Of Minds and Language (46 page)

In contrast, words that describe unobservable mental states and acts are cued almost exclusively by information that resides in the semantics of syntactic structures (see
Fig. 16.8
, from a verb-identification task with adult subjects, which shows this effect). These adults identify action verbs largely by examining the scenes in which these are uttered, but they identify mental verbs largely from hearing nonsense-containing structures in which these occur (Snedeker and Gleitman 2004). Because children acquire the requisite (language-particular) aspects of the grammar only during the second and third years of life, they are limited until then in their word learning largely to
lexical items whose meaning can be wrested more or less directly from transactions with the referential world.

Fig. 16.8. Different verbs require different kinds of information to acquire: Referential information (the visual–situational context) provides the lion's share of information for identifying action verbs such as
jump
or
put
, but syntactic information is far more informative for mental verbs such as
think
,
see
, and
want
.

Source
: Snedeker and Gleitman 2004

More generally, my colleagues and I have tried to explain word learning as a mapping process, one which matches sounds to their meanings. To be sure, the mapping procedure is a complex one, requiring the recruitment and integration of several kinds of linguistic and extralinguistic information. Word learners, in the special case where they are young children, may also be undergoing significant conceptual change. Even if so, these changes in mentality do not seem to be the chief limiting factors in vocabulary growth.

P
ARTICIPANT
: I have two questions. I think there is an important difference between saying that you need a particular structural context, a sentential complement, to solve the mapping problem for propositional attitude verbs, and saying that you need particular kinds of structural arrangements to acquire the concepts. So there are two problems: first, to solve the mapping problem for propositional attitude verbs, and for this you need a particular structural context (syntax, you said, is needed). So that is one problem. The other problem is to ask to what extent you need sentential complements – a certain structural context – to have propositional attitude concepts in the first place. To what extent is the structure actually instrumental for having the concepts in the first place? I of course would go for the latter alternative, and I was wondering about your view on that. Related to that, if you go against the conceptual change view
of Gopnik (Gopnik and Meltzoff 1997) and Carey,
²
so you posit belief-type verbs in the biology (the evolution), it is obviously just pushing the problem. It's not solving any problem, I would say. So here my question is: how little Platonism do you get away with?

G
LEITMAN
: You are correct that there are two problems here. One has to do with where the concept
think
(or any other) comes from, how these ideas get into the mind. The second has to do with identifying the word in the exposure language that encodes each such idea; for instance, learning that
think
is pronounced /think/ in English. You, along with many others, find it congenial to suppose that hearing the sound /think/ (in some sentential context) is what – or part of what – causes the concept to grow in the mind. I myself find that position hard to understand, it seems to imbue words with some magical property. But we can't argue from what is a congenial or intuitively plausible approach on these matters, at least we won't get far that way. So, congeniality aside, what I tried to do in my talk is to show you some evidence to the effect that infant and adult word learning look very much alike. This suggests that both populations are solving the same problem, namely the mapping problem (which sound encodes the concept
think
) rather than one population solving this problem (the adults) while the other (the children) is solving two problems at once – the mapping problem and the concept acquisition problem. I tried to show you that when by experimental artifice one reduces the information that the adult has – his or her evidentiary sources for word learning – the learning trajectory and contents for child and adult look much alike. By exhibiting such laboratory effects, I invited you to consider whether information availability rather than concept availability might not hold the major key in explaining word learning.

No one doubts that there are conceptual-sophistication differences between, say, an average 3-year-old and Noam Chomsky or even the college sophomores to whom we teach
gorping
and
pilking
in the laboratory. It is the sameness in learning properties, once the task is equated for information, across these individuals and populations, that suggests that the mapping problem rather than the concept-learning problem is the chief limiting factor in word learning. But what I most wanted to show you is that observation of “the world” is insufficient as the input basis for acquiring the word /think/ – for anyone, child or adult. Even to use the situation as a constraint, one needs to narrow the search space by being told the argument-taking properties of the novel predicate. That is what the syntax does for you, and it does so for 2- and 3-year olds as well.

P
ARTICIPANT
: I'm not a linguist and I really want to comment on the question of language acquisition. From an interdisciplinary approach, I wanted to offer a possible alternative way of thinking about it. When for example a parent gives a child a stuffed animal, and the parent utters the sound
elephant
, the child has an experience of the joy of the moment, of possibly understanding that they are getting something and it's a toy and it's fun. Later, the parent sees an elephant on television and utters the same sound. So at this point, the child has to negotiate for a distinction. Now in another theory you look for distinctions between phenomena, but you also want to find the categories of representation. In the first case, the stuffed animal resembles an elephant – to the parent. To the child, those distinctions don't yet exist, so it could be a cat, it could be a puppy, it's a stuffed animal to the child, whatever that means to the child. The television representation actually points to an elephant in the world somewhere. So there you have this index to something in the world. Then you have a third scenario: the parent takes the child to the zoo and suddenly the child hears this same utterance while experiencing this huge object in front of him, the actual elephant in nature. It is at this point that I believe Peirce used the term
abduction
. The child is confronted with a sign, the sound
elephant
, which is used in three different contexts as a reference to an object in the world, and the child then has to negotiate the initial meaning of the sound associated with this stuffed animal, with the TV image, and now this massive object in nature. So this is where this abductive reasoning is a partial explanation of what I believe Peirce meant by abduction. This is a partial explanation of abduction where the child then has to negotiate the semantics.

