Skip to content
BEST·BOOKS
+ MENU
Books Authors Tags Endorsers Sources
← Back to Learnability and Cognition: The Acquisition of Argument Structure

AI Study Notebook AI-generated

Study Guide: Learnability and Cognition: The Acquisition of Argument Structure

Steven Pinker

By Best Books

This AI-generated study guide is a reading aid. The source-backed recommendation record and evidence for this book live on the book page.

Key points Not available Flashcards Not available
On this page

Learnability and Cognition: The Acquisition of Argument Structure — Chapter-by-Chapter Outline

Author: Steven Pinker First published: 1989 (MIT Press) Edition covered: New edition, 2013 (MIT Press, 512 pp.), which is a reprint of the original 1989 text with a new preface by Pinker titled "The Secret Life of Verbs." All chapter content refers to the original 1989 text, which is identical in the new edition. The 1989 first edition is 411 pp. The 2013 edition adds approximately 100 pages of new prefatory material.

Central thesis

Children learning language face a profound puzzle: they must acquire not just the words of their language but the intricate system of rules governing how verbs combine with their arguments — who does what to whom. Yet children routinely generalize these rules beyond the verbs they have heard in a given construction, and they also learn, without explicit correction, that certain verbs that seem like candidates for a construction in fact cannot appear in it. How do they know when to generalize and when to stop?

Pinker's central claim is that children solve this puzzle — Baker's Paradox, the seeming impossibility of learning verb-argument alternations without negative evidence — through a two-part mechanism: a rich system of lexicosemantic representations that encodes verb meanings at a fine-grained conceptual level, combined with semantically conditioned linking rules (both broad-range and narrow-range) that restrict which verbs may enter which syntactic constructions. Children do not need to be told that "donate" cannot appear in the double-object dative (He donated her a book) because the semantic features of "donate" fail to satisfy the narrow conditions under which the dative alternation applies. Meaning, in effect, is the constraint.

The book simultaneously advances a second, deeper thesis: that studying how children acquire argument structure is a window into the nature of mental representation itself — into how the human mind categorizes events, encodes causation, agency, motion, and contact, and maps those conceptual structures onto the grammatical machinery of language.

How can children learn which verbs can and cannot appear in which argument structures, given that they hear only grammatical sentences and are rarely corrected?

Chapter 1 — A Learnability Paradox

Central question

Why does learning argument-structure alternations appear logically impossible, and what form does the paradox take?

Main argument

Argument structure and the lexicon

Verbs are not passive vocabulary items; they carry detailed information about how many participants an event involves, what roles those participants play, and which grammatical positions they must occupy. This information — argument structure — is stored in the mental lexicon alongside phonological and semantic information. The dative alternation illustrates the issue: some verbs freely alternate between the prepositional dative (She gave a book to Mary) and the double-object dative (She gave Mary a book), while superficially similar verbs do not (She donated a book to the library / *She donated the library a book). The pattern is not random, but its logic is not obvious from surface syntax alone.

The logical problem of language acquisition

Pinker situates the puzzle within the broader logical problem of language acquisition: children must converge on the adult grammar of their language despite receiving only positive evidence (sentences that are grammatical) and very little, if any, reliable negative evidence (information about what is ungrammatical). This is Gold's (1967) theorem applied to natural language: for any class of languages learnable only from positive data, the learner must be conservative enough to avoid overgeneralizing. But children are not uniformly conservative — they do overgeneralize in well-documented ways.

Baker's Paradox

The paradox is named after the linguist Mark C. Baker, who noted its sharpness for lexical rules. Children must learn that verbs like "give," "hand," and "send" can appear in the double-object construction but verbs like "donate," "carry," and "say" cannot. They cannot learn this simply by observing that they have never heard "donate" in the double-object frame — absence of evidence from a positive-evidence-only input is not evidence of absence. They could simply not have heard the sentence yet. If children freely generalize verbs into constructions they have not yet heard those verbs in, they will systematically overgeneralize and produce forms the adult grammar rejects. But if they never generalize, they will fail to learn the productivity of the rules at all.

Attempted solutions to Baker's Paradox

Pinker surveys and systematically dismisses a range of proposed solutions:

  • Strict lexical conservatism: children only use verbs in constructions they have explicitly heard them in. Empirically false — children demonstrably overgeneralize (e.g., "Don't giggle me" for "Don't make me giggle").
  • Indirect or subtle negative evidence: children use statistical regularities in input — if a verb is heard frequently without appearing in a construction, they might infer it cannot appear there. Pinker shows this is implausible given the sparse and noisy input children receive.
  • Syntactic representations as constraints: appeal to abstract syntactic features that distinguish dative-alternating from non-alternating verbs. This displaces the problem: how do children acquire those syntactic features?
  • Semantic constraints: the correct direction, but previous accounts are too vague to be empirically testable. This preview sets up the book's positive program.

