2.1 Background

The A-over-A constraint proposed by Chomsky (Reference Chomsky and Lunt1962) had the effect that any category X could not be extracted across a node of the same category. For example, this prevented NP extraction from NPs as in *What did he know [someone who has __]? and in *Who did you hear [the rumor [that Mary kissed __]]? Footnote ³ But it was soon observed that such a constraint was too strong as it ruled out perfectly acceptable extractions like Who did you read [a book about __]? Subsequent accounts like Chomsky (Reference Chomsky, Anderson and Kiparsky1973, Reference Chomsky1986) were more permissive, allowing (1a) while ruling out (1b).

1. (1)

   1. (a) Who did you take a photograph of __?

   2. (b) *??Who did you take a photograph of a statue of __?

      (Fodor Reference Fodor1983: 190)

In the Barriers framework (Chomsky Reference Chomsky1986), island effects were considered additive in proportion to the number of ‘cyclic blocking categories’ that were crossed (in English, NP and IP nodes). This approach also blocked the extraction of NPs from subject phrases as in (2), because under transformational assumptions extraction crosses two blocking categories in such cases, which ‘should yield a considerable decrement in acceptability’ (Chomsky Reference Chomsky1986: 86).

1. (2)

   1. (a) Your interest in him seemed to me rather strange.

   2. (b) *Whom did your interest in seem to me rather strange?

      (Chomsky Reference Chomsky and Lunt1962: 437)

   3. (c) The hoods of these cars were damaged by the explosion.

   4. (d) *Which cars were the hoods of damaged by the explosion?

      (Ross Reference Ross1967, ex. (4.252))

There are a variety of problems with this kind of account. First, it has never been clear why such ‘cyclic bounding nodes’ should interfere with extraction in the first place. Second, this view assumes that there are various degrees of grammaticality. Not only is it unclear what this means for a derivational grammatical theory (where, presumably, a grammatical violation would simply abort the parse) but it is also empirically hard to justify since all speaker judgments are necessarily acceptability judgments rather than grammaticality judgments. And if (1b) is odd due to performance rather than grammar, then the grammar should remain silent about extraction constraints on bounding nodes. For Fodor (Reference Fodor1983) and others, the fact that multiple embeddings often make object extraction more difficult, as in (1), suggests an explanation based on processing complexity.

A third problem was noted by Ross (Reference Ross1967) and Deane (Reference Deane1991). The existence of deep extraction cases like (3) suggests that there is no syntactic depth limitation to NP extraction from object NPs.Footnote ⁴ To my knowledge, these facts have never been taken into consideration in the formulation of subjacency-based accounts of islands.

1. (3)

   1. (a) Nixon was one president that [they had [no trouble [getting votes for [the impeachment of __ ]]]].

   2. (b) Which committee did you have [aspirations for [an appointment to [the chairmanship of __ ]]]?

      (Deane Reference Deane1991: 11)

Deaene (Reference Deane1991) and other authors (Chung & McCloskey Reference Chung and McCloskey1983, Hegarty Reference Hegarty, Cheng and Demirdash1990, and Pollard & Sag Reference Pollard and Sag1994: 191, 206) noted further counterexamples like (4), which show that complex NPs are not impermeable to extraction.

1. (4)

   1. (a) How much money are you making [the claim [that the company squandered]]?

   2. (b) Which rebel leader did you hear [rumors [that the CIA assassinated]]?

   3. (c) Which Middle Eastern country did you hear [rumors [that we had infiltrated]]?

   4. (d) Which diamond ring did you say there was [nobody in the world [who could buy]]?

The evidence above suggests that the oddness of (1b) may not result from purely configurational factors. A promising avenue of inquiry should consider semantic, pragmatic, and performance factors as well. As an analogy, it would be misguided to postulate a ban on NP extraction from multiple embedded clauses in order to capture the oddness of ?*Who did you say Tom thinks Mary denied Tom believes Fred trusts?. Although humans routinely process longer sentences, the oddness of such examples cannot be seen as a matter of competence.

Matters are further complicated by the fact that extraction out of subject phrases is generally held to be impossible and therefore much more restricted than extraction from objects, as the sample in (5) shows. The judgments reported below are the cited author's.Footnote ⁵

1. (5)

   1. (a) *Who did stories about terrify John? (Chomsky Reference Chomsky1977: 106)

   2. (b) ??Which tree did John see the leaves of turn color?

      (Kuno Reference Kuno1973a: 381)

   3. (c) *Who did pictures of lay on the table? (Postal Reference Postal1974: 189)

   4. (d) *Who was a picture of lying there? (Kayne Reference Kayne1981: 114)

   5. (e) *Who do you think pictures of would please John?

      (Huang Reference Huang1982: 497)

   6. (f) ?Who did John's talking to bother you most? (Engdahl Reference Engdahl1983: 14)

   7. (g) *Which books did talking about become difficult? (Cinque Reference Cinque1990: 1)

   8. (h) *Who does the claim that Mary likes upset Bill?

      (Lasnik & Saito Reference Lasnik and Saito1992: 42)

   9. (i) *Which politician did pictures of upset the voters?

      (Nunes & Uriagereka Reference Nunes and Uriagereka2000: 21)

   10. (j) *Which book did a review of appear in the Times?

       (Jackendoff Reference Jackendoff2002: 48)

   11. (k) *Which candidate were posters of all over the town?

       (Lasnik & Park Reference Lasnik and Park2003: 651)

   12. (l) ?*Who was a friend of arrested? (Stepanov Reference Stepanov2007: 85)

The puzzle is not only to explain why extraction from objects is possible in some cases but not in others, but to explain why extraction from subjects is more restrictive. Standard accounts categorically prohibit movement from the subject phrase. In Takahashi (Reference Takahashi1994) such islands are the consequence of the requirement that movement must be as short as possible and the stipulation that nothing can adjoin to a chain (subjects move from [Spec,vP] to [Spec,IP], thus creating an A-chain with the in situ copy of the subject). Any extraction from a subject would be required to adjoin to the closest possible site, which is the subject DP, hence violating the ban on chain adjunction. Conversely, a longer movement is banned by the closest move condition. As a consequence, the subject island effect arises. In Uriagereka (Reference Uriagereka, Epstein and Hornstein1999) and Nunes & Uriagereka (Reference Nunes and Uriagereka2000), phrases are assumed to be linearized before they can be merged with other phrases. Thus, subjects must be linearized before combining with the V′. Once linearized, the DP is considered a morphological unit, and hence, no extraction from it can occur. Syntactic accounts like the ones above predict that extraction from an extracted phrase is banned in general, as a freezing effect (Wexler & Culicover Reference Wexler and Culicover1980). A conclusion that may be excessive, as the acceptable freezing violations in (6) suggest.

