Linguistic Bootstrapping¶

The language-acquisition family in which a learner uses an available cue system—meaning, syntax, prosody, or communicative intent—to constrain and acquire otherwise underdetermined structure in another linguistic domain.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2190
Origin domain: linguistics
Subdomain: developmental psycholinguistics and first language acquisition
Aliases: Bootstrapping (linguistics)

Core Idea¶

Linguistic bootstrapping is the family of first-language-acquisition accounts in which information already available to a learner in one representational domain constrains a learning problem in another. The learner does not recover words and grammar from an undifferentiated speech stream in one step. An accessible cue—conceptual meaning, an emerging syntactic frame, phrasal prosody, a function word, or a speaker's communicative intention—reduces the target hypothesis space enough for further learning to proceed. The defining relation is therefore not simply “small knowledge becomes larger knowledge.” It is a developmentally available source cue linked by a defensible linguistic regularity to a target representation that the cue makes more learnable.

The term names a family rather than one mechanism. In semantic bootstrapping, prelinguistic concepts and initially recognized meanings help the child identify syntactic categories and grammatical relations; Pinker's acquisition theory is the historically central formulation.^[1] In syntactic bootstrapping, the direction reverses: already learned sentence frames and argument structures constrain the likely meaning of a novel word, especially a verb.^[2] In prosodic or phonological bootstrapping, rhythm, stress, intonation, and phrase boundaries in the speech signal help segment words or locate syntactically relevant units.^[3] In pragmatic bootstrapping, evidence that an action is communicative and evidence about a speaker's intention select more reliable word–referent mappings from noisy context.^[4]

The subtype labels identify the source of leverage, not four mutually exclusive schools. A learner may use prosody to find phrase boundaries, function words to classify a neighboring content word, and a partial syntactic analysis to infer that word's meaning. Modern work therefore treats cue reliability, cross-linguistic variation, and integration of multiple information sources as central questions, rather than assuming a single universally decisive bootstrap.^[5] The family includes historically innatist accounts but is not defined by an innate language-acquisition device. Computational models show that statistical joint learning can implement bootstrapping without making any one model's representational assumptions defining of the whole linguistic-bootstrapping family.^[6]

A valid bootstrapping claim must do more than point to a correlation. It must specify what the learner can know at the relevant developmental time, why that cue predicts the target structure in the language being acquired, how the learner could exploit the mapping, and what observable acquisition effect follows. The concept earns its specialist identity from those source–target and evidence obligations.

Structural Signature¶

Sig role-phrases:

the developing language learner — an infant, child, or explicit acquisition model whose knowledge state changes over exposure
the source cue domain — semantic, syntactic, prosodic/phonological, lexical/distributional, or pragmatic information already accessible at the tested stage
the target linguistic domain — word boundary, lexical meaning, grammatical category, argument structure, phrase structure, or another representation not yet independently available
the linking regularity — a language-general or language-specific dependency that makes states of the source cue informative about states of the target
the input episode — speech, scene, interaction, or paired form–meaning observation in which the cue is encountered under ambiguity and noise
the constrained hypothesis space — the set of target analyses remaining after the cue rules out alternatives
the learner-side inference or update — a process that uses the cue, rather than merely exposing the learner to it, and produces a durable change in prediction or representation
the developmental availability condition — evidence that source knowledge precedes or is jointly learnable with the target rather than presupposing the very structure it is meant to explain
the cue-reliability and integration regime — the cross-linguistic validity, relative weighting, convergence, or conflict of several cue sources
the acquisition readout — behavior or model performance that differs from an appropriate no-cue, mismatched-cue, or alternative-cue baseline

Recognition test. Name the learner, source cue, target structure, and proposed mapping. Show that the cue is available before or jointly with the target; show that the mapping is reliable enough in the learner's input; and show that changing the cue changes acquisition or interpretation in the predicted direction. A correlation in adult grammar, a cue that only an already competent speaker can recognize, or an improvement attributable only to frequency does not by itself establish linguistic bootstrapping. The decisive evidence is learner-side use of one constrained information source to unlock another linguistic analysis.

