1. Introduction

The fundamental work on islands is Ross 1967/1986. Ross distinguished between transformations that move a constituent and leave behind a pro-form (i.e., resumptive pronoun) and ones that "chop" a constituent and move it without leaving anything behind. An example of the former is left dislocation (that is, on Ross's analysis, left dislocation is not base generated). An example of chopping is topicalization. Of the two, only chopping operations are constrained by islands. Islands will be enclosed in angled brackets:

Diagnosing islands has become more complex since Ross, because it is assumed that there exist invisible pro-forms (empty resumptive pronouns), in addition to overt ones. I will use "gap" as a pretheoretical term that covers both traces and empty pronouns.

Islands come in two main varieties: "weak" and "strong". Note that the weak/strong qualification does not correspond to degrees of ungrammaticality.

The focus of this chapter is weak islands (WI). Strong islands are considered only to set the stage for the discussion of WIs.

Up until the late Eighties, nothing much beyond wh-islands had been thought to be a weak (selective) island. Beginning with Relativized Minimality, however, an ever growing range of weak islands have been recognized. Thus, theories of weak islands have mushroomed, each coming with a significant set of new data and important new connections to other domains. The review of weak islands will emphasize the correspondence between data sets and theories. Less attention will be given to proposals that primarily recast the theoretical account of some narrow range of data.

2. Absolute versus selective islands

Although the notion of strong versus weak islands as defined in (3) and (4) seems like a simple descriptive one, it is not. Consider the following examples involving wh-islands and adjunct islands:

These two sets might suggest that both wh-clauses and adjunct clauses are strong islands: the extraction of three different types of phrases was attempted and failed. On the other hand, it is easy to construct other examples in which some, but not other, phrases can extract. The percentage sign % indicates that there is variation among the speakers of English:

The difference between the two sets involves Tense in the first place. While tensed clauses are themselves not islands, cf.

we see that the presence of Tense may considerably strengthen other islands. This fact must be systematically controlled for in the evaluation of various islands. (Chomsky 1986 briefly touches upon the role of Tense. Cinque 1990 uses it as an important test, and Manzini 1992 proposes a theory of its behavior.)

Thus, the specific constructions in (5)-(6) may preclude the presence of any gaps, but this in itself does not mean that wh-complements and adjuncts per se are totally impenetrable.

Shall we conclude now that wh-clauses and adjuncts per se are merely weak islands? Note a fine distinction: while adverb extraction is blocked in both, the infinitival whether-clause allows the extraction of both DPs and PPs (at least for some speakers), while the gerundival adjunct clause allows only DP-extraction.

The interpretation of this situation is a theoretical, rather than descriptive, matter. Making the implicit assumption that the PP versus DP facts are highly consistent in each island type, Cinque 1990 argues that this difference warrants radically different analyses, because a DP gap may be an empty resumptive pronoun but a PP gap cannot be. Wh-islands are indeed weak (selective); adjunct-islands on the other hand are strong (absolute), and the DP-gap that they may contain is an empty resumptive pronoun.

Cinque's distinction will serve as a basic organizing principle in this chapter, because it is typically taken for granted in the weak island literature: theories of weak islands do not try to account for the islands that Cinque classifies as strong.

This is not the only possible way to cut the cake. For instance, Postal 1997,1998, who uses the terms locked (absolute) versus unlocked (selective) island, does not attribute this much significance to PP versus DP. Following Ross, he assumes that extraction from unlocked islands always involves empty resumptive pronouns and subsumes Cinque's weak islands under this rubric.

A note on the data. Question formation, relativization, and topicalization are all instances of wh-movement. Because the bulk of the literature to be surveyed focuses on examples involving question formation, this chapter follows the same practice, unless the discussion of a particular work requires us to do otherwise.

3. Strong islands

3.1 Classical strong islands

While recent literature has identified a host of weak islands, the classical inventory, with the exception of certain wh-islands, consists of strong ones. What follows is a list, sometimes interrupted by comments.

