Island Constraints

How an island blocks extraction

Wh-movement is normally permissive. You can ask "What did Mary buy?" from a one-clause sentence; you can ask "What does John think Mary bought?" from a two-clause one; you can ask "What does John think Bill said Mary bought?" from a three-clause one. Long-distance extraction is fine, even across many clause boundaries, as long as each step is local enough.

But certain configurations refuse to let the wh-element escape. Compare the grammatical (1) with the ungrammatical (2):

(1) What does John know that Mary bought ___?    OK
(2) *What does John know the claim that Mary bought ___?    *

In (1), "that Mary bought" is a finite clause complement of "know". Wh-movement can target the embedded SpecCP as an intermediate stop and then continue up. In (2), the same string is wrapped in the noun "the claim", making the whole thing a complex NP. Now the wh-word is trapped: the noun phrase is an island. No matter how the listener tries to interpret the question, the sentence falls apart.

This is the core puzzle Ross identified. The intuition that "long extraction is hard" doesn't explain it — the ungrammatical (2) is no longer than the grammatical (1) — and a purely semantic explanation fails because the meaning is recoverable. The block must be syntactic: certain phrase-structure configurations seal their interior against movement.

Ross's original islands (1967)

Ross's dissertation is a 600-page catalogue of the configurations. The four most cited:

Island	Configuration	Grammatical baseline	Island violation
Complex NP (relative clause)	NP containing a relative clause	What did John meet the man who bought?	*What did John meet [the man who bought ___]?
Complex NP (noun complement)	NP containing a finite clause complement	What did John believe that Mary bought?	*What did John believe [the claim that Mary bought ___]?
Wh-island	Embedded clause with filled SpecCP	What did Mary say John bought?	?What did Mary wonder [whether John bought ___]?
Sentential subject	Clause functioning as a subject	—	*What is [that John bought ___] surprising?
Coordinate structure	One conjunct of a coordination	—	*What did John buy [a hat and ___]?
Adjunct (Huang 1982)	Adjunct clause	—	*What did John leave [before buying ___]?
Subject (Huang 1982)	Subject phrase, including derived subjects	—	*Who did [a picture of ___] surprise Mary?
Left Branch (Ross 1967)	Pre-nominal modifier	—	*Whose did you read ___ book?

The Coordinate Structure Constraint has a famous exception, the "across-the-board" movement: What did John buy and Mary sell? is grammatical because the gap appears in both conjuncts. Williams (1978) gave the canonical analysis.

Worked example: "What does John know that Mary bought?" vs the violation

Take the legal example first:

What does John know that Mary bought ___?

The derivation:

1. Base: John knows that Mary bought what.
2. Move "what" to embedded SpecCP: John knows [_CP what (that) Mary bought ___].
3. Move "what" from embedded SpecCP to matrix SpecCP: [_CP What does John know [_CP ___ that Mary bought ___]].

Each step is local — between adjacent CP edges, no more than one bounding node crossed. Subjacency is satisfied, and the sentence is fine.

Now the violation:

*What does John know the claim that Mary bought ___?

The same wh-element starts in the same place, but its containing structure is now the claim that Mary bought what — a noun phrase with a clause-complement modifier. To leave the embedded clause, the wh-word must escape both the inner CP and the dominating NP simultaneously. There is no intermediate landing site within the NP for it to stop at. Subjacency (Chomsky 1973) is violated: the movement crosses both an S and an NP node in a single step. The result is sharply ungrammatical for English speakers — the canonical complex NP island violation.

Strong vs weak islands

Cinque (1990, Types of A-bar Dependencies) refined the picture. Some islands block all extraction; others only block certain kinds:

Island type	Argument extraction (who/what)	Adjunct extraction (how/why/how-many)	Class
Complex NP	Blocked	Blocked	Strong
Adjunct	Blocked	Blocked	Strong
Subject	Blocked	Blocked	Strong
Wh-island	Marginal: ?Who do you wonder whether Mary saw?	Blocked: *Why do you wonder whether Mary left?	Weak
Factive island	Marginal	Blocked	Weak
Negative island	Marginal	Blocked	Weak

The argument/adjunct asymmetry is the diagnostic for weakness. Adjuncts (how, why, how-much) carry less referential content and are more easily blocked by intervening operators; arguments (who, what) carry more, and can sometimes squeeze past.