1. (6)

   1. (a) This is the handout_i that I can't remember [how many copies of __ _i]_j we have to print __ _j.

   2. (b) [How many videos __ _i]_j are there __ _j on the web [of Mitt Romney getting booed]_i ?

A competing line of research holds that subject islands are due to pragmatic factors, such as Erteschik-Shir (Reference Erteschik-Shir1973, Reference Erteschik-Shir2006, Reference Erteschik-Shir2007), Van Valin (Reference Van Valin, Farley, Farley and McCullough1986, Reference Van Valin, Devriendt, Goossens and van der Auwera1995), and Goldberg (Reference Goldberg2006)inter alia. Basically, the claim is that subjects (and their adnominal structures) are typically part of the common ground and therefore cannot be felicitously questioned. For example, Erteschik-Shir (Reference Erteschik-Shir2006: 325) states that extraction out of subjects is always ungrammatical and cannot be contextually ameliorated. This categorical prohibition follows from her assumption that subjects must be assigned topic and that extraction never targets topics. Hence, Erteschik-Shir obtains the prediction that such extractions are always blocked.

Adjunct phrases are another type of phrase that is usually regarded as being particularly difficult to extract from, as (7) illustrates.

1. (7)

   1. (a) John met a lot of girls without going to the club.

   2. (b) *Which club did John meet a lot of girls [without going to __ ]? (Cattell Reference Cattell1976: 38)

   3. (c) Who cried after John hit Mary?

   4. (d) *Who did Mary cry [after John hit __ ]? (Huang Reference Huang1982: 503)

Huang (Reference Huang1982) argued that subject and adjunct islands ought to be unified under a broader universal condition called the Constraint on Extraction Domain (CED). This rule stipulates that a phrase P may be extracted out of domain D, iff D is properly governed. An α is said to be properly governed by β if α is c-commanded by a lexical head β and no major category or major category boundary appears between α and β. Many authors have sought to derive the CED from independent principles (Kayne Reference Kayne1983, Chomsky Reference Chomsky1986, Rizzi Reference Rizzi1990, Lasnik & Saito Reference Lasnik and Saito1992, Uriagereka Reference Uriagereka, Epstein and Hornstein1999, Sabel Reference Sabel2002, and Müller Reference Müller2010).

2.2 Counterindications

There are good reasons to doubt that subject islands are so intimately related to adjunct islands, and more importantly, to doubt that such islands are simply due to grammatical conditions. The first problem for the grammatical status of strong islands comes from the observation made by Ross (Reference Ross1967), Taraldsen (Reference Taraldsen, Belleti, Brandi and Rizzi1980) and Engdahl (Reference Engdahl1983) that extraction from subjects and adjuncts is usually more acceptable if the gap is ‘parasitic’ on another gap, located in a non-island environment, as in (8)–(10).

1. (8)

   1. (a) *Who did [the rivals of __ ] shoot Castro?

   2. (b) Who did [the rivals of __ ] shoot __?

   3. (c) Who did [the rivals of Castro] shoot __?

1. (9)

   1. (a) *What did [the attempt to repair __ ] ultimately damage the car?

   2. (b) What did [the attempt to repair __ ] ultimately damage __?

   3. (c) What did [the attempt to repair the car] ultimately damage __?

1. (10)

   1. (a) *Which paper did John [read his email [before filing __ ]]?

   2. (b) Which paper did John [read__[before filing __ ]]?

   3. (c) Which paper did John [read__[before filing a complaint]]?

Instead of taking (8b), (9b) and (10b) as counterexamples, these island-violating gaps have been seen as some kind of null resumptive pronoun. As such, they would pose no challenge to the view that subjects and adjuncts are strong extraction islands. For example, Cinque (Reference Cinque1990) and Postal (Reference Postal1993, Reference Postal1994, Reference Postal1998) draw a distinction between ‘A-extractions’ and ‘B-extractions’. While the former are canonical wh-extractions, the latter arise by insertion of a null resumptive pronoun at the extraction site. In both accounts, B-extractions are identified as parasitic extractions.

But, as Steedman (Reference Steedman1996: 98, fn.41), Levine (Reference Levine2001), Levine, Hukari & Calcagno (Reference Levine, Hukari, Calcagno, Culicover and Postal2001), and Levine & Hukari (Reference Levine and Hukari2006: 256) show, parasitic gaps can be non-nominal, as illustrated in (11).

1. (11)

   1. (a) [How harshly] do you think we can treat them __ without in turn being treated __ ourselves?

   2. (b) I wonder [how nasty] you can pretend to be __ without actually becoming __?

   3. (c) This is a cause [to which] many people are attracted __ without ever becoming seriously devoted __.

   4. (d) This is the table [on which] anyone who puts some books __ must subsequently put some magazines __ as well.

This casts doubts on the assumption that parasitic extractions can be explained away in terms of null resumptive pronouns. Furthermore, the notion that only one gap is ‘real’ and all others are null pronouns is refuted by cases where neither gap occurs in a parasitic environment, as in Who did you inform __ that the police was about to arrest __? or in Who did you hug __ and kiss __?.

The view that adjuncts are islands and that the only gaps that such phrases allow are of the ‘parasitic’ type is problematic. There is robust evidence that adjuncts do in fact allow extraction, as illustrated in (12), from Ross (Reference Ross1967), Chomsky (Reference Chomsky1982), Engdahl (Reference Engdahl1983), Kayne (Reference Kayne1983), Hegarty (Reference Hegarty, Cheng and Demirdash1990), Cinque (Reference Cinque1990), Pollard & Sag (Reference Pollard and Sag1994), Borgonovo & Neeleman (Reference Borgonovo and Neeleman2000) .

1. (12)

   1. (a) That's the symphony that Schubert [died [without finishing _ ]].

   2. (b) Which report did Kim [go to lunch [without reading _ ]]?