Complex DP (with relative clause):

Complex DP (with complement clause):

But just as it was possible to arrange for a DP-gap in adjuncts by removing Tense, the same is possible with Complex DPs if the definiteness of the "head" is removed. (Pollard & Sag (1994, vol.2) assume, for this reason, that there is no need for a Complex DP constraint.) PP-gaps remain practically excluded.

In distinction to Tense, Definiteness may create an island by itself (Fiengo and Higginbotham 1981, Manzini 1992):

Definites:

See also the Case of Extraction from NP and complex NPs.

Subjects:

Adjuncts (see the discussion of Tense above):

Authier 1991 argued that adjuncts can contain CP gaps besides DP gaps; for an alternative analysis, see Postal 1994:

Coordinate structures, unless extraction is across-the-board:

However, DP-gaps are acceptable under various circumstances (see Postal 1998 for detailed discussion):

See also the Case of ATB phenomena. <<#10>>

The status of tensed constituent wh-complements seems to vary cross-linguistically. Some speakers of English reject extractions from them entirely, whereas others find them tolerable. Likewise, they are rejected by speakers of Dutch, whereas they are acceptable in Scandinavian languages (see Engdahl and Ejerhed 1982) and in Hungarian. (Hungarian has object pro-drop, but only for singulars, thus we know that the gap in (c) is not a dropped pronoun.) This variation is not well-understood (though see Bayer 1984, 1996:6.6.4).

Tensed constituent wh-complements (in some languages/dialects)

See also the Case of Extraction from wh-islands. <<#??>>

Left branches:
As (35) shows, it is often difficult to extract from the left branch of a structure (Ross 1967/1986):

But comparable examples are perfect in other languages:

Whether all the counterexamples are well-understood, the Left Branch condition seems like a less plausible generalization than the others. See Corver (1990).

Finally, some islands constrain left-extraction but not Right Node Raising, indicating that the latter may not be an extraction (Bouma 1987, Postal 1998):

We will not discuss the relation between wh-in-situ and islands, see the Case of Wh-in-situ.

3.2 Types of explanation

The standard explanation of the islandhood of Complex DPs, Subjects, Wh-complements, and Left Branches is in terms of Subjacency: they are out because movement crosses more than one bounding node (barrier), see 3.4 and 7.1.

Adjunct islands are standardly explained by Huang's 1982 Condition on Extraction Domains, hence by the Empty Category Principle: an extraction domain needs to be properly governed. Likewise, Pesetsky 1982 subsumes the Coordinate Structure Constraint under the path containment version of the ECP, see 7.1..

Manzini 1992 is the only attempt to unify the effect of Definiteness and Tense with other locality phenomena: D and T block dependencies based on Case-addresses, which, in her theory, DPs otherwise rely on to escape from islands.

3.3 Ways of salvaging strong island violations

As has been mentioned, a strong island violation can sometimes be salvaged by resorting to the resumptive pronoun strategy (whether the pronoun is overt or empty).

Second, the gap in the island may be parasitic on a well-behaved trace of movement, e.g.:

See Engdahl 1983, Kayne 1983, Chomsky 1982,1986, Cinque 1990, among others.

Third, the island itself may be pied piped, overtly or at Logical Form:

Pied piping at LF has been proposed for apparent violations of Complex DPs in wh-in-situ languages in Pesetsky 1987, and for apparent violations of a single adjunct island in Cinque 1990, for instance.

These matters are not discussed here further; see the Cases of Resumptive pronouns, Parasitic gaps, and Pied piping.

3.4 More on Subjacency

Subjacency is classically understood as a condition on movement and requires that movement not cross more than one bounding node. Bounding nodes were originally defined as a list: NP and S (=DP and IP) in English (Chomsky 1973, 1977), NP and S' (=DP and CP) in Italian (Rizzi 1978).

Chomsky 1986 redefines bounding nodes as barriers. An XP is a blocking category for α iff it is not theta-marked by a sister lexical head and dominates α. β is a barrier for α iff (i) β is a blocking category for α but not IP, or (ii) β is the first XP that dominates a blocking category for α. Chain-formation (which takes the place of movement) requires 1-subjacency: no more than one barrier may be crossed. (For an excellent introduction to "Barriers" as well as other theories of locality, see Roberts 1997.)

