The argument takes the form of an inference: (1) the child masters a specific grammatical target; (2) the primary linguistic data available are limited โ mostly positive examples, with speech errors and no systematic correction; (3) the data combined with general, non-language-specific learning mechanisms are insufficient to derive the target; (4) therefore there must be an innate, language-specific factor that narrows the space of possible grammars. The crux is showing that the data really are "impoverished" relative to the attained competence, and that the missing factor is language-specific rather than domain-general. In its computational form the hypothesis is tested by training models on data comparable in scale to a child's exposure and checking whether they acquire the target grammatical generalizations (e.g. structure-dependence): success is a proof of concept that the target is learnable without innate linguistic biases, while failure can support the nativist claim.
It explains an apparent paradox of language acquisition: how children master a complex, structure-dependent grammar despite finite, impoverished input lacking negative feedback. It highlights the gap between the poverty of the stimulus and the richness of attained linguistic competence, which โ according to nativists โ must be filled by innate constraints (inductive biases) that narrow the space of possible grammars.
Chomsky critiques the behaviorist, empiricist account of language learning โ an intellectual precursor to the poverty of the stimulus argument.
Development of the argument for an innate language faculty and a language acquisition device (LAD).
In "Rules and Representations" Chomsky introduces the term naming the argument.
Chomsky links the poverty of the stimulus to "Plato's Problem": how knowledge exceeds what experience provides.
Geoffrey Pullum and Barbara Scholz challenge the empirical foundations of poverty-of-the-stimulus arguments.
A systematic assessment of syntactic generalization in neural LMs shows architecture matters more than dataset size.
Warstadt and Bowman propose using artificial neural networks trained on human-scale data to rigorously test claims about innate linguistic knowledge.
The BabyLM Challenge operationalizes the poverty of the stimulus: pretraining models on 10M- and 100M-word budgets over developmentally plausible corpora.
