Home | Schedule | Register | Directions | Accommodations | Contact

I focus on three closely related questions:

• Why do natural languages have extraction islands?
• Why do extraction islands take the form that we observe?
• Why is there variability in the acceptability of extraction island violations?

The Simpler Syntax (SS) perspective of Culicover and Jackendoff (2005, 2006) suggests a particular way of answering these questions. On the SS view, syntactic representations are as simple as possible, as long as the correspondence between form and meaning is accounted for. Maximizing the simplicity of the syntax leads to the hypothesis that that violations due to extraction from islands are not part of syntax, but have to do with unacceptability that arises from processing complexity. Configurations contribute a certain degree of complexity, and other factors contribute to complexity as well, producing variability in acceptability judgments.

Hence SS fits comfortably with recent work that argues that the complexity of processing sentences contributes in essential ways to judgments of unacceptability, including cases that have traditionally been categorized as "island violations" (Arnon et al. in press; Featherston 2005; Hawkins 1994, 2004). The key idea is that unacceptability judgments arise when the complexity of mapping the syntactic representation into the corresponding conceptual structure representation exceeds a certain threshold (Culicover and Nowak 2002; see also Keller 2000 and Kluender 1998, 2004). I apply Kluender.s proposals about how to measure complexity to a number of classical island phenomena. It appears that complexity beyond the threshold produces the subjective experience of unacceptability, and that grammaticality per se is not at issue. Moreover, relative complexity appears to determine the acceptability of extraction and the acceptability of parasitic gaps in similar ways, albeit with different consequences, suggesting as expected that complexity beyond the threshold blocks the identification of both true and parasitic gaps (cf. Phillips 2006).

To a first approximation, minimalism seeks to deduce GB principles from good design considerations revolving around elegance, economy, efficiency, symmetry and the like. It has proven quite difficult to recast locality/island conditions in these terms. This is in part due to the fact that locality is one domain where our empirical knowledge, gathered over the EST/GB period, is quite detailed. It is hard to unify so many fine-grained generalizations. But the difficulty faced by minimalists may also be due to the fact that the standard approach to locality, epitomized by Chomsky 1973, 1977, attempts to characterize islands in derivational terms built into the essence of narrow syntax. Within minimalism, the basic operations defining narrow syntax (Merge, Agree) are so general (read also: unrestricted) that it seems very hard to see how islands could emerge from these. Perhaps for this reason, it is not surprising to find some recent works (including my own [Boeckx 2003, 2008]) going back to pre-1973 works such as Ross 1967, 1969, and push the idea that islands are best characterized as interface phenomena. Contexts where islands are circumvented (resumption, sluicing) have proven very rich sources of information in this respect. All in all, I think it is fair to say that we now have several interesting, partially converging hypotheses regarding one aspect of the CED (the one pertaining to derived specifiers), but the other aspect of the CED . the adjunct condition . remains more elusive. The phenomenon of pied-piping also remains very poorly understood. In this talk, I would like to do three things: (i) insist that most locality conditions are interface phenomena, and that narrow-syntactic devices (including phases) are only indirectly related to islands, (ii) examine various conceptions of adjunction (late Merge, Pair Merge, Merge-with-no-label) and explore the consequences of each for islandhood, and (iii) use resumption to shed light on other island- circumvention contexts involving ellipsis and pied-piping.

Current work on islands raises the question of whether a unified account can be offered for all island effects. The answer to this question clearly depends on one's view of what counts as an island effect.

A large array of island effects can be accounted for (e.g., .extraction. islands, superiority) under the view that such effects follow from the sensitivity of processing to the linear alignment of Information Structure (topic-focus) with syntactic stru cture (subject-predicate). A particular strength of this approach is its ability to predict the context sensitivity and gradedness of some of the effects and the absoluteness of others. Another advantage of the approach is that it naturally accounts for t he repair of islands by resumptive pronouns and sluicing.

One of the problems with a processing account of islands is the existence of island effects in wh-in-situ languages: If the processing account is based on the idea that it is the processing of the filler-gap relationship that is at stake, then it would predict that island effects would be lacking in wh-in-situ languages (Lasnik, 1999, Boeckx, 2007). The proposed processing account.s requirement that Information Structure (IS) be aligned with syntactic structure does, however, nicely accommodate these f acts and predicts that the island effects in these languages mirror those of wh-movement languages (Bayer, 2005).

