Of Minds and Language (15 page)

A question that will be at least somewhat debated, perhaps in the corner where Randy, Rochel, and I sit, is what the nonlinguistic quantificational systems are in animals and humans. One system that certainly is not questioned is the one that Randy and Rochel have worked on for many years, and is often called the “analog magnitude system.” This is a system whose signature or definitional property is that it computes approximate number estimation with no absolute limit on number, but with discrimination limited by Weber (logarithmic) ratios. There is abundant evidence for this in the animal world, shown by studies that involve training animals, and studies that involve spontaneous methods. Such studies are complementary in the sense that they both reveal the signature of the system in animals like chimpanzees, rhesus monkeys, tamarins, lemurs, rats, pigeons, and so forth. A second system, which is perhaps more heatedly debated in terms of whether it should count as something numerical is a system that some of us have called the “parallel individuation system,” or the “object file system.” This system has a different kind of signature. It seems to be very precise, but it is limited in terms of the numbers that it is precise for – specifically in a range of 3 to 4. So discrimination is limited by how many individuals can be tracked at the same time in parallel. Here as well, there is evidence from some training studies and some spontaneous methods, in both human adults and infants, as well as in primates.

I want to take you now to one of my labs, the beautiful island of Cayo Santiago, off the coast of Puerto Rico, which is the sole location for 1,000 rhesus monkeys. What's beautiful about this island is that, in contrast to most studies of primates, this island has a very large number of individuals, about a thousand at a given time. They are perfectly habituated to our presence, allowing us to observe them at very close range, safely, and carry out experiments with them in a naturalistic setting. What I want to tell you about today is one kind of experiment that lends itself to asking about the capacity for numerical quantification in a functionally significant, ecologically relevant foraging task. Here is the basic nature of the design, which you will hear about over and over again in the next few pages. We find an animal who is by himself or herself; we place two boxes in front of the animal; we show them they're empty, and then we proceed to lower objects into the boxes. In most cases, what we are lowering are food objects that we know they're highly motivated to go find. In the typical experiment we are in effect asking them, “Do you prefer the box with more food
or the one with less food?” Since we can assume that they are going to try to go for more food, the experiment should work.

So here is the idea for the basic experiment, counterbalancing for all sorts of necessary things. We load into the first box one apple followed by a second apple (the boxes are opaque so the monkeys can't see inside) and then we load one apple into the second box; we walk away and let the animal choose. This is one trial per animal, we don't repeat the individuals, so we are going to be comparing across conditions where every condition has 20–24 different individuals. We don't train them, we don't even cue them into what the task is until we walk away. We place the apples in the box, walk away, and let them choose a box. When we do that, here are the results we get. If we compare one piece of apple going into a box and nothing in the other, they prefer 1 vs. 0, 2 vs. 1, 3 vs. 2, and 4 vs. 3, but they fail to show a successful discrimination of 5 vs. 4, 6 vs. 4, 8 vs. 4, and 8 vs. 3. So although the ratios are favorable here relative to what they can do with 2 vs. 1, they are not using ratios to make discrimination. The discrimination is falling out precisely at 4 vs. 3. They can do no more. So under these conditions (no training, one trial per individual), this is the level of discrimination that we find, and this pattern cannot be explained by the analog magnitude system. It is, however, entirely consistent with the signature of the parallel individuation system.

Now, let us turn to a conceptual domain that might appear to be privileged for language, morpho-syntax in particular – namely the singular–plural distinction – and ask the question whether the conceptual roots upon which language was constructed over evolutionary time and in development built upon some conceptual primitives that may be seen in nonlinguistic creatures and in pre-linguistic human infants. The basic idea is that if we have one cat, or we have two, or millions of cats, we simply take the noun and add a terminal -s. The result that opens the door to the comparative angle comes from a recent study by Dave Barner, Susan Carey, and their colleagues (Barner et al. 2005). They presented infants with a version of the box-choice study I just described for rhesus monkeys. When infants in the age range of 12–20 months were tested, Barner and Colleagues found that subjects could discriminate 1 cracker from 2, as well as 3 from 2, but they failed with 4 vs. 3, 2 vs. 4, and surprisingly, even 1 vs. 4. As soon as the number of items going into one box exceeds 3, infants at this age fail the discrimination task. Of interest is that at the age of around 22 months, when infants are producing, in English, the singular–plural morphology, they now succeed on the 1 vs. 4 task. Barner and Colleagues explain these results by suggesting that the explicit formulation of the singular–plural morphology, in terms of its representational structure, enables a new form of numerical discrimination, specifically, one between singular and plural entities.
Therefore, in ontogeny we see a linguistic distinction first, and then a conceptual distinction second. Now if this interpretation is correct, and numerical discrimination of this kind depends on the singular–plural morphology, then of course animals lacking this morphology will fail on a comparable task.

