Page images
PDF
EPUB

semantic net (as in several other programs, such as SIR (Raphael, 1964); and (2) a complex structure encodes the definition (as in dictionary definition) of the word, thus relating it to the other concepts used in its definition. In his original work he used the task of giving the system two words, e.g., FIRE and BURN, and having it state the relationship between these concepts; e.g., FIRE IS CONDITION WHICH BURN, also TO BURN CAN BE TO DESTROY SOMETHING ON FIRE.

Now this program is an example of sufficiency analysis, as we have used the phrase. For the system is not intended as a detailed model of human memory and it was never tested as such. But it is relevant to psychology, because he was able to make (and demonstrate via the living program) conceptual progress in how human memory might be structured for tasks where we understand by general experience what performance can typically be expected of humans. Indeed, the work was a Ph.D. dissertation in Psychology at Carnegie-Mellon (Quillian, 1966).

69

There is a sequel to this work and it makes my point. Quillian is, indeed, interested in the psychology of human memory. Thus, he followed up this work in sufficiency analysis with an attempt to explore whether human memory could be modeled by such a structure (Collins and Quillian, 1969). The essential feature of a semantic net is that information about a concept is not all localized at the node corresponding to that concept, but is distributed through the network. Thus, that a canary can sing, might be located at canary, but that a canary can fly is probably not located at canary, but at bird, since it is a property of all birds. Similarly, that a canary has skin is probably not even located at bird, but rather an animal. If this were the case, then it should take longer for such a system to answer yes or no to such questions (when embedded in a population of other questions, such as "Does a house sing?", "Does a cat fly," etc.). Further, if the net is homogeneous in its structure, then there should be a constant operation time to go from node to node in the net.

Figure 8 shows the results of asking these questions experimentally of humans, using reactions times. The points are averages over populations of questions of similar type. The quantity of interest is the difference between points, as

[blocks in formation]

(SO)A CANARY IS A CANARY

1000

THE NUMBER OF SUBJECT MEANS FOR
EACH POINT ARE IN PARENTHESIS
SENTENCES SHOWN ARE ONLY ILLUSTRATIVE

(27)

900

LEVELS

2

FIGURE 8: Average reaction times for different types of true sentences (from Collins and Quillian, 1969, Fig. 2).

indicated above. Three of them are essentially identical at the 80 ms. The fourth, from "A canary is a canary" to "A canary is a bird" is too large, due, it appears, to the reaction to the former being too fast. But this is the one question that admits of an answer by perceptual matching, avoiding the meaning of the word "canary" altogether (hence the time to go from image node in the net). There is other evidence in the literature that indicates the same phenomena, e.g., it takes longer to recognize that A and a are the same than that A and A are the same, since (apparently) the latter can be done by a perceptual match and the other not (Posner, 1968).

Before leaving Quillian's work, let me note that a number of assumptions are embedded in Figure 8. Thus, nothing guarantees that the particular words are related as the experimental analysis assumes, even if the structure of memory is precisely of this postulated type, i.e., the information that a canary can fly could be as close to "canary" as that it could sing. Thus, Collins and Quillian attempted to select a population of words that had a high prior plausibility of being this way and were successful, as you can see.

13

To make the point of this work for us explicit: (1) if you want to assert the relevance of a theoretical information structure to psychology, then you had better build a bridge from it making use of experimental data; (2) it can be done. The bridge built by Collins and Quillian is a frail one, of course, since it only addresses one tiny aspect of the total memory system and it is compatible with many other similar structures. Indeed, as you would expect, neither Quillian nor anyone else thought the original structure was right in detail, and his second iteration is a much modified system (Quillian, 1969). But with the experimental work, it has passed well beyond metaphor.

Even more briefly, let me present one more example. This is taken from my own work with H.A. Simon on problem solving. I insert it, both because I'm always inclined to mention some of my favority psychology, especially when it fits the story so well, and because none of our examples, with the possible exception of Quillian's memory structure, are from artificial intelligence.

We work intensively with cryptarithmetic tasks, such as the SEND+MORE=MONEY that Fikes used to describe the operation of his program, REF-ARF in this conference. It is a puzzle that still retains modest challenges as a task for various problem solving programs. We could simply work with programs for solving this task, say like REF-ARF, or even some that are more like we imagine a human processor to be constructed, e.g., with a short term memory; a long term memory, etc. From these we could draw various conclusions about the general character of human problem solving. In fact, we did exactly this with the Logic Theorist, our original program (Newell, Shaw and Simon, 1958). But for some time now, since the first work with GPS (Newell, Shaw and Simon, 1960, 1961), we have taken an attitude much as I am trying to prescribe here.

Thus, our typical operation is to present a subject with the task, asking him to talk aloud as he works on it. A fragment of the result, called a protocol, is shown in Figure 9, where the task is DONALD+GERALD ROBERT and D-5 is given as initial information (Newell, 1966, 1967). We then attempt to construct a processing system that mirrors the behavior of the subject and agrees with what we know about human processing capabilities.

Typical collegiate subjects in this task can be described with excellent fidelity as working in a problem space, whose elements are the states of knowledge the subject can have about the task, and whose structure is given by a small set of operators that work on a given state of knowledge to produce a new state of knowledge. Problem solving is search through this problem space. This search, what we call the problem behavior graph (PBG), is shown in Figure 10 for the fragment of protocol shown in Figure 9. The four operators used by this subject are Process a column (PC), Assign a digit to a letter (AV), Generate the possible values for a letter (GN), and Test if a digit - letter assignment is legal (TD). (Actually, these are four specific variants of processes that meet these four general functional descriptions.)

The problem space, with its operators, is a reasonable description only if there exists an information processing system that describes the way the subject's

[blocks in formation]
« PreviousContinue »