Adopting Rizzi's 1990 Relativized Minimality (the pertinent aspects of which will be discussed when we turn to Weak Islands), Cinque 1990 proposes two important changes in the understanding of strong islands. (i) Strong islands constrain binding chains, not necessarily movement, and (ii) Not only government but also binding chains require 0-subjacency (no barrier may be crossed).

3.4.1 Cinque 1990

We take a closer look at some aspects of Cinque's proposal, because they are fundamental in defining the division of labor between theories of strong and weak islands, as assumed in most of the literature.

Following Obenauer (1984/85), Cinque observes that both parasitic gaps and gaps inside strong islands are restricted to the category DP and takes this to indicate that these gaps are not variables, but A-bar bound empty pronominals: pro. This pro is unmoved in syntax but must move at LF (like some kind of abstract wh-phrase). But instead of moving on its own, it pied pipes the minimal island it is contained in. This accounts for the fact that, if all goes well, one strong island can be evaded.

Now recall the contrast between tensed and gerundival adjunct islands:

Bona fide cases of pied piping are also constrained by Tense (Nanni and Stillings 1978):

Cinque stipulates that Tensed Inflection weakly blocks the upward percolation of features. Thus, even one island cannot be pied piped by pro if it is tensed.

Cinque predicts that at most one island can be evaded: strong islands cannot be compounded:

The reason is that the movement of the pied piped domain continues to be sensitive to islands (can at best be parasitic on overt movement). Postal (1998) notes, though, that an adjunct island can be compounded with a complex DP (see further in 3.4.3):

To summarize, the observation that certain islands only contain DP gaps leads Cinque to a theory according to which these islands are strong, and the DP gap is not a trace of movement but A-bar bound pro. This contrasts with Weak Islands, which may also contain PP gaps. The latter must be traces since, according to Cinque, human languages generally lack resumptive pronouns of category PP. The behavior of weak islands is explained by considerations that do not apply to strong islands, in Cinque's work and in the many other proposals this chapter will review. Thus, the DP versus PP distinction carries a great burden in deciding which data are to be accounted for by each theory.

3.4.2 Some intriguing similarities between SIs and WIs

While the division of labor so determined has proven very useful in the literature, it may be worth pointing out some intriguing similarities between the phrases that may or may not escape from the two types of islands.

First, there is dialectal variation among speakers of English regarding the acceptability of PP-extraction out of wh-islands (one of the weak islands). Unfortunately, no systematic empirical study of this variation exists, to my knowledge, wherefore it is difficult to assess its significance for Cinque's theory.

Second, consider the following contrast:

Wh-phrase associated with pro in a strong island:

Wh-phrase associated with a variable in a weak island:

In both cases, the wh-phrase needs to be referential in some sense. But the reasons are different. In the case of strong islands, the wh-phrase needs to be of the kind that resumptive pronouns tend to associate with; in the case of weak islands, it has to carry a referential index and corefer with its variable (see in 7.2). Cinque 1990 characterizes the suitable wh-phrases rather similarly in the two cases, but neither this book nor any subsequent work known to me addresses the question whether the two requirements are exactly identical. Cinque (pers. comm., 1998) has kindly suggested the following contrast between a weak and a strong island,

but the exact generalization remains an open question. One proposal that attempts to unify weak and strong islands is Starke 2001.

3.4.3 Postal 1998

Perlmutter 1972 proposed that all extractions leave invisible resumptive pronouns, whether they are from islands or not. Obenauer (1984/85) proposed that all extractions from islands involve empty resumptive pronouns. Cinque (1990) developed a specific resumptive pronoun theory of apparent strong island violations. Postal 1998 proposes to generalize Cinque's idea to all islands and introduces many novel pieces of data.

Postal 1997 asks the fundamental question whether constituents are by default islands or by default not islands. The standard answer is that they are by default not, whence theories attempt to characterize islands as a natural class. In contrast, he argues that non-islands form a small natural class, characterizable in Arc-Pair Grammar.

