Comments

Wednesday, July 9, 2014

Guest Post: The [Spec,TP]-agreement fallacy

Omer Preminger sent me this interesting post commenting on Chomsky's lecture 3 proposal that Spec-TP agreement might circumvent problems for the MLA. The gist is that Chomsky's proposal faces some well-known empirical challenges, especially evident in languages like Icelandic (what Gert Webelhuth once called the super conducting super collider of linguistics).  I hope that this generates some useful discussion, especially among those partial to Chomsky's take on the labeling issues he raises. Given that Sp-X agreement lies at the chart of Chomsky's analysis of successive cyclic movement, EPP and Fixed Subject Constraint effects, Omer's challenge needs addressing if this proposal is to fly. So, let the games begin!

*******

DISCLAIMER: None of what I am about to write draws on my own research. These are results that, in one form or another, have been around for decades.

In Part 3 of Norbert’s comments on Chomsky’s third lecture, he discussed Chomsky’s suggestion for why it is that movement can stop at (what we ruffians call) the [Spec,TP] position. Why is this a question? Because Chomsky is assuming the Minimal Labeling Algorithm (henceforth, MLA), which normally cannot assign a label to a structure {X, Y} if both X and Y are internally complex; and under the MLA, an unlabelable structure needs to be broken up via movement of one of its terms. One loophole for this (see Norbert’s discussion for some others) is if X and Y enter into some agreement relation; then, the feature (or perhaps set of features) F that has undergone agreement can serve as the label of {X, Y}. What is the F, then, that allows – and often times, forces – subjects to remain in [Spec,TP]? Chomsky’s answer: phi (i.e., the familiar set of person, number, gender/noun-class).

The point of this post is to show that this is not, and cannot be, the answer (for reasons that have been known for quite a while now). It starts with Icelandic, but as I will note at the very end, we could have perhaps made the point even based on English alone (though perhaps in a somewhat more tenuous fashion).

Icelandic, as is well known, has non-nominative subjects. These are not merely noun phrases bearing non-nominative case that have come to c-command the other noun phrases in their clause (cf. German); everything that Chomsky wants to say about subjects in English holds of these non-nominative subjects as well, save for two properties: their case (obviously), and the fact that they don’t control agreement (crucially).

So you get, e.g., sentences of the form in (1), where the finite verb agrees with the nominative object (which also passes a series of direct-object diagnostics), not with the dative subject:

(1)  SUBJ(dative)  FINITE-VERB(agr-with-obj)  OBJ(nominative)

[There are other complications, as there are bound to be – in this case, concerning what happens when OBJ is 1st/2nd person. But if everything is 3rd person, things work as shown in (1). And, importantly, even if the OBJ is 1st/2nd person, agreement is not with the person features of SUBJ (i.e., choosing a 1st/2nd person SUBJ does not make possible 1st/2nd person agreement on the verb).]

So, the short version of the story: there are subjects, that show all the subjecthood properties (e.g. landing and staying in subject position), and yet they are not what enters into agreement in phi-features with T. Not only that, but T in fact enters into overt phi agreement with something else (in this case, the nominative direct object). Tying subjecthood properties (e.g. the ability to move to and stay in [Spec,TP]) to agreement in phi-features is wrong. Fin.

But there is a slightly longer version of this story. Norbert, for one, is partial to the idea that what someone like me would call “probe-goal agreement” (as in, for example, the relation between T and the direct object in (1)) is really a movement relation, one where both LF and PF privilege the lower copy for interpretation/pronunciation, and the consequences of this movement can only be seen via the effects it has on the formal features of the landing site (TP). I have suggested we refer to this kind of movement as “interface-vacuous” movement, since the interfaces ignore its having occurred.

Suppose, then, that the OBJ in (1) has a second merge position in [Spec,TP], but is pronounced and interpreted in its lower position within the verb phrase. This second position of OBJ enters into agreement in phi-features with T, allowing all of (what we would call) TP to be labeled by these phi-features, as discussed above. Does this salvage Chomsky’s story?

The answer is “no.” That is because Icelandic is not a null-subject language; Icelandic clauses need subjects, in a way that this “interface-vacuous” movement (if it actually exists) does not seem to satisfy. To put it another way, even if OBJ has a second unpronounced and uninterpreted merge position in [Spec,TP], the facts are that this does not absolve the clause of its need to have a(nother) subject. To see why that’s a problem for Chomsky, let’s consider how the need to have a subject arises in his system. In the proposed system, the difference between a null-subject language (say, Italian) and a non-null-subject language (say, English), is in the capacity of T to serve as the label of a {T, XP} structure (say, for XP=vP). In a non-null-subject language, it cannot (“T is weak”) – and so in fact the only way to assign TP a label is to move something to [Spec,TP] (as a sister of the {T, XP} node), have it agree with {T, XP} in phi-features, and have those phi-features label the resulting complex object (what we would call “TP”). In a null-subject language, T can serve as the label of {T, XP}, and thus movement to [Spec,TP] is not required (if such movement were to nevertheless occur, the English-style story just described could still kick-in).

