Structure-Mapping vs.

High-level Perception:
The Mistaken Fight Over The Explanation of Analogy

Clayton T. Morrison and Eric Dietrich

Philosophy And Computers & Cognitive Science (PACCS)
State University of New York
Binghamton, NY 13902-6000
[email protected]
[email protected]

1993; Kotovsky & Gentner, 1990). SMT has two key

Abstract
strengths: 1) it makes a clean distinction between analogies
There is currently a competition between two theories that and other types of similarity comparisons (abstraction,
propose to explain the cognitive phenomenon of analogy: anomaly, literal similarity, and mere appearance), both in
Dedre Gentner’s Structure-Mapping Theory and Douglas theory and as evidenced in psychological examination; and
Hofstadter’s theory of Analogy as High-level Perception. 2) it is generally applicable — rather than requiring a
We argue that the competition between the two theories is specific algorithm for each potential analogy, or even a
ill-founded because they are after two different aspects of collection of algorithms for each domain of comparison, the
analogy: structure-mapping seeks a “horizontal” view of generalized structure of knowledge representation and the
analogy where the phenomena is examined at the level of structure-mapping algorithm makes it possible for any
already existing psychological representations, and where the properly constructed knowledge structure to be compared
task is to identify what processes are common to all or most and considered for structure-mapping.
analogy function; High-level Perception, on the other hand,
seeks a “vertical” view of analogy in which the goal is to
The Structure-Mapping Engine (SME) (Falkenhainer,
explain the processes that make up the construction of Forbus & Gentner, 1986, 1989; Gentner, Falkenhainer &
representations. An integrated theory of analogy should Skorstad, 1987) is a computer model intended to simulate
encompass both horizontal and vertical views. the structure-mapping process of Gentner’s theory. When
given a properly constructed representation, SME can find
a mapping between the appropriate relations for each
Introduction representation. A system of analogical retrieval, called
In this paper, we argue that there is a mistaken MAC/FAC (Gentner, 1989; Gentner & Forbus, 1991;
competition between two theories of analogy: Dedre Gentner, Ratterman & Forbus, 1993), provides a plausible
Gentner’s structure-mapping theory (SMT), and Douglas method by which a huge database of structured knowledge
Hofstadter’s theory of analogy as high-level perception representations may be searched and an item retrieved from
(HLP). Proponents of either theory have developed it with little computational strain. MAC/FAC includes
arguments claiming that their theory captures more of SME as the key component that performs the structure-
analogy than the other. The task of this paper is to separate mapping.
these combatants and show that their theories are after Douglas Hofstadter and several of his graduate students
explanations of different aspects of analogy phenomena, not (Hofstadter, 1984; Hofstadter, Mitchell & French, 1987;
necessarily pitted against one another. Chalmers et al., 1992; Mitchell, 1993; Hofstadter et al.,
1995) propose a different approach to the explanation of
Summary Of The Positions analogy by which an analogy is conceived of as the product
Gentner’s SMT (Gentner, 1980, 1983, 1989) describes of a more general cognitive function called high-level
analogy as a product of structure-mapping. A basic perception. HLP is the process by which an organism’s
assumption of SMT is that our psychological concepts have representation of a situation at a conceptual level is
a structure to them. According to Gentner, these structures constructed based on an interaction between high-level
are the psychological representations of relations between concepts and low-level perceptual processes: high-level
perceptual and conceptual objects. According to SMT, an concepts influence low-level perceptual processing, while
analogy, the ability to recognize that “one thing is like what is perceived at a low level affects the activation of
another,” is a mapping of one structure onto another high-level concepts as a representation of the situation is
according to a similarity comparison based on the relations constructed.1 The Copycat project, designed by Hofstadter
represented in the concept structures. Gentner and others and Mitchell (1992; Mitchell, 1990, 1993; Mitchell &
have set out to empirically test the explanatory power of
this conception with respect to human analogical
1The conceptual level is the level at which concepts begin to
production (Gentner, Falkenhainer & Skorstad, 1987;
Gentner & Imai, 1992; Gentner & Landers, 1985; Gentner play a role; a concept is anything from object recognition (e.g.,
& Rattermann, 1991; Gentner, Rattermann & Forbus, recognizing an apple) to the ability to grasp complex relations and
situations (e.g., that Bill Clinton is in the Democratic Party).
