Of Minds and Language (45 page)

It remains to say that no one of these cues (situation or syntax) can be sufficient. Obviously the subjects couldn't have learned (and therefore didn't) the meaning of “gorping” solely by hearing it used in an appropriately structured sentence, any more than they could have disambiguated, say,
chase
from
flee
solely by observing the puppet shows. What does the trick for learning is the two cues working conjointly. The argument structure is revealed by the syntax, to be sure, but simultaneously the sentence is interpreted against the world to which it refers. This use of multiple cues lies at the heart of the syntactic bootstrapping procedure. With acquisition of the language-specific grammar, the learner is able to bring to bear a linguistic representation that matches in sophistication, and dovetails with, his or her natural ability to impose a predicate–argument interpretation on events. Given this narrowing of the hypothesis space to fit the argument-structure framework, the observed world more efficiently fills in the richer semantic content of the novel predicate.

I mentioned back at the beginning of this paper that I was going to motivate the syntactic bootstrapping approach in terms of two kinds of lexical item that pose a principled difficulty for lexical learning models that rely solely, or even very heavily, on word-to-world pairing. The first were these perspective verb pairs. Now I want to turn to the second case, which looks even harder. This is acquisition of verbs that describe unobservable acts and events, such as
think
and
believe
. Here the world is of very little value, or so it seems at first glance. You can't see thinking. And the literature tells us that these items indeed appear relatively late in the infant's verb-learning career. Even though children produce verbs describing actions or physical motion very early, often before the second birthday (Bloom et al.1975), and appear to understand them well, they do not use mental verbs as such until about two and a half years of age (Bretherton and Beeghly 1982; Shatz, Wellman and Silber 1983) and do not fully distinguish them from one another in comprehension until around age 4. These facts are often adduced as rather straightforward indices of concept attainment (e.g., Dromi 1987;
Huttenlocher, Smiley and Charney 1983), put forward to support the view that conceptual change is what's accounting for the trajectory and contents of early vocabularies. In particular, the late learning of credal (“belief”) terms is taken as evidence that the child doesn't have control of the relevant concepts, in this case the ability to entertain concepts that refer to one's own or others' mind, aka Theory of Mind. As Gopnik and Meltzoff (1997: 121) put this:

the emergence of belief words like “know” and “think” during the fourth year of life, after “see”, is well established. In this case …changes in the children's spontaneous extensions of these terms parallel changes in their predictions and explanations. The developing theory of mind is apparent both in semantic change and in conceptual change.

And in this case too I'm going to try to convince you that there is another potential explanation for why these terms are late acquired, short of saying that they are too “abstract” for young ears and minds. Specifically, I suggest that the child's problem isn't the inability to think about thinking, but only to find the evidence that the sound/word
think
is the item that expresses the concept ‘think' in English: a mapping problem rather than a conceptual problem. It simply is harder to glean, by observation alone, that thinkers are thinking than that, say, jumpers are jumping. Not only is thinking invisible in the first place. Even more important, the difficulty is that it is actions that people, young and old, are inclined to think about when they interpret the world, rather than the thoughts of those performing the actions (alas, perhaps, but true nonetheless).

Now here is a parade case to introduce this topic: When one shows Rodin's famous statue, The Thinker, even to museum-knowledgeable adults and asks “What's going on here?” the respondents are disinclined to say “That's a thinker thinking.” Even though, if
anything
is, this
is
a thinker thinking. They are inclined to respond instead: “He's resting his head,” or “He's scratching his chin,” in short to offer just about any overt act in preference to an internal one, in describing this scene– though I grant that Rodin himself was an exception to this generalization. In short this is a case of massive insalience of a concept. Nobody thinks about thinking even though it's always going on when people are around. Even here in this room, most of you are emphatically thinking, but thinking is not what you're thinking
about
. If children are to learn the word
think
, there must be circumstances in which the concept it encodes comes readily to mind.

I'll now describe just one experiment in a line we have been pursuing, focusing on this vexing class of words (Papafragou, Cassidy and Gleitman 2007; and for a theoretical review, Gleitman et al. 2005). The idea again is to assess the contribution both of syntactic cues and cues from observation. Pilot findings had provided us with the intuition that for the case of mental verbs,
it's not the truth that sets you free
. Instead, people think about thinking under
circumstances where someone is in a state of
false
belief. Moreover, just as was the case for the perspective verbs, there are characteristic structural environments in which such verbs leap immediately to subjects' minds.
Fig. 16.6
shows an example from a study by Snedeker and Gleitman (2004). It is constructed from a random sample of mothers' natural usages of a credal verb in sentences uttered to their 18–24-month-old children, but the experimental version of these that you see here is doctored and disguised. We leave enough of the closed-class material in place so that the subject can recover the structure spontaneously, that is, without explicit instruction from us. All the other words are replaced by nonsense words. The “mystery word” (the verb) is also replaced by nonsense (in caps) and the subject's task is to recover and report its meaning, given these half-dozen
Jabberwocky
-like exemplars. People are very good at this task, evidently using the appearance of sentential complements as a giveaway clue for a credal verb interpretation.