   3. (c) What did he [fall asleep [complaining about _ ]]?

   4. (d) What did John [drive Mary crazy [trying to fix _ ]]?

   5. (e) What did John [come back [addicted to _ ]]?

   6. (f) Who did you [go to Girona [in order to meet _ ]]?

   7. (g) Who would you rather [sing [with _ ]]?

   8. (h) What temperature should I [wash my jeans [at _ ]]?

See Szabolcsi (Reference Szabolcsi, Everaert and Riemsdijk2006) and Truswell (Reference Truswell2011) for overviews of adjunct islands and their exceptions. To my knowledge, there is no direct empirical evidence that can justify the assumption that the examples in (12) are grammatical because the adjuncts have some special syntactic status that does not lead to an island violation. The natural conclusion is that adjunct islands are probably not a syntactic phenomenon.

Although untensed adjuncts do not always block extraction, it has been nearly universally held that tensed adjuncts, as in (13), are truly exceptionless islands.

1. (13)

   1. (a) John came back before I had a chance to talk to you.

   2. (b) *Who did John [come back [before I had a chance to talk to _ ]]?

      (Huang Reference Huang1982: 491)

But a small number of authors have occasionally noted that it is in principle possible to extract from finite adjuncts. The first cases were noted by Grosu (Reference Grosu1981: 88). The examples in (14) are adapted from Grosu's original data.Footnote ⁶

1. (14)

   1. (a) These are the pills that Mary died before she could take.

   2. (b) These is the house that Mary died before she sold.

Taylor (Reference Taylor, Conroy, Jing, Nakao and Takahashi2007) claims that extraction from tensed conditional adjuncts is allowed as long as the adverbial clause is fronted, as in (15a) (her judgments).

1. (15)

   1. (a) Which book does John believe if Kim reads, she will understand linguistics better?

   2. (b) *Which book will Kim understand linguistics better if she reads?

      (Taylor Reference Taylor, Conroy, Jing, Nakao and Takahashi2007: 192)

Taylor (Reference Taylor, Conroy, Jing, Nakao and Takahashi2007) argues that such fronted conditionals allow extraction because they are base-generated, but (14) and (16) refute the generality of this view.

1. (16)

   1. (a) This is the watch that I got upset when I lost.

      (attributed to Ivan Sag (p.c.) by Truswell Reference Truswell2011: 175, fn. 1)

   2. (b) Robin, Pat and Terry were the people who I hung around at home all day without realizing were coming for dinner.

      (Levine & Hukari Reference Levine and Hukari2006: 287)

Further counterexamples are provided in (17), drawn from Chaves (Reference Chaves2012: 471).

1. (17)

   1. (a) Which email account would you be in trouble if someone broke into?

   2. (b) Which problem would you be impressed if someone had already solved?

   3. (c) Which costume would mom freak out the most if I wore on Halloween?

   4. (d) This is the type of problem that I would be relieved if someone had already solved.

   5. (e) Which toy did Tom throw a tantrum because somebody broke?

Some of these sentences may be complex and difficult to process, which in turn can lead speakers to prefer the insertion of an ‘intrusive’ resumptive pronoun at the gap site, but they are more acceptable than the classic examples in (7) and (13) above.

The data discussed so far suggest that adjunct islands are not absolute and have acceptable non-parasitic violations, contrary to widespread belief. The problems for the standard notion of parasitism are underlined by the fact that the ‘parasitic’ gap need not be semantically linked to the ‘real’ gap. In (18) – attributed to Polly Jacobson (p.c.) in Pollard & Sag (Reference Pollard and Sag1994: 199, fn. 35) – the subject-internal gap is not co-referenced with the object gap.

1. (18)

   1. (a) There are certain heroes that [long stories about __ ] are always very easy to listen to __.

   2. (b) There are certain heroes that [long stories about __ ] are too boring to listen to __.

If the two gaps are not related, then in what grammatical sense is one licensing the other? Pollard & Sag (Reference Pollard and Sag1994: 199) note other cases, such as (19), in which the subject-internal gap is not linked to the object-internal gap. Although the extraction of the phrase which topics causes low acceptability in (19a), the existence of a second gap in (19b) improves the sentence. Again, it is most unclear in what grammatical sense a ‘parasitic’ gap must be licensed by a ‘real’ gap.

1. (19)

   1. (a) ??I never know [which topics]_j [jokes about __ _j] are likely to offend people.

   2. (b) [People that sensitive]_i, I never know [which topics]_j [jokes about __ _j] are likely to offend __ _i.

This pattern also occurs in adjunct gaps as shown by Chaves (Reference Chaves2012: 471). In (20) the gap located in the adjunct phrase is not semantically linked to the preceding gap, and hence would not be expected to be parasitic.

1. (20)

   1. (a) [A project this complex]_i, [how much time]_j could they spend __ _j before finishing __ _i?

   2. (b) This was [the kind of person]_i that even [the simplest problem]_j became difficult to solve __ _j without shouting at __ _i.

The standard view on parasitic gap licensing is also challenged by acceptable cases noted by Horvath (Reference Horvath1992:201), in which a parasitic gap is c-commanded by co-indexed non-parasitic gap (e.g. Who do you expect __ to withdraw his candidacy [before the Committee has a chance to interview __ ]?). See Culicover & Postal (Reference Culicover and Postal2001) for more discussion.

Finally, Levine & Sag (Reference Levine, Sag and Müller2003) and Levine & Hukari (Reference Levine and Hukari2006: 256) note what they call ‘symbiotic gaps’. In such cases, one gap is located in the subject and the other in the adjunct, as shown in (21).

1. (21)

   1. (a) *What kinds of books do authors of __ argue about royalties after writing malicious pamphlets?

   2. (b) ??What kinds of books do authors of malicious pamphlets argue about royalties after writing __?

   3. (c) What kinds of books do the authors of __ argue about royalties after writing __?

The crucial point of these data is that (21a, b) are less acceptable than (21c). This is unexpected, given that the sentence arguably has no legitimate non-parasitic gap. Although (21c) is by no means impeccable, it should be at least as bad as (21a, b), contrary to fact.Footnote ⁷ Given the evidence discussed above, I follow Levine et al. (Reference Levine, Hukari, Calcagno, Culicover and Postal2001) and Levine & Hukari (Reference Levine and Hukari2006: 292) in assuming that there is no grammatical distinction between ‘parasitic’ and ‘real’ gaps. Furthermore, I conclude that any account of parasitism based on stipulations about c-command, if any exists, would shed no light on the phenomena.