Hofstadter, 1990), is a model of analogy as high-level For these reasons, CFH believe that perception must be
perception. Copycat is intended to embody the principles of accounted for in a model of analogical thought —
high-level perception by building it’s own representation of something they claim has not been taken into account by
an analogy situation in a simplified letter-domain. In order current models of analogy. CFH make use of three
to really get at HLP, we must now turn to Hofstadter’s technical terms in characterizing the pitfalls of separating
criticisms of Gentner, and thus the confrontation between perception from high-level cognition. The first term is that
structure-mapping and high-level perception. of hand-coding representations. This refers to the
structuring of representations by humans. Hand-coding
The Fight becomes a fallacy when an implementer’s representations
The theory of high-level perception is first introduced in beg all the interesting cognitive questions; i.e., when the
print in the paper, High-level perception: representation implementer encodes all the information needed to
and analogy, by CFH2 (Chalmers, et al., 1992). The main complete the task (e.g., BACON3). The second term is
thesis of their paper is that HLP is deeply interwoven with rigidity, a feature of representations in which they are
other cognitive processes. They argue that much of the inflexible or unable to change. This, CFH claim, may be a
work in AI has attempted to model conceptual processes result of hand-coding representations rather than having
independently of perceptual processes, but that this the mechanism build them through some sort of
approach cannot lead to a satisfactory understanding of the constructive process. The third term is 20-20 hindsight,
human mind. Therefore, they argue, researchers in AI which occurs when a researcher develops a mechanism that
must integrate perceptual processing into their modeling of relies on hand-coded concepts rather than a mechanism
cognition. that independently arrives at a representational structure for
As an example of high-level cognition that depends on a particular concept. The mechanism using the hand-coded
high-level perception (and vice versa), CFH take up the representations is said to be guilty of 20-20 hindsight.4
cognitive process of analogy. They posit that, "when CFH accuse Gentner's model of analogy of bypassing the
people make analogies, they are perceiving some aspects of process of perception. They say that Gentner's approach
the structures of two situations — the essences of those involves starting with fixed, hand-coded representations
situations, in some sense — as identical" (Chalmers et al., which are compared and a mapping between them is
1992, p.193). The structures of the representations of these performed. CFH's criticisms are focused on the difficulties
situations are, they propose, the product of high-level that arise out of the rigidity of SME's representations, the
perception. Analogical thought also provides an foremost of which is the inability to change representations
illustration of the flexible nature of our perceptual abilities. as needed during processing. According to Gentner’s
For instance, making an analogy requires highlighting model, CFH claim, mapping is the only process involved in
various aspects of a situation, and the aspects that are analogy. In SME, representations are hand-coded, and thus
highlighted are often not the most obvious features from the building is ignored — it is assumed that "correct
beginning. The perception of the situation can change representations" will be available through some external
radically, depending on the analogy we are making. process. The result of SME's dependence on rigid
CFH divide the processes involved in analogical thought representations is that decisions in representation must be
into two basic components. One is situation perception, precisely the right one's before processing starts. But such
the filtering and organizing of data involved in a given information as to which one's are correct is not available
situation according to a given context. And the other is ahead of time unless (even to some extent) the analogy to
mapping, the taking of the representations of two situations be made is known ahead of time. And this is what makes
and finding appropriate correspondences between SME guilty of 20-20 hindsight.
components of one representation with components of the
other to produce the analogy match-up. CFH claim that, "it 3The creators (Langley et al.) of BACON, a model of scientific
is by no means apparent that these processes are cleanly discovery, claim that BACON is able to make scientific
separable; they seem to interact in a deep way" (Chalmers discoveries, such as Kepler's third law of planetary motion.
et al., 1992, p.195). They propose that because perception However, the model is given precisely the data required to derive
underlies analogy, we are tempted to divide the process of the law, so that it's "discovery" is reduced to a rather simple
analogy sequentially into situation perception, followed by deduction that any beginning physics student should be able to
mapping (as Gentner's analogy architecture suggests). deduce. For this reason, Chalmers et al. accuse BACON of
However, analogy is deeply involved in the situation having 20-20 hindsight in being given only what is needed, when
perception stage as well; perceptions of many situations are in fact the actual cognitive processes involved in such scientific
possible because of analogical mappings. Thus, CFH discoveries are faced with the much more monumental task of
conclude that situation perception and mapping processes paring down what is relevant and making careful hypotheses
go hand-in-hand. paired with testing, all part of an intricate process of induction.