To test this hypothesis, I now want to run through a series of experiments that ask the following question. If we consider the two nonlinguistic systems that I have described, the parallel individuation system, which is precise (less than 4 in rhesus monkeys), and the analog magnitude system, which is approximate but with no absolute limit, both will predict success at singular vs. plural, and for plural–plural as long as there are favorable ratios or fewer than four objects. So if both systems are operative, which we know they are, then singular–plural should work fine and so should plural–plural, as long as it has these conditions are satisfied. So we are back to the box-choice experiment, but we are going to do it in a slightly different way. Now, rather than presenting the items one by one, we present them as sets. So we show them five apples; those five apples go into the box all at once and disappear; next we show them one apple and this one apple disappears into the box; and then we allow subjects to approach and choose one box. What we do therefore is present plural sets, presented all at once as opposed to presenting individuals, and we counterbalance the order in which they go into the boxes. We test for singular–plural (1 vs. 2 and 1 vs. 5), as well as plural–plural (2 vs. 4 and 2 vs. 5). Now recall that if either the system of parallel individuation or analog magnitudes is operative, subjects will be able to discriminate values of 4 or less.

What we find in terms of the proportion of subjects picking the larger number of objects, in this case apples, is success on 1 vs. 2 and 1 vs. 5. Now this is an uninformative result, at least for analog magnitude or set-based quantification, because both could work. But here is where it gets interesting: subjects fail at 2 vs. 4, 2 vs. 3, and 2 vs. 5. These results cannot be explained on the basis of the analog magnitude system, and certainly the 2 vs. 3 and 2 vs. 4 failures cannot be explained on the basis of parallel individuation. How, then, can we explain these data? These data do not force a rejection of the systems for parallel individuation or analog magnitude. Rather, they simply indicate that under the testing conditions carried out, these mechanisms are not recruited or expressed. Why?

Let's now run the same exact experiment, but carry it out as individuals going into the box. For example, we show them five apples going into a box one at a time, followed by two apples going into another box one at a time. So now it is still 5 vs. 2, but this time presented as individuals as opposed to sets. They succeed again on 1 vs. 2, 1 vs. 5, but also on 2 vs. 3 and 2 vs. 4, while failing on 2 vs. 5. Remember that this pattern is consistent with the parallel individuation system, but inconsistent with analog magnitude. We therefore recover the
pattern of results obtained in the original experiment, a pattern that is entirely consistent with the system of parallel individuation. But we can do better. We can actually turn the system on and off.

If we start out with individual apples, but we load them in as sets, what happens? Here, subjects succeed on 1 vs. 2 and 1 vs. 5, but they fail 2 vs. 3 on 2 vs. 4 and 2 vs. 5. In other words, when sets go in last, they are back to set-based quantification, even though they see them individuated. If we start out with sets, but we load them in as individuals, they succeed on 1 vs. 2, 1 vs. 5, 1 vs. 5 and 2 vs. 4, but they fail on 2 vs. 5. In other words, what is driving the system is the set-based quantificational system. If they see objects as sets as the last thing, then they use a set-based system to quantify which has more; if they see things going in as individuals, then discrimination is based on the system of parallel individuation.

What I would like to argue, therefore, is that rhesus monkeys seem to be making a conceptual distinction between singular and plural. The results I have presented today cannot be explained by the currently available mechanisms that have been discussed, either analog magnitude or parallel individuation. Again, this is not to reject those mechanisms as viable mechanisms for quantification, but they simply cannot account for the pattern of data we see today. Therefore, as a working hypothesis, what I would like to argue is that this system of set-based quantification is part of the faculty of language in the broad sense (FLB), but it is not something specific to language and is not therefore part of FLN.