The claim that the CED is a universal syntactic constraint is also problematic. It is known since Ross (Reference Ross1967) that there is a variety of languages that seem to allow extraction from subjects. According to the survey by Stepanov (Reference Stepanov2007), these include Swedish, German, Russian, Spanish Palauan, Malagasy, Tagalog, Armenian, Hungarian, Turkish, Hindi, Japanese and Navajo. See also Chung (Reference Chung, Georgopoulos and Ishihara1991: 71) for acceptable sentential subject island violations in Chamorro. For illustration, consider the Swedish example in (22) from Sells (Reference Sells1984: 304ff.).

1. (22) den deckare som de sista sidorna i__ hade kommit bort

   that detective.novel that the last pages in had come away

   ‘that detective novel that [the last pages in had come away] … ’

Because there are many languages that allow extraction from subjects, but not from adjuncts, Stepanov (Reference Stepanov2007) concludes that the subject and adjunct islands are due to different mechanisms. In Stepanov's view, English just happens to be one of those languages where both subjects and adjuncts are opaque to extraction.

In spite of this robust typological evidence, there is still resistance in abandoning the universality of the CED, as illustrated in Hornstein et al. (Reference Hornstein, Lasnik and Uriagereka2006), Boeckx (Reference Boeckx2008: 161), and Sheehan (Reference Sheehan, Kan, Moore-Cantwell and Staubs2009). For instance, Jurka, Nakao & Omaki (Reference Jurka, Nakao, Omaki, Washburn, McKinney-Bock, Varis, Sawyer and Tomaszewicz2011) conducted acceptability studies which show that extraction out of specifiers in German, English, Japanese and Serbian are consistently less acceptable than extraction out of complements. The authors of the study take this to show the CED is a valid constraint and that syntactic accounts like Uriagereka (Reference Uriagereka, Epstein and Hornstein1999) are still tenable. I do not dispute the experimental findings, nor the oddness of the subject island violations reported in the standard literature. Rather, I dispute the conclusion that the CED is empirically viable as a grammatical phenomenon. When a more representative data are taken into consideration, it becomes highly unlikely that a parsimonious grammatical generalization can be formulated to capture the full range of facts.

2.3 On a processing account

Given the empirical problems with the notions of strong islands, parasitism and null resumption, it is possible that the configurational nature of subject and adjunct islands has been overstated. Perhaps such extractions are syntactically well-formed, but constrained by more cognitive and pragmatic general factors. In such a view, there would be no need to stipulate a grammatical distinction between parasitic and real gaps, nor the existence of null resumption in English.

Kluender (Reference Kluender, Goodluck and Rochemont1992, Reference Kluender, Culicover and McNally1998), Kluender & Kutas (Reference Kluender and Kutas1993), Hofmeister (Reference Hofmeister and Elliott2007), Sag et al. (Reference Sag, Hofmeister, Snider and Elliot2007), and Hofmeister & Sag (Reference Hofmeister and Sag2010) argue that at least some object islands may be best viewed as cases where grammatical sentences obtain low acceptability because they are difficult to process. For example, infrequent expressions, semantically rich phrases, and discourse-new referents are known to consume more cognitive resources than highly frequent (or collocational) expressions, semantically light phrases and discourse-old referents. The processing cost is also inflated when sentences are long and ambiguous, contain phrases that do not cohere or are stylistically deviant, or include filler–gap dependencies. It is known that reading and response times to various kinds of tasks increase inside long-distance dependencies (e.g. Wanner & Maratsos Reference Wanner, Maratsos, Hale, Bresnan and Miller1978, Chen, Gibson & Wolf Reference Chen, Gibson and Wolf2005), which shows that maintaining an extracted phrase in memory, searching for possible gap sites, and linking the gap to the filler significantly exhausts cognitive processing resources. Such an account would allow for the existence of apparent counterexamples because two isomorphic sentences can incur different processing costs depending on the particular words they contain.

There is independent evidence that processing difficulty can cause low acceptability. For example, doubly center-embedded sentences are known to become easier to process and more acceptable if there are certain cues that help the processor link the predicates to their respective arguments: overt relative pronouns, verbs that cohere well with their respective arguments, semantically light expressions (e.g. pronouns), and a prosodic break between the subject and the VP. This explains why the double center-embedding sentences in (23) obtain completely different degrees of acceptability: whereas (23a) is unparsable, and (23b–d) are graded, (23e) is easy to parse and fully acceptable.

1. (23)

   1. (a) *People people people left left left. (Rogers & Pullum Reference Rogers and Pullum2011: 330)

   2. (b) ?*The boy the cat the dog bit scratched started crying.

   3. (c) ?The guy whom the secretary we fired slept with is a real lucky dog.

      (Kimball Reference Kimball and Kimball1975)

   4. (d) ?A syntax book that some Italian that I had never heard of wrote was published by MIT Press. (Frank Reference Frank1992)

   5. (e) The movie everyone I know loved was Inception.

      (Chaves Reference Chaves2012: 480)

This shows that the language processor can be influenced by various kinds of linguistic information, such as complexity, prosody, context and pragmatics. Many examples abound, as Sag (Reference Sag and Martin-Vide1992) notes. For example, the ambiguous string I forgot how good beer tastes is preferentially parsed as [I forgot [[how good] [beer tastes]]] if it is known that the speaker just returned from a place where beer is banned, and preferentially parsed as [I forgot [how [good beer tastes]]] if the speaker just returned from a place that only has poor quality beer. In this paper it is argued that the same factors that facilitate the processing of double center-embeddings like (23) also facilitate the processing of subject islands.

