Certain syntactic structures are islands; i.e, they are effectively isolated from the rest of the sentence they are in, for the purpose of various kinds of rules, typically those which involve licensing the appearance of sentence elements in unexpected places in the structure, while suppressing them in their "original" (i.e, expected) position.
Viewed in a temporal perspective on a derivation, this produces a metaphor of moving them from one place to another. Hence, these rules are called "Movement" rules (actually, linguists rarely use scare quotes around movement; we know it's just a metaphor. But I retain them here for the uninitiated). Prototypic examples are Wh-Question Formation (where the "Movement" is of the wh- interrogative pronoun to the beginning of the sentence) and Relative Clause Formation (which moves the relative pronoun to the beginning of the clause).
There are basically two types of such "Movement" rules:
This kind of rule is sometimes called a cyclic rule. Cyclic rules are governed by the predicate in the relevant clause, and often have lexical exceptions. Equi-NP-Deletion and Subject-Raising are other examples of cyclic rules. Islands are not a relevant concept to cyclic rules, because the scope of cyclic rules is already limited.
Some examples:
What language_i does Bill want to try
(Wh-Question Formation)
to begin to learn
to speak ____i this year?
Mary_i is the person_i who_i Sue
thinks you said that
I heard that Fran claimed that Bill
likes ____i.
(Relative Clause Formation)
The metaphor of "island" comes from the fact that Type 2 "Movement" rules cannot move a constituent grammatically over an island boundary. Application of these rules over island boundaries produces ungrammaticality.
Islands are described in terms of constraints on Type 2 "Movement" rules, each constraint describing an island category. Below are a few examples of such island constraints; the complete dissertation can be downloaded.
Coordinate Structures are those joined with a Coordinating Conjunction, like and, or, but, etc. An element from one conjunct cannot be moved out of that structure:
I.e, coordinate structures are islands.
There are some interesting
exceptions to this, such as
A Complex NP is one with a noun head and a modifying
clause.
There
are two types:
[ the dog_i [(which_i) Mary saw ___i] ] NP S S NP
[ the report_i [that Quayle sleeps with a Teddy Bear] ] NP S_i S_i NP
Violations of Ross Constraints are very ungrammatical. Most people never encounter them. We appear to formulate our discourse to avoid them. Occasionally, we get in a bind and see one looming at the end of the clause, and have to do something quick. What we do is often illuminating about the relative importance of syntactic rules.
For instance, consider the following:
Neither sentence is terrifically grammatical, but the first seems more appropriate (and common as a type) than the second, though the last word in the first sentence still feels strange. The ordinary rule of relative clause formation operating on the last clause should result in its deletion at the end of the clause (and thus the sentence). However, it appears inside another relative, an island, and is thus safe from such "Movement" by the Complex NP Constraint.
Sentences like the first one are generated when, at the last minute, the speaker realizes what is going to result, and cancels the deletion, substituting an alternative relative-formation rule (called a Resumptive Pronoun in the trade), which merely pronominalizes the coreferential NP, instead of deleting it in the object position.
This is not the way English forms its relative clauses (though other languages use it frequently, e.g, Hebrew), and the sentence is thus ungrammatical. But this turns out to be a venial syntactic sin by comparison with a violation of a Ross constraint, which typically produces extreme ungrammaticality.