Key ideas

  • Argument structure is a distinct level of linguistic representation, neither purely syntactic nor purely semantic.
  • Baker's Paradox is not a peripheral puzzle but a central test case for any theory of language acquisition.
  • The absence of reliable negative evidence is an empirical fact, not just a theoretical assumption — parental corrections of grammar are rare and children do not reliably use them.
  • Any successful solution must explain both the productivity of lexical rules (children do generalize) and the constraints on that productivity (they eventually converge on adult patterns).
  • The paradox reveals a deep requirement on the architecture of the language acquisition device: the learner must have some way to represent not just which constructions a verb has appeared in but which constructions it could appear in.

Key takeaway

Baker's Paradox demonstrates that learning argument structure from positive evidence alone is logically impossible unless the learner brings to the task some prior constraints on the hypothesis space — the book's mission is to identify exactly what those constraints are.

Chapter 2 — Constraints on Lexical Rules

Central question

What restrictions govern lexical rules — the rules that map a verb's canonical argument structure to alternate syntactic frames — and can those restrictions solve Baker's Paradox?

Main argument

Morphological and phonological constraints

Lexical rules are not fully general. Many are sensitive to the morphophonological shape of the verb: English dative alternation applies overwhelmingly to monosyllabic, non-Latinate verbs ("give," "hand," "send") but not to polysyllabic, Latinate verbs ("donate," "transmit," "communicate"). Similarly, the causative alternation (intransitive-to-transitive: "The bottle broke" / "She broke the bottle") applies to native-Germanic verbs but not to Latinate ones ("The vase shattered" / "She shattered the vase" vs. the marginal status of "The substance liquefied" / "They liquefied the substance"). Pinker argues that these morphophonological constraints are real but insufficient on their own — they are correlates of, not the cause of, the semantic distinctions that do the true constraining work.

Semantic constraints

The main content of this chapter is a detailed survey of the semantic conditions under which major lexical alternations apply or fail to apply:

  • Dative alternation: double-object dative is licensed only for verbs of caused possession/transfer (a narrow class including "give," "send," "bring," "hand") and verbs of future having (a narrow class including "promise," "owe," "guarantee"). Verbs expressing transfer via a trajectory but without caused possession ("carry," "pull") do not alternate. Verbs encoding manner of transfer (e.g., "schlep") are borderline. Pinker maps out these verb classes in fine-grained semantic detail.
  • Causative alternation: the intransitive-to-transitive causative alternation (the "inchoative-causative alternation," as in "The door opened" / "She opened the door") applies to internally caused change-of-state verbs but not to externally caused ones. Verbs of sound emission (intransitive) do not freely causativize.
  • Locative alternation: the alternation between "load hay onto the wagon" (theme-object) and "load the wagon with hay" (location-object) applies only to verbs that entail holistic change of state of the goal/location (filling, covering). "Pour water into the glass" does not alternate to *"pour the glass with water" because pouring does not entail that the glass ends up full.
  • Passive: the passive is far more general but is also semantically conditioned — stative verbs like "resemble" resist agentive passivization.

How semantic and morphological constraints might resolve Baker's Paradox

The emerging proposal: if children can identify the semantic class of a verb (what kind of event-type it denotes), they can use semantic class membership as the basis for predicting alternation behavior. A verb is learnable as dative-alternating or not on the basis of its meaning, without needing explicit negative evidence about its syntactic distribution. The semantic class of "donate" (asymmetric transfer to a beneficiary but without direct physical bestowal) differs from the semantic class of "give" in exactly the ways that predict non-alternation.

Key ideas

  • Morphophonological constraints on lexical rules are real but derivative — they reflect the historical etymology of verb classes rather than the causal mechanism.
  • Semantic constraints are systematic, not idiosyncratic: each major alternation is tied to a specific semantic property that all and only the alternating verbs share.
  • The dative alternation applies to verbs of caused possession and verbs of future having, but not to verbs of manner-of-motion or abstract/metaphorical transfer.
  • The locative alternation requires holistic change of state of the goal; pouring, dripping, and spooning do not meet this condition.
  • These semantic conditions are potentially knowable to a child from the meaning of the verb itself, without observing the verb's syntactic distribution.
  • The program raises the next question: what is the cognitive architecture underlying these fine-grained semantic conditions?

Key takeaway

Lexical rules are semantically conditioned in precise, reconstructible ways, suggesting that a child who correctly represents verb meanings could, in principle, determine the alternation potential of any verb from its semantics alone.

Chapter 3 — Constraints and the Nature of Argument Structure

Central question

Why do lexical rules carry semantic constraints, and what theory of argument structure explains the systematic relationship between verb meaning and syntactic behavior?

Main argument

Why lexical rules carry semantic constraints

The pattern from Chapter 2 cries out for explanation: why should the dative alternation care whether a verb denotes caused possession rather than caused motion? Pinker argues that the semantic conditions on alternations are not arbitrary historical accidents but follow from the underlying semantics of the constructions themselves. The double-object dative construction as a whole encodes a meaning — roughly, "X causes Y to have Z" — and this construction-level meaning constrains which verbs can appear in it. A verb that does not entail caused possession is semantically incompatible with the double-object frame. The semantic condition on the rule is the semantics of the target construction.