4This fallacy is close to the fallacy of begging the question: for
example, claiming that a program has independently arrived at a
"discovery," when the "discovery" was already coded by the
2The authors of this paper are David J. Chalmers, Robert M. programmer into the program. For an excellent discussion of this
French and Douglas Hofstadter. We will refer to the authors as sort of fallacy concerning two other programs, AM and
CFH for abbreviation. EURISKO, see Koza 1992, pp.232-236.
 Talking Cross Purposes that structure-mapping is ultimately after. The situation
While these characterizations are true, CFH may have Gentner is explaining is consistently presented as an agent
inaccurately construed what Gentner et al. claim SME can being given an analogy and the agent then performs the
do. CFH claim that, "the SME program is said to discover mapping to create a new representation, thus high-lighting
an analogy between an atom and the solar system" how the two situations are analogous. This is to be
(Chalmers et al., 1992, p.196; emphasis added). This is not distinguished from the case of novel, independent, and
true. SME is intended to explain what happens in the unprompted analogy creation.
comprehension of an analogy between an atom and the It seems that CFH may have gotten their mistaken
solar system (Gentner, 1983; Gentner, 1989). It is not impression of Gentner from this ambiguity between analogy
being proposed as a model of discovery through analogy comprehension and production. This also demonstrates
(unlike BACON, where CFH's criticisms do seem apt). that CFH probably don't make a distinction between the
It seems there is an important distinction which both two. A model may exist in which these two are in one-and-
Gentner and CFH have missed. It begins with an approach the-same mechanism, but as Gentner's model appears to
that is not made explicit by Gentner concerning what the demonstrate, in not accounting for representation
structure-mapping theory is an explanation of. First construction, analogy comprehension might be separated
outlined in Gentner 1983, the Structure-Mapping Theory is from analogy production. SME is a model of analogy
described as an explanation of how analogies are produced. understanding or comprehension, where the cognitive
In order to get this explanation, however, Gentner claims agent is given the analogy and must understand it given its
that the theory is aimed at answering the following current knowledge database (i.e., construct a representation
question: how is it that one derives meaning from an from existing ones). In this case, there may be no need for
analogy? (Gentner, 1983, p.155) This latter construction of novel representations, and this limits its
characterization makes structure-mapping an explanation capacity to capture the dynamic aspects of natural analogy
of how analogies are understood or comprehended — the production. Any model that doesn't take this into account
analogy has already been given, but the agent must produce will be dependent on someone (or some module) to
a mapping in order to interpret the analogy. It is not clear organize the information in a suitable manner so that a
that the processes of production and comprehension are the mapping can take place. So the dependence on hand-
same. On the one hand, there are the situations where one coding representations has severely limited the ability of
may hear an analogy or be presented with one and then SME to capture creative analogical production in the wild.
come to understand it. In most examples of how structure- However, as SME is a model of analogy understanding, it is
mapping works, and especially in the operation of SME, immune to CFH's criticisms based on creating an analogy
the theory is presented as explaining this kind of situation. because in understanding, the hand-codings they look for
Production, on the other hand, seems to involve a different (hand-coding the representations) may be legitimate.
kind of situation: an agent observes events or situations and In light of this, it is no surprise that SME doesn’t do
is able, on its own, to construct a similarity comparison what CFH would like to see — they have mistaken what
between two situations (one observed and one recalled) and SME is explaining. This mistaken accusation as to what
thus produce a novel analogy. While both understanding Gentner ultimately claims SME can do is very telling. The
and production situations may share the common aspect of key criticisms that CFH bring to bear, with respect to what
structure mapping, the conditions that lead to and influence is missed in not accounting for representation construction,
the comparison are very different. All the conditions that depend on SME being intended to explain how
surround the mapping are vital to understanding whether representations are constructed. HLP seems to be after the
an analogy is being understood or whether a novel analogy an explanation of novel analogy production. Their
is being produced — to leave the conditions of the mapping explanatory task, from which their criticisms against
ambiguous seems to be a mistake. To claim that they are structure-mapping are derived, are aimed at how the
the same process is a profound statement about cognition. analogy situation is perceived, and thus, how
Whether structure-mapping is claiming this, or if it is only representations are constructed in the first place. But this
a model of analogy comprehension, is not made clear in seems to be something that SME isn’t trying to do. CFH
Gentner’s central papers (1980, 1983, 1989). and the SME researchers are talking at cross purposes.