Continuing to adopt (for the time being) the “interface-vacuous” movement wrinkle, the OBJ in (1) has moved to [Spec,TP], agreed with {T, vP} in phi features, and thus labeled the resulting object; why does this clause still need an overt subject? Or more to the point, why is the equivalent of “arrived.PL some people.NOM” (a VS-order unaccusative with no expletive) not grammatical in Icelandic? After all, the nominative will have “interface-vacuously” moved to [Spec,TP], satisfying all apparent labeling needs.


So what has gone wrong here? My answer would be: the [Spec,TP]-agreement connection is a red herring, and this is what happens when you build your edifice on a red herring (fish are slippery!). Yes, in many languages many of things that end up in [Spec,TP] were also the things that T agreed with (or as Norbert would have it: many of the things that end up in [Spec,TP] overtly, turn out to obviate the need for another, separate thing to undergo “interface-vacuous” movement to [Spec,TP]). But taking that to be a fundamental fact about the computational system is just wrong, for there are languages where that’s just not how it works. Icelandic is one such language – but depending on your analysis of expletive-associate constructions and of Locative Inversion, English may very well be such a language, as well.



19 comments:

  1. why is the equivalent of “arrived.PL some people.NOM” (a VS-order unaccusative with no expletive) not grammatical in Icelandic?
    From my point of view this is an interesting question only if there's a typological gap. There are four logical possiblities based on [+/- only Spec,TP agr] and [+/- null-subject] (under the simplifying assumption that null-subject configurations all have the same underlying structure, otherwise the number of descriptive parameters is higher). If each option is realized in some language (which I believe is the case, no?), then you can have two completely independent stories for these two parameters.

    You see, the Merge-MSO equivalence means that Merge by itself is already powerful enough to give rise to all agreement phenomena we see in natural language if you start out with the null assumption that there are no restrictions on what the lexicon may look like. The curious thing are not the phenomena we find, it's the phenomena we do not find (in the case of agreement only a few PCCs, which also show DO-IO asymmetries, and only a few types of resolved agreement).

    So if you look at it from this different point of view, interface vacuous movement does what you need in order to satisfy the MLA, and the availability of null-subjects is an independent issue (e.g. a parameter in the mapping from derivations to PFs). Not the most satisfying story, but a tenable story. Still, I agree with your conclusion that Spec,TP --> Subj Agreement is false and should not be an integral part of any formalism's architecture.

    ReplyDelete
    Replies
    1. We are in agreement, Thomas (no pun intended). And I think the [-only SpecTP can agree, +null subject] cell, which is the only one we haven't yet discussed explicitly, is indeed instantiated – e.g. in certain Indo-Aryan languages where only absolutives determine agreement on finite verbs, but "subjecthood diagnostics" happily pick out the ergative external argument of a transitive (at least in the perfective aspect), and, crucially, subject pro-drop is available.

      Delete
  2. Hmm. Isn't Omers argument an argument about whether the DP in spec TP *agrees* with T. Imagine T bears phi and probes the lower nominative. But the Dative subject then raises to spec TP because it's higher. Both T and the dative bear phi (just not matching phi) and one could then say that the MLA sees both phi on DP and phi on TP, and is satisfied. Omer's argument is that spec head agreement cannot be what stops the DP in spec, but it could be the presence if phi on both. Now, I'm not sure if that's what Chomsky says, but I think there's an interpretation of the workings of the MLA that doesn't fall down here.

    ReplyDelete
    Replies
    1. I'm not sure. Norbert alluded earlier to the fact that it's not enough that XP and YP have a feature (or set of features) in common in an {XP, YP} structure; they have to undergo agreement in that feature (or set of features). If memory serves, the reason is that you don't want "spuriously stable" structures – structures that the MLA is supposed to want to break up due to being unlabelable, but turn out to be labelable because the two terms happen to share some features. One example that occurs to me right now is:

      (1)
      a. Obama is the president.
      b. The president is Obama.

      (2) *There is Obama the president.
      (except as an item in a list, with "the president" as an appositive)

      The MLA is supposed to build on Moro's idea that [DP DP] structures are unstable; but both of these terms have phi features, and so if it is enough to have some features in common, the MLA shouldn't need to break up this object.

      Delete
    2. Chomsky does insist that the phi features have the same value in virtue of being *assigned* the same value by an operation like Agree. Whther he needs this, I take it, is what you are questioning. But, if you do need this then do you agree that Omer's point is valid?