This argument is developed through a detailed analysis of event structure. Each verb meaning can be decomposed into a causative structure, a change-of-state predicate, and thematic roles. The lexical rule that maps the canonical structure to an alternating structure is possible only when the semantics of the alternating construction is consistent with (a sub-event of) the original verb meaning. This is the semantic coherence condition on lexical rules.

A theory of argument structure: universality

Pinker draws on cross-linguistic evidence to argue that the linking of semantic roles to syntactic positions is governed by universal principles — linking rules — that map thematic roles (agent, patient, goal, theme, source) to grammatical relations (subject, object, oblique). These linking rules are:

  • Agent → Subject (the default subject linking)
  • Patient/Theme → Direct Object (the default object linking)
  • Goal → Indirect Object or Oblique (depending on construction)

These defaults explain why passives, causatives, and argument-structure alternations are cross-linguistically similar in their semantic conditioning: the same underlying semantic properties trigger the same mapping options across languages.

Thematic tiers and event structure

Pinker adopts a framework in which argument structure is organized around thematic tiers — the action tier (who acts on whom: agent and patient) and the thematic tier (what moves where: theme, goal, source). This two-tier analysis explains why certain alternations affect one tier but not the other, and why some verbs — particularly psychological verbs (e.g., "fear," "frighten") — have anomalous linking patterns: the experiencer occupies a grammatically subject-like position but a semantically patient-like position.

Key ideas

  • Semantic constraints on lexical rules are not arbitrary but follow from the semantics of the construction itself; a verb and a construction are compatible only when the construction's meaning is a sub-event of the verb's meaning.
  • Universal linking rules map thematic roles to grammatical relations, providing the language-universal foundation on which each language's specific rules operate.
  • Event structure is organized into action-tier and thematic-tier components, each governing a different aspect of argument realization.
  • Psychological verbs illustrate the complexity of the mapping: "frighten" is causative and maps the experiencer to object position, while "fear" is stative and maps the experiencer to subject position, despite both describing similar psychological states.
  • The theory predicts the directionality of alternations: which form is "basic" and which is "derived."
  • Argument structure is a level of representation intermediate between full lexical semantics and surface syntax.

Key takeaway

Lexical rules carry semantic constraints because they are meaning-to-meaning mappings, not form-to-form mappings; a child learning argument structure is, at bottom, learning how event-conceptual structure connects to syntactic structure.

Chapter 4 — Possible and Actual Forms

Central question

What is the inventory of actually occurring verb-argument structures, and how do we distinguish the possible from the merely actual?

Main argument

The problem of negative exceptions

Even granting a rich theory of semantic constraints on lexical rules, there remain verbs that appear to have the right semantic properties for an alternation yet fail to alternate in practice. Pinker calls these negative exceptions — verbs that are semantically eligible but syntactically restricted. How are they learned, and how does the learning theory handle them?

The answer lies in the difference between broad-range rules and narrow-range rules. Broad-range rules apply to wide, semantically general verb classes (e.g., "all verbs of caused possession can appear in the double-object dative"). These rules produce the default behavior of large verb classes. Narrow-range rules apply to semantically specific sub-classes, carving out smaller groups of verbs that share a more tightly defined semantic property and that participate in (or are excluded from) a construction for reasons tied to that fine-grained property.

Narrow sub-classes: transitive action verbs as evidence

Pinker develops extended case studies showing that verb classes are organized into a hierarchical structure of increasingly specific sub-classes. For example, verbs of motion can be divided into verbs of manner-of-motion (run, walk, fly) and verbs of directed motion (go, come, arrive). Only directed motion verbs freely participate in certain deictic constructions. Within verbs of caused-change-of-state, there are sub-classes based on instrument, result, manner, and means — each with slightly different alternation behavior.

This hierarchical organization is not arbitrary: it reflects the conceptual structure of the event-types the verbs describe. Children, as they acquire more detailed knowledge of the semantic sub-classes of their language's verbs, simultaneously acquire more narrow-range constraints on lexical rules.

The relation between narrow-range and broad-range rules

Broad-range rules provide the productive default; narrow-range rules implement the exceptions. A verb like "donate" starts out as a potential candidate for the dative alternation under the broad-range rule but is excluded by the narrow-range condition that the bestowal must be physically direct rather than institutionally mediated. The learning theory predicts that children should initially overgeneralize at the broad level, then progressively retreat as they acquire the narrow-range semantic classes that divide the productive from the unproductive cases.

