Harnad, S. (1996) Experimental Analysis of Naming Behavior Cannot Explain Naming Capacity. Journal of the Experimental Analysis of Behavior 65: 262-264.

[Invited Commentary on Horne & Lowe: "On the Origins of Naming and Other Symbolic Behavior" Journal for the Experimental Analysis of Behavior"]


Experimental Analysis of Naming Behavior Cannot Explain Naming Capacity

Stevan Harnad
Department of Psychology
University of Southampton
Highfield, Southampton
SO17 1BJ UNITED KINGDOM
harnad@soton.ac.uk
phone: +44 1703 592582
fax: +44 1703 594597
http://cogsci.soton.ac.uk/harnad/
ftp://ftp.princeton.edu/pub/harnad/

The experimental analysis of naming behavior can tell us exactly the kinds of things Horne & Lowe (H & L) report here: (1) the conditions under which people and animals succeed or fail in naming things and (2) the conditions under which bidirectional associations are formed between inputs (objects, pictures of objects, seen or heard names of objects) and outputs (spoken names of objects, multimodal operations on objects). The "stimulus equivalence" that H & L single out is really just the reflexive, symmetric and transitive property of pairwise associations among the above. This is real and of some interest, but it unfortunately casts very little light on symbolization and language in general, and naming capacity in particular. The associative equivalence between name and object is trivial in relation to the real question, which is: How do we (or any system that can do it) manage to connect names to things correctly (Harnad 1987, 1990, 1992)? The experimental analysis of naming behavior begs this question entirely, simply taking it for granted that the connection is somehow successfully accomplished.

Note that I am not talking about the (equally trivial) mechanism of pairwise association between particular inputs (I), or between particular inputs and outputs (O). There's no mystery there. And unfortunately, the kinds of arbitrary association between specific pictures and written and spoken words examined in H & L's Figure 1 and the arbitrary unique "class members" examined in Figures 15-17, simply take us further away from what makes language special and send us off instead on the more general and less informative path of the experimental analysis of associative directionality.

What makes linguistic naming (as opposed to pairwise I-I or I-O associations) special is that it requires categorization (Harnad 1987): The abstraction of an invariance shared by the variable sensory projections of the members of the category designated by the name and not shared by the likewise variable sensory projections of members of other categories with which they might be confused. This is not the mere association of a specific X with a specific Y (whether X and Y be input or output, picture or name). And categorization does involve an equivalence, but it is unfortunately not H & L's associative equivalence, but the input equivalence among all sensory projections of the members of the category named by the name. To realize that there is nothing in the analysis of associative directionality that can explain this kind of input equivalence, we must first stop thinking about tasks like pairing pictures of cars and spoken and written names of cars in the laboratory, because those are overlearned categories we already have. Even less relevant are the arbitrary "classes" in the Saunders and Green (1992) or the Lowe and Beasty (1987) kind of study cited by H & L; those are not categories but associative clusters.

Think instead of a nontrivial naming task, and one that the subject has not already overlearned before entering the experiment (which would beg the question): Think of cancer cell identification, or chicken sexing (Biederman & Shiffrar 1987; Livingston et al., in prep.): What good does it do me if you tell me that the way I learned which kinds of cells were and weren't cancerous, or which newborn chicks were male and female, was by "associating the name with the stimulus"! This might be conceivable if there were only a small number of stimuli that I was shown over and over again, till I memorized the pairwise connections. But that's not what happens with cancer cell identification or chicken-sexing: I learn to recognize what kind of stimuli fall under which name; i.e., I learn the invariance underlying the variability and interconfusability. Such learning is not easy; it takes time. I cannot say how I am successfully doing it, once I can (so the question cannot be answered by introspection either). But what is certain is that there are some invariant properties that inputs from the members of the category designated by the name share, and that reliably distinguish them from members of other categories with which they could be confused, and that I have somehow managed to learn them. The naming problem requires an answer to the question of how I (or any system that can do it) manage to do that.