      Delete
    3. Yes, I think Omer is right that Icelandic EPP isn't related to agreement between T and the subject, so unless there's something else going on, the argument seems good.

      Delete
  3. But actually it seems like an extra stipulation to have them match in value, especially if valuing takes place at transfer. The MLA just wants an unambiguous signal as to the label. The issue if the distribution of phi features on nominal predication structures is pretty unclear, I think, and of course all right thinking people assume a Pred in (1)!

    ReplyDelete
    Replies
    1. Fair enough, David. How about successive-cyclic movement out of DPs –

      (3) What did you read an article about?

      Wouldn't the shared phi-features here between [what] and the DP [an article ...] incorrectly "freeze" (in the criterial sense) the wh-element in SpecDP?

      (I guess I'm not 100% how freezing works in this system, yet. For example, is the whole "Edge Features" thing still in play, or is this MLA-driven successive-cyclic movement supposed to replace that? If it's the latter, then I think (3) might be a problem.)

      Delete
  4. I think Chomsky thinks indefinite DPs aren't phases, so no edge here and no need to stop off. See discussion at the end of approaching ug from below.

    ReplyDelete
    Replies
    1. Okay – last try for now: successive-cyclic wh-movement out of vP in languages where v triggers object agreement? So, the counterpart of (4) in a lg. with object agreement:

      (4) Who did John persuade Mary Bill should fire?

      Delete
    2. If clitic doubling is analyzed as object agreement, it is incompatible with wh-movement in the languages I am familiar with. At least doubling of accusative objects is incompatible with wh-movmement.

      Delete
  5. :-), ok, but I think Chomsky would say that uphi on v is inherited by the root, so isn't at the edge! But maybe possessor raising might be a case.

    ReplyDelete
    Replies
    1. If the Mirror Principle still has currency, then there are certainly languages where uphi can't be "inherited" from v to V, since obj-agreement occurs farther away from the root than transitivity/causativity morphemes.

      (Besides – and speaking of right thinking people – didn't Haegeman & van Koppen already demonstrate that this whole feature-inheritance idea is off-base, anyway...?)

      Delete
  6. I don't think anyone really buys the idea that the mirror principle applies to uninterpretable features. It appears to be related to the semantic scope of functional categories. I'm pretty sure that Chomsky would take the morphological spell out of phi to be independent of its syntactic position, so the Haegeman argument wouldn't go through. Not that I'm a big fan of feature inheritance myself, but think the system Chomsky presents is consistent with the phenomena you mention.

    ReplyDelete
    Replies
    1. But wait, the Haegeman-van Koppen work isn't about the position where features are spelled out, it's about the fact that C and T can agree with separate things (and thus, express contradictory phi values).

      Delete
    2. yeah, but the argument doesn't quite go through. Again it depends on when phi is valued. So you could say that phi is on both C and T, but is valued at transfer under agree. phi on C agrees with the left conjunct (in their examples) and phi on T with the whole coordinate subject. On transfer, the relevant values are filled in and spelled out in the morphology. So in the syntax, it's the same unvalued phi bundle, but the values end up being important not in the syntax but in the morphological component (whatever that is). I think this actually makes sense, as syntax usually doesn't care about values of features (especially if, as I think, all features are actually just binary, so when we say phi we are talking about a set of features that give you person and number (+/- participant, +/-author +/-atomic etc)). Syntax certainly doesn't care about the plusses and minuses, just about the fact of the specification. Mind you, that approach does assume that there is a copy of phi on C and T, and that would be in conflict with Richards (Marc) proposal that the uphi has to really move to T, with no copy, to get phasal transfer timing to work out. Another approach to the Haegeman/van Koppen data would be to say that the C-agreement isn't really in the syntax at all, it's just a morphological copy rule applying under adjacency. I wasn't overwhelmed by their arguments against this view since we know that some adjuncts don't figure into the calculation of adjacency (cf Bobaljik's thesis on do-support and adjuncts).

      Delete
    3. This comment has been removed by the author.

      Delete
    4. This comment has been removed by the author.

      Delete
    5. I think this is the juncture where we declare we agree to disagree.

      In my view, there is good reason to think there is no such thing as [-group] (your [+atomic]; I'm using McGinnis' 2005 labels, here), [-author], or [-participant] in the syntax (cf. in the morphology, for example, where there is reason to think that [-participant] can be made reference to). In my view, saying that a single feature bundle can exhibit overt agreement with two different targets, expressed in two different places, and still be thought of as a single feature bundle teeters on non-falsifiability. And in my view, a notion like "PF-adjacency that ignores adjuncts" is incoherent. (Not saying there aren't adjacency effects that ignore adjuncts, just that putting those at PF seems incoherent to me.)

      [re-edited to clarify labeling of singular-plural distinguishing feature]

      Delete