Key ideas

  • Negative exceptions are not truly arbitrary: they reflect the mismatch between a verb's meaning and the fine-grained semantic requirements of the construction.
  • Verb meaning is organized hierarchically into broad classes and semantic sub-classes; alternation behavior is governed at multiple levels of this hierarchy.
  • Broad-range rules generate productivity; narrow-range rules limit it.
  • Acquisition of argument structure is the acquisition of both levels: broad rules come first, narrow rules emerge as the child's semantic analysis of verb meanings becomes more refined.
  • Cross-linguistic comparison reveals that what counts as a "sub-class" is often universal, grounded in conceptual structure rather than language-specific convention.
  • The theory predicts developmental trajectories: early overgeneralization followed by retreat, driven by the acquisition of narrow semantic classes.

Key takeaway

The possible space of verb-argument structures is structured hierarchically; children converge on the adult system not by rote memorization but by acquiring the nested system of semantic sub-classes that governs which verbs belong in which constructions at each level.

Chapter 5 — Representation

Central question

What is the nature of lexicosemantic representation — the mental encoding of verb meaning — and how does it interface with syntactic structure?

Main argument

The need for a theory of lexicosemantic representation

The account developed in earlier chapters rests on the claim that children represent verb meanings at a sufficiently fine-grained conceptual level to distinguish, say, "give" from "donate" in ways relevant to dative alternation. This requires a theory of what verb meanings actually consist of — not just what they refer to, but how they are mentally structured. Pinker argues that semantic representations are decompositional: verb meanings can be analyzed into primitive conceptual constituents and functions.

Is a theory of lexical semantics feasible?

Pinker confronts the skeptical position — prominent in philosophy of language (Fodor's arguments against decomposition) — that word meanings are atomic, unanalyzable units. He argues that the evidence from argument-structure alternations is precisely evidence for decomposition: the fact that alternation behavior is predictable from verb meaning shows that meaning has internal structure that is systematically related to syntactic behavior. Meaning cannot be atomic if it makes structural predictions.

Conceptual constituents for motion and change-of-state events

Drawing on the framework of Ray Jackendoff, Pinker proposes that verb meanings are built from a small inventory of conceptual primitives: predicates like CAUSE, GO, BE, STAY, ACT, and functions like TO, FROM, VIA, AT, FOR. These combine into event representations using conceptual functions. For example:

  • "give" ≈ CAUSE([x], GO([y], TO([z]))) — x causes y to go to z
  • "fill" ≈ CAUSE([x], GO([y], TO(IN([z])))) — x causes substance y to go into z, with the resultant state that z is full

The semantic representation of "fill" entails the holistic change-of-state of the container z, which is why "fill the glass with water" but not *"pour the glass with water" — pour lacks the resultant-state component.

Interfacing semantic structures with syntax: linking arguments

Pinker specifies how the conceptual-structure representation interfaces with syntactic argument structure via the linking rules established in Chapter 3. The thematic role of each argument is read off its position in the conceptual structure, and the universal linking rules then map those thematic roles to syntactic positions. Crucially, it is the conceptual structure — not a list of thematic role labels — that drives linking, making the system compositional and extendable to novel verbs.

Key ideas

  • Verb meanings are decompositional: they can be analyzed into conceptual primitives (CAUSE, GO, BE, AT, etc.) and their structured combinations.
  • Skepticism about lexical decomposition is undercut by the evidence that argument-structure alternation behavior is systematically predicted by semantic structure.
  • The conceptual primitives are not language-specific; they reflect general categories of event cognition — motion, change of state, causation, agency, possession.
  • Jackendoff's Lexical Conceptual Structure framework provides the formal language for encoding these representations.
  • Thematic roles (agent, patient, theme, goal) are not primitive features added to verbs but rather derivative labels assigned based on a constituent's position within the conceptual structure.
  • The same representational framework handles the semantics of spatial prepositions, aspect, and causation, suggesting it describes a single, integrated system of event cognition rather than a language-specific module.

Key takeaway

Verb meaning is compositionally structured using a small set of conceptual primitives; this structure, not a list of thematic role labels, drives the mapping to syntax and explains why children who correctly acquire verb semantics automatically acquire its argument structure.

Chapter 6 — Structures

Central question

How is the full system of argument-structure representations — including both canonical and alternating frames — organized, and what is the relationship between lexical entries and syntactic phrase structure?

Main argument

Syntactic frames and their semantic correlates

This chapter lays out the full inventory of argument-structure types that English (and languages generally) use, describing the systematic relationship between each frame and its semantic interpretation. The frames include:

  • Unaccusative intransitives (the vase broke): the single argument is semantically a theme/patient, occupying a deep-object position.
  • Unergative intransitives (the dog barked): the single argument is an agent/actor.
  • Simple transitives (she opened the door): agent subject, patient/theme object.
  • Ditransitives / double object (she gave Mary a book): agent subject, goal/beneficiary first object, theme second object.
  • Prepositional-object constructions (she gave a book to Mary): agent subject, theme object, goal oblique.
  • Resultatives (she wiped the table clean): agent subject, theme object, result predicate.

