Of Minds and Language (50 page)

R
IZZI
: I was wondering if it would be possible, or desirable, to incorporate into your computational model certain data from the empirical study of development which strongly suggest that it is not the case that parameters are all fixed at the same time. There is a temporal order, it seems, though we are far from having a precise temporal chart of what happens. That is a big gap in our knowledge. But at least certain things are known, particularly if one considers the critical moment when the child starts to produce syntactically relevant structures; that is to say, when he puts together at least two words so that some syntax can be observed. It is clear that some parameters have already been fixed, and others have not, at least on the basis of the productions we hear. For instance, major word-order parameters have been fixed, like Head Initial / Head Final (e.g., the verb precedes the object or follows it). That seems to have been established, because as soon as the child produces two-word sentences, if he is exposed to Japanese he'll say
Sandwich eat
, while if he's exposed to French, he'll say
Eat sandwich
. Similarly for other major word-order parameters. However for other facts it is not the case; we see a phase in production in which certain parametric properties apparently have not been determined. One relevant case that directly bears on certain things you have said is that certain types of scrambling are acquired relatively late (at least they manifest themselves relatively late). Similarly for certain kinds of grammatically determined ellipsis. There is more ellipsis (grammatically determined ellipsis, I think) in early productions, and you see a developmental effect in production. So this suggests that some parameters are fixed earlier than other parameters. And there are many stories one might propose about that. It could be that some parameters are easier; it could be that certain critical parameters come with very specific and easy triggers, as in the phonological bootstrapping hypothesis (Mazuka 1996). The infant just listens to the stress pattern and determines whether the language is head-initial or head-final. And there are other stories around. But
I wondered if you would be interested in incorporating these observations into the computational model.

F
ODOR
: Yes, I wish we were there, but you can see we are still at a fairly primitive stage in the project of modeling real child learners. In fact I would add to the factors you mention that might determine the order of events. There is also information structure (topic, focus, etc.), which children may not be very good at; they are not very good in general at pragmatic aspects of conversation. Consider scrambling, for example. There is a study by Otsu (1994) of children's comprehension of scrambling (object before subject) in Japanese. They are very poor at it if an isolated sentence is just scrambled out of the blue. But if the scrambled sentence occurs in a conversational context where scrambling is appropriate, due to previous mention of the object, then the children perform much better. So it may be that it is not so much the syntax itself, as the work it is doing in the language. I am ashamed to admit that none of our simulation studies take meaning into account at all. We have obviously got to do so eventually, because clearly what is being learned by children is relationships between sentence forms and sentence meanings. But so far we have no interfaces in the parameter-setting models. They are treated as a pure computational system, which is interesting for us to study as psycholinguists and linguists, but in the real world the interfaces are extremely important.

I want to add one more point concerning the order in which parameters are set. There is a game that we can play with the CoLAG system, though it will be very laborious and we are just waiting until we can entice a graduate student into doing this for a dissertation. When we run our simulations, the computer keeps a record of every grammar the learning model hypothesizes along the way to final convergence on the target. Now we can do the following research project. We can order the hypothesis sequences by their length, which will tell us which sequences converged (terminated) first, and which took much longer. There is a lot of variability. So we can look to see which parameters were set first in the most efficient learning sequences, compared with which were set first in the least efficient ones. This can reveal whether (at least on purely structural grounds, not topic, focus, and so forth), there is some optimum order in which the parameters should be set. This is a huge data-crunching task, but we actually could set about doing it, and I would really like to see how it comes out.

P
ARTICIPANT
: How do you decide, when you are going to run a simulation experiment, what a possible parameter is? I mean, doesn't it depends on your theoretical assumptions?

F
ODOR
: Yes, you are right, the parameters are dependent on the underlying principles that you assume. There is a very interesting paper on re-parameterization of the linguistic facts, by Frank and Kapur (1996). That is, if you find that there is a learning problem somewhere, you might consider that it's because the parameterization is wrong. So you might try re-describing the facts as falling under different parameters and perhaps the learning problem disappears. For us the choice of parameters was largely a practical question. We needed to be conservative, with quite old-fashioned parameters, because we needed them to stay stable. It took three years to build the language domain, and if the syntacticians change their minds about what the parameters are tomorrow, we can't re-engineer our 3,000 languages. So we kept to very traditional sorts of parameters that any linguist would recognize (e.g., wh-movement, verb-raising, pied-piping, etc.). You are absolutely right that the results of our experiments could change if we were to shift to a different linguistic theory with different parameters. What I don't think will change are the fundamental problems that I was talking about today. I suspect those will still be with us, even when the linguistic details differ. The one thing that would make a significant difference for us is if something like the Manzini and Wexler defaults system that I mentioned for
generating
the grammar relations in the lattice could be made to square with the linguistic facts; then we could implement the Subset Principle with no lattice representation at all.

P
IATTELLI-PALMARINI
: The idea of subsets is one of the most interesting, I think, in the history of learning theories. It was very clear, when we had the idea of E-languages (languages as things out there), that there could be a smaller language contained in a bigger language. I am wondering what the idea of “subset” becomes in I-language.