Now I move to a second line of experiments that plays on the mass–count distinction, a topic of considerable interest to both semanticists and syntacticians. The question is: could this distinction, and its ontological commitments, be rooted in a nonlinguistic conceptual format, and therefore be present in other animals? We have count nouns, things that can be enumerated (cup, shovel, apple), and we have mass nouns, things that cannot be enumerated unless there is a preceding classifier or packaging term (e.g. not *
waters
, but
cups of water
, not *
sands
but
piles of sand
), so we don't say, for example, *
three sands
. The question is: does this kind of distinction, which appears in natural languages (not all, but many), translate into conceptual resources that are nonlinguistic, present early in evolution and ontogeny? Consider the experiments on enumeration in human infants, and specifically the classic studies by Karen Wynn (Wynn 1990, 1992) that were done initially with solid objects (e.g. Mickey Mouse dolls), using the violation-of-expectancy looking time method.
⁴
Wynn's results, and the many replications that followed, show that if you place one object behind a screen followed by a second one, and you pull the screen away,
babies will look longer at violations of those numbers. So if you place two objects behind the screen but then reveal one or three, babies look longer at these outcomes than at an outcome of two. But if you run the exact same experiment, but pour sand (one pour of sand followed by a second pour of sand) and reveal one, two, or three piles of sand, babies do not look longer at these different outcomes. This suggests that in order for enumeration to proceed, infants require individuals, discrete items that can be enumerated. There is something fundamentally different between solid objects and nonsolid masses.

To address the evolutionary or phylogenetic aspect of this problem, we (Wood et al., 2008) ran a similar experiment, using the box-choice experiment I described earlier. To motivate the animals, we used small pieces of carrot, poured out of a bucket. We filled up beakers with carrot pieces and then poured them into the opaque buckets, walked away, and gave the monkeys a choice between two buckets that had different quantities of carrot pieces. We presented I vs. I, 3 vs. 2, and so forth, pouring pieces of carrot out of a beaker. The monkeys picked 2 vs. I beaker pours, 3 vs. 2, and 4 vs. 3, but they failed at 5 vs. 4 and 6 vs. 3. This is exactly the pattern of results I presented for objects, but now the computation is carried out over pouring of quantities or masses of carrot pieces. Now, this confounds many things including volume, so can we control for these factors and see if they are actually enumerating? To find out we poured 1 big quantity of carrot pieces vs. 2 medium ones, where volume is now equated but the actions are different. Here they picked 2 medium over 1 big, so now quantity is preferred over the number of actual pours. We showed them the identical number of actions, 1 vs. 1, but where one beaker was a full volume of carrot pieces and one a small volume, they pick the one big over small, showing they're paying attention to the volume. Regarding all the previous conditions, they could actually see the amount of carrot pieces in the beaker, because the beaker was transparent, but if we make it opaque so they actually have to attend to what is falling out of the beaker, they still picked 2 vs. I. So they are actually tracking the amount of stuff falling out of the beaker. Together, these results suggest that rhesus monkeys are computing numerosities over solid and nonsolid entities, tapping, in these conditions into the system of parallel individuation. These patterns stand in contrast to those presented thus far for infants, where the enumerative capacities tapped for objects falls apart for masses.

Let me now end by returning to the questions I posed at the beginning. First,
to what extent are the conceptual representations that appear to uniquely enter into linguistic computation built from nonlinguistic resources
? This question is, to me, only beginning to be addressed, but the problem of quantifiers and their representational format seems ideally suited for further exploration. Can we get to the point where we can ask about whether animals have some notion of
many
vs.
all or some
? Are the kinds of logical quantifiers that enter into language built upon conceptual resources that have a much more ancient evolutionary trajectory? We are only beginning to ask questions such as this, and we have few answers. Secondly,
to what extent have linguistic conceptual representations transformed in evolution and ontogeny some of our ontological commitments
? The speculation I'd like to leave you with is this. If you consider the results I just presented, involving rhesus monkeys enumerating carrot pieces, and you contrast these with the baby results on pouring sand, I think there is an interesting proposal with respect to the relationship between language and onto-logical commitments. Specifically, although infants do not yet have, in their production or comprehension, anything like a mass–count distinction, the evolution of that distinction within language has actually transformed our ontological commitments such that infants see the world differently than do rhesus monkeys, who are happily enumerating masses in a way that at least babies seem not to. In other words, humans uniquely evolved the mass–count distinction as a parametric setting, initially set as a default, but then modifiable by the local language, leading some natural languages to make the distinction, but only optionally.