What It Is Not¶

Not unqualified bootstrapping. The live prime requires self-construction from a minimal seed using the system's own products and no external scaffold. Linguistic bootstrapping necessarily consumes external speech and interaction; its invariant is cross-domain cue leverage, not external-resource independence.
Not one unified acquisition algorithm. Semantic, syntactic, prosodic, and pragmatic bootstrapping reverse directions, use different cues, and support different target claims. The umbrella relation unifies them without erasing their empirical differences.
Not statistical learning in general. Distributional learning can discover regularities within a signal and can implement or support a bootstrap. It becomes linguistic bootstrapping only when an available cue or learned representation constrains acquisition of a distinct target representation.
Not transfer of learning alone. Transfer applies previously acquired knowledge in a new task or context. Linguistic bootstrapping explains the coupled acquisition of interdependent linguistic representations, often before either counts as stable mastery.
Not instructional scaffolding. A caregiver may simplify speech or direct attention, but the construct does not require a teacher who temporarily supplies support and fades it. The leverage can reside in regularities of ordinary input.
Not proof of universal grammar. Some historical semantic-bootstrapping accounts posit innate linking rules; prosodic, usage-based, and computational accounts may instead emphasize learned cue weights and general inference. Evidence for a cue effect does not settle the origin of the learner's priors.^[5]^[6]
Not circular explanation. Saying “syntax teaches semantics and semantics teaches syntax” is incomplete unless a temporally available seed, a noncircular mapping, and an empirical update path are specified.
Not any NLP procedure called bootstrapping. Iteratively expanding a seed lexicon or labeled set may instantiate the broad prime, but it is not this developmental-linguistic family unless it models cue-guided acquisition of linguistic representations under the stated validity conditions.

Scope of Application¶

The abstraction belongs primarily to developmental psycholinguistics and theories of first-language acquisition. It recurs through different cue and target domains while retaining the same specialist test.

Syntactic category acquisition. Known meanings, function words, morphology, and local distributional contexts can constrain whether an unfamiliar item is a noun, verb, or other category. Semantic bootstrapping uses conceptual categories as an initial route into syntax, while later lexical and syntactic cues refine the result.^[1]
Verb and word meaning. Syntactic frames constrain participant number and causal structure, narrowing meanings that a visual scene alone leaves ambiguous. Gleitman's formulation and Naigles's experiments are canonical cases.^[2]^[7]
Speech segmentation. Stress patterns, rhythm, phonotactics, and prosodic boundaries help infants locate word and phrase units in continuous speech. These units become restricted domains for later distributional and lexical learning.^[3]^[5]
Phrase structure and word order. Phrasal prosody and function words can constrain an infant's analysis of constituency and the category of adjacent content words. Evidence summarized by Brusini and colleagues shows such cues supporting lexical and syntactic judgments by about eighteen months.^[8]
Pragmatic word learning. Joint attention, communicative action, speaker intention, and discourse context can distinguish intended referents from co-present distractors; the cues modulate when an associative mapping should be trusted.^[4]
Joint and computational acquisition. Probabilistic models can learn lexical mappings and syntax together from child-directed utterances paired with noisy possible meanings. Their value is to show that mutually informative representations need not be learned in a rigid one-way order.^[6]
Cross-linguistic acquisition. A proposed cue must be retested in languages where prosody, word order, case marking, or argument realization differ. A cue that is reliable in English may be weak, reversed, or redundant elsewhere.^[5]
Developmental disorder and atypical acquisition research. Separating cue availability, perception, integration, and use helps locate whether difficulty lies in the signal, the mapping, or the learner's weighting of evidence. Such application requires direct evidence and must not treat the construct as a clinical diagnosis.

The unqualified node does not cover adult second-language strategies, teaching techniques, or language-model training merely because prior knowledge assists later learning. Those may be analogies or qualified extensions; the core scope is the acquisition of linguistic structure by a learner whose component representations are still being built.

Clarity¶

The concept becomes clear when written as a directed relation:

source cue available now → constrained alternatives → target linguistic representation learned next.

For semantic bootstrapping, the source may be a conceptual representation of an event and the target a syntactic category or linking rule. For syntactic bootstrapping, the source is an emerging sentence frame and the target is a novel word's meaning. For prosodic bootstrapping, the source is rhythm, stress, or a boundary cue and the target is segmentation or phrase structure. For pragmatic bootstrapping, the source is evidence of communicative intention and the target is a word–concept mapping. The direction must be stated for each claim; “syntax and semantics interact” is too weak.