Kluender (Reference Kluender, Chand, Kelleher, Rodríguez and Schmeiser2004) argues that complex subject phrases place particularly heavy demands on working memory, which may in turn lead to low acceptability when these demands are increased by processing a filler–gap dependency. This view receives experimental support from Clausen (Reference Clausen2010, Reference Clausen2011), which shows that clausal subjects cause a measurable increase in processing load, with and without extraction. Kluender notes that there are several kinds of evidence in favor of his claim. First, the acquisition data from Bloom (Reference Bloom1990) indicate that there is a trade-off between subject length and verb phrase length in early acquisition: the longer the subject, the shorter the verb phrase, and vice-versa. This trade-off appears to also exist in the grammar of adults, and remains constant with age in spontaneous written production as well (Kemper Reference Kemper1987). Bloom (Reference Bloom1990) also found that children reduce subjects by whatever means possible. For example, lexical noun subjects are significantly shorter than lexical noun objects, and pronouns are used more often in subjects than in object position. Kemper (Reference Kemper1986) also shows that adults between 70 and 90 years of age have far more difficulty repeating sentences with complex subjects than sentences with complex objects. In fact, in 97% (69/75) of the cases, the elderly eliminated the complex subjects from the repetition, whereas there was no such tendency to eliminate complex objects. A similar difficulty was found in timed reading comprehension tasks by Kynette & Kemper (Reference Kemper1986). The same pattern is observed in disfluencies in non-elderly adults. For example, Clark & Wasow (Reference Clark and Wasow1998) show that in spontaneous unmonitored speech, adult speakers of English are more likely to experience a disfluency with subjects than with objects. Interestingly, disfluency rates for both simple and complex NPs show a monotonic decrease across the course of the sentence, in the order topic >subject >prepositional object. Kluender argues that this trend mirrors his results of adult comprehension studies using N400 amplitude as a measure of referential and lexical processing cost. Finally, Garnsey (Reference Garnsey1985), Kutas, Besson & Van Petten (Reference Kutas, Besson and Petten1988), and Van Petten & Kutas (Reference Van Petten and Kutas1991) show that the processing of open-class words, particularly at the beginning of sentences, cause a significantly greater processing effort than closed-class words.

I note that there are other sources of support for Kluender's claim that complex subjects impose a specially heavy processing burden. Ferreira (Reference Ferreira1991) and Tsiamtsiouris & Cairns (Reference Tsiamtsiouris and Cairns2009) found that speech initiation times for sentences with complex subjects were longer than for sentences with simple subjects, and Amy & Noziet (Reference Amy and Noziet1978) and Eady & Fodor (Reference Eady and Fodor1981) provide evidence that sentences with center-embedding in subjects are harder to process than sentences with center-embedding in objects.

It is intuitively plausible that a sentence that starts with a complex structure will consume more processing resources than one that starts with a simple structure and ends with a complex structure. This is because the speaker's memory resources are strained sooner in the sentence, and longer, since those resources are not available for processing the remainder of the sentence. And the fact that complex subjects are harder to process than complex objects is consistent with the fact that extraction from objects is easier than extraction from subjects.

Consequently, speakers will prefer to use sentences with complex structures toward the end of the sentence, thus creating a temporally shorter strain on memory resources (see Yngve Reference Yngve1960; Hawkins Reference Hawkins1994; Wasow Reference Wasow1997, Reference Wasow2002; and Gibson Reference Gibson1998, among others). This account can shed light on subject island effects as well as on the existence of graded exceptions. Because subjects are inherently more difficult to process, filler–gap dependencies in them are harder to maintain in working memory without additional support. If the sentence in question happens to contain other elements that are independently harder to process (e.g. infrequent words, complex structures, semantically unexpected constructions, misleading prosody, or stylistically awkward phrasings), then the combined effect causes difficulty. Conversely, sentences with highly frequent words (e.g. collocations), discourse-old nominal phrases, and semantically simpler expressions (such as pronouns), and prosodic phrasings that cue the location of the gap will cause less strain on the cognitive resources needed to process the filler–gap dependency.

Another factor that may play a role in the asymmetry between subject-internal gaps and object-internal gaps is that the former gaps are sentence-medial and the latter tend to be sentence-final. This can make a difference. If there is no material after the gap then there is no risk of creating a garden-path. Conversely, if the gap is sentence-medial, there is a good chance that the gap is not recognized. This is illustrated by the contrasts in (24), where the brackets indicate prosodic boundaries.

1. (24)

   1. (a) ?*[Which cars did the explosion bulge the hoods of a good four inches]?

   2. (b) [Which cars] [did the explosion bulge the hoods of] [a good four inches]?

   3. (c) ?*[Who did you give to the canary that I brought home yesterday]?

   4. (d) ?[Who did you give to] [the canary that I brought home yesterday]?

In all of these data, the object-internal gap is sentence-medial and strands the preposition. Without the proper intonation, these sentences obtain low acceptability because there is a tendency for the preposition to combine with the following phrase. The oddness of such examples is related to the Clause Non-final Incomplete Constituent Constraint (Kuno Reference Kuno1973a), which was originally intended to be a more general version of Ross's (Reference Ross1967) Sentential Subject Constraint. Interestingly, there is evidence that Kuno's constraint is at least partially a performance effect, since it can be weakened with the proper intonation and a judicious choice of words. Consider the acceptable example in (25), where the brackets indicate prosodic phrasing.

1. (25) [Who_i did you offer to __ _i] [the chance to win $1,000]?

Although this sentence is isomorphic to (24c, d), it is much easier to process. Hukari & Levine (Reference Hukari and Levine1991) and Fodor (Reference Fodor, Goodluck and Rochemont1992) note various counterexamples to Kuno's Clause Non-final Incomplete Constituent Constraint, such as (26).

1. (26)

   1. (a) Who did Kim argue with about politics?

   2. (b) Who did Kim have an argument with about politics?

   3. (c) Who did you appeal to to get the requirement waived?

   4. (d) Which company did you persuade the director of to make an appearance?

What these data suggest is that there is a general difficulty in identifying the location of a gap when it is sentence-medial (be it subject-internal or not). This is a likely contributing factor for why it is usually harder to extract from subjects than from sentence-final objects.

Although the proposal in Kluender (Reference Kluender, Chand, Kelleher, Rodríguez and Schmeiser2004) is plausible, there are two problems which indicate that there are additional factors at work. First, Kluender admits that nothing explains parasitic gap effects (e.g. in which case the existence of a second gap outside the subject causes the sentence to be acceptable). Second, it is not clear how simple sentences like (27) can lead to sufficient strain on memory and processing resources, causing the parser to crash.

