B. M. (2024). Grounded language acquisition through the eyes and ears of a single child. Science 383(6682), 504–511. 10.1126/science.adi1374 · Checked 15 Oct 2026
- Kallini, J., Papadimitriou, I., Futrell, R., Mahowald, K. & Potts, C. (2024). Mission: Impossible Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14691–14714. 10.18653/v1/2024.acl-long.787 · Checked 15 Oct 2026
Source
Open, as listed by Wikipedia · Unsolved problems in neuroscience. Checked 08 Oct 2026.
The status is the source’s. Papex does not decide whether a problem is solved.
Records on this problem
No public record addresses this problem yet.