Four evidential layers should not be collapsed. Cue presence asks whether the input contains a correlation. Cue accessibility asks whether the learner can perceive and represent it at the tested age. Cue use asks whether manipulating it changes the learner's interpretation or learning. Acquisition consequence asks whether the resulting update carries forward beyond the immediate trial. Adult corpus statistics can establish presence, but not child use. A preference in one experiment can establish use, but not a full developmental account.

The family label also prevents false competition. Semantic and syntactic bootstrapping can form a coupled loop: some meanings seed grammatical analysis, which later constrains new meanings. Prosody can bootstrap the syntactic representation that then bootstraps vocabulary. The loop is explanatory only if at least one entry point is independently available and each transition has evidence. Otherwise, mutual support merely relocates the learnability problem.

Finally, the source cue need not determine a unique target. A useful bootstrap can be probabilistic: it may shrink a large hypothesis set, change posterior weights, or combine with other cues. Höhle's review makes cue reliability and integration across cross-linguistic variation central rather than optional qualifications.^[5]

Manages Complexity¶

Language acquisition presents coupled “chicken-and-egg” problems. Recognizing a word can require knowing where words begin; learning segmentation can benefit from recognizing familiar words. Learning verb meaning can require knowing sentence structure; learning structure can benefit from understanding event meaning. Linguistic bootstrapping decomposes these apparently simultaneous demands into asymmetric, testable transitions.

The source–target matrix reduces explanatory ambiguity. Instead of claiming that “context helps,” an account states which context, which representation, which learner state, and which alternatives are removed. This lets researchers design controlled contrasts: hold a scene constant while varying syntax, hold a sentence frame constant while varying the scene, preserve segmental material while moving a prosodic boundary, or preserve co-occurrence while changing communicative intention.

Cue integration turns a brittle single-key account into a weighted system. When prosody, distribution, and semantics converge, confidence can rise; when they conflict, behavior reveals relative weights and developmental change. Cross-linguistic comparison then becomes a structural stress test: if a language weakens one correlation, successful learners should reweight other cues rather than mechanically apply an English-specific rule.^[5]

Computational implementations add a sufficiency check. Abend and colleagues paired real child-directed utterances with noisy candidate meanings, jointly learned a probabilistic lexicon and grammar, and reproduced several developmental phenomena, including syntactic-bootstrapping effects.^[6] Such a model does not prove that children use the identical algorithm, but it demonstrates that a stated seed, mapping, and update rule can jointly solve more of the acquisition problem than verbal circularity alone.

The abstraction also localizes failure. If a prediction fails, the analyst can ask whether the cue is absent, perceptually inaccessible, unreliable in this language, incorrectly mapped, outweighed by another cue, or measured with a readout too immediate to count as learning. Each diagnosis implies a different next experiment.

Abstract Reasoning¶

Direction audit. Mark the source and target explicitly. Reversing them changes semantic bootstrapping into syntactic bootstrapping rather than restating the same claim.

Seed audit. Identify what is available without presupposing the target. A semantic account must not require a syntactic parse to recover the very meaning used to learn syntax; a syntactic account must not require knowing the novel verb's meaning to recognize its argument frame.

Reliability audit. Estimate whether the source–target correlation holds in the actual input distribution and language. A cue can be perceptually salient yet systematically misleading.

Accessibility audit. Separate evidence available to an analyst from evidence available to a learner. Adult judgments and annotated corpora do not establish infant-accessible representation.

Manipulation and ablation. Compare matched conditions with and without the cue, or remove the cue from a computational learner. If performance is unchanged, the cue has not been shown to bootstrap the target.

Temporal-order test. Establish that source knowledge precedes the target, or fit a joint-learning account in which both can update without circular initialization. Developmental timing is part of causal identification.

Cue-conflict test. Put two cues in opposition and observe which controls the response. Agreement trials demonstrate possibility; conflict trials reveal weighting.

Cross-linguistic stress test. Seek a language where the proposed correlation differs. A genuine mechanism should either predict a changed acquisition path or specify how other cues compensate.

Alternative-mechanism test. Ask whether raw frequency, associative co-occurrence, perceptual salience, teaching, or already acquired target knowledge explains the effect. Bootstrapping survives only when the cross-representational cue adds explanatory work.

Knowledge Transfer¶

Within language-acquisition research, transfer is literal. The same analysis can be carried from verb meanings to noun/verb classification, from English word order to another language's case marking, or from prosodic segmentation to pragmatic reference resolution: identify a source cue, target representation, linking regularity, developmental entry condition, and acquisition readout. The vocabulary changes with the cue, but the specialist validation procedure remains intact.

