Emotional Analogies and Analogical Inference


Paul Thagard and Cameron Shelley

1. Introduction

Despite the growing appreciation of the relevance of affect to cognition, analogy researchers have paid remarkably little attention to emotion. This paper discusses three general classes of analogy that involve emotions. The most straightforward are analogies and metaphors about emotions, for example "Love is a rose and you better not pick it." Much more interesting are analogies that involve the transfer of emotions, for example in empathy in which people understand the emotions of others by imagining their own emotional reactions in similar situations. Finally, there are analogies that generate emotions, for example analogical jokes that generate emotions such as surprise and amusement.

Understanding emotional analogies requires a more complex theory of analogical inference than has been available. The next section presents a new account that shows how analogical inference can be defeasible, holistic, multiple, and emotional, in ways to be described. Analogies about emotions can to some extent be explained using the standard models such as ACME and SME, but analogies that transfer emotions require an extended treatment that takes into account the special character of emotional states. We describe HOTCO, a new model of emotional coherence, that simulates transfer of emotions, and show how HOTCO models the generation of emotions such as reactions to humorous analogies. Finally, we supplement our anecdotal collection of emotional analogies by discussing a more comprehensive sample culled from Internet wire services.

In logic books, analogical inference is usually presented by a schema such as the following (Salmon, 1984, p. 105):

Objects of type X have properties G, H, etc.

Objects of type Y have properties G, H, etc.

Objects of type X have property F.

Therefore: Objects of type Y have property F.

For example, when experiments determined that large quantities of the artificial sweetener saccharine caused bladder cancer in rats, scientists analogized that it might also be carcinogenic in humans. Logicians additionally point out that analogical arguments may be strong or week depending on the extent to which the properties in the premises are relevant to the property in the conclusion.

This characterization of analogical inference, which dates back at least to John Stuart Mill's nineteenth-century System of Logic, is flawed in several respects. First, logicians rarely spell out what "relevant" means, so the schema provides little help in distinguishing strong analogies from weak. Second, the schema is stated in terms of objects and their properties, obscuring the fact that the strongest and most useful analogies involve relations, in particular causal relations (Gentner, 1983; Holyoak and Thagard, 1995). Such causal relations are usually the key to determining relevance: if, in the above schema, G and H together cause F in X, then analogically they may cause F in Y, producing a much stronger inference than just counting properties. Third, logicians typically discuss analogical arguments and tend to ignore the complexity of analogical inference, which requires a more holistic assessment of a potential conclusion with respect to other information. There is no point in inferring that objects of type Y have property F if you already know of many such objects that lack F, or if a different analogy suggests that they do not have F. Analogical inference must be defeasible, in that the potential conclusion can be overturned by other information, and it must be holistic in that everything the inference maker knows is potentially relevant to overturning or enhancing the inference.

Compared to the logician's schema, much richer accounts of the structure of analogies have been provided by computational models of analogical mapping such as SME (Falkenhainer, Forbus, and Gentner, 1989) and ACME (Holyoak and Thagard, 1989). SME uses relational structure to generate candidate inferences, and ACME transfers information from a source analog to a target analog using a process that Holyoak, Novick and Melz (1994) called copying with substitution and generation (CWSG). Similar processes are used in case-based reasoning (Kolodner, 1993), and in many other computational models of analogy.

But all of these computational models are inadequate for understanding analogical inference in general and emotional analogies in particular. They do not show how analogical inference can be defeasible and holistic, or how it can make use of multiple source analogs to support or defeat a conclusion. Moreover, the prevalent models of analogy encode information symbolically and assume that what is inferred is verbal information that can be represented in propositional form by predicate calculus or some similar representational system. (One of the few attempts to deal with nonverbal analogies is the VAMP system for visual analogical mapping: Thagard, Gochfeld, and Hardy, 1992.) As section 5 documents, analogical inference often serves to transfer an emotion, not just the verbal representation of an emotion. We will now describe how a new model of emotional coherence, HOTCO, can perform analogical inferences that are defeasible, holistic, multiple, and emotional.

Thagard (forthcoming, ch. 6) proposes a theory of emotional coherence that has applications to numerous important psychological phenomena such as trust. This theory makes the following assumptions about inference and emotions:

From this coherentist perspective, all inference is defeasible and holistic and differs markedly from logical deduction in formal systems. Philosophers who have advocated coherentist accounts of inference include Bosanquet (1920) and Harman (1986).

