Syntactic Structures appeared in 1957 from the Dutch publisher Mouton — a slim book, about ninety pages of text. Its author was twenty-eight: Noam Chomsky (1928– ), newly a teacher in MIT’s Department of Modern Languages. Born in Philadelphia, he studied under Zellig Harris at the University of Pennsylvania, and, by his own account in the preface, spent 1951–55 as a Junior Fellow of Harvard’s Society of Fellows. The book reworks material largely written during those years into lecture notes for an introductory course at MIT. It appeared as No. 4 in Mouton’s “Janua Linguarum, Series Minor.” A story often repeated is that the manuscript was turned down by more than one publisher before landing at Mouton — though the preface itself says nothing about this.
It is a small book that traveled far. Reprinted for decades (the fourteenth printing, which this piece follows, dates to 1985) and widely described as the opening move of a “generative revolution” in linguistics.
1. What the Book Sets Out to Do
The preface states the aim: to build “a formalized general theory of linguistic structure” and explore its foundations. Syntax is the study of the principles by which sentences are constructed in a given language, and its goal, the Introduction says, is “the construction of a grammar that can be viewed as a device… for producing the sentences of the language under analysis.”
The key term is “linguistic level.” Phonemics, morphology, phrase structure — each is a set of descriptive devices available for building grammars. Chomsky’s project reaches beyond the description of English alone to ask, in purely formal terms, what levels a theory of grammar must contain.
2. “Grammatical” Is Neither “Meaningful” Nor “Frequent”
Chapter 2 opens with a strict definition: a language is “a set (finite or infinite) of sentences, each finite in length and constructed out of a finite set of elements.” The definition sounds dry, but the conclusions drawn from it are sharp.
First, “grammatical” cannot be identified with any particular observed corpus. Every grammar projects a finite, somewhat accidental body of observed utterances onto a (presumably infinite) set of grammatical ones — mirroring a speaker who, from finite experience, produces and understands indefinitely many new sentences.
Second — and this is the book’s most famous argument — “grammatical” cannot be identified with “meaningful.”
Colorless green ideas sleep furiously. Furiously sleep ideas green colorless.
Both are nonsense, but any English speaker recognizes only the first as grammatical. Likewise, “have you a book on modern music?” and “the book seems interesting” are grammatical while “read you a book on modern music?” and “the child seems sleeping” are not — and there is no semantic reason for the difference. This second puzzle, about have and be, is quietly solved later in the book (§7) as a byproduct of a rule built to explain something else entirely. It is a planted clue.
Chomsky further argues that “grammatical” cannot be identified with “a high order of statistical approximation to English” either. Neither (1) nor (2) has likely ever occurred in an English discourse, so a purely statistical model would rank them as equally remote from English. Yet speakers read (1) with ordinary sentence intonation, recall it easily, and reject (2). Grammaticalness reduces to neither semantics nor statistics: “grammar is autonomous and independent of meaning.”
3. Why the Finite-State Model Fails
How, then, should the set of grammatical sentences be generated? Chomsky first examines a finite-state (Markov process) model drawn from contemporary communication theory: a machine with finitely many “states,” producing one word at each transition, until it reaches a final state. Drawn as a “state diagram,” such a machine — with loops added — can generate “the old man comes,” “the old old man comes,” “the old old old man comes,” and so on forever.
The model is attractive. But Chomsky proves that English is not a finite-state language — not merely hard to describe this way, but impossible to describe this way, even in principle. The key is a “mirror-image” style of dependency. English has constructions such as
- “If S₁, then S₂”
- “Either S₃, or S₄”
- “The man who said that S₅ is arriving today”
where “if” demands “then,” “either” demands “or,” and “man” demands “is” rather than “are” — a dependency between words on either side of the comma. And between the two dependent words, an entire further sentence of the same kind can be embedded. Put another sentence from (11) in the S₅ slot, and another inside that one, and so on: nesting with no fixed limit. No finite-state machine, however large, can handle this in principle — for the same reason that no finite-state grammar can generate the artificial language “ab, aabb, aaabbb, …,” where the count of a’s must always match the count of b’s.