Gentner’s theory is a theory of understanding
(specifically, for good analogies) but she doesn't have a The View Of Analogy From The Horizontal
theory of how analogies are created in the first place. For And The Vertical
example, she has a theory of what happened when The fight between Hofstadter and Gentner appears
Rutherford thought "the atom is like that solar system" and unwarranted, particularly in light of the fact that they are
of what may go on when we are given this information, but after different aspects of analogical reasoning. And CFH
she does not have a theory of how he managed to create this have in some sense been unfair in their criticisms of SME.
analogy in the first place. In other words, she doesn't have Their 1992 paper is written as though SMT and approaches
a theory of where structures come from (but she is working like it had completely missed what analogy was about.
on it; see her treatment of unpacking in Kotovsky & Instead, they addressed a new issue in analogy, and
Gentner, 1990). However, she does have a theory of what likewise failed to appreciate the perspective SMT addresses.
analogies are, given already existing structures. Thus, it
appears that it is an explanation of analogy comprehension
 But what has happened here? Both explanations seem a complete story using both horizontal and vertical
important to an understanding of analogy, yet we are perspectives; and it should be able to explain how analogies
arguing that they are talking past one another. We claim are produced as well as understood, according to the same
that this a result of the two different sets of perspectives and model (this issue is addressed in Morrison, 1994). While
goals in considering analogy phenomena. This is best we do not have a theory to offer, we do think that we have
described using a spatial metaphor of horizontal and added perspective to the “fight” that is taking place —
vertical perspectives. On the one hand, we have SMT, perspective that gives us a deeper understanding of the
which looks at analogy from a horizontal perspective: SMT theoretical terrain to be covered in a comprehensive theory
views analogy in it’s variety of forms and psychological of analogy, and what current approaches have to offer to
manifestations, over a variety of different comparison such theory.
domains, and searches for the common mechanism(s)
involved in all these cases. This perspective is likewise Table 1: Summary Comparison of
seen in SME, which embodies the positive aspects of SMT SME and Copycat:Pros and Cons
by employing explicit structure-mapping and being a SME
general mechanism for all kinds of possible comparison PROS:
domains. HLP, on the other hand, looks at analogy from a Employs explicit and well-defined structure-
vertical perspective: HLP views analogy as a process from mapping - if knowledge is structured (and we believe it
the bottom up; as a representation-building process based is), then there must be some sort of structure-mapping
on low-level perceptual processes interacting with high- that takes place to link up base-knowledge to target-
level concepts. Copycat embodies the positive aspects of knowledge
HLP by demonstrating how representations might be Based solidly on a robust psychological theory
constructed, and by not depending on human hand-coding developed over almost two decades of empirical
once set up in it’s micro-domain. investigation
This metaphor also helps to make sense out of the Generally applicable to all domains of analogical
criticisms that each theory has of the other in that we can comparison
now see what questions they don’t answer. First, SME does CONS:
not have an answer for the hand-coding problem. CFH and Hand-coding - can't produce it’s own
Mitchell are correct in that SME does not account for a representations. Thus, depends on work done by
very compelling problem: where do the representations humans (or implausible separate representation
upon which mapping is performed come from? In light of module).
this question, SME appears artificial. Copycat, on the other
hand, is constrained by domain specificity: it can only Copycat
produce analogies in a limited letter-domain. And it is PROS:
unclear as to how the Copycat-style architecture can be In a vague way, captures the notion of representation
extended to other, more psychologically plausible domains construction - closer to novel representation production
(Morrison, 1994). Thus, a similarly compelling problem is Once set up by humans in a specific domain (the
left unanswered with Copycat: How do we account for the letter-domain), it operates independently of humans
ability to produce analogies between practically any CONS:
domain? Both of these are daunting problems, and while Domain specificity - can only produce analogies in
each theory proposes to answer one, it has failed to limited letter-domain
captured the other. Table 1 is a summary comparison of
SME and Copycat to highlight the two perspectives, what
they are successful in explaining, and where they don’t. References
Chalmers, D. J., French, R. M., & Hofstadter, D. R.