Another problem with processing accounts in general is that no language variation is predicted. In fact, languages vary with respect to both superiority and extraction. Here these differences are made to follow from different options available to mark topics and foci in these languages.

The account proposed here also has architectural consequences: IS processing explains linear order as well as islandhood. Therefore these phenomena should be removed from the domain of grammar proper and be handled by the interfaces instead. One possib le way of reconciling the two points of view, one forcing IS outside of grammar, the other keeping it inside, is to consider IS part of the phonological computation.

Like many languages, Greek morphologically differentiates between the marker of the standard of comparison in phrasal comparatives such as (1a) and in clausal comparatives such as (1b).

(1) O Kostas          exei perissotera vivlia  'Kostas has more books than Giannis has.' 
    the Kostas.NOM    has  more        books
   
    a. apo          ton Gianni. 	b. apoti        o    Giannis. 
       than.PHRASAL the Giannis.ACC       than.CLAUSAL the  Giannis.NOM 

In phrasal comparatives, the standard marker is the preposition apo ('than, from, of'), which assigns accusative case to its necessarily DP successor; in clausal comparatives, the marker is apoti and what follows can be of whatever category or case is required by the syntax internal to the standard clause.

Traditional analyses of phrasal vs. clausal comparatives posit that phrasal comparatives are simple PP structures, while clausal comparatives such as (1b) embed an unpronounced clause in which the remnant (o Giannis in (1b)) has moved to a clause-external position followed by clausal ellipsis (Hankamer 1973, Kennedy 1997, and many others). For such analyses, the contrast in is the reverse of what is expected: the DP following the standard marker can correspond to a DP embedded in an island (here a relative clause) in a clausal comparative, as in (2b), but not in the phrasal one of (2a).

(2) Perissoteri anthropi menun sto     kratos pu   kivernai o    Putin  
    more        people   live  in.the  state  that governs  the  Putin.NOM 
    
    a. *apo            ton Bush.  	b. apoti             o    Bush. 
    than.PHRASAL       the Bush.ACC        than.CLAUSAL      the  Bush.NOM 

    'More people live in the country that Putin governs than live in the country 
    that Bush governs.' (lit. ... 'than Bush') 

I argue that this surprising pattern shows that syntactic structures must be able to be phonologically abstract: either there is overt movement of the remnant followed by ellipsis (using a PF-theory of islands), or there is covert movement of the correlate (and such LF movement is island-sensitive). The first alternative makes sense from the perspective of some recent approaches to variable island behavior under ellipsis (Merchant 2004) and to ellipsis in comparatives (Lechner 2004). Lechner 2004 (in the spirit of early generative work on comparatives) proposes that at least some phrasal comparatives have clausal, but elliptical, syntax; here, in particular, that the DP ton Bush moves from within the standard clause to position in the projection of the preposition apo. The contrast between this movement, which gives rise to an island effect, and the movement of the remnant o Bush in the clausal case, which is not anomalous, can be assimilated to the contrast found between some stripping or fragment answer structures (which are sensitive to islands) and some sluicing structures (which are insensitive to islands). Island-repair occurs when no island-violating intermediate traces (marked t*) survive ellipsis (here, of TP), as in sluicing and in clausal comparatives, given in (3); when such traces occur external to the ellipsis site (here, in specFP), as in fragment answers and phrasal comparatives as in (4), they trigger a PF crash.

(3)  apoti                 [CP [FP o Bush1 <[TP menun sto kratos pu kivernai t1]>]] 
(4)  apo [PP ton Bush1 tapo [CP [FP t''1* <[TP  menun sto kratos pu kivernai t1]>]]] 
     than    the Bush.ACC      the Bush.NOM  live  in.the state that governs  

The second approach, following Reinhart 1991 and Bhatt and Takahashi 2008, posits a 3- place semantics for the comparative morpheme, requiring that the correlate and the DP in the phrasal comparative outscope the comparative morpheme at LF: for the correlate, this requires a non-quantificational syntactic displacement at LF.

I argue that such facts cannot easily be accounted for in grammar architectures such as Culicover and Jackendoff 2005, which eschew unpronounced syntactic structures of the kind needed to state the island constraints over. Instead, these facts support traditional approaches to the syntax of ellipsis, and are consistent only with the phonologically abstract syntactic structures.