Fig. 16.6. What does PILK mean? The range of syntactic environments is revealing of the verb interpretation.

Source
: Snedeker and Gleitman 2004

In the Papafragou, Cassidy and Gleitman (2007) study, 4-year-old children and adults watched a series of videotaped stories with a pre-recorded narrative. At the end of each clip, one of the story characters described what happened in the scene with a sentence in which the verb was replaced by a nonsense word. The participants' task was to identify the meaning of this mystery word. The stories fully crossed type of situation (true vs. false belief) with syntactic frame (transitive frame with direct object vs. clausal that-complement) as shown in the design diagram (
Fig. 16.7
). For instance, in one of the false-belief stories inspired by the adventures of Little Red Riding Hood, a boy named Matt brought food to his grandmother (who in reality was a big bad cat in disguise); in the true-belief variant of the story, Matt accompanied by the big cat brought food to his real grandmother. At the end of the story, the cat offered one of these two statements:

(a) [Complement Clause condition] “Did you see that? Matt GORPS that his grandmother is under the covers!”

(b) [Transitive condition] “Did you see that? Matt GORPS a basket of food!”

Fig. 16.7. Scene type X syntax type: This illustrates the design of an experiment in which the child hears a story in which a true or a false belief figures prominently, crossed by a verbal description in the form of a transitive construction (e.g., “The boy is eating his snack”) or in a sentence-complement construction (e.g., “The boy thinks that this is his snack”).

Source
: Papafragou, Cassidy and Gleitman 2007

It was hypothesized that false-belief situations would increase the salience of belief states and acts and would make these more probable topics for conversation, thereby promoting mentalistic conjectures for the novel verb. It was also hypothesized that sentential complements would prompt mentalistic interpretations for the target verb. It was expected that situations where both types of cues cooperate (i.e., in the false belief scenes with a sentential complement) would be particularly supportive of mentalistic guesses. Finally, syntactic cues were expected to overwhelm observational biases when the two conflicted (e.g., in false-belief scenes with a transitive frame).

These predictions were borne out. Scene type had a major effect on the verb guesses produced by both children and adults. Specifically, false-belief scenes increased the percentage of belief verbs guessed by the experimental subjects, compared to true-belief scenes (from 7.4% to 26.5% in children's responses and from 23.5% to 46.3% in adults' responses). The effects of syntax were even more striking. Transitive frames almost never occurred with belief verbs, while complement clauses strongly prompted belief verbs (27.2% and 66.2% of all responses in children and adults, respectively). When both types of supportive cue were present (i.e., in false-belief scenes with complement clause syntax), nearly half (41.2%) of children's responses and an overwhelming majority (85.5%) of adults' responses were belief verbs.

Similar effects were obtained in a further experiment with adults, which assessed “pure” effects of syntactic environment (minus supporting content words) in the identification of mental verbs. True- and false-belief scenes were paired up with transitive or complement clause structures from which all content words had been removed and replaced with nonsense words (e.g.
He glorps the fleep
vs.
He glorps that the fleep is glexing
). Again syntax proved
a more reliable cue than even the most suggestive extra-linguistic contexts. Furthermore, the combination of clausal and scene (false belief) information again resulted in an overwhelming proportion of mental-verb guesses.

Taken together, these experiments demonstrate that the syntactic type of a verb's argument (e.g., whether the object of a transitive verb is a noun phrase or a tensed sentence complement) helps word learners narrow down their hypotheses about the possible meaning of a new word. Furthermore, this type of syntactic cue interacts over additively with cues from the extralinguistic environment (e.g., the salience of a mental state). We interpret these findings to support the presence of a learning procedure with three crucial properties: (1) it is sensitive to different types of information in hypothesizing the meaning of novel words; (2) it is especially responsive to the presence of multiple conspiring cues; (3) it especially weights the language-internal cues when faced with unreliable extralinguistic cues to the meaning of the verb.

To summarize some of the effects I've been discussing, the first general finding is that not all words are learned from the same kind of information. Certain items, including words encoding the basic-level object terms, appear early. This is one of the most robust effects in the literature of language learning, and is seen again and again cross-culturally and cross-linguistically. A popular explanation for why these items are so rapidly and uniformly learned is that they instantiate just about the only concepts that infant minds can entertain. But I have argued instead that it is these words' tractability to the first-available property of the learning procedure, word-to-world pairing, that explains why they are learned first. As support for this view, we have shown in several experiments that when adults are by experimental artifice reduced to this same information – roughly, if they are exposed to single “mystery words” in context, rather than to whole sentences in context – they too are capable of little lexical learning beyond the basic-level nominals. The information for acquiring the noun
apple
and such physical-action verbs as
jump
or
hit
resides largely in the observable world, as interpreted by both adults and very young children.