Conclusion (1992). High-level perception, representation, and
We are left with a key question: can the two problems analogy: A critique of artificial intelligence methodology.
above be solved and in a unified way? As things stand Journal of Experimental & Theoretical Artificial
now, it seems as though we are in a dilemma in that the Intelligence 4:185--211.
attempts of SMT and HLP to model one aspect of analogy Falkenhainer, B., Forbus, K. D., & Gentner, D. (1986). The
resulted in missing another. And this is particularly structure-mapping engine. In Proceedings of the
compelling when one considers the extent and complexity American Association for Artificial Intelligence (pp. 272-
of each — these models and the theories behind them are -277), Philadelphia.
among the most respectable cognitive science has to offer. Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989). The
And while they may not have gotten the whole picture, they Structure-Mapping Engine: algorithm and examples.
do offer important perspectives of analogy phenomena. In Artificial Intelligence 41:1-63.
spite of this dilemma, we do wish to keep the faith and Gentner, D. (1980). The Structure of analogical models in
continue to believe that a comprehensive theory of analogy science (BBN Tech. Rep. 4451). Cambridge, MA: Bolt,
does exist. A comprehensive theory of analogy should be Beranek & Newman.
able to view analogy from all sides; it should be able to tell
Gentner, D. (1983). Structure-mapping: A theoretical Mitchell, M. & Hofstadter, D. R. (1990). The emergence
framework for analogy. Cognitive Science 7:155--170. of understanding in a computer model of concepts and
Gentner, D. (1989). The mechanisms of analogical analogy-making. Physica D 42:322--334.
learning. In Vosniadou, S. & Ortony, A. (Eds.), Morrison, C. T. (1994). Analogy and Constructive
Similarity and Analogical Reasoning (pp. 199--241). Representation. Master’s Thesis. Dept. of Philosophy,
Cambridge: Cambridge University Press. SUNY Binghamton, Binghamton, NY.
Gentner, D., Falkenhainer, B., & Skorstad, J. (1987,
January). Metaphor: The good, the bad, and the ugly. In
Proceedings of the Third Conference on Theoretical
Issues in Natural Language Processing (pp. 155--159).
Las Cruces, NM.
Gentner, D. & Forbus, K. (1991). MAC/FAC: A model of
similarity-based retrieval. In Proceedings of the
Thirteenth Annual Conference of the Cognitive Science
Society. Hillsdale, NJ: Lawrence Earlbaum Associates.
Gentner, D., & Imai, M. (1992). Is the future always
ahead? Evidence for system-mappings in understanding
space-time metaphors. In Proceedings of the Fourteenth
Annual Conference of the Cognitive Science Society.
Chicago, IL.
Gentner, D. & Landers, R. (1985, November). Analogical
remindings: A good match is hard to find. In
Proceedings of the International Conference on Systems,
Man, and Cybernetics. Tucson, AZ.
Gentner, D., & Rattermann, M. J. (1991). Language and
the career of similarity. In Gelman, S. A. & Byrnes, J. P.
(Eds.), Perspectives on thought and language:
Interrelations in development. London: Cambridge
University Press.
Gentner, D., Ratterman, M., & Forbus, K. (1993). The
roles of similarity in transfer: Separating retrievability
from inferential soundness. Cognitive Psychology 25:
524--575.
Hofstadter, D. R. (1984). The Copycat project: An
experiment in nondeterministic and creative analogies
(Memo 755). Cambridge, MA: MIT, Artificial
Intelligence Laboratory.
Hofstadter, D. R. & Mitchell, M. (1992). An overview of
the Copycat project. In Holyoak, K.J. & Barnden, J.
(Eds.) Connectionist Approaches to Analogy, Metaphor,
and Case-Based Reasoning (Norwood, NJ: Ablex).
Hofstadter, D. R. & Mitchell, M., & French, R. M. (1987).
Fluid concepts and creative analogies: A theory and its
computer implementation. Technical report 10,
Cognitive Science and Machine Intelligence Laboratory,
University of Michigan, Ann Arbor.
Hofstadter, D. R. et al. (1995). Fluid Concepts and
Creative Analogies. New York: Basic Books.
Kotovsky, L. & Gentner, D. (1990). Pack Light: You will
go farther. In Proceedings of the Second Midwest
Artificial Intelligence and Cognitive Science Society
Conference. (pp.60--72).
Mitchell, M. (1990). Copycat: A computer model of high-
level perception and conceptual slippage in analogy-
making. Doctoral dissertation, University of Michigan.
Mitchell, M. (1993). Analogy-Making as Perception: A
Computer Model. Cambridge: The MIT Press.