1. (27)

   1. (a) *What did the owner of sneeze? (the owner of the cat sneezed)

   2. (b) *Who did that Bill married surprise you? (that Bill married Mia surprised you) (Huang Reference Huang1982: 495)

It seems implausible that sentences that are so short can cause such a cognitive breakdown, given that speakers regularly process much larger sentences with much larger subject phrases. Although I agree with Kluender (Reference Kluender, Chand, Kelleher, Rodríguez and Schmeiser2004) that the processing strain caused by complex subjects has a central role in subject island phenomena, it is in my view unlikely that the mere presence of a filler–gap dependency in the subject phrase exhausts memory resources so severely as to cause subject islands effects.

In what follows I concentrate on subject islands and hope to convince the reader that the range of acceptable subject island violations is substantially wider than previously held. This means that the strength of so-called ‘strong islands’ has been overstated in the literature, and that no previous account discussed so far can cope with the full range of facts. I then proceed to show how the processing account put forth by Kluender (Reference Kluender, Chand, Kelleher, Rodríguez and Schmeiser2004) can be expanded so that all the subject island phenomena under discussion are predicted.

3.1 Factors that facilitate the processing of subject island violations

In a performance-based account of subject islands like Kluender (Reference Kluender, Chand, Kelleher, Rodríguez and Schmeiser2004), it is expected that subject island violations involving non-finite subjects are easier to process than those with finite or nominal subjects. This is because the former are the weakest, semantically, as they carry no referentiality (Kluender Reference Kluender, Goodluck and Rochemont1992, Reference Kluender, Culicover and McNally1998). This is independently supported. First, extraction from NPs is known to often be difficult even in object position (see Section 2 above). Second, finite object relative clauses are also well-known to impose stronger extraction limitations than non-finite object relatives (Ross Reference Ross1967, Engdahl Reference Engdahl1983), and are harder to process than non-finite clauses (Kluender Reference Kluender, Goodluck and Rochemont1992, Gibson Reference Gibson, Marantz, Miyashita and O'Neil2000). Third, non-finite subjects are the only environment where ‘filled-gap’ effects have been detected (Stowe Reference Stowe1986, Ellis Reference Ellis1991, Pickering et al. Reference Pickering, Barton, Shillcock, Clifton, Frazier and Rayner1994, Phillips Reference Phillips2006).

In what follows, I focus on three independently motivated factors that seem to contribute to the relative acceptability of subject island violations. These factors are particularly evident in NP extractions from subject NPs like the ones in (31) and (32): (i) the specificity of the wh-phrase; (ii) the highly coherent pragmatic relevance that the extracted phrase bears to the subject and the predicate; and (iii) the syntactic cues introduced by prosodic phrasing. To be clear, none of these factors is directly connected to any grammatical condition on subject islands, nor is it by itself sufficient to completely mitigate a subject island violation. Collectively, however, they can cue the correct parse and aid the processor in roughly the same way that double center-embeddings can be made easier to process with similar cues, as discussed in Section 2.3 above. Crucially, the ameliorative role of specificity, relevance and prosody is independently motivated, since it is known to aid the processing of a wide range of other complex constructions.

4.1 Cognitive and pragmatic reasons for avoiding subject-internal gaps

As discussed in Section 2.3 above, there is evidence that sentences with complex subjects are significantly harder to process than sentences with complex objects, and that sentences with medial gaps are more difficult to process than sentences with sentence-final gaps. Hence, there are cognitive reasons for such structures to be avoided. Below, I note other factors that may further discourage speakers from using complex sentences with subject-internal gaps, one cognitive and one pragmatic.

As (60) illustrates, sentences with subject-internal gaps require discourse contexts which are fairly complex as well. This added complexity is bound to strain the processor further.

1. (60) [Biology exam question]

   For quite some time now, medical science has been seeking cures for the flu and for tuberculosis. Current research indicates that the cure for one of these diseases may be possible to find in the near future, but not for the other.

   Which disease do you think the cure for will not be available in the near future?

As a consequence, there is an additional reason to avoid complex sentences that have subject-internal gaps. Hence, speakers will plan and package their discourse differently, using paraphrases that are simpler to plan and produce, and avoid discourse like (60).

Moreover, it is often the case that a subject-internal gap serves no special communicative purpose. Consider, for example, the possible alternative continuations in (61).

1. (61)

   1. (a) I asked Fred to write me two lines of poetry. But instead of that...

   2. (b) … he wrote me a letter, and every single line of it was a piece of poetry.

   3. (c) (?) … he wrote me a letter of which every single line was a piece of poetry.

   4. (d) ?* … he wrote me a letter which every single line of was a piece of poetry.

Although both (61c, d) are suitable continuations for (61a), they compete with the much simpler paraphrase in (61b), which involves a discourse-old pronominal and no extraction. Consequently, (61a) is not only easier to process, but also pragmatically more felicitous than (61c, d). Note also that the PP extraction in (61c) is more helpful for the processor than the NP extraction in (61d): in the latter the extracted phrase is maximally uninformative about its grammatical role.

There is no semantic or pragmatic motivation to use (61c) instead of (61b), nor to use (61d) instead of (61c). To be clear, my claim is not that the existence of a simpler paraphrase causes the more complex paraphrase to be unacceptable. Rather, the point is that simpler paraphrases are preferred over more complex ones, which in this case adds further pressure for the lower frequency of sentences with subject-internal gaps. In sum, the evidence above suggests that there may be Gricean and stylistic reasons for speakers to avoid continuation paraphrases like (61c, d).

Let us take stock. There are cognitive and pragmatic reasons for speakers to avoid complex subjects that contain filler–gap dependencies. From this it follows that under real-time communicative pressure speakers will package discourse in a simpler way and resort to sentences with simpler (gapless) subject phrases. This leads to extremely low (near zero) frequency of sentences with subject-internal gaps. As a consequence of this, it is possible that the language processor develops a parsing heuristic in which subjects are not expected to contain gaps. Such a heuristic is useful because it efficiently prunes the search space of filler–gap dependencies and speeds up processing. I now turn to this hypothesis in more detail.

4.2 Extremely low frequency leads to parsing expectations

Fodor (Reference Fodor1978, Reference Fodor1983), Berwick & Weinberg (Reference Berwick and Weinberg1984), and Hawkins (Reference Hawkins1999) propose that processing difficulties might lead to the grammaticization of islands. This position is excessive for subject islands given that there are various grammatical exceptions (both parasitic and non-parasitic) as shown in Sections 2 and 3 above. An alternative is to assume that islands are extremely heterogenous phenomena that depend on the particular lexical and constructional types, as in Postal (Reference Postal1998), but such an account would lack explanatory power since it would merely stipulate which specific words and constructions block extraction.