Postal 1998 distinguishes locked (absolute) and unlocked (selective) islands. Locked islands permit only left dislocation (or operations that are like LD in the relevant respect). Unlocked islands permit some gaps and comprise both strong and weak islands in Cinque's terms: interrogative clauses, complements of factive verbs like regret, rationale clauses (go home (in order) to ...), clausal complements of certain nouns (formulate a plan to prove ...), adjuncts, relative clauses with quantifier heads, subjects, and certain coordinate structures (Postal's list is not intended to be exhaustive).

On Postal's analysis, every extraction from an unlocked island contains an invisible resumptive pronoun (RP). Such islands are unlocked because they permit the extraction of an RP. The mechanism is in many respects like Cinque's. The invisible RP in question requires control, hence it extracts to a position sister to the fronted phrase. This derives island-sensitivity. The reason why limited island violations are possible is that an RP controls a secondary invisible resumptive pronoun, notated as RPx, which extracts only to the left boundary of the lowest island from which the primary RP extracts. Tertiary RPs are excluded.

Postal assumes that RPs in unlocked islands are DPs. The fact that some selective islands permit PP gaps is acknowledged (fn. 15 to Ch. 3) but not addressed in any depth. But, unlike Cinque, Postal does not derive his reason for adopting an RP-analysis from category restrictions. Instead, the differential ability of phrases to occur in what he calls "anti-pronominal contexts" plays a major role.

Antipronominal contexts are ones that do not allow weak definite pronouns. One antipronominal context is the object of tell (as opposed to determine):

A pseudo-clefted phrase can, whereas a topicalized one cannot, be linked to a gap in an anti-pronominal context:

This is explained if topicalization requires an RP, independently of whether it is from an island, whereas pseudo-clefting does not require one. Assume further that pseudo-clefting allows, but comparative extraction forbids an RP. Then, if extraction from an island invariably involves an RP, one arrives at the following typology:

Note that A1-extractions are linked to an RP when escaping from an unlocked island, and they are not linked to an RP when occurring in an antipronominal context.

This typology predicts that when an antipronominal context is embedded in an otherwise unlocked island, no extraction is possible, since the two ingredients impose contradictory requirements. Consider, for instance, the existential there context. The examples indicate that it is antipronominal; it disallows topicalization (B), it permits negative fronting (A1), but when embedded in a factive island, it comes to disallow negative fronting, even though the latter is otherwise possible from a factive island:

Postal identifies the following further antipronominal contexts (illustrated with the impossibility of topicalization):

Postal remains non-committal as to why each extraction type allows, forbids, or requires an RP. It would clearly be desirable to identify syntactic or semantic properties that entail either that an RP is forbidden or that an RP is required, and to analyze the permissive type as syntactically or semantically ambiguous. This would be especially tempting in connection with the wh-phrases how big a N (RP forbidden), whose N (RP required), and what size N (RP permitted), whose divergent behavior Postal notes.

Postal’s data, whether well-understood or not, provide some novel counterexamples to certain claims in the literature. Although many of the anti-pronominal contexts involve adjuncts or non-referential expressions, many others are argumental and/or referential, i.e. their island sensitivity is surprising. On the other hand, while Cinque takes the island-sensitivity of Italian negative phrases to support the referentiality generalization, (64) shows that negative fronting in English is not blocked by weak islands.

As was mentioned at the outset, Postal lumps all unlocked islands together and does not distinguish between Cinque's two subtypes, viz. weak islands and single strong islands. The restricted viability of such a position is in fact implicit in Cinque's work. The DPs that can escape from weak islands are described quite similarly to those that can escape from strong ones. If this is correct, then, as long as only DP-extraction is considered, we do not expect to be able, or need, to distinguish between the two subtypes. Recall that Cinque distinguishes the two subtypes with reference to the possibility of PP-extraction.

