Monday, December 29, 2014

If God is in the details then Evans fills a badly needed gap in the literature

If God is in the details, Vyvyan Evans’ writings on generative linguistics are profane. What follows is a little illustrative example and how bad the arguments are.

As mentioned before (see here, here, and here), Evans, like many before him, seems to have difficulty understanding the concept of a universal as used by generative grammarians of the Chomsky stripe. Evidence? Consider this discussion of the wh-island constraint (76-7 in his book; in a chapter entitled “Are There Linguistic Universals?”). Evans illustrates the constraint using the following contrasting English examples:

(1)  a. Where did the supermodel say that the window cleaner had to get off the train to meet her
b. * Where did the supermodel say whether the window cleaner had to get off the train to meet her

Evans argues that typological considerations prove that the contrast above cannot be attributed to a universal “rule” because it fails to hold in “other Indo-European languages such as Italian and Russian.”

Let me start with what may seem a persnickety comment: The wh-island constraint is not often thought of as a “rule” but as a condition on rules. This may sound like an innocuous distinction, but it actually isn’t. It reflects what I think is the basic underlying confusion permeating Evans’ discussions: the inability to distinguish Greenberg Universals (which aim to describe surface patterns) and Chomsky Universals (which aim to describe the generative properties of Gs and FL).  The former are intended to be surface true, the latter cannot be. While there is a direct relationship between a surface pattern and a rule in the case of Greenberg Universals, no such direct relation exists in the second.  Thus, as a matter of logic, it takes more than a review of contrasting surface patterns to debunk a Chomsky universal. It requires a discussion of the rules that generate said surface forms (i.e. a discussion of the generative procedures (i.e. Gs)). As we shall see, Evans’ discussion is entirely oblivious to this, and this makes his critique entirely worthless, as will become clear as we proceed.

So, let’s return to Evans’ discussion. What’s the “invalid” universal rule that Evans aims to debunk? It’s the following: A wh-word cannot intervene between the two clauses in a question. What’s the problem according to Evans? In some languages [Italian and Russian-NH], “a wh-word can intervene between the two clauses in a question” (77). So, according to Evans the universal rule says “no intervening wh-words between clauses in questions” and Italian and Russian allow such intervening wh-words. Thus, the universal rule is invalid.  In other words, by examining the surface forms in three different languages Evans concludes that a proposed universal, motivated by English data, is not universal after all.

Evans’ argument here is very instructive.  For it fails in almost every way imaginable, as I will show. But before deconstructing it, let me start by saying that there is nothing wrong at looking at more than one language to investigate linguistic universals.[1] Generativists do this all the time. As David Pesetsky notes here  even big bad Chomsky thinks that this is a very good thing to do (Evans seems loath to concede this point despite his evident misquotation (please take a look at the Facebook discussion linked to above. This is really egregiously bad behavior on Evans’ part)). So, the idea that cross linguistic investigations are relevant to establishing universal properties of Gs and FL is as close to a truism as there is today among practicing generativists.  However, if you do this, you need to do this right, and, sad to say, Evans’s discussion is seriously defective. Let’s count the ways.

First, as already noted, Evans’ criticism understands the universal to be one about surface forms (viz. can a wh-word intervene between “two clauses in a question”). Sadly, as it stands, this is an incorrect description of the relevant phenomenon. How do I know? Because sentences like (2) are fine in English even though “a wh-word intervenes between two clauses in a question”:

(2)  Who did the supermodel ask whether the window cleaner had to get off the train to meet her

The correct description of the phenomenon Evans is interested in requires fixing where the wh comes from and this requires more than mere surface description. In particular, it requires determining the underlying source of the wh-word (i.e. roughly its DS-position). The generalization that covers (1) and (2) is that it the wh that is sentence initial cannot hail from the embedded clause. Note that in (1b) the sentence is only unacceptable if where is querying the getting-off. The sentence is fine if we are questioning where the saying occurred.  So the structure that is ungrammatical is (3a) (corresponding to (1a)) with the indicated trace annotating the relevant illicit dependency while (3b) and (3c) are fully grammatical.[2] This is why the matrix reading of (1b) where where modifies the saying is fully acceptable. I leave as an exercise for you (and Evans) to explain why (2) is also fine and fully acceptable (hint: note where the trace of where sits in (3c)).