F
ODOR
: You have put your finger on a central problem that we face in modeling. We assume children learn
grammars
(I-languages), but the subset principle is about
languages
(E-languages). As you say, it is about one E-language being included in another one. If we had a neat translation system from grammars to languages, we could manage the SP problem a great deal better. We would love to be able to look at the grammar and say, “The language this grammar generates is going to be a subset of the language generated by this other grammar.” But in fact, there doesn't appear to be a transparent correspondence between grammars and languages. Noam emphasized that a small change in a grammar can make a great change in the set of sentences generated. The Manzini and Wexler system which assumes an independently contributing default value for each parameter (which we call the ‘Simple Defaults Model') does offer a transparent translation. Every subset relation between languages
is due to the default value of one or more specifiable parameters in their grammars. Now that is not true of our CoLAG language domain, and so we suspect it's not true of the natural language domain at large. And we haven't yet found any alternative system for going back and forth between grammars and languages. As far we now know, the relationship between languages is not projectable from the relationship between grammars. We wish it were.

Thomas G. Bever

This paper reviews an approach to the enterprise of paring away universals of attested languages to reveal the essential universals that require their own explanation. An example, discussed at this conference, is the long-standing puzzle presented by the Extended Projection Principle (EPP, Chomsky 1981). I am suggesting an explanation for the EPP based on the learner's need for constructions to have a common superficial form, with common thematic relations, the hallmark of EPP. If one treats EPP phenomena as the result of normal processes of language acquisition, the phenomena not only receive an independently motivated explanation, they also no longer constitute a structural anomaly in syntactic theory.
¹

EPP was initially proposed as the structural/configurational requirement that sentences must always have a subject NP, even without semantic content (cf. Chomsky 1981, Lasnik 2001, Epstein and Seely 2002, Richards 2002; see Svenonious 2002, McGinnis and Richards, in press, for general reviews). This principle was first proposed to account for subject-like phrases in sentences, so called expletives (e.g., “it”):

(1) a. “it” is raining

b. “there” are three men in the room

c. “it” surprised us that john left

d. “es” geht mir gut

e. “il” pleut

The EPP was initially proposed as a universal syntactic constraint that all languages must respect. While roughly correct for English, a number of troubling facts have emerged:

(2) a. EPP may not be universal (e.g., Irish as analyzed by McCloskey 1996, 2001).

b. Different languages express it differently: e.g., via focus as opposed to subject, in intonation patterns, with different and inconsistent agreement patterns.

c. It generally corresponds to the statistically dominant form in each language.

d. It has not found a formal derivation within current syntactic theory – it must be stipulated.

Accordingly, the EPP may be a “configurational” constraint on derivations – it requires that sentences all conform to
some
typical surface pattern. Epstein and Seeley (2002: 82) note the problem this poses for the minimalist program:

If (as many argue) EPP is in fact “configurational,” then it seems to us to undermine the entire Minimalist theory of movement based on feature interpretability at the interfaces. More generally, “configurational” requirements represent a retreat to the stipulation of molecular tree properties … It amounts to the reincorporation of…principles of GB…that gave rise to the quest for Minimalist explanation…

In other words, the EPP is a structural constraint stipulated in the minimalist framework (as well as others), which violates its structural principles and simplicity. Yet EPP-like phenomena exist.

Below I outline a language acquisition model which requires that languages exhibit a canonical form, the Canonical Form Constraint (CFC) – which renders EPP phenomena in attested languages. Thus, there are two potential explanations of EPP phenomena. Either it is indeed a syntactic constraint, part of universal syntax in the narrow faculty of language; or it is a constraint on learnable languages: Sentences have to conform to the CFC – they must sound like they are sentences of the language to afford the individual child a statistical entrée into acquiring it. How can we decide between these two explanations? First, the EPP adds a stipulated constraint to grammars, undercutting their simplicity. Second, the EPP is a heterogeneous constraint, with
different kinds of expressions in different languages. Third, the CFC, as we will see, is independently motivated: it explains statistical properties of language, stages of acquisition, and significant facts about adult language processing. Thus, I argue that the phenomena that motivated the EPP are actually expressions of the Canonical Form Constraint (CFC).

Syntacticians may object that this line of reasoning is circular. In many languages, the EPP constraint does not merely exert “stylistic” preferences on sentence constructions, it dictates syntactic requirements on grammatical derivations. But the issue is the source of the constraint that results in processes that conform to the EPP. On my view, the child tends to learn sentence constructions that conform to the canonical form constraint, and not other constructions. The notion of “learn” can be glossed as “discovers derivations for statistically frequent meaning/form pairs, using its available repertoire of structural devices.” Thus, in individual languages the child accesses and learns specific derivational processes that conform descriptively to the EPP. But the EPP itself is merely a descriptive generalization reflecting acquisition constraints as its true cause. In the sense of Boeckx (this volume), EPP-like phenomena are among the set of E-universals (corresponding to E-language), not I-universals (corresponding to I-language). In the sense of Hauser et al. (2002), it is a property of the interface between the narrow faculty of language and the acquisition interface.

The following discussion will serve as an outline of how a simplified model of what individuals do during language acquisition, based on a general model of human learning, can explain universal properties of attested languages, such as the EPP. My argumentation strategy here is the following:

(a) a general method of paring down universals, with some non-syntactic examples

(b) a comprehension model showing how the linguistic structures are implemented in an analysis-by-synthesis comprehension model

(c) an application of the analysis-by-synthesis model as a model of acquisition

(d) implications for the Canonical Form Constraint (CFC) as a language universal

(e) implications of the CFC for a correct interpretation of EPP phenomena