A practical research workflow is:

define the target structure that is otherwise underdetermined;
inventory candidate cues available in the learner's natural input;
verify cue reliability in the target language;
establish perceptual and representational availability at the relevant age;
manipulate or ablate the cue while holding alternatives constant;
test whether the update persists or supports later learning; and
repeat under cue conflict and cross-linguistic variation.

Outside linguistics, the honest transfer is a more general pattern already represented by learning, inductive_reasoning, transfer_of_learning, or the broad bootstrapping prime. Calling a security system's seed-key expansion or a firm's early revenue “linguistic bootstrapping” imports domain vocabulary without importing infant speech, semantic/syntactic representations, or developmental evidence. Even natural-language-processing seed expansion normally instantiates algorithmic bootstrapping, not this acquisition family. The domain node earns autonomy precisely because its source–target taxonomy and evidence bar do not survive arbitrary substrate substitution.

Examples¶

Canonical: syntax constrains a novel verb meaning¶

Naigles showed two-year-old children an ambiguous event containing both a causative action between two participants and a simultaneous action performed by both. One group heard a novel verb in a transitive frame naming one participant as subject and the other as object; another heard a frame compatible with the participants acting separately. At test, the children hearing the transitive syntax preferred the causative interpretation more than those hearing the alternative frame.^[7] The scene alone supported both meanings. The syntactic frame reduced the target meaning space.

Mapped back to the signature: the two-year-old is the learner; the argument frame is the source cue; the novel verb meaning is the target; the learned mapping between transitivity and participant/causal structure is the linking regularity; the ambiguous scene is the input episode; and the interpretation preference is the acquisition readout. The example supports cue use, not the claim that syntax uniquely determines a complete lexical meaning.

Failure modes: if the visual events were not equally available, perceptual salience could masquerade as syntactic bootstrapping. If children already knew the novel form, the target would not be novel. If the response vanished when syntax was the only manipulated factor, the cue-use claim would fail.

Applied research: joint lexical and grammatical learning from naturalistic input¶

Abend and colleagues built an incremental Bayesian learner that received syntactically complex child-directed utterances from the CHILDES corpus paired with several contextually plausible logical forms, including distractors. It jointly learned mappings between words and components of meaning and a probabilistic grammar. Previously learned constructions accelerated novel-word learning, while emerging lexical knowledge supported grammar induction; the model also produced developmental patterns such as rapid word-order learning and vocabulary acceleration.^[6]

Mapped back to the signature: the model is the learner; structured possible meanings and emerging constructions are alternating source cues; lexical meanings and syntax are alternating targets; probabilistic form–meaning correspondence is the linking regularity; corpus utterances with noisy interpretations are the input; posterior restriction and improved generalization are the update and readout.

Failure modes: the result is a computational sufficiency demonstration, not direct evidence that children represent logical forms or run this algorithm. If clean meaning annotations leaked the correct target, or if performance did not exceed no-cue and shuffled-cue baselines, the claimed bootstrap would be overstated.

Structural Tensions¶

Semantic-to-syntax vs. syntax-to-semantics. The two directions can reinforce one another, but each has a different seed and different risk of circularity.
Innate linking rules vs. learned cue weights. Historical accounts often posit language-specific prior structure; statistical and usage-based accounts explain more of the mapping from experience. Cue effects alone do not adjudicate the origin question.^[5]
Single decisive cue vs. integrated evidence. Clean experiments isolate one cue, whereas natural acquisition combines prosody, distribution, meaning, and pragmatics. Isolation improves causal clarity but can misstate ecological operation.
Universal regularity vs. language-specific reliability. A cue useful in one language may be absent or differently aligned in another. Universality must be demonstrated, not inherited from English.
Cue presence vs. learner use. Corpus correlations show information availability; only developmental behavior or an explicit learning model shows exploitation.
Early seed vs. hidden target knowledge. The stronger the source representation, the easier the learning problem—and the greater the danger that the account has smuggled in what it promised to explain.
Subtype autonomy vs. family coherence. Prosodic, pragmatic, semantic, and syntactic accounts warrant distinct experiments, yet all instantiate cue-guided cross-representational acquisition.
Immediate interpretation vs. durable acquisition. A cue can bias looking in one trial without producing a lasting lexical or grammatical representation.
Mechanistic adequacy vs. behavioral fit. A model can reproduce an acquisition curve using representations children may not possess; empirical fit and psychologically plausible mechanism are separate obligations.
Autonomy vs. reduction. The construct shares learning, inference, transfer, compositionality, and broad bootstrapping structure. Reduction to those primes loses the linguistic cue taxonomy, developmental entry condition, cross-linguistic reliability test, and acquisition-specific evidence surface; complete autonomy would ignore those real structural dependencies.