The unaccusative hypothesis

A major theoretical commitment of the chapter is Pinker's treatment of the unaccusative hypothesis (Perlmutter, Burzio): certain intransitive verbs place their single argument in underlying object position rather than subject position. Verbs of motion and change of state ("arrive," "break," "fall," "melt") are unaccusative; verbs of agent activity ("run," "bark," "sneeze") are unergative. This syntactic difference correlates with a semantic one: unaccusative verbs denote events whose single participant is affected or moved; unergative verbs denote activities performed by an agent. The unaccusative/unergative distinction is directly derivable from the conceptual-structure representation described in Chapter 5.

Resultative constructions and the limits of lexical rules

The resultative construction (She cried her eyes red; He ran himself tired) provides a test case for the interaction of lexical rules and conceptual structure. Resultatives show that new argument-structure configurations can be assembled compositionally from event structure: the manner-of-action component (running) and the result component (becoming tired) combine into a single event description with a predictable argument-structure pattern. Pinker uses resultatives to show that the conceptual-structure formalism is productive — it generates new event descriptions, not just reorganizes existing ones.

Key ideas

  • The major argument-structure types of English follow systematically from the possible configurations of conceptual-structure representations.
  • Unaccusative verbs have their sole argument projected from the deep object position; this explains their behavior in auxiliary selection (Italian "essere" vs. "avere"), ne-cliticization, and other diagnostics.
  • The semantic distinction between affected themes (unaccusative) and actor agents (unergative) is universally relevant, appearing in morphological and syntactic asymmetries across languages.
  • Resultative constructions demonstrate that conceptual structure is compositional — new event descriptions can be built, with predictable syntactic consequences.
  • The mapping from conceptual structure to argument structure is mediated by a small set of universal linking rules; language-specific variation comes in at a later stage of the mapping.
  • Passives, antipassives, and applicatives across languages represent systematic re-configurations of the mapping between conceptual-structure roles and syntactic positions.

Key takeaway

The full inventory of argument-structure types is generated by a finite system of conceptual representations and linking rules; what looks like syntactic stipulation is largely the reflex of conceptual organization.

Chapter 7 — Learning

Central question

What is the computational mechanism by which children acquire argument structure, and how does the learning procedure interact with the representational system described in earlier chapters?

Main argument

The computational problem

Pinker frames language acquisition as an inference problem: the child must, from a sequence of sentence-meaning pairs, induce the mapping from semantic representations to syntactic structures that constitutes the adult grammar. The key insight is that the acquisition device is not a blank slate performing unconstrained inductive inference over all possible mappings; it is constrained by the representational system (conceptual structures, thematic roles, linking rules) developed in previous chapters. The search space is structured, so the learning problem is tractable.

Semantic bootstrapping

The acquisition of the initial syntactic categories and phrase-structure rules is addressed through semantic bootstrapping (Pinker 1984): children use their pre-linguistic conceptual categories to identify exemplary instances of nouns (words for objects), verbs (words for actions), and adjectives (words for properties), and then use distributional evidence to extend these categories to non-prototypical members. Semantic bootstrapping provides the initial entry point into the syntax without requiring innate knowledge of language-specific phrase structures.

The acquisition of lexical semantic representations

The central learning proposal is that children acquire verb meanings by constructing lexicosemantic representations (conceptual structures) from the semantics of the observed events concurrent with the utterance. They begin with broad-range hypotheses about a verb's semantic class and refine these as they encounter more evidence. The narrowing of hypotheses is guided by the same conceptual primitives that structure the representations themselves: a child who observes that a verb is used only for events involving holistic change of state of a container will narrow its semantic class accordingly, and the linking rules will then automatically predict its syntactic behavior.

The acquisition of narrow-range rules

A specific proposal for how children retreat from overgeneralization: as children acquire more fine-grained semantic sub-classes for verbs in a domain, they progressively restrict their use of broad-range rules to the sub-class of verbs that satisfy the narrow semantic conditions. Some predictions about the acquisition of narrow-range rules:

  • Children will initially accept alternating uses of verbs that adults reject (overgeneralization).
  • The retreat from overgeneralization will correlate with acquisition of the relevant semantic distinctions.
  • Children will be more conservative (less willing to generalize) for verbs from semantically heterogeneous classes than for verbs from semantically homogeneous classes.

Key ideas

  • Language acquisition is a structured inference problem, not unconstrained induction; the representational system reduces the hypothesis space to a tractable size.
  • Semantic bootstrapping provides the entry point for learning syntactic categories without requiring syntactic primitives to be innate in their specific, language-particular form.
  • Children acquire lexicosemantic representations for verbs from semantic analysis of observed events, not from distributional analysis of surface syntax alone.
  • Broad-range rules are learned first; narrow-range rules are acquired as the child's semantic analysis of verb classes becomes more refined.
  • The theory predicts measurable developmental asymmetries: verbs whose semantic class is clear and unambiguous should be mastered earlier than verbs from semantically heterogeneous classes.
  • The learning mechanism is conservative at the narrow level but productive at the broad level — exactly the combination needed to solve Baker's Paradox.

