Reply by Kenneth Wexler ("On Realizing External Arguments")

The following post is a reply by Prof. Kenneth Wexler to Chris Collins' blog post reviewing Koring et. al. ("On Realizing External Arguments", Linguistic Inquiry, forthcoming).

The article is found here:

Koring et. al. ("On Realizing External Arguments")

The review is found here:

Review of Koring et. al. (forthcoming)

I am grateful for Chris Collins’ review of Koring, Reuland, Sanger and Wexler (in press, henceforth referred OREA), which considers our attempt to integrate acquisitional and judgmental evidence in helping to construct some portion of linguistic theory. I will start by totally agreeing with Chris’ observation that the fact that two different approaches, with two very different data sets, converge on the same conclusion – that the external argument of short (i.e. no audible external argument) verbal passives is syntactically projected, lends strong support to the conclusion. It was one of the points of our paper that that was in fact the case. And Chris’ well-known 2005 paper and recent book (in press, MIT Press) also reaches this conclusion. Yet they argue in a completely different way.

In the penultimate paragraph of his review, Chris writes that OREA would have been “strengthened considerably” had we noted how its conclusions decided between “current theories of the passive,” namely between Bruening (2013), who argued that the implicit external argument of the verbal passive is not syntactically projected and theories in which the argument is syntactically projected. But I believe that’s exactly what we did do, though perhaps in a different form.  We argued against Bruening’s view that verbal and adjectival passives project exactly the same functional structure. We argued that a major difference was that we claimed that verbal passives have a syntactically projected EA and adjectival passives don’t. Bruening claims neither does.

At one point in the reviewing history of OREA, a reviewer thought that too much was devoted to a critique of Bruening’s conclusions, which we had carried out since that is a current approach that gets attention, and we thought that it had to be wrong. In response, we greatly shortened our discussion. Possibly too much. Nevertheless it is obvious from our conclusions that our results disagree with Bruening’s theory of the verbal passive. We stressed that the EA in verbal short passives is syntactically projected.

Chris argues that implicature analysis of the disjointness effect in short passives has a problem. Following Fox and Katzir’s grammatical implicature model, which unites implicatures and focus theory, we argue that the reference of the external argument of the implicit argument in a short verbal passive is determined to some extent by implicatures.  In the sentence Mary was nominated EA, the theory says to substitute other salient constituents (there are other alternatives, not relevant here) for the EA, and if this new sentence is strictly stronger logically than the original sentence, then that new sentence is negated. So Mary was nominated implies that it’s not the case that Mary nominated Mary. This is the disjointness effect.

Chris argues that we could substitute any DP x that is somehow in the context into the EA position and derive the implicature that it’s not the case that x nominated Mary. In particular, in our experiments, we could have done just that. So when we show a picture of Homer washing Bart another one of Bart washing Homer, and have the kid choose a picture as the meaning of Bart was washed, not only is “It’s not the case that Bart washed Bart” is derived, but so is “It’s not the case that Homer washed Bart.”

However, despite that fact that early in the experiment, kids were made familiar with the characters (usually in a minute or so), there is no reason to take any character who hasn’t been mentioned in the sentence as salient. Clearly, it’s the mentioned character who is salient when we create alternatives. So no such incorrect implicature is made.

Chris argues that OREA’s argument against a Principle B analysis of the missing disjointness in verbal passives in children is wrong. At first, I thought that Principle B might be the explanation. But with further thinking, it didn’t seem pan out. We followed (our footnote 11) Bhatt and Pancheva (2006) in accepting that the implicit external argument is an existentially bound variable and thus would not allow coreference. Chris accepts this claim for existentially bound implicit EAs but argues that there are different kinds of bound variables besides existentially bound ones.  See his in-press book, chapters 2 and 4. One of them is pro. If the argument were pro, Principle B would prevent coreference with the surface subject of the passive, yielding disjointness. If children were “more forgiving” of Principle B errors than adults then they would violate disjointness, yielding our experimental results.