Competence-based theories of island effects are abundant in generative grammar, yet the graded nature of many syntactic islands (Ross 1986; Deane 1991; Kluender 1992) has never been properly accounted for. We argue that this is primarily because the processing (performance) factors that interact with the grammar of filler-gap constructions have never been properly controlled in the datasets island theories seek to explain. Processing factors are known to play a significant role in the perception of (un)acceptability (Miller & Chomsky, 1963; Bever 1970; Osterhout, Holcomb, Swinney, 1994) -- an excessive sentence processing load can lead to lower acceptability, as well as slowed reading times, slowed response times, and decreased response accuracy.

We have examined the role of processing factors in a range of CNPC/`Subjacency'/Wh-island and Superiority effects. I will summarize data from self-paced reading-time experiments and controlled acceptability studies that isolate the effects of various factors that have not been properly controlled in generative research. Most notable among these is the informativity of the filler (which-NP vs. who/what). Other factors include the accessibility of referential elements intervening between the filler and the gap, the finiteness of intervening verbs, lexical frequency effects, and various semantic and pragmatic issues that contribute to processing difficulty.

We argue that once these diverse effects are properly isolated, competence grammars have no need for such principles as CNPC, Subjacency, WH-Island Condition, Superiority or the Minimal Link Condition. This `minimalist' approach truly simplifies grammar, while maximizing the explanatory effect of factors known independently to contribute to processing difficulty, some of which may ultimately be derived from more general considerations of processing architecture (e.g. Gibson 2000) or memory (Hofmeister 2007).

The traditional approach to extraction islands in research on nonnative language (L2) has focused on differences between the native language (L1) and the Target Language (TL), typically asking whether adult L2ers (come to) have the same sensitivity as mature natives (e.g. Bley-Vroman, Felix & Ioup 1988; Hawkins & Chan 1997; Johnson & Newport 1991; Martohardjono 1993; Schachter 1989; White 1992). These studies, where the TL is the more restrictive, point to three findings: (1) (in)sensitivity to extraction islands in the L1 exerts effects, from early adult L2 development onward; (2) as L2 proficiency increases, adult L2ers exhibit patterns of sensitivity similar to those of mature natives (e.g. differentiating 'weak' from 'strong' islands--Martohardjono 1993); and (3) nevertheless, even 'advanced' adult L2ers do not display the exact sensitivity of mature natives. In their study on (ostensible) L2-endstates, Johnson & Newport (J&N 1991) sought to determine whether there are L2 critical period effects in this domain by testing the L2 English of L1-Chinese speakers whose English-onset age ranged from 4 to 38 (all adults at time of testing). J&N found that oral acceptability judgments of adult L2 starters (> age 18), but not (early) child L2 starters (between age 4 and 7), fell (far) below native levels across 3 types of island violations: extraction from Relative Clauses (RCs), wh-islands and Complex NPs (CNPs). These findings suggest to J&N that "adult learners of a language will sometimes form hypotheses or rules ... unnatural to human languages" (p. 245).

Our study revisits (in)sensitivity to extraction islands but from a non-endstate perspective. Following Schwartz (e.g. 2003, 2004), we make a three-way comparison of L2 adults, L2 children (holding L1 constant) and L1 youths. Assuming that (early) child L2ers--in the course of development--create natural language grammars, we hypothesize that if adult L2ers and child L2ers--as well as L1 youths--exhibit similar features (specifically, divergence from mature natives' sensitivity to extraction islands), this would argue, contra J&N, that L2 adult Interlanguage is similarly constituted, and evinces non-targetlike insensitivity to islands for whatever reasons L2 child Interlanguage does.