It will not do to reply that most of our named categories are not as difficult as cancer cell identification or chicken sexing. First, cancer cell identification and chicken sexing are not difficult for the successfully trained expert; they are only difficult for us, who have not learned how to do it. Well, the child is in that position initially, with respect to all the objects, events, and states of affairs in the world. The fact that he eventually succeeds in naming the ones he does is thanks to the same capacity that allows adults to learn to do chicken-sexing. Anything in between is simply begging the question: taking for granted the vast but still unexplained naming capacity of the child (and the adult), or focusing on trivial associations between arbitrary stimuli and "names."

Renaming the capacity as the exercise of "listener behavior" or "echoic behavior" does not explain anything either: If the child already knows what category of things names name before ever uttering them, then the critical connection was made through hearing names and interacting with objects in the presence of naming and pointing by others, but it still does not explain how. Given that the connection is made, all the associative equivalences can come to be useful (of course I think of objects when I name them, and think of their names when I see or operate on or think about objects), but what can I thank for this capacity in the first place?

Why do I emphasize what the experimental analysis takes for granted (and characteristically relegates to another specialty, such as innate perceptual mechanisms, or brain function in general)? Because whatever turns out to be the true substrate of naming capacity will in turn cast light on both the substrate of language and its unique symbolic/propositional power (Steklis & Harnad 1976) -- which is decidedly not just the power of associative equivalence between name and object! Yes, perceptual processes are involved in categorization, and some of them may well be innate, but most of them are not, because in naming we are dealing mainly with learned perceptual categories: How many of the entries in a dictionary do you think we were born with specific built-in invariance detectors for? No, most of our lexicon was purchased by perceptual learning (analogous to the learning in cancer cell identification and chicken-sexing) rather than by inborn Darwinian detectors. If inborn detectors were responsible for most of our naming capacity, then the question of the origin and nature of language would become the question of the origin and nature of those detectors (Harnad 1976) -- but it is not: Most of our named categories are learned, rather than inborn, that is, they are not acquired by (phylogenetic) "theft" but by (ontogenetic) "honest toil": The honest toil consists of sampling instances of members and nonmembers and laboriously learning [attention Skinnerians], from the consequences of miscategorization, which are which (Skinner 1984a).

But experimental analysts cannot take too much heart from the familiar "shaping by consequences" phenomenon at work here (Catania and Harnad 1988), because the fact that successful categorization performance can be shaped by trial and error learning with feedback is not an explanation; it is what calls for explanation: What internal properties must a system have in order to be capable of learning the categories we can learn? To see that the behavior-analytic approach is nonexplanatory in this regard, consider how much better off a roboticist (Harnad 1995a) would be if, knowing the categorization skills people were capable of, but clueless as to how to build a system that could do that, he were informed by a behavior-analyst that it's accomplished on the basis of feedback from consequences! And that there is a relation of "equivalence" between the category names and the objects they designate!

Nor does the unfulfilled explanatory agenda stop there, for there is another mystery about language (over and above the mystery of how we manage to name things), and it too is a form of "theft." If "Darwinian theft" is the source of the categories we are born already able to detect, and "honest toil" is new category learning through trial and error with feedback, then once we have a repertoire of category names earned by honest toil, language gives us the unique further symbolic capacity to acquire new category names from strings of prior names alone: To use an example I have used many times before (Harnad 1990), if you have learned, by honest toil, to call horses "horses" when you see them, reliably distinguishing them from members of other categories with which they might be confused, and you have learned, likewise by honest toil, to call "stripes" stripes, then, even though you have never encountered one, you are in a position to correctly name your first zebra upon merely being told that a "zebra" is a "horse" with "stripes." That's (symbolic/propositional) theft; it can spare you an awful lot of honest toil; and it is the true power of language. Nor is it explained by (or equivalent to) associative equivalence. It will be explained by a successful explanation of what internal structures and processes give us the capacity to learn to categorize and name classes of inputs by detecting the invariance in their sensory projections (Harnad 1992; Harnad et al. 1991, 1995), and then how strings of names in the form of propositions about category membership can give us the capacity to name new members of categories we have not encountered before.