G
LEITMAN
: Your suggested solution is a very sensible one. Your idea is to redress the insufficiencies of any one situational observation by comparing across many such observations, parsing out of scenes in which, say, /elephant/ is uttered, that which is common to all these otherwise quite variable scenarios. This cross-situational observation solution has commended itself to everyone from John Locke and David Hume to modern connectionist modelers. And as I mentioned, surely such a procedure
must
play a role, your various elephant-scenarios are probably a good sample of how this goes. Yet among the many problems of trying to do the whole job of word learning using this situation-observing procedure are the ones I concentrated on in my talk – you can't easily tell
chase
from
flee
this way because they map onto about the same scenes, and it is hard to “observe” thinking in any literal or straightforward way, no matter how many thinking scenes/utterances you are exposed to.

But there is a greater problem and that is the infeasibility of your suggested model given the rate and relative errorlessness of actual word learning. The
child is learning about ten words a day. This is a very, very large number. In light of it, there doesn't seem to be enough time and varied, yet systematic, scene-observations for such a model to work,
unaided
. In fact there's considerable evidence that children are correctly inducing the meanings of words from one or a very few instances, rather than pursuing a compare/contrast procedure across many observations. And this “fast mapping” of new words goes on for a long time, probably until you're about 30–35 years of age, so you get a vocabulary of maybe 75,000 words. Though then, as we elders can tell you, it plummets [laughter]. Luckily Noam and I started with a big vocabulary [laughter].

But seriously: the speed and accuracy and persistence of word learning is something which I think influences how plausible various models should look to you. Another feature of acquisition that might influence you in this regard is the sameness of word meanings acquired by learners whose observational circumstances are wildly different, for instance, deaf, blind, and even deaf-blind persons. I and my many colleagues have offered a different solution. Though of course it involves information gleaned from word-to-world correspondences, it is not limited to this evidentiary source, at least not after the child is 18 or 24 months old and has gained some principled linguistic (as well as world) experience. What this model substitutes for sole use of a multitude of cross-cutting situational observations is a small set of exposures to a novel word, but with most such exposures
simultaneously offering evidence of different kinds
. Observations of a word's fit with the passing scene, yes, but also observations of its structural environment, its morphology, and its cooccurrence with other words (e.g.,
cake
occurs more often with
bake
than with
wake
). These cues trade and conspire to overdetermine interpretation based on very small numbers of incidents during which a novel word is heard.

Janet Dean Fodor

First I would like to acknowledge the contributions of my collaborators, especially my colleague William Sakas, and our graduate students. We are all part of the CUNY Computational Language Acquisition Group (CUNY-CoLAG), whose mission is the computational simulation of syntax acquisition. We have created a large domain of languages, similar to natural languages though simplified, which we use to test the accuracy and speed of different models of child language acquisition.

I will start today by taking you back to 1965, to
Chapter 1
of Noam Chomsky's
Aspects of the Theory of Syntax
, which I recommend to you all. It is, I think, one of the most important fifty pages of all of the important fifty pages that Noam has written, and it is still very germane today. So that will be our beginning point, but it won't be our ending point. We are going to look at Noam's outline of a program for how to set about modeling language acquisition, and then I will tell you why we haven't actually fulfilled it. The past few decades have seen many excellent acquisition studies of real children, studies of what they know and when they know it. But our job is modeling
how
children come to know these things, and that hasn't yet progressed very far at all. I thought that this conference would be a wonderful occasion to bring a gift to Noam, so that I could say “Here, in this box wrapped up with ribbons, is the learning model that you called for in 1965.” But I don't have anything to give. I'm sorry. I can offer only an apology to Noam and an excuse, which is that the problems turned out to be really difficult, much more difficult than
could have been anticipated. Why that is so is what I want to explain to you today.

What Noam asked us to do back then was to consider what must be involved in any acquisition model for language. He said there must be a
representation of the input signal
(the sound waves coming to the child's ears)
in terms of linguistic derivations
. Secondly, there has to be a
specification of the class of possible grammars
, that is, all the candidate grammar hypotheses that the learner might contemplate. Third, there has to be a
method for selecting one of these grammars on the basis of the child's input
, that is, an
evaluation measure
. And that turns out to be particularly difficult. The class of possible grammars is what linguists work on, but the evaluation measure (EM) determines the sequence in which learners try out different grammar hypotheses, so it is something that psycholinguists and computational linguists should have contributed to. But we still don't have it under control. EM is important, though, as a means of explaining why all children exposed to the same language make much the same choices and arrive at much the same grammar, and why they don't get confused along the way in the vast maze of alternatives. In addition, Aspects
Chapter 1
notes that there must be a
strategy for finding hypotheses
. Even in a tightly constrained theory, there are many, many possible grammars. (Estimating how many is easier to do in terms of parameters: if there were just thirty binary parameters, there would be more than a billion possible grammars, and that is probably an underestimate.) Because it is a huge search space, there has to be a method, as Noam observed, for finding hypotheses that fit the particular input sentences a child hears.