A 72-item, contextualized oral acceptability-judgment task combined with an elicited-production task was administered in English to L1-Korean adult L2ers (n=7, English-onset age>20, mean age at testing=29;4), L1-Korean child L2ers (n=9, English-onset age=4;0-7;9, mean age at testing=9;4), L1-English youths (n=8, mean age at testing=10;1) and L1-English adults (n=9). Each item comprises a series of pictures which the researcher narrates (in Korean to L2ers); after this, an English-learning puppet 'says' 3 things (each recorded twice) in English--sometimes questions, sometimes statements, sometimes well-formed, sometimes ill-formed--and the participant tells the puppet whether each sounds 'ok' or 'strange' (acceptability judgment), and if 'strange', how to say it 'better' (elicited production). General (prerequisite) grammar relevant to extraction from islands was also tested (RCs, do-support, simple questions, subject-auinversion (k=3 each)), but critical items target 3 island violation types (k=4 each):

1. *Who did Daddy send the card to the policeman that found? (RC)
2. *What did Pooh ask Piglet where Daddy put? (wh-island)
3. *What did Pooh remember the fact that Mickey liked? (CNP)

Results: (a) all 4 groups show sufficient knowledge of general (prerequisite) English grammar; (b) L1-English adults are at ceiling on all 3 types of island violations (e.g. mean accuracy=98%); (c) adult L2ers exhibited (some) sensitivity to island violations (e.g. mean accuracy=80%), as did child L2ers (e.g. mean accuracy=71%) and L1-English youths (e.g. mean accuracy=83%); and (d) whereas our adult L2ers differed little across the 3 island violation types, our child L2ers evinced a pattern similar to that of our L1-English youths: most targetlike in rejecting extraction from RCs and least targetlike in rejecting extraction from CNPs--paralleling, moreover, J&N's adult 'endstate' L2ers.

These findings pose a challenge to J&N's conclusion about adult Interlanguage grammars. Yet, they also seem to pose a question to processing-based accounts of island violations: If strains on processing give rise to the unacceptability of sentences like (1)-(3), why then do populations with reduced sentence-processing resover-accept rather than over-reject?

The classical example of a weak, i.e. selective island violation is the extraction of an adjunct out of an non-tensed wh-complement, and the classical account, purely syntactic, is that the intervening wh-phrase blocks one escape hatch with its bulk. I nterestingly, work in the past two decades has demonstrated that many further relations are sensitive to weak islands, i.e. are blocked by roughly the same interveners that block the adjunct­gap relation, and that many further interveners induce such weak islands.

A sample of weak island sensitive constructions includes the extraction of non-referential expressions in general (whether adjuncts or arguments), the extraction of obligatorily collective arguments, event-related readings, functional questions, split constructions, partial wh-movement, some cases of wh-in-situ, comparative constructions and, more strikingly, NPI-licensing and cross-sentential anaphora.

A sample of weak island inducers includes besides wh-phrases downward entailing operators, response stance and non-stance predicates, VP-adverbs, any distributive quantifiers scoping between two terms of a sensitive relation, and focus-sensitive operators.

Some though not all of these are summarized in Szabolcsi 2006 (in: The Blackwell Companion to Syntax). That article furthermore observes that expressions that escape from strong islands (by binding a silent resumptive pronoun) are not unl ike those ones that escape from weak ones: they tend to be referential individual-denoters. If not purely coincidental, this is intriguing.

The widening of the data sets has been paralleled by the emergence of new theories of weak islands. The first major innovation, Relativized Minimality, was largely though not purely syntactic. Due to the heterogeneity of the data, subsequent the ories have gone more in the direction of semantics and/or pragmatics. They fall into two main categories. Kiss 1993, Szabolcsi & Zwarts 1993, Fox & Hackl 2006, and Abrusán 2007 are squarely semantic; according to these proposals, weak island violations a re essentially incoherent. In contrast,de Swart 1992, Cresti 1995, Honcoop 1998, Pesetsky 2000, Butler & Mathieu 2004, and Beck 2006 localize the problem in syntax or at the syntax/semantics interface; likewise Starke 2002, in the only systematic a ttempt to unify strong and weak islands in a new version of Relativized Minimality.

How is this relevant to linearization accounts of islands? To my mind, the task starts with identifying the natural classes of phenomena; then the question becomes, do syntactic accounts, linearization among them, constitute a good theory of some na tural class? It is not clear to me whether the question has been asked this way.

It is to be stressed that although the constructions listed above are tantalizingly similar, it is by no means certain that they are instances of the same thing. In fact, the theories mentioned overlap in coverage, but none of them covers all the data. So the jury is out on exactly what the natural classes are, although it seems clear that those classes are fairly large, and therefore fairly challenging. Experimental studies could also provide important new insights here.