An explanation like this is impossible from just the experimental analysis of behavior: one must also hypothesize and then analyze the internal structures and processes that generate the capacity to exhibit the behavior (Harnad 1982, 1984). It is only from such research that we will come to understand the origins of naming and symbolic behavior (Harnad 1995b). -- An internal representational analysis of the honest-toil vs. symbolic-theft distinction might even cast some explanatory light on Skinner's named but unexplicated distinction between contingency-based and ruled-governed behavior (Skinner 1984b).

REFERENCES

Biederman, I. & Shiffrar, M. M. (1987) Sexing day-old chicks: A case study and expert systems analysis of a difficult perceptual-learning task. Journal of Experimental Psychology: Learning, Memory, & Cognition 13: 640 - 645.

Catania, A.C. & Harnad, S. (eds.) (1988) The Selection of Behavior. The Operant Behaviorism of BF Skinner: Comments and Consequences. New York: Cambridge University Press.

Harnad, S. (1976) Induction, evolution and accountability. In: Harnad, S., Steklis, H. D. & Lancaster, J. B. (eds.) (1976) Origins and Evolution of Language and Speech. Annals of the New York Academy of Sciences 280: 58 - 60.

Harnad, S. (1982) Neoconstructivism: A unifying theme for the cognitive sciences. In: Language, mind and brain (T. Simon & R. Scholes, eds., Hillsdale NJ: Erlbaum), 1 - 11.

Harnad, S. (1984) What are the scope and limits of radical behaviorist theory? The Behavioral and Brain Sciences 7: 720 -721.

Harnad, S. (1987) The induction and representation of categories. In: Harnad, S. (ed.) (1987) Categorical Perception: The Groundwork of Cognition. New York: Cambridge University Press.

Harnad, S. (1990) The Symbol Grounding Problem. Physica D 42: 335-346.

Harnad, S. (1992) Connecting Object to Symbol in Modeling Cognition. In: A. Clarke and R. Lutz (Eds) Connectionism in Context Springer Verlag.

Harnad, S. (1995a) Grounding Symbolic Capacity in Robotic Capacity. In: Steels, L. and R. Brooks (eds.) The "artificial life" route to "artificial intelligence." Building Situated Embodied Agents. New Haven: Lawrence Erlbaum

Harnad, S. (1995b) The Origin of Words: A Psychophysical Hypothesis In Durham, W & Velichkovsky B (Eds.) "Naturally Human: Origins and Destiny of Language." Muenster: Nodus Pub.

Harnad, S., Hanson, S.J. & Lubin, J. (1991) Categorical Perception and the Evolution of Supervised Learning in Neural Nets. In: Working Papers of the AAAI Spring Symposium on Machine Learning of Natural Language and Ontology (DW Powers & L Reeker, Eds.) pp. 65-74. Presented at Symposium on Symbol Grounding: Problems and Practice, Stanford University, March 1991; also reprinted as Document D91-09, Deutsches Forschungszentrum fur Kuenstliche Intelligenz GmbH Kaiserslautern FRG.

Harnad, S. Hanson, S.J. & Lubin, J. (1995) Learned Categorical Perception in Neural Nets: Implications for Symbol Grounding. In: V. Honavar & L. Uhr (eds) Symbol Processors and Connectionist Network Models in Artificial Intelligence and Cognitive Modelling: Steps Toward Principled Integration. pp. 191-206. Acadamic Press.

Livingston, K., Andrews, J. & Harnad, S. (in preparation) Categorical Perception Induced by Learning.

Skinner, B.F. (1984a) Selection by Consequences. Behavioral and Brain Sciences 7: 477-510.

Skinner, B.F. (1984b) An Operant Analysis of Problem-Solving. Behavioral and Brain Sciences 7: 583-614.

Steklis, H.D. & Harnad, S. (1976) From hand to mouth: Some critical stages in the evolution of language. In: Harnad, S., Steklis, H. D. & Lancaster, J. B. (eds.) (1976) Origins and Evolution of Language and Speech. Annals of the New York Academy of Sciences 280: 445-455.