(3)  a. Where1 did the supermodel say that the window cleaner had to get off the train t1 to meet her
b. Where1 did the supermodel say t1 that the window cleaner had to get off the train to meet her
c. Who1 did the supermodel ask t1 whether the window cleaner had to get off the train to meet her

So, first conclusion, Evans’ discussion does not really describe the data correctly and furthermore to do the English data descriptive justice requires looking at more than the surface (string) features of the sentence. We need, at least, sound-meaning pairs to get the data described correctly and this requires some conception of underlying form, something that is not string visible. This is what good typological work currently does and Evans’ discussion does not.

Second problem: Evans’ discussion gets the facts wrong even allowing for the first adjustment. The contrast he claims to hold between English on the one hand and Italian/Russian on the other does not hold in questions.  Sentences analogous to (1) with structure (3a) are unacceptable in Italian/Russian too. Question formation in both languages appears to yield unacceptability when extracting one question ­wh over another question wh.[3] What Evans probably meant to report is that the wh-island condition fails to appear in the Italian analogues of (4):

(4)  The book that John asked whether to review

This was first noted by Rizzi in his justly celebrated paper on island effect variations. Rizzi noted that wrt relativization (not question formation) it appears to be possible to extract the relative operator out of the embedded question with acceptable results.  As Grimshaw noted not long afterwards, examples like (4) also seem pretty acceptable in English, so Rizzi’s noted contrast between the two languages might be inaccurate (I for example find (4) quite acceptable).  At any rate, this is the kind of counterexample to the wh-island constraint that Generativists started studying in the mid 1980s. Research led to a proposal that largely saved the universal principle. We return to this in a moment, but first another problem with Evans’ set up of the discussion.

Note that in the example in (4) the head of the relative clause controls an argument position inside the relative (the object of review). In the examples in (1), where is an adjunct. There are well known differences between arguments and adjuncts as they relate to islands, viz. adjuncts are far more susceptible to the wh-island condition than arguments are. Thus, (5) is considerably less acceptable than (4) (again with the head modifying the place of the interview (viz. roughly, a relativized version of “John asked whether to interview MD in NYC):

(5)  ?*The city that/where John asked whether to interview Moby Dick

So, Evans’ illustrative examples are triply unfortunate: they mis-describe the typological contrast (questions are uniform across the three cited languages wrt unacceptability), they mis-describe the generalization (the underlying source of the wh is critical), and they focus on the wrong cases as the contrast of interest emerges largely with argument extraction, adjuncts being more recalcitrant and quite uniform in their behavior cross-linguistically.

This noted, let’s put these details aside and simply assume that Evans’ discussion does not go off the rails from the get go (though it does and this should tell you something about whether Evans’ criticism is serious (which, of course it isn’t) given that even the simple description of the generalization it “debunks” is so inaccurate), though It should make you wonder how trustworthy the critic is if he can’t get the basic descriptive facts right.

Ok, where are we? We have a purported difference between English on the one hand versus Italian (and Russian) on the other concerning extraction out of embedded questions in relative clauses. Does this debunk the universal as Evans’ discussion claims? Not really. The whole discussion, as I noted, was initiated by Rizzi in the context of grounding Ross-like Island generalizations more deeply in a more general theory of locality.[4] Here’s a slightly ahistorical reconstruction.

Rizzi noted the contrast between English and French regarding extraction out of wh-islands. He offered an explanation for this that preserved an important linguistic universal (the subjacency condition) by allowing the category of bounding nodes to differ across Gs (CP and DP for Italian, TP and DP for English).  Given this parametric variation, both English and Italian (and Russian) Gs obeyed the same universal subjacency condition (i.e. movement across more than a single bounding node is illicit in all Gs). In other words, the relevant generalization due to Rizzi is that there exists a universal structural condition on movement that is not in any way undermined by the observed differences between Italian and English that Rizzi reported. As you can see, this universal is very abstract (it relates to G processes and structures, not output forms) and cannot be contested by citing surface differences the way that Evans’ discussion does.[5]

In short, even when corrected for the evident mis-descriptions, Evans’ discussion is simply irrelevant to what generative grammarians have understood universals to be. Thus, Evans’ discussion is just another example of the confusion rampant in his writing between structural universals of the Chomsky variety (that have to do with properties of Gs and FL) and surface typological universals of the Greenberg variety (that mainly describes the string properties of surface forms). And this is a very big deal. It indicates that Evans’ discussions (aside from indicating a lack of fluency with the relevant literature) is simply beside the point logically. His criticisms miss the mark because they are not targeted at the right concept of universal.