Parametric variation

Most strong islands hold across languages, but the details vary:

• Italian (Rizzi 1982). Allows extraction from some wh-islands that English blocks: Tuo fratello, a cui mi domando che storie abbiano raccontato — "Your brother, to whom I wonder what stories they told" is fine. Rizzi attributed this to Italian taking IP rather than CP as a bounding node.
• Swedish (Engdahl 1985). Allows extraction from some relative clauses: Det huset känner jag en man som har målat — "That house I know a man who has painted". The acceptable cases involve definite, presuppositional structures.
• Hebrew, Irish, Welsh. Use resumptive pronouns to bypass islands: The man that Mary saw him — the pronoun fills the gap and avoids the violation. Resumption is the most common cross-linguistic island-rescue strategy.
• Mandarin and Japanese. Show island effects on in-situ wh-phrases too, evidence for covert wh-movement at LF (Huang 1982, Watanabe 1992).

Theoretical accounts

Account	Author / year	Key idea
Subjacency	Chomsky 1973	Movement cannot cross two bounding nodes (NP, S) in one step
ECP	Chomsky 1981	Empty Category Principle: traces must be properly governed
Barriers	Chomsky 1986	Movement-blocking phrases (barriers) defined relative to L-marking
Relativised Minimality	Rizzi 1990	Movement of α blocked by intervening β of the same feature class
Phases	Chomsky 2000, 2001	vP and CP are phases; their interiors become opaque after Spell-Out
Processing	Hofmeister & Sag 2010	Some islands reduce to memory load and integration cost, not grammar

The grammar-vs-processing debate has been particularly active since the 2000s. Sprouse, Wagers, and Phillips (2012) used factorial designs to argue that at least some islands (whether-islands, complex-NP) cannot be explained by working-memory load alone — extraction is harder than predicted by the additive cost of length plus complexity.

Variants and edge cases

• Parasitic gaps. A second gap inside an adjunct island becomes acceptable when there is a primary gap outside it: Which paper did you file ___ without reading ___? The licit gap "parasitises" on the legitimate one (Engdahl 1983).
• Across-the-board movement. Coordinate structures permit extraction if all conjuncts share the gap: What did John buy ___ and Mary sell ___?
• Sluicing as island repair. Ross (1969) noted that sluicing — John bought something but I don't know what — can repair some island violations, suggesting that the violation is a PF rather than LF effect.
• Relativisation tracks wh-movement. The same islands constrain relative-clause formation; resumptive pronouns are the main escape hatch in languages that allow them.
• Negative islands. "How didn't John fix it?" is degraded compared with the positive counterpart; negation creates a weak island for adjuncts.

Common pitfalls

• Treating every long extraction as island-violating. Long-distance wh-movement across multiple complement clauses is fine; islands are specific configurations, not a function of length.
• Confusing complex NP with relative clause. The Complex NP Constraint covers two distinct cases: noun-complement clauses ("the claim that ...") and relative clauses ("the man who ..."). Both are islands, but the explanations sometimes diverge.
• Assuming Italian-style wh-island violations exist in English. Italian and other Romance languages allow some wh-island extraction; English does not. Translation problems arise from the difference.
• Ignoring resumption. Hebrew and Irish embed pronouns where English would have a gap, and the result is grammatical. This is not an English-style violation, but a different repair strategy.
• Reducing all island effects to processing. Hofmeister and Sag (2010) make a real point, but Sprouse and colleagues have shown that working memory alone undergenerates the data. The grammar-only and processing-only positions are both extreme.