Key takeaway

Children acquire argument structure through a semantically guided process of hypothesis refinement: broad-range rules provide early productivity, and the gradual acquisition of semantic sub-classes drives the progressive retreat from overgeneralization without the need for negative evidence.

Chapter 8 — Development

Central question

What does the empirical developmental record show about children's acquisition of argument structure, and does it confirm the theoretical predictions of the model?

Main argument

Early argument structure in children's language

Children's spontaneous speech reveals a striking pattern: very young children (around age 2) use verbs in fairly limited argument-structure configurations, adhering fairly closely to the constructions they have heard. This early conservatism is often cited as evidence for strict lexical conservatism, but Pinker argues it is better explained by the small size of the child's verb lexicon and the limited range of semantic classes represented in that lexicon. As the vocabulary grows, so does the productivity of the rules.

Overgeneralization errors

Between approximately age 2;6 and age 5, children produce systematic overgeneralization errors in argument structure: causative overgeneralizations (She disappeared the rabbit; Don't fall me down), dative overgeneralizations (She suggested me a book), and locative overgeneralizations (She filled water into the glass). These errors are documented in diary studies, elicited production studies, and corpus analyses. Crucially, the errors are not random — they are systematically constrained to constructions and verbs where the semantic conditions on the broad-range rule are met but the narrow-range condition is violated. This pattern is exactly what the theory predicts.

The retreat from overgeneralization

The developmental evidence also shows that children gradually retreat from overgeneralization as they grow older, converging on adult patterns without receiving explicit correction. Pinker argues that this retreat is driven by the acquisition of narrow-range semantic classes: as children come to represent the semantic distinctions between, say, "pour" and "fill" or between "give" and "donate" at a sufficiently fine-grained level, the narrow-range rules automatically exclude the overgeneralized forms. The retreat requires no negative evidence — only more detailed semantic analysis of verb meanings.

Comparison with accounts appealing to input statistics

Pinker considers and argues against explanations that rely entirely on token frequency in the input: children retreat from overgeneralization even for high-frequency verbs, and they maintain correct restrictions on low-frequency verbs they have heard rarely. The pattern of errors and corrections tracks semantic class membership, not token frequency, which is what the semantic-class account predicts.

Key ideas

  • Early conservatism in children's argument-structure use reflects vocabulary limitations, not a principled avoidance of generalization.
  • Overgeneralization errors are systematic: they occur in constructions where broad-range rules apply but narrow-range conditions are not met.
  • The retreat from overgeneralization is driven by the acquisition of semantic sub-classes, not by frequency effects or negative evidence.
  • The developmental timeline — early conservatism, then overgeneralization, then retreat — is exactly the trajectory predicted by a model combining broad-range and narrow-range rules.
  • Cross-linguistic data show that children learning languages with different argument-structure configurations (e.g., causative morphology in Japanese, locative alternation in Hebrew) exhibit analogous patterns, suggesting the underlying learning mechanism is universal.
  • Experimental studies using novel verbs (nonce words) confirm that children generalize new verbs into argument-structure alternations consistent with their semantic class, even when they have never heard the nonce verb in the alternating construction.

Key takeaway

The developmental record closely fits the theoretical predictions: overgeneralization errors reflect broad-range rules, retreat reflects narrow-range class acquisition, and neither errors nor retreat require negative evidence, confirming the central mechanism of the book.

Chapter 9 — Conclusions

Central question

What are the broader implications of the theory for the nature of language, cognition, and the mind?

Main argument

The solution to Baker's Paradox

Pinker summarizes the proposed solution: children acquire argument structure without negative evidence because verb meaning — encoded as a conceptual structure — constrains which alternations are available. The constraint is not arbitrary but follows from the semantics of the constructions themselves. Narrow-range semantic classes implement the exceptions. The result is a learnability-theoretically tractable system.

Implications for the language-thought relationship

The argument-structure system is not a linguistic curiosity; it is evidence about the nature of conceptual representation. The fact that the semantic primitives underlying argument-structure alternations (CAUSE, GO, BE, AT, FOR, the action/thematic tier distinction) appear in the same form across languages suggests that these are features of universal conceptual structure, not language-specific conventions. Languages differ in how they package conceptual structure into morphosyntax, but the underlying conceptual distinctions are broadly shared.

This has a direct consequence for the relationship between language and thought: the cognitive categories that determine argument structure — categories of causation, motion, change of state, agency, possession — are not derived from language. They are pre-linguistic conceptual categories that language exploits. The Whorfian hypothesis that language shapes thought is, at least in this domain, reversed: conceptual structure shapes the grammar of argument structure.

Implications for linguistic theory

The semantic-constraint theory favors a lexicalist approach to syntax: the information needed to predict argument-structure behavior is stored in the lexicon, in the form of conceptual structures and semantic class memberships, rather than in abstract syntactic rules. This aligns with work in Lexical Functional Grammar and Head-Driven Phrase Structure Grammar against purely syntactic transformational accounts.