I have attempted to understand Chapter 2’s arguments about the different types of implicit arguments, and I just haven’t had enough time to unpack them, so I want to be careful here. In fact, I quite admire Chris’ dogged and detailed search for exactly what type of entity the implicit argument is, it’s very sophisticated and grapples with much data. I’m not sure how we know that the relevant implicit EAs in our example is pro, not an existentially bound variable. Furthermore, one thought I have had is that if in fact pro is bound by the subject, not just co-referential with it, then experimental results starting with Wexler and Chien (1985) and Chien and Wexler (1990) through Thornton and Wexler (2000) and many more papers in several languages and many labs show that bound variables do not induce much in the way of errors in children (e.g. every duck pointed to her can’t mean that every duck pointed to herself, for children as well as adults.) So if this binding of pro is true variable binding, we wouldn’t get the errors in disjointness in children. Now it may be that coreference rather than binding could be involved with pro, in which case my argument against it is wrong, and the typical “binding” (that is referential) errors would induce the disjointness error. It’s my lack of understanding, my fault I am sure, that prevents me from coming to a more explicit conclusion here.

Chris goes on to argue that the advantage of his Principle B proposal is that it wouldn’t require such a “drastic difference” between the way children and adults represent verbal passives. And in fact, the perfect result would obtain; children learn what the verbal passive is on their first hearing one.

Several comments are worth making. First, we have no idea how it is that a child, before knowing how a form of words is syntactically decomposed, takes a sentence (string of words? partially analyzed?) in the input and maps it into a derivation. One might think, well, the child knows from the context that the first DP is the theme rather than the agent, so it can’t be the typical active sentence. Fair enough; I have long argued that some semantic information from context is necessary to allow development of grammar (Wexler and Hamburger, Wexler and Culicover, many other papers). But I have also pointed out how difficult it is to know how information is available and reliable and how it works. Even with many years of useful work on syntactic bootstrapping, we don’t have as yet a good substitute for semantic information. We do want the data to be what Chomsky calls “epistemologically available” to the child. The simpler and more surface transparent the better. But let’s say that the child without passive grammar knows enough about thematic roles and has enough cognitive capacity to figure out that the thematic role of the surface subject in a short verbal passive is a theme/patient, not an agent. So not the external argument.  Perhaps, and this is even more difficult to know happens, they even know that the sentence is intended as eventive rather than stative, e.g. perhaps it includes a progressive, so that it’s not an adjectival passive (though consider that even this is problematic given what we’ve seen about get-passives). Can the kid really tell from context whether the verb is eventive or whether it’s an aux (like get) with a resultant-state adjectival passive? Even if the child can figure out the verb is in fact eventive, couldn’t the sentence be a topicalization? We know that it’s not topicalization because we have judgmental data about all sorts of phenomena, about how passives and topicalizations work differently. But the child has no such access, by a long shot. So there’s not a perfect result. This is unknown.

That doesn’t mean that we must believe that the child represents verbal passives as adjectival passives. (Not that it’s actually a drastic difference between the representations of the adult and child for verbal passives. Rather the child simply thinks that verbal passives are ungrammatical. I’ve suggested, provocatively to be sure, in my Development of Phases paper in MITWPPL, that the child is a more perfect minimalist grammarian than the adult, who seems to deviate in certain ways, in particular in allowing defective phases. In general the theme of much of my work in acquisition is that children have more tightly constrained UG than adults.

It's not a prima facie argument that tells us that kids analyze verbal passives as adjectival Rather, it’s years of empirical research. The starting point for our research for OREA was an attempt to find yet another empirical test of the conclusion reached first in Borer and Wexler (1987) that children performed well on a subclass of verbal passives (especially activity verbs versus subject experiencer verbs) because they analyzed these verbs as adjectival passives. The syntax of verbal passives was taken to be ungrammatical for young children, but the adjectival interpretation allowed the children to at least compute the theta roles correctly, despite considering verbal passive syntax as ungrammatical. Children from ages 4 to 5 performed well on verbal passive, but only several years later on subject experiencer passives (despite doing well on subject experiencer actives.)