Structural–Framed Character¶

The abstraction is framed-leaning. Its directed cue-to-target relation is structural enough to formalize, experimentally manipulate, and implement computationally. It is value-neutral: a bootstrap can succeed, fail, or mislead. It also does not depend on a school, law, or administrative institution.

Yet the identity is inseparable from a human learner entering a natural-language system. “Syntax,” “semantic category,” “prosodic phrase,” and “communicative intention” are not merely interchangeable labels for generic variables; they determine what counts as a cue, target, valid mapping, and acquisition effect. The term also inherits theoretical history from generative, prosodic, connectionist, social-pragmatic, and statistical accounts. Cross-domain use therefore imports a linguistic frame rather than recognizing this full construct in a new substrate.

The five-criterion result is consequently asymmetric: vocabulary does not travel without loss; the concept carries no necessary evaluative weight; its institutional dependence is modest and disciplinary rather than constitutive; it is human-practice-bound; and invoking it outside acquisition usually reframes rather than discovers. The structural core is real, but the specialist frame supplies most of its diagnostic force.

Structural Core vs. Domain Accent¶

Structural core: a learner has a partial representation; a correlated source cue reduces uncertainty over a target representation; an update converts that reduction into later capability; cue reliability and conflict govern success. This core explains why the family is more than a list of named theories.

Domain accent: the learner is acquiring natural language; source cues are semantics, syntax, prosody, phonology, distribution, or pragmatics; targets are words, categories, boundaries, argument structures, or grammar; mappings must be developmentally available and linguistically valid; evidence comes from infant/child behavior, naturalistic input analysis, cross-linguistic comparison, and psychologically interpretable models.

Removing the domain accent leaves a generic cue-guided learning or bootstrapping pattern. It loses the subtype directions, the circularity problem between lexicon and grammar, the special evidential distinction between adult input statistics and infant use, and the requirement to survive variation among human languages. Those losses are material: without them, one cannot classify a study as semantic rather than syntactic bootstrapping, diagnose an unavailable seed, or design the canonical cue-conflict and developmental tests.

learning — strict subsumption. Every linguistic-bootstrap event is a durable, experience-driven update in a learner that changes later interpretation or production. The node specializes learning by requiring a linguistic source cue to unlock another linguistic representation.
bootstrapping — related structural namesake, not strict subsumption under the current live definition. Both use limited initial capability to obtain more capability. The live prime additionally requires internal sufficiency, products of prior stages, and absence of an external scaffold; linguistic acquisition depends on external speech and interaction, so those defining conditions need not hold.
transfer_of_learning — related/presupposed mechanism, not a clean genus. A learned representation can inform a new linguistic problem, but linguistic bootstrapping often jointly acquires the source and target before source mastery or a distinct transfer task exists.
inductive_reasoning — composition, presupposes. The learner moves from finite, noisy cues to defeasible generalizations about unseen words and constructions. Induction does not specify which linguistic domain supplies the cue or target.
compositionality — conditional component. Some semantic and computational accounts rely on decomposing utterance forms and meanings into reusable parts, but prosodic segmentation or pragmatic reference learning need not presuppose full compositional semantics.
pattern_recognition — related mechanism. Detecting distributional, prosodic, or frame regularities can supply a cue. Recognition alone does not establish cross-domain acquisition or durable use.

The minimal prospective DAC placement is therefore strict subsumption under prime:learning plus strict composition/presupposition under prime:inductive_reasoning. The broad namesake remains an explicit semantic neighbor rather than a forced parent because its live structural signature would make ordinary speech input a disqualifying external scaffold.

Relationships to Other Abstractions¶

Current abstraction Linguistic Bootstrapping Domain-specific

Parents (2) — more general patterns this builds on

Linguistic Bootstrapping is a kind of Learning Prime

learning — strict subsumption. Every linguistic-bootstrap event is a durable, experience-driven update in a learner that changes later interpretation or production.
Linguistic Bootstrapping presupposes Inductive Reasoning Prime

inductive_reasoning — composition, presupposes. The learner moves from finite, noisy cues to defeasible generalizations about unseen words and constructions.