I propose that the extremely low frequency of extractions from subjects causes speakers to create a processing heuristic about the absence of gaps in subject phrases. The rationale is the following. If a given construction is hard to process for cognitive reasons, this makes it less frequent, and therefore less expected, which in turn may cause it to be harder to process. This in turn makes the construction even less frequent, less expected, and so on. The acceptability of subject-internal extractions is possible, however, in the presence of cues that identify the correct parse and aid the filler-gap processing mechanism in coping with the unexpected location of the gap. I conjecture that different languages can create different parsing heuristics for filler–gap dependencies, informed by processing considerations, frequency, and pragmatic conditions, in the spirit of Hawkins (Reference Hawkins2004). For example, Engdahl (Reference Engdahl1983: 9–11, 29) notes that there is a hierarchy for the acceptability of parasitic gaps, and that the break between acceptable and unacceptable extraction domains seems to be language-specific. Engdahl (Reference Engdahl1983: 29) plausibly conjectures that these facts reflect parsing principles.

Given the view that I am advocating here, subject island violations are grammatical structures that have different degrees of acceptability in proportion to the amount of reprocessing needed for the correct parse to recover from the expectations that it has about the location of the gap. Suppose that the parser has encountered a filler, a wh-phrase of some kind. This causes it to actively seek a gap in the remainder of the sentence. When a subject phrase is encountered, a problem presents itself: either the subject contains the gap for that filler or it does not. An efficient decision based on the frequency and distribution of gaps in English is to assume the latter. Subject-internal gaps are so rare that this processing strategy will reliably speed up processing in the vast majority of cases, and therefore help the processor operate efficiently under real-time communicative constraints. The proposed processing strategy is formulated in (62).

1. (62) Gapless subject heuristic

   Expect subject phrases to be gapless.

It is important to stress that this is not part of the grammar. In other words, this is not a constraint on well-formed English syntactic structures, but rather, the preferential route taken by the parser when seeking a gap for a filler stored in memory. Crucially, (62) has the corollary in (63) given that the absence of a subject-internal gap logically leads to the expectation that the gap is located in the subject's sister phrase.

1. (63) Corollary

   Expect any gap to reside in the sister of the subject phrase (i.e. the VP).

Violating any or both of these expectations does not lead to ungrammaticality, since these are parsing rules of thumb rather than grammar rules. The combined product of (62) and (63) is a digging-in effect: the parser is committed to the prediction that no gap will be found in the subject and that a gap will be found in the VP. The longer the parser remains committed to these expectations the stronger the digging-in. In turn, speakers will experience a strong garden-path when the two expectations about the gap location turn out to be incorrect. When there are prosodic, pragmatic and contextual cues which boost the grammatical parse – as discussed in Section 3 – the digging-in effect is naturally weaker and therefore it is easier for the parser to backtrack. But in the absence of such cues, the parser experiences a strong digging-in effect which hampers backtracking.

Let us consider some examples. Uttered with neutral prosody, (64a) violates two expectations: the expectation that the subject is gapless and the expectation that the verb phrase contains a gap. As a consequence of the expectations, the wh-phrase is retrieved in the VP but the parse crashes because the legal gap is located in the subject instead.

1. (64)

   1. (a) *Who did the rivals of __ shoot Castro?

   2. (b) Who did the rivals of __ shoot?

   3. (c) Who did the rivals of Castro shoot __?

Violating two expectations leads to a stronger digging-in effect than violating just one, and therefore to a stronger garden-path effect. Consequently, (64a) is harder to process than (64b). The relative acceptability of (64b) follows from the fact that the two gaps are in very close proximity (which causes less digging-in) and are referentially linked to the same filler. It is known that referents that have been recently accessed in memory are easier to reactivate and process (see e.g. Vasishth & Lewis Reference Vasishth and Lewis2006). Hence, the filler is reactivated as the parser finds each gap, and processing the filler–gap dependency is facilitated. The overall effect is the illusion that the subject gap in (64a) is somehow grammatically different from the subject gap in (64b).

Cases like (19b), repeated as (65), are harder than (64b) because the two gaps are not referentially linked, but are still easier than (64a) because only one expectation is violated.

1. (65) [People that sensitive]_i, I never know [which topics]_j jokes about ___j are likely to offend __ _i. (=ex. (19b) above)

The same explanation applies to pairs of sentences like (66).

1. (66)

   1. (a) What did the attempt to repair __ ultimately damage __?

   2. (b) What did the attempt to repair the car ultimately damage __?

In (66a) only one expectation is violated and the two co-referential gaps are in very close succession. As a consequence, this structure cannot yield a very strong digging-in effect, and therefore it is easy for the parser to overcome the gapless subject expectation and consider the possibility that there is a gap in the non-finite phrase. In (66b) there is more material between the filler and the gap but since the structure is not excessively complex and both gap expectations are met, no major difficulty arises.

It should be clear that the function of the rule in (62) is to aid processing. This heuristic is successful in the vast majority of cases, since subject-internal gaps are virtually absent from normal discourse. In the presence of a subject-internal gap, the heuristic backfires and misleads the parser. Digging-in effects and the extra processing cost incurred by reprocessing cause a complex sentence to be even harder to process, which leads to lower acceptability. If there are no strong cues for the existence of a subject gap as discussed above, then the expectations that the subject is gapless and that the gap is located in the VP are too strong for the parser to overcome.

A heuristic-based analysis might also offer an account of adjunct islands and their parasitic and non-parasitic exceptions. If gaps are expected to reside in verb phrases rather than in their adverbial adjuncts, then we predict that non-parasitic extraction from adjuncts is rather difficult (but not impossible, under ideal conditions, which may depend on how well the structure coheres pragmatically). As in the case of complex subject phrases, the reprocessing costs caused by violating such a heuristic compound with the overall processing costs of the sentence, and create a complex landscape of graded acceptability. This is left for future study.