4. Phrase Structure: Rewrite Rules, a Stronger Device
Chomsky next examines a formalization of immediate constituent analysis, then the mainstream method in structural linguistics: rules that rewrite one symbol as others — “Sentence → NP + VP,” “NP → T + N,” “VP → Verb + NP,” and so on, a system [Σ, F]. Applying such rules in sequence derives a terminal string such as “the man hit the ball,” and the derivation is summarized by a tree diagram — a phrase marker.
This model is clearly more powerful than the finite-state one — it can generate languages like “ab, aabb, …” — but Chomsky shows with three examples that it still falls short of English.
- The conjunction test. From “the scene of the movie was in Chicago” and “the scene of the play was in Chicago” we can form “the scene of the movie and of the play was in Chicago,” but conjoining across constituent boundaries generally fails (“the liner sailed down the and tugboat chugged up the river” is not a sentence). Stated in terms of “conjoin only like constituents,” the rule is dramatically simple — but applying it requires knowing how a string was derived, its history, not merely its present shape. A rewrite rule can only see the string it applies to.
- Discontinuous elements. A form like “has been reading” is built from have…en and be…ing — two discontinuous pieces that only cohere once something is inserted between them. This resists a straightforward rewrite-rule treatment.
- Active and passive. Deriving “sincerity is admired by John” from “John admires sincerity” requires swapping subject and object, inserting “by,” and adding be+en to the auxiliary. One such rule (the passive) automatically explains why “sincerity admires John” is odd, why be+en only follows a transitive verb, and more — but again, it doesn’t fit inside a single rewrite rule.
5. Transformations: The Kernel, and a Chain of Rules
This is where Chomsky introduces the book’s central proposal: the transformation — a more powerful rule that rewrites a string, along with its constituent structure, into a new one.
From here the grammar takes on three layers.
A sentence produced using only obligatory transformations is a kernel sentence. Every other sentence is derived from a kernel string by one or more optional transformations — the rule that forms negatives (T-not), the rule that forms questions (T-q), and the rule that inserts “do” as the “bearer” of a tense or agreement affix left stranded with nothing to attach to. That last rule is where the “do” in “John doesn’t come” or “did they arrive” comes from — and the very same rule serves negation, questions, and sentences like “so do I” alike.
And this framework also resolves the puzzle about have from §2 — why “have you a book…?” is grammatical while “read you a book…?” is not. Because have happens to share a shape with the auxiliary have, it admits two structural analyses that other verbs do not. A rule required for entirely independent reasons quietly settles an unrelated puzzle — this is what Chomsky means by “simplification,” and it is the book’s central demonstration.
6. Explanatory Power: Where Does Ambiguity Live?
Chomsky proposes a further test of a grammar’s adequacy: if a string of sounds can be analyzed in more than one way, it should in fact be ambiguous (and conversely, real ambiguity should show up as multiple analyses somewhere in the grammar). He calls this constructional homonymity.
“The shooting of the hunters” can mean either that the hunters did the shooting, or that they were shot. At the level of phrase structure both readings look identical — “the - V+ing - of + NP.” But at the transformational level they diverge sharply: one derives from the kernel sentence “the hunters shoot,” the other from “they shoot the hunters,” by two different nominalizing transformations. The ambiguity is the difference in transformational origin.
The same logic applies to “the picture was painted by a new technique” versus “the picture was painted by a real artist” — identical in phrase structure, understood quite differently. The first keeps the instrumental “by” of an active sentence; the second is the agentive “by” that the passive transformation supplies. Their transformational histories differ.
The test also runs the other way: “John played tennis” and “my friend likes music” are entirely distinct at the phonemic and morphemic levels, yet identical in phrase structure — “NP - Verb - NP” — which is exactly why they feel like “the same kind of sentence.” The claim that “to understand a sentence… it is necessary to reconstruct its analysis on each linguistic level” shows that the syntax of this book is not merely a test for grammaticality; it is a hypothesis about understanding itself.