In support of his unitary analysis, P. Postal (p.c.) notes that the (b,c) examples below pose a problem for Cinque's dual analysis. Only one of the gaps being inside an island, the two members of the conjunction in (b) and the two individuals that make up the plural in (c) need to participate in two different filler--gap relations:

The island sensitivity of an extraction may not be fully equated with its inability to link to an RP, however, whether the latter property be a primitive or a derived one. As will be detailed in 5.6, collectively interpreted arguments of "one time only" predicates are sensitive to weak islands (in the sense that extraction out of a WI forces a multiple events reading on the naturally "one time only" predicate):

The same collective arguments can nevertheless antecede resumptive pronouns, in contrast to some amount phrases, for example:

Whether or not the unification of all unlocked islands that Postal proposes is eventually viable, in the rest of this chapter, I follow standard practice in excluding the islands that Cinque calls strong from the discussion of weak (selective) islands.

4. Weak Islands: A preview

The historical starting point is the assumption, made in Huang 1982, Lasnik and Saito 1984, 1992, Chomsky 1986, that the paradigmatic (if not the only) case of weak (selective) islands is wh-islands, and the expressions whose extraction is sensitive to WIs are adjuncts, as opposed to arguments.

Since "Barriers," a number of new theories of weak islands have appeared in quick succession. What makes this process especially interesting is the fact that practically each theory comes with a significant new set of data. The survey below will reflect this spirit. The critical data come in two dimensions:

5. (A) What extractions are sensitive to WIs?

5.1 (A1) Arguments versus adjuncts

Huang 1982, Lasnik and Saito 1984, 1992, and Chomsky 1986 draw the distinction between arguments (claimed not to be sensitive to WIs) and adjuncts (claimed to be sensitive to them), e.g.

Besides how, why and, to a somewhat lesser extent, when are WI-sensitive adjuncts.

Where does not fit the picture very well, because its WI-sensitivity does not depend much on whether it is subcategorized for (as with put) or not (as with read):

The morphological constitution of these wh-words may also play a role. It has been noted that when counterparts of why and when have an articulated PP structure in a language, they extract better. Korean, Japanese, and Hungarian are cases in point.

5.2 (A2) Referential versus non-referential / Existential presupposition

A major revision of the argument/adjunct distinction is prompted by observations by Ross 1984, Kroch 1989, Comorovski 1989, Rizzi 1990, Cinque 1990, Obenauer 1992, and Kiss 1993. In a nutshell, the claim is that originating in an argument position is not enough: a successful extractee must also be referential in some sense.

Some subcategorized for, and thus presumably argumental, XPs are unexpectedly WI-sensitive:

According to Rizzi 1990, amount and manner phrases may be arguments but they do not have the theta-roles of event-participants (referential theta-roles). On the other hand, as events take place in time and space, Rizzi surmises that the event specification may license a temporal and locative index that accounts for the fact, noted above, that such phrases are less sensitive to WIs than manners and reasons.

Based on the fact that expletive wh-phrases do not have referential theta-roles, Rizzi 1992 extends this reasoning to partial wh-movement, which is blocked by a negative island (for cross-linguistic variation, see Dayal 1994):

            'With whom do(n't) you think Jacob is talking now?'

Drawing from work by Kroch and Comorovski, Cinque 1990 makes a finer distinction that involves pragmatics: a referential wh-phrase, in addition to having a referential theta-role, needs to be Discourse-linked, i.e. drawn from a pre-established set.

Some of the most persuasive most persuasive examples involve how many-phrases:

The good examples involve a contextually established set of books or a specific range of points that figure skating jurors conventionally assign to mistakes in the program.

Similar contrasts are easy to construct with other wh-phrases, including adjuncts. For instance, although how-extraction out of a wh-island or a negative island is by default bad, it becomes rather acceptable given a contextually specified checklist of ways to solve the problem:

In a similar spirit, Kiss 1993 assumes that specificity in the sense of Enç 1991 is a prerequisite for extraction out of a wh-island.

We might say that these proposals define WI-sensitivity in terms that combine syntax with pragmatics.

In the name of philological correctness, it should be mentioned that Kroch's influential paper apparently had more than one version, and the "official unpublished" version gives a different explanation from what Cinque relies on. According to Kroch, what saves (91) is not that the how many-phrase is D-linked but, rather, the fact that the context licenses the existential presupposition that there is a particular amount such that you are wondering whether to publish that amount.