There is a wide range of independent psycholinguistic motivation for the existence of frequency-based parsing preferences of the kind that I advocate here. Direct evidence is provided by Vasishth & Lewis (Reference Vasishth and Lewis2006), who show that sentence processing can speed up when the parser encounters certain expressions that reinforce already existing expectations. Language-specific parsing preferences also seem to be at work in relative clause attachment biases: in languages like Spanish, Dutch, French, German, and Japanese there is a preference of ‘high attachment’ over ‘low attachment’ for relative clauses (Hemforth et al. Reference Hemforth, Konieczny, Scheepers, Strube and Hillert1985, Cuetos & Mitchell Reference Cuetos and Mitchell1988, Brysbaert & Mitchell Reference Brysbaert and Mitchell1996, Kamide & Mitchell Reference Kamide and Mitchell1997, Zagar, Pynte & Rativeau Reference Zagar, Pynte and Rativeau1997). The opposite preference is observed in Arabic, English, Norwegian, Romanian and Swedish (Frazier & Clifton Reference Frazier and Clifton1996, Ehrlich et al. Reference Ehrlich, Fernández, Fodor, Stenshoel and Vinereanu1999, Quinn, Abdelghany & Fodor Reference Quinn, Abdelghany and Fodor2000, Fodor Reference Fodor and Hirotani2002a). The strength of this preference also seems to vary across languages. See Fodor (Reference Fodor and Hirotani2002a) for an account of these phenomena in terms of prosodic phrasing preferences.

Further experimental evidence for parsing preferences in language is provided by Norcliffe (Reference Norcliffe2009), who shows that the asymmetric frequency distributions of gaps and resumptive pronouns within and across various languages can be explained by the conventionalization of processing preferences, along the lines of Hawkins (Reference Hawkins1999). Another example of parsing preferences and conventionalized probabilistic information in language can be observed in the ordering of NP and PP in English VPs. The canonical ordering is NP–PP, but it is often reversed if the NP is significantly longer than the PP. Wasow (Reference Wasow and Elliot2007) reasons that this canonical but violable ordering tendency should be seen as part of the grammar of English. Crucially, this is gradient information: the preference for NP–PP ordering has a magnitude. By adopting the NP–PP ordering as a default, the grammar usually adopts the ordering that allows the earliest possible identification of the categories of the daughters of the node being constructed. Conversely, if the grammar were categorical rather than gradient, then it either would have to stipulate one ordering (incorrectly ruling out the other) or say nothing about the ordering (incorrectly predicting that shorter constituents will consistently precede longer ones). For further evidence of conventionalized probabilistic information in English dative alternations see Bresnan et al. (Reference Bresnan, Cueni, Nikitina, Baayen, Boume, Kraemer and Zwarts2007) and Bresnan & Ford (Reference Bresnan and Ford2010). See also Jurafsky (Reference Jurafsky1996) for an overview of the role of probabilistically-enriched grammars in models of human language and in models of language processing.

4.3 Formalization

There are at least two ways to implement the heuristic proposed in (62). One possibility is to extend grammar rules with frequency information that corresponds to predictive information. In other words, the grammar rules that introduce subject phrases would be augmented with a convention specifying that gapped subject phrases have virtually zero frequency (or that gapless-subject phrases are the norm). This can be stated in a straightforward way in non-transformational frameworks such as Head-driven Phrase Structure Grammar (HPSG; Pollard & Sag Reference Pollard and Sag1994). For example, one can augment the phrasal rules that introduce subject phrases and state that the likelihood of such phrases being [gap {}] (i.e. gapless) is close to 1. Or, alternatively, that the likelihood of such phrases being specified as [gap {… X …}] (i.e. containing gaps) is close to 0. See Ginzburg & Sag (Reference Ginzburg and Sag2000) and Sag (Reference Sag2010a) for discussion about how filler–gap dependencies are modeled in HPSG. As discussed above, such a convention follows from the extremely low distributional frequency of subject-internal gaps, which in turn follows from processing and pragmatic factors. This account would allow for constructionally-specific differential biases (i.e. stronger subject island violations in some constructions and weaker in others).

Another possibility is that (62) is a parsing rule completely separate from the grammar. This approach seems to be more in line with Engdahl (Reference Engdahl1983: 29), and would not allow constructionally-specific differential biases. Again, this could be formalized straightforwardly in theories like HPSG by stating a rule like (67), where the slash ‘/’ operator is a persistent default constraint in the sense of Lascarides & Copestake (Reference Lascarides and Copestake1999) .

1. (67) verb-lxm → [valence 〈XP[gap/{ }] …]

Basically, this rule states that for every verbal lexeme, the subject valent is by default expected to be [gap {}]. By stating the constraint over the valence feature (Sag Reference Sag2010b) rather than over the arg-st feature, we allow extraction from ‘subjects’ of there-insertion, as in examples like (68).

1. (68) Who was there a picture of __ on the wall? (Stepanov Reference Stepanov2007: 32)

As before, the default parsing constraint in (67) could be associated with frequency information, if needed. I illustrate the effect of either account in Figure 1. The filler–gap expectations are highlighted, for perspicuity.

Figure 1 Application of parsing heuristics during sentence processing.

As the parser encounters the filler who, it creates the unbounded dependency recorded in gap. Informally, this states that there is an NP_i gap embedded in the sentence. The information about the missing phrase percolates down in the structure as specified in Ginzburg & Sag (Reference Ginzburg and Sag2000) and Sag (Reference Sag2010a). As the parser encounters the auxiliary verb, the (top–down) expectation that the gap is not located in the subject phrase is created. This in turn leads to the expectation that the gap must be located in the verb phrase, since there is no other logical possibility. If the subject NP turns out to not contain a gap, and there is a gap in the VP, then the expectations lead to the correct parse. This is by far the most frequent situation, which is why (62) is a useful rule-of-thumb for sentence processing. If both the NP and the VP each contain a gap, then some extra processing is needed in order to overcome the digging-in effect and backtrack enough to link the filler to the subject- internal gap. If two gaps are co-referential and in close proximity then this backtracking is facilitated. Finally, if the subject contains a gap but the VP does not, then stronger digging-in effects arise. However, by employing prosodic cues and fillers that are highly relevant for the predicate and the subject, it is possible to boost the subject-internal gap parse and to reduce the digging-in effects, in which case backtracking is not so heavily preempted. As a consequence, the island violation can be relatively acceptable.