Implications for cognitive science

The book closes by observing that the argument-structure system provides an unusually clear window into mental representation because it is productive (allowing novel generalizations), constrained (not everything is possible), and linked to observable behavior (children's language and their errors). The theory argues for rich, structured, decompositional representations in the mental lexicon — against both nativist-atomist accounts (all meanings are innate unanalyzed primitives) and behaviorist-empiricist accounts (meanings are learned from distributional statistics alone).

Key ideas

  • Baker's Paradox is solved by the combination of rich lexicosemantic representations and semantically conditioned linking rules, at both broad and narrow levels.
  • The conceptual primitives underlying argument-structure semantics appear to be universal, pointing to a layer of event cognition that is shared across human languages.
  • Language does not create conceptual categories in the argument-structure domain; it maps pre-existing conceptual categories onto syntactic structures.
  • The lexicalist architecture of the grammar — with rich semantic information stored in lexical entries — is supported over purely syntactic-transformational architectures.
  • The method of studying learnability — asking what representational assumptions make language learnable from the available input — is a powerful tool for discovering the structure of mental representations.
  • The new preface (2013) draws connections between these findings and Pinker's later work on irregular verbs, the language of politics, the semantics of sexual innuendo, and the metaphorical structure of everyday thought.

Key takeaway

Solving the puzzle of argument-structure acquisition requires and reveals a rich system of conceptual representation: studying what children must know to learn their language is a direct route to understanding the organization of the human mind.

The book's overall argument

  1. Chapter 1 (A Learnability Paradox) — establishes that children face a logical impossibility: they must learn which verbs can and cannot appear in argument-structure alternations without receiving negative evidence, and prior proposed solutions all fail.
  2. Chapter 2 (Constraints on Lexical Rules) — demonstrates that lexical rules are systematically constrained by verb semantics, with detailed documentation of dative, causative, locative, and passive alternations, establishing that meaning predicts syntactic behavior.
  3. Chapter 3 (Constraints and the Nature of Argument Structure) — explains why semantic constraints exist by showing that lexical rules are meaning-to-meaning mappings, and introduces universal linking rules and thematic-tier structure as the formal backbone.
  4. Chapter 4 (Possible and Actual Forms) — introduces the distinction between broad-range and narrow-range rules, and shows that negative exceptions reflect semantically conditioned narrow classes rather than arbitrary lexical idiosyncrasy.
  5. Chapter 5 (Representation) — provides the formal framework for lexicosemantic representation, arguing that verb meanings are composed from a small set of conceptual primitives (CAUSE, GO, BE, AT, etc.) in a Jackendoff-style Lexical Conceptual Structure.
  6. Chapter 6 (Structures) — maps out the full inventory of argument-structure types, derives them from the representational framework, and handles unaccusativity, resultatives, and cross-linguistic variation.
  7. Chapter 7 (Learning) — proposes the acquisition mechanism: semantic bootstrapping for initial category learning, plus semantically-guided hypothesis refinement for verb-specific argument structure, predicting the broad-then-narrow developmental trajectory.
  8. Chapter 8 (Development) — tests the predictions against the empirical record of child language: overgeneralization errors, the retreat from overgeneralization, novel-verb generalization experiments, and cross-linguistic developmental data.
  9. Chapter 9 (Conclusions) — draws out implications for the language-thought relationship (conceptual structure drives grammar, not the reverse), for linguistic theory (lexicalism), and for cognitive science (rich structured mental representations).

Common misunderstandings

Misunderstanding: Pinker claims children never overgeneralize argument structure.

The opposite is true. Pinker's theory predicts and accounts for overgeneralization errors: they arise from the productive application of broad-range rules to verbs that satisfy the broad semantic conditions but fail the narrow ones. The book's developmental account centrally depends on the existence of overgeneralization.

Misunderstanding: The book argues that all of grammar is semantic.

Pinker does not claim that syntax reduces to semantics. He claims that argument-structure alternations — a specific subset of grammatical phenomena — are semantically conditioned. Purely formal syntactic principles govern other aspects of grammar. The claim is domain-specific, not a global semantic theory of syntax.

Misunderstanding: Baker's Paradox shows that language acquisition requires innate knowledge of specific argument-structure rules.

Pinker's solution does not require innate knowledge of dative alternation, causative alternation, or any other specific lexical rule. What is innate is the system of conceptual primitives (CAUSE, GO, etc.) and linking rules — a much more abstract and general representational capacity, not a list of construction-specific rules.

Misunderstanding: Children solve Baker's Paradox by attending to input frequency (hearing "donate" rarely in the double-object frame).

Pinker argues explicitly against frequency-based accounts. The retreat from overgeneralization tracks semantic class membership, not token frequency; and children maintain correct restrictions on verbs they hear rarely, which a pure frequency account cannot explain.

Misunderstanding: The book is exclusively about English.