Hierarchy paths (3) — routes to 3 parentless roots

Linguistic Bootstrapping → Learning → Adaptation

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Linguistic Bootstrapping sits in a sparse region of the domain-specific corpus (73^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Memory Encoding & Retrieval Effects (15 abstractions)

Nearest neighbors

Generation Effect — 0.85
Innateness Hypothesis — 0.85
Core Vocabulary — 0.83
Agglutination — 0.83
Allophone — 0.83

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Bootstrapping in computing: staged system initialization; usually a literal instance of the broad prime, not language acquisition.
Statistical bootstrap: resampling from an empirical sample to estimate uncertainty; no linguistic source–target representation is required.
Seed-set bootstrapping in NLP: iterative expansion of labels, patterns, or lexicons; an algorithmic family that may be used to model acquisition but is not identical to developmental linguistic bootstrapping.
Semantic bootstrapping: one subtype, conventionally meaning semantics-to-syntax leverage; not an alias for every linguistic bootstrap.
Syntactic bootstrapping: one subtype, conventionally meaning syntax-to-word-meaning leverage; not the whole family.
Prosodic or phonological bootstrapping: a signal-to-segmentation or signal-to-grammar subtype; not merely any use of auditory information.
Pragmatic bootstrapping: communicative intention or social-pragmatic evidence guiding lexical acquisition; not generic context use.
Lexical bootstrapping: an overlapping label for known words or lexical distributions enabling further grammatical or lexical learning; its source and target must be stated.
Scaffolding: temporary, contingent support supplied by another actor and later withdrawn.
Connectionism or statistical learning: possible implementation frameworks; neither entails the cross-domain cue relation by itself.
Language acquisition device or universal grammar: proposed sources of initial constraints in some accounts; not required by the umbrella identity.

References¶

[1] Pinker, Steven. Language Learnability and Language Development: With New Commentary by the Author. Harvard University Press, 1996; original edition 1984. Foundational acquisition theory associated with semantic bootstrapping and learnability constraints. registry ↩a ↩b

[2] Gleitman, Lila. “The Structural Sources of Verb Meanings”. Language Acquisition 1, no. 1 (1990): 3–55. Foundational account of syntactic structure as evidence for verb meaning. registry ↩a ↩b

[3] Morgan, James L., and Katherine Demuth, eds. Signal to Syntax: Bootstrapping from Speech to Grammar in Early Acquisition. Psychology Press, 1996. Authoritative interdisciplinary volume on phonological, prosodic, distributional, and statistical information in speech as cues for early language learning. registry ↩a ↩b

[4] Caza, Gregory A., and Alistair Knott. “Pragmatic Bootstrapping: A Neural Network Model of Vocabulary Acquisition”. Language Learning and Development 8, no. 2 (2012): 113–135. Primary model in which recognition of communicative action gates reliable word–concept learning. registry ↩a ↩b

[5] Höhle, Barbara. “Bootstrapping Mechanisms in First Language Acquisition”. Linguistics 47, no. 2 (2009): 359–382. Authoritative review distinguishing historical principles-and-parameters and prosodic cue-processing accounts and emphasizing cue reliability, cross-linguistic variation, and multi-cue integration. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g

[6] Abend, Omri, Tom Kwiatkowski, Nathaniel J. Smith, Sharon Goldwater, and Mark Steedman. “Bootstrapping Language Acquisition”. Cognition 164 (2017): 116–143. Primary Bayesian model jointly learning words and syntax from naturalistic child-directed utterances paired with noisy possible meanings. registry ↩a ↩b ↩c ↩d ↩e

[7] Naigles, Letitia. “Children Use Syntax to Learn Verb Meanings”. Journal of Child Language 17, no. 2 (1990): 357–374. Primary experimental demonstration that two-year-olds use syntactic frames to constrain interpretations of a novel verb. registry ↩a ↩b

[8] Brusini, Perrine, et al. “Bootstrapping Lexical and Syntactic Acquisition”. In Sources of Variation in First Language Acquisition, 63–80. John Benjamins, 2018. Synthesis of phrasal prosody, function-word cues, semantic seeds, and their use in early lexical and syntactic acquisition. registry ↩