A similar interpretive contrast is pointed out in Obenauer 1992. He characterizes it in terms of specificity and correlates it with a difference in agreement with the past participle. The split construction (combien... de fautes 'how many ... errors') never allows agreement and only the non-specific reading is available. In the non-split construction, on the specific reading agreement is optional: Dis-moi combien de fautes tu as fait / faites 'Tell me how many (of the typical, expected) errors you made', and on the non-specific reading, which asks about the number of errors, participle agreement is excluded.)

5.3 (A3) Re-evaluating the role of D-linking

The pragmatic argument is very powerful: indeed, almost any wh-phrase (save for why, perhaps) can be made immune to WIs by D-linking. But it has been argued that D-linking is not the discriminating factor.

Szabolcsi and Zwarts 1990, 1993 argue that the moral of the "salient checklist" examples is different from what Kroch and Cinque draw from them. The checklist effectively turns elements of a non-individuated domain into discrete individuals. D-linking may thus play an important role in transforming a domain but, they claim, it is not D-linking itself but the emergent set of individuals that is decisive.

Likewise, the pragmatic approach predicts that wh-the-hell can never be extracted out of a weak island, because it is "aggressively non-D-linked," in the words of Pesetsky 1987. Szabolcsi and Zwarts 1993 submit that it is an independent property of wh-the-hell that makes it difficult to extract from WIs. E.g.,

Unless a rhetorical question, the above is felicitous only if we have unquestionable evidence that someone saw John and merely wish to identify the person. The requirement of unquestionable evidence is often difficult to fulfil in the complex situations described by WI-violations:

On the other hand, when such evidence is available, a WI-violation by wh-the-hell is acceptable. E.g., seeing someone rifling through a dictionary, we may felicitously ask,

Perhaps the clearest example demonstrating the significance of individuals, in distinction to D-linking, comes from Dobrovie-Sorin 1994. Clitic doubling in Romanian signals D-linking, and indeed, it enables a how many-phrase to extract from a factive island:

On the other hand, cîte femei 'how many women' can be extracted even if it is not doubled by le and, consequently, is not D-linked. It turns out that the critical factor is whether it is interpreted as quantifying over numbers of women (case (i), which is bad) or over individual women (case (ii), which is okay):

These data lead to the conclusion that semantics, rather than pragmatics, plays the real role in the characterization of WI-sensitivity. (On the other hand, the WI-sensitivity of certain expressions in anti-pronominal contexts, e.g. place names in (76) remains a mystery.)

5.4 (A4) Individual versus non-individual, and how many-phrases

Szabolcsi and Zwarts argue that the individual / non-individual distinction is what sets apart WI-escaping which/what-phrases from manners, reasons, amounts, and other WI-sensitive expressions (when the latter are not individuated by contextual brute force). A very similar conclusion is reached by many authors specifically in connection with the ambiguity of how many-phrases.

Dobrovie-Sorin proposes to split QR into two Logical Form operations: NPR (the raising of noun phrases with a quantifier feature) and DR (the raising of determiners). The first leaves an individual variable and is immune to WIs. The second leaves a higher order, determiner-type variable and is sensitive to WIs. (NPR may be followed by DR but that does not matter in relation to island escaping.)

The distinction is of course highly reminiscent of the contrast between overt combien extraction and combien de N extraction (Obenauer 1984/85):

The overt extraction of combien 'how many' is blocked by WIs, whereas the overt extraction of the full noun phrase is (generally) not; see also Rizzi 2000 for a similar case of "splitting" in Italian, involving wh - d'altro 'wh else'. The finer point that Dobrovie-Sorin makes is that combien de livres is in itself ambiguous between an amount and an individual reading, and the amount reading is absent when combien de livres is (grammatically) extracted out of a WI, exactly as was observed in connection with cîte femei:

In fact, even overt combien de N extraction out of a WI may be ungrammatical when, as the object of a verb of creation, the phrase can only have an amount reading (Szabolcsi and Zwarts 1993). The contrast between this and grammaticality of (102) is taken to show that combien de livres is ambiguous:

The same interpretive contrast holds for non-wh numeral phrases, whose amount reading is produced by DR, and whose individual reading is produced by NPR, at Logical Form (Dobrovie-Sorin 1994):