7. Grammar Doesn’t Rest on Meaning — But Isn’t Unrelated to It Either
In §9, Chomsky reinforces the argument from the other side. He takes up claims then common — that morphemes are “the smallest meaning-bearing units,” that the subject-verb relation corresponds to “actor-action,” and so on — and shows each fails by counterexample. “Bachelor” and “unmarried man” are synonyms with different sounds; “bank” (of a river) and “bank” (for savings) share a sound but not a meaning — “same sound iff same meaning” fails in both directions. In “John received a letter” or “the fighting stopped,” the subject is no “actor.” Even active and passive pairs are not always synonymous: “everyone in the room knows at least two languages” and its passive can differ in truth value, because the scope of the quantifier shifts.
None of this means form and meaning have nothing to do with each other — “undeniable, though only imperfect correspondences hold between formal and semantic features in language.” The correspondence simply isn’t precise enough to found grammar on. The discussion of structural ambiguity in §8 is, in Chomsky’s view, the right way to pursue that correspondence. Syntax should be built up autonomously from semantics, and only then is it worth studying how the two relate — a two-step stance often flattened, in later retellings, into “Chomsky ignored meaning,” which is not quite what the book says.
8. What Kind of Theory? The Modest Goal of an “Evaluation Procedure”
In §6, Chomsky draws a methodological distinction that turns out to matter enormously. What should a linguistic theory be asked to provide?
- A discovery procedure — a mechanical method for constructing a grammar from a corpus.
- A decision procedure — a mechanical method for judging whether a proposed grammar is the correct one.
- An evaluation procedure — a method for judging which of two proposed grammars is the better one.
Chomsky rejects the first goal — which much of contemporary structural linguistics (and his own earlier work) had implicitly pursued — and argues that a theory need only meet the third, weakest requirement. How a linguist happens to arrive at a grammar (intuition, guesswork, whatever works) is not the theory’s business; the theory’s job is only to compare candidate grammars by a standard of simplicity and say which is better. This modest goal is the methodological floor the whole book stands on, and it is what later licensed generative grammar to take grammatical intuition itself as a starting point.
Summary
| Question | Syntactic Structures’ answer |
|---|---|
| What is “grammatical”? | Neither meaningful nor statistically likely. “Colorless green ideas sleep furiously” is grammatical; its reversed word order is not |
| The finite-state model | Fails — it cannot, even in principle, handle nested dependencies like “if… then” |
| The phrase-structure model | Stronger than finite-state, but still not enough — conjunction, discontinuous auxiliaries, and passive inversion resist simple rewrite rules |
| The fix | Transformations. A three-layer grammar: kernel (obligatory transformations only) plus chains of optional transformations |
| Ambiguity | The formal correlate of “being understood in more than one way” is being analyzed in more than one way at some level |
| Syntax and semantics | Syntax should be built autonomously, but the correspondence between form and meaning is worth studying on its own |
| The theory’s goal | Not a discovery procedure — an evaluation procedure. It is enough to compare candidate grammars by simplicity |
Why It’s Still Read
Syntactic Structures reshaped the second half of twentieth-century linguistics — the opening move of what came to be called the “Chomskyan revolution.” The book itself is only the first systematic statement of transformational grammar; Chomsky went on to recast the framework in Aspects of the Theory of Syntax (1965) with “deep structure” and “surface structure,” and later to the broader programs of “universal grammar” and generative grammar more widely. In 1959 his review of B. F. Skinner’s Verbal Behavior (1957) attacked behaviorist accounts of language learning — a separate piece of work from the same period, often credited with real influence on the rise of cognitive science.
The book names no ancestors, but in 1966’s Cartesian Linguistics Chomsky would claim Arnauld and Lancelot’s General and Rational Grammar as one — the idea that the structure of mind determines the structure of grammar. The phrase “infinite use of finite means” has a precedent in Wilhelm von Humboldt, whom Chomsky himself quotes. The “statistical approximation” line the book firmly rejects would later resurface, in a different form, as corpus linguistics and statistical natural-language processing — and the tension between the two outlooks continues in debates about today’s AI language models.
Opinion remains divided. Generative grammar’s later turns — deep structure, binding theory, minimalism — have been argued over within linguistics ever since, and cognitive and functionalist linguistics start from different premises altogether. But the way the book poses its question — how to generate the set of grammatical sentences from a finite set of rules — and its methodological choice to make simplicity the arbiter between theories, remain a shared starting point for supporters and critics alike.
.jpg){kind=link}