Over the years, there has been a great deal of empirical evidence supporting the conclusion that children before about 4 or 5 parse verbal passives as adjectival and very little counterevidence. This has even reached to direct interpretive experiments, where the children very often interpret verbal passives as adjectival, but not vice-versa. The languages in the experiments were Catalan and Spanish, in which the verbal and adjectival passives are both auxiliary + participle, but are not homophonous, unlike English, since the auxiliaries for verbal and adjectival passives differ.

We thought that Dutch would be another language in which to test the adjectival analysis assumption without homophony. The auxiliaries of the verbal and adjectival passives differ. But we wanted to test yet another, very different, phenomenon, the disjointness effect in verbal passives, which doesn’t hold in adjectival passives. We predicted that kids at a young enough age would violate disjointness because they had misanalysed verbal passives as adjectival. It’s always a useful idea to find totally different types of phenomena to support a theoretical position.

Our thinking, of course, would not have been accepted by Bruening’s analysis. He argued that disjointness holds or doesn’t hold in either verbal or adjectival passives, and that when it appeared that disjointness held, it was for reasons other than the existence of an implicit external argument, which we took as central. So we argued against Bruening, since it seemed to us that his article was getting attention. 

Perhaps we should have referenced the literature that argued for the existence of a syntactically projected external argument in verbal passives. But the point seemed to us to have been accepted in the literature for more than 70 years. In (1a) (examples like this from Kratzer) there is simply a strongly felt intuition that there is an external argument: there is an agent or instrument that is securing the mountaineer. In (1b), which is adjectival (the non-progressive present tense verbal passive is ungrammatical or, perhaps, infelicitous, in English), there is no such intuition.

(1)

a.  The mountaineer is being secured

b.  The mountaineer is secured

Of course, there is a disjoint reference intuition in (1a) but not (1b). In (1a) we strongly feel that somebody or something other than the mountaineer is doing the securing. In (1b) we don’t – it could be the mountaineer herself who is doing it. Secure is not in the class of verbs that Bruening claims are the only ones that shows the relevant effects – the verbs which alternative between transitive and intransitive, like dress, wash. In fact, we would claim that they exist for all activity verbs, the ones that Kratzer shows all have a resultant-state adjectival passive. One more example.

(2)

a. Mary is being established (in her law practice)

b. Mary is established (in her law practice)

Establish is not an alternating one or two argument verb, but native speakers have a strong intuition that in (2a), the person establishing Mary is not Mary, but in (2b), it might very well be (probably the preferred reading, in fact, although of course (2b) is ambiguous as to whether or not the establisher is Mary.)

These judgments have been accepted as so strong that they are used in baby syntax or even in intro to linguistics or even in discussions with the non-linguistic general public to illustrate the syntactic and semantic role of arguments for which here are no words, phonetically empty arguments. We took this position as simply classical and didn’t reference (much of?) the literature. Perhaps it would have been better to do so. We did say that the strong intuition was foundational.

OREA does 2 experiments (one Dutch, one English) which show that young kids in fact violate disjointness – 3-year-olds are completely at chance, displaying no knowledge. This is yet strong further evidence that follows from our analysis that disjointness follows from the syntactic projection of the external argument in verbal but not adjectival passives. The young kids have no external argument in their comprehension of the passive – it is adjectival for them. At the same time, the experimental evidence shows yet again and with totally different phenomena that the classical and already amply confirmed result that kids take verbal passives as adjectival is true.

The major empirical counterargument to kids’ considering verbal passives as adjectival is Steve Crain’s (and somebody else, was it Janet Fodor?) that elicited verbal passives fairly easily from young children. They didn't notice one property of their data, that was immediately obvious (they provided many examples on their handout at the time, a BUCLD presentation, probably printed someplace by now), and as I immediately pointed out after their talk. Most of the elicited passives were get-passives, like (3) not be passives.