Let me put this another way: Evans writes as if differing typological patterns are in and of themselves problems for the generative conception of universals. But this is to misunderstand what a grammatical universal is. It is not the description of patterns in the data, but principles of grammatical organization (descriptions of generative procedures). In other words, though Greenberg universals can be relevant to evaluating Chomsky universals, it takes a lot of grammatical work to relate them. Evans’ argument does not do any of this work. Why? Because it fails to recognize the difference between the two kinds of universals and hence fails to understand what is logically required to show that a Chomsky (grammatical) is incorrect. This makes Evans’ critique similar to Emily Litela’s confusion about “soviet jewelry” (here), though Emily’s misunderstanding is far more amusing (and, deliberate, unlike Evans’ critique).
So, here’s the bottom line of our little illustration: Evans’ specific “criticism” of the work on wh-island effects in generative grammar is deeply misguided. How deep? Well, the discussion wins the junk argument trifecta: it is inadequate descriptively, theoretically and logically. In other words, this is intellectual garbage, pure and simple. His discussion here is not serious and the charitable should simply ignore it.  I would have done so myself (indeed, I really want to ignore it) had it remained justly obscure. But it did not. Rather, Evans’ critique has come to fill a badly needed hole in the literature.[6] If only that hole were still unfilled. Make sure you mention this to anyone that suggests otherwise.[7]

[1] I should add, perhaps playing into Evans’ hands, that I personally do think that one can argue for universals based on investigations of a single G (note, G, not language). This is what POS arguments do all the time. Of course, no single argument need be dispositive and it is always worth finding other kinds of evidence for a proposed universal. But logically speaking, investigating one G in depth can serve to ground a universal, not unlike studying just one organisms in depth, say the fruit fly or e-coli or pea plants, can serve to ground biochemical or genetic generalizations that hold across many phyla.
[2] Structures are ‘grammatical’ or not, sentences are ‘acceptable’ or not. Linguists explain unacceptability in part via the grammaticality of the structures they supervene on. But the two notions are distinct and must be kept conceptually separate.
[3] Of course, whether these derivations are ungrammatical in Italian/Russian Gs is a further question.
[4] I say “Ross-like” as Ross himself did not think that the wh-island condition was a true island. It was added by Chomsky later on based on the mechanics of the theory of subjacency.
[5] There are other accounts for exceptions to wh-island effects involving the number of “escape” hatches in a given G. Reinhart initiated this line of analysis and it is still much with us (under the name of ‘phase edges’). At any rate, this line of inquiry also preserved the subjacency condition by parameterizing the number of “escape” hatches in CP a given G allows.
[6] I comment I heard attributed to a review by Quine. Great line!
[7] Last point: the wh-island condition has been widely discussed in the generative literature. Exactly how to formulate it is still subject to lots of discussion.  The above is not intended to defend nor debunk it. My sole intention has been to show that whatever the right answer turns out to be, Evans’ kinds of considerations are conceptually incapable of furthering the discussion.

Saturday, December 27, 2014

The 4th Hilbert Question: Is There Repair by Ellipsis

There has been a little Evans induced hiatus from the Hilbert Project. This is too bad, for unlike the Evans' stuff, this is both interesting and valuable. At any rate, I will try to be more diligent in posting these more frequently. Today Craig Sailor and Carson Schütze discuss Ellipsis and one of the leading theories we have about it.  Hopefully, this will generate some vigorous discussion.


Is there repair by ellipsis?*

Craig SailorUniversity of Groningen

Carson T. SchützeUCLA
     Ross (1969) observed that an island violation could apparently be overcome or “repaired” if the island were deleted, as in sluicing, where wh-movement occurs out of an elided structure. This could be superficially described as an instance of ellipsis “feeding” an otherwise-illicit application of movement. Over the years, this phenomenon has been thought to extend beyond islands: ellipsis can apparently feed or repair a variety of illicit movements, including unlicensed instances of (multiple) focus movement, head movement, etc. (Merchant 2010).
     Recently, though, several authors have leveled convincing arguments against the original claim that ellipsis can repair island violations, showing apparent examples to be illusory (Barros to appear, a.o.). With the foundation of elliptical repair in doubt, the following question arises: To what extent, if any, can ellipsis make an ill-formed structure acceptable?

        The structure in (1), in which an XP has moved out of some elided YP, has been ascribed to several ellipsis phenomena (some involving more than one moved XP):

(1)       XP … [YP … tXP …]

We refer to this as the “move-and-delete” derivation.
Sluicing was the first phenomenon to be assigned this derivation, following Ross (1969). Sluicing (2) is widely thought to involve wh-movement out of an elided constituent (3):

(2)      Who?
(3)      Whoi [ did he see ti ] ?