While the empirical evidence is heaviest from English, the theoretical framework is explicitly cross-linguistic. Pinker discusses causative alternations in many languages, the universal linking rules, and cross-linguistic developmental data throughout.

Central paradox / key insight

The book's central paradox is that children learn not to say certain things — they learn restrictions on generalization — yet they do this without ever being told what is wrong. The standard assumption is that you can only learn a prohibition if someone tells you about it (negative evidence), and children do not reliably receive such feedback.

The key insight resolving the paradox is that the restriction is not, at bottom, a grammatical rule at all. It is a semantic fact. "Donate them a book" is not ungrammatical for some arbitrary syntactic reason; it is unacceptable because the double-object construction means something (caused direct possession) that "donate" does not entail (donations are institutionally mediated transfers). The child who correctly understands what "donate" means will not put it in the double-object construction, for the same reason an adult who knows that pouring does not entail holistic filling will not say "pour the glass with water."

The child does not need to be told that "donate them a book" is wrong; she needs only to know what "donate" means.

This reframes the entire learnability problem: the puzzle of learning argument structure without negative evidence dissolves once we recognize that verb meaning encodes the relevant constraints. The acquisition task is not memorizing a list of permitted and forbidden constructions; it is acquiring accurate semantic representations of verb meanings.

Important concepts

Argument structure

The specification, stored in a verb's lexical entry, of how many participants the verb requires, what semantic role each plays, and which grammatical position each occupies. Distinct from surface syntax: the same argument structure can be realized in different syntactic frames (dative alternation).

Baker's Paradox

The learnability problem posed by argument-structure alternations: children must learn which verbs alternate and which do not, without negative evidence, yet they also must generalize productively. Named after linguist Mark C. Baker.

Lexical Conceptual Structure (LCS)

A formal notation, developed by Ray Jackendoff, for representing verb meanings using a small inventory of conceptual primitives (CAUSE, GO, BE, STAY, ACT, AT, TO, FROM, VIA, FOR) arranged into hierarchical event representations. Pinker adopts LCS as the representational framework for verb meanings.

Linking rules

Principles that map thematic roles (positions in the conceptual structure) to syntactic positions (subject, object, oblique). Universal linking rules include: Agent → Subject, Patient/Theme → Object, Goal → Indirect Object or oblique. Language-specific rules can modify these defaults.

Thematic roles

Semantic participant roles: Agent (volitional causer), Patient (entity undergoing change), Theme (entity in motion or the content of a transfer), Goal (endpoint of transfer or motion), Source (starting point), Experiencer (sentient participant in a mental state). In Pinker's framework, roles are derivative — read off from a constituent's position in the conceptual structure rather than stipulated as features.

Thematic tiers

The division of event structure into two tiers: the action tier (who acts on whom: agent and patient) and the thematic tier (what goes where: theme, goal, source). This structure explains asymmetries in alternation behavior and in the linking of psychological verbs.

Broad-range rules

Lexical rules that apply to semantically general verb classes — the productive defaults of argument-structure alternations. Example: the broad-range dative rule applies to all verbs that denote caused possession transfer.

Narrow-range rules

Lexical rules restricted to semantically specific sub-classes of verbs. These implement the exceptions to broad-range rules and account for negative exceptions (verbs that are semantically eligible at the broad level but excluded at the narrow level). The acquisition of narrow-range rules drives the retreat from overgeneralization.

Semantic bootstrapping

The hypothesis (Pinker 1984) that children use pre-linguistic semantic knowledge — categories like object, action, property — to identify the referents of the grammatical categories noun, verb, adjective, and thereby gain initial leverage on syntactic structure from positive evidence alone.

Negative exceptions

Verbs that appear to satisfy the semantic conditions of an alternation at the broad level yet fail to alternate. Explained by narrow-range conditions that identify a more specific semantic property they lack.

The locative alternation

The alternation between a theme-object construction (load hay onto the wagon) and a location-object construction (load the wagon with hay). The location-object form implies holistic change of state of the goal (the wagon ends up fully loaded); the theme-object form does not. This semantic difference is the constraint on the alternation.

Unaccusativity

The hypothesis that certain intransitive verbs (verbs of motion and change of state: "arrive," "fall," "break") underlyingly place their single argument in object position, while others (verbs of agentive activity: "run," "bark," "sneeze") place their argument in subject position. The distinction correlates directly with semantic properties encoded in the conceptual structure.

Primary book and edition information

Background and overview

Baker's Paradox and the logical problem of language acquisition

Jackendoff's Lexical Conceptual Structure (the representational framework)

  • Jackendoff, Ray. Semantics and Cognition. Cambridge, MA: MIT Press, 1983.
  • Jackendoff, Ray. Semantic Structures. Cambridge, MA: MIT Press, 1990.

Semantic bootstrapping and language acquisition theory

  • Pinker, Steven. Language Learnability and Language Development. Cambridge, MA: Harvard University Press, 1984.

Reviews and academic commentary

Additional resources

Send feedback

Optional. We'll only use this if you want a reply.