(3)  

Luke Skywalker got zapped. 

Much judgmental data would be most consistent with the conclusion that that these get-passives use the aux get meaning something like become (though with different aspectual/time properties) with a resultant-state adjectival passive participle complement, meaning as Kratzer has demonstrated: the state such that there was a previous time t at which an event of zapping took place on Luke. So there was a transition of Luke into a state such there was an event of zapping Luke that took place earlier. We might think that (3) is verbal because there is an activity going on, a transition, but it’s not a verbal passive, simply a transition into a state.

Let’s test disjointness in get-passives. Since the complement is adjectival there is no reason for disjointness to hold. For comparison let’s use the same verb as in (2).

(4)  

Mary is getting established (in her law practice)

Disjointness is not required in (4), strongly contrasting it with (2a). In fact, the preferred out of the blue reading of (4) is that Mary herself did the establishing. Of course, there might be another establisher. Adjectival passives simply don't have a disjointness requirement; they are ambiguous in whether or not the external argument is disjoint.

Given this very brief summary, with most data not even mentioned, we have very good evidence to support the claim that children often analyze verbal passives as adjectival. So the suggestion that children represent verbal passives identically as adults is not an advantage, it actually contradicts established data.

Here we have to be very clear. We cannot give precedence to any one type of data on conceptual grounds. As Chomsky has often written, data don’t come with labels saying, “I am useful data, pay attention” or, “I am data that doesn’t have to be paid attention to.” In particular, as Chomsky has argued, there is no prima facie reason to give pride of place to traditional judgmental data of adults, which of course has been by far the most frequent type of useful data in generative linguistics. Supplemented more and more frequently by experimentally induced judgments of grammaticality, meaning, felicitousness, etc. Child data fits in in exactly the same way, it’s another type of data. And if we want to obtain explanatory adequacy and feasibility, we will want to understand child data. If the data look good that verbal passives in kids are often analyzed as adjectival passives, then we have to understand this fact; it’s not an advance to dismiss it, it’s a loss.

In the course of my career, I and my colleagues have often ended papers by pointing out what we’ve learned about the nature of UG from child data. Not just about child grammar but about UG. Others have done this we well. Of course adult judgmental data is hugely important. But it doesn’t hold pride of place if there are also other data that are useful. I continue to hope, perhaps too optimistically, that this doesn’t have to be said in every paper, that it will just be as natural in pursuing linguistic theory as the use of adult judgmental data. In OREA we argue for exactly that – that the kid data tells us a good deal about UG, not just when it confirms that children know everything that there is to know about a piece of UG, but even more when it has to assume that some pieces are missing. A different system to test and to help break down components and yield natural experiments.

Chris mentions at the conclusion of his review that he leaves to future work a minimalist merge derivation for adjectival passives. This would be a pleasure to see, especially given how useful and stimulating his very careful and very detailed merge analysis of verbal passives is. OREA doesn’t attempt as detailed a merge analysis, for verbal or adjectival passives, but does suggest some merge properties that account for the difference, for example a difference in timing for certain operations. Language acquisition results clearly establish that the needed operations for verbal passives arise considerably later in development than for adjectival passives. Given the theories on offer for the delay of verbal passives (especially the Universal Phase Constraint (UPR)), it is clear that there has to be some fundamentally different syntax between verbal and adjectival passives, in particular involving the relation between the underlying and surface positions of the theme. In studies of semantic properties of resultant-state adjectival passives together with their syntax, it is perhaps mostly concluded that the theme of adjectival passives does not originate in object position, but either directly in subject position or some close by position, e.g. Embick, David. 2004. On the structure of resultative participles in English. Linguistic Inquiry, 35(3), 355-392. I look forward with great anticipation to what Chris’ endeavors will yield. If they take consideration clear results in language acquisition, so much the better.