English would require this movement of who independent of the ellipsis; thus, sluicing appears to involve incidental co-occurrence of two discrete syntactic operations (movement and deletion).
     Ross noted, though, that the situation might not be this simple. In unreduced wh- questions, wh-movement is prohibited out of islands (e.g. coordinate-structure islands (4)); however, sluiced analogues of such questions are often perfectly acceptable (5).

[Context: John saw Mary and someone else.]
(4)      *Who did he see her and?
(5)        Who?

According to Ross, (4) and (5) differ only in what is pronounced, meaning the two share a structural description containing an island violation (but see below). This led to the proposal that island violations could be mitigated if the offending islands were deleted, which came to be known as “elliptical repair” of island violations (see Merchant 2001 for extensive discussion, and Fox & Pesetsky 2005 for a generalized approach based on “cyclic linearization”).
Since Ross (1969), the scope of elliptical repair has expanded beyond amelioration of apparent island violations. In the ellipsis literature, move-and-delete analyses of many different phenomena invoke movements that would yield ungrammatical sentences if ellipsis were not applied, even without an island present. Consider Merchant’s (2004) influential move-and-delete analysis of fragment answers: Merchant argues convincingly that fragment XPs originate within clausal answers, but escape ellipsis of these clauses via movement (6). Crucially, (7) shows that this movement is ungrammatical without ellipsis.

[Context: How is John feeling?]
(6)       Sicki [ he is feeling ti ].
(7)      *Sick he is feeling.

This ellipsis dependency is common among analyses of move-and-delete phenomena, including pseudogapping (8) (Jayaseelan 2001, echoed by Merchant a.o.) and multiple fragment answers (10) (adapted from Merchant 2004:711):

(8)        John won’t read magazines, but he will booksi [ read ti ].
(9)      *John won’t read magazines, but he will books read.

(10)   Q:    Wer         hat  gestern    wen        gesehen?            German
                      who.nom  has yesterday who.acc seen
                     ‘Who saw whom yesterday?’

              A1:   Der        Manni den       Jungenj  ti  hat gestern      tj  gesehen ].
                            the.nom man     the.acc boy            has yesterday     seen
              A2:  *Der       Mann den       Jungen hat gestern     gesehen.
                      the.nom man    the.acc boy      has yesterday  seen

The same state of affairs arises in well-motivated proposals for gapping (Coppock 2001), stripping (Depiante 2000), apparent non-constituent coordination (Sailor & Thoms, to appear), and other ellipsis phenomena (see Thoms, to appear, and Merchant 2010). Space restrictions preclude exemplifying each phenomenon, but they can all be shown to involve the derivation in (1), and, in each, ellipsis behaves like a well-formedness condition on the ellipsis-dodging movement: it facilitates convergence of a structure that is otherwise ill-formed. This is strongly reminiscent of apparent island repair in sluicing, except that examples such as (6) do not involve islands in the familiar sense.

     There are, however, good reasons to question Ross’s (1969) initial claim that ellipsis can repair island violations (cf. Merchant 2001:ch. 4 and references therein). Recent work by Barros (to appear) and others provides compelling arguments that apparent cases of island repair in sluicing (qua TP ellipsis), e.g. (11), are actually illusory: they always and only arise when the missing material is recoverable either as some subpart of the island in the antecedent (the “short source” strategy: (11a)) or as a simple cleft (the “pseudosluicing”/“pseudofragment” strategy: (11b)), neither of which involves an island violation, as full recovery would (11c) (see Merchant 2001; example adapted from Barros (51)):

(11)   They hired someone who speaks a Balkan language, but I don’t know which one.
a.     …which onei [ they speak ti ].                                                                    Short source
b.     …which onei [ it was ti ].                                                                Pseudosluice
c.     …which onei [ they hired someone who speaks ti ].                                 Full recovery

Given that the appearance of island repair only arises in environments where (11a) or (11b) is an available parse, there is no reason to believe that the parse in (11c) is ever available. As Barros and others point out, ellipsis sites are widely believed to contain silent structure, which in turn predicts that ellipsis should be unable to repair island violations. This prediction is maintained if (11c) is simply ruled out for the same reason its non-elided counterpart is.
That being said, if ellipsis is incapable of repairing island violations, then the analytical foundation of elliptical repair in general is called into question. Thus, our contribution to this volume—the core set of open questions we wish to pose—is:

(12)   Can ellipsis ever rescue an illicit derivation, or are all apparent cases illusory?
a.     If elliptical repair is real:
 i.   Is it a uniform phenomenon, perhaps operating on a natural class of structures or movement types within the general schema of (1), or is the appearance of uniformity accidental, and (1) perhaps too restrictive?
1.   If it is a uniform phenomenon, what is the proper analysis of it?
2.   If it is non-uniform, how can each case be accounted for without egregious additions to the grammar?
ii.   Why is ellipsis able to repair the underlying deviance of move-and-delete derivations such as (6), (8) and (10), but unable to repair island violations?
b.     If elliptical repair is illusory:
i.   What mechanisms do the apparent cases reduce to?
ii.   If some or all of the move-and-delete approach is to be maintained:
1.   Are the movements indeed illicit (and therefore need repair), but repaired by something other than ellipsis?
2.   Or are the movements actually not illicit, and we simply do not understand the underlying structure that is obscured by ellipsis (cf. pseudosluicing)?

We close with commentary on some of these questions.

It is commonly held that islands are not uniform phenomena, meaning any successful approach to (12a.ii) would presumably require uniformity of (some subpart of) the move-and-delete phenomena, part of the open question in (12a.i). Thus, those two questions may be implicationally related. That islands cannot be repaired is significant: it reins in the theory of repair, and potentially makes predictions about the nature of the repairable movement(s).
Regarding (12a.i), a comparative inversion phenomenon described in Merchant (2003) poses a challenge for unification of elliptical repair (13). It exhibits the same ellipsis dependency as the move-and-delete phenomena discussed above, but it differs from them by not obviously involving the move-and-delete derivation in (1). First, the illicit movement being repaired is head movement (T-to-C), not phrasal movement; second, the trace of this illicit movement is apparently outside the elided constituent:

(13)    Abby can speak more languages than cani her father ti [ speak ].
(14)   *Abby can speak more languages than can her father speak.

Perhaps such cases (and others, including as-clauses: Merchant 2003) can be related to the move-and-delete phenomena we have been discussing; if so, a uniform approach to elliptical repair may be achievable.

Barros, Matthew. To appear. “A non-repair approach to island sensitivity in contrastive TP ellipsis.” In Proceedings from the forty-eighth Annual Meeting of the Chicago Linguistic Society.
Coppock, Elizabeth. 2001. “Gapping: In defense of deletion.” In Mary Andronis, Christopher Ball, Heidi Elston & Sylvain Neuvel (eds.), Proceedings from the thirty-seventh Annual Meeting of the Chicago Linguistic Society, CLS 37–1: The Main Session, 133–148.
Depiante, Marcela. 2000. The Syntax of Deep and Surface Anaphora. Doctoral dissertation, University of Connecticut.
Fox, Danny and David Pesetsky. 2005. “Cyclic linearization of syntactic structure.” Theoretical Linguistics 31: 1-45.
Jayaseelan, K. A. 2001. “IP-internal topic and focus phrases.” Studia Linguistica 55: 39–75.
Merchant, Jason. 2001. The Syntax of Silence. Oxford University Press.
Merchant, Jason. 2003. “Subject-auxiliary inversion in comparatives and PF output constraints.” In The Interfaces: Deriving and Interpreting Omitted Structures, eds. Kerstin Schwabe and Susanne Winkler, 55–77. John Benjamins.
Merchant, Jason. 2004. “Fragments and ellipsis.” Linguistics and Philosophy 27: 661–738.
Merchant, Jason. 2006. “A taxonomy of elliptical repair.” Handout from École d’Automne de Linguistique 2006, École Normale Supérieure, Paris.
Ross, John Robert. 1969. “Guess who?” In Robert I. Binnick, Alice Davison, Georgia M. Green, Jerry L. Morgan et. al. (eds.), Papers from the fifth Regional Meeting of the Chicago Linguistic Society, 252–286.
Sailor, Craig and Gary Thoms. To appear. “On the non-existence of non-constituent coordination and non-constituent ellipsis.” In Proceedings of the 31st West Coast Conference on Formal Linguistics. Cascadilla Press.
Thoms, Gary. To appear. “Constraints on exceptional ellipsis are only parallelism effects.” In NELS 43: Proceedings of the forty-third Annual Meeting of the North East Linguistic Society. GLSA.
* Future updates at lingbuzz/002181

No comments:

Post a Comment