The computational model HOTCO (for "hot coherence") implements these theoretical assumptions. It amalgamates the following previous models of coherence:

Amalgamation is natural, because all of these models use a similar connectionist algorithm for maximizing constraint satisfaction, although they employ different constraints operating on different kinds of representation. What is novel about HOTCO is that representational elements possess not only activations that represent their acceptance and rejection, but also valences that represent a judgment of their positive or negative emotional appeal. In HOTCO, as in its component models, inferences about what to accept are made by a holistic process in which activation spreads through a network of units with excitatory and inhibitory links, representing elements with positive and negative constraints. But HOTCO spreads valences as well as activations in a similar holistic fashion, using the same system of excitatory and inhibitory links. For example, HOTCO models the decision of whether to hire a particular person as a babysitter as in part a matter of "cold" deliberative, explanatory, conceptual, and analogical coherence, but also as a matter of generating an emotional reaction to the candidate. The emotional reaction derives from a combination of the cold inferences made about the person and the valences attached to what is inferred. If you infer that a babysitting candidate is responsible, intelligent, and likes children, the positive valence of these attributes will spread to him or her; whereas if coherence leads to you infer that the candidate is lazy, dumb, and psychopathic, he or she will acquire a negative valence. In HOTCO, valences spread through the constraint network in much the same way that activation does (see Thagard, forthcoming, for technical details). The result is an emotional Gestalt that provides an overall "gut reaction" to the potential babysitter.

Now we can describe how HOTCO performs analogical inference in a way that is defeasible, holistic, and multiple. HOTCO uses ACME to perform analogical mapping between a source and a target, and uses copying with substitution and generation to produce new propositions to be inferred. It can operate either in a broad mode in which everything about the source is transferred to the target, or in a more specific mode in which a query is used to enhance the target using a particular proposition in the source. Here, in predicate calculus formalization where each proposition has the structure (predicate (objects) proposition-name), is an example of a scientific analogy concerning the coelacanth, a rare fish that is hard to study directly (Shelley, 1999).

Source 1: centroscymnus

(have (centroscymnus rod-pigment-1) have-1

(absorb (rod-pigment-1 472nm-light) absorb-1)

(penetrate (472nm-light deep-ocean-water) penetrate-1)

(see-in (centroscymnus deep-ocean-water) see-in-1)

(inhabit (centroscymnus deep-ocean-water) inhabit-1)

(enable (have-1 see-in-1) enable-1)

(because (absorb-1 penetrate-1) because-1)

(adapt (see-in-1 inhabit-1) adapt-1)

Target: coelacanth-3

(have (coelacanth rod-pigment-3) have-3)

(absorb (rod-pigment-3 473nm-light) absorb-3)

(penetrate (473nm-light deep-ocean-water) penetrate-3)

(see-in (coelacanth deep-ocean-water) see-in-3)

(enable (have-3 see-in-3) enable-3)

(because (absorb-3 penetrate-3) because-3)

Operating in specific mode, HOTCO is asked what depth the coelacanth inhabits, and uses the proposition INHABIT-1 in the source to construct for the target the proposition

(inhabit (coelacanth deep-ocean-water) inhabit-new)

Operating in broad mode and doing general copying with substitution and generation, HOTCO can analogically transfer everything about the source to the target, in this case generating the proposition that coelacanths inhabit deep water as a candidate to be inferred.

However, HOTCO does not actually infer the new proposition, because analogical inference is defeasible. Rather, it simply establishes an excitatory link between the unit representing the source proposition INHABIT-1 and the target proposition INHABIT-NEW. This link represents a positive constraint between the two propositions, so that coherence maximization will encourage them to be accepted together or rejected together. The source proposition INHABIT-1 is presumably accepted, so in the HOTCO model it will have positive activation which will spread to provide positive activation to INHABIT-NEW, unless INHABIT-NEW is incompatible with other accepted propositions that will tend to suppress its activation. Thus analogical inference is defeasible, because all HOTCO does is to create a link representing a new constraint for overall coherence judgment, and it is holistic, because the entire constraint network can potentially contribute to the final acceptance or rejection of the inferred proposition.

Within this framework, it is easy to see how analogical inference can employ multiple analogies, because more than one source can be used to create new constraints. Shelley (1999) describes how biologists do not simply use the centroscymnus analog as a source to infer that coelacanths inhabit deep water, but also use the following different source:

Source 2: ruvettus-2

(have (ruvettus rod-pigment-2) have-2)

(absorb (rod-pigment-2 474nm-light) absorb-2)

(penetrate (474nm-light deep-ocean-water) penetrate-2)

(see-in (ruvettus deep-ocean-water) see-in-2)

(inhabit (ruvettus deep-ocean-water) inhabit-2)

(enable (have-2 see-in-2) enable-2)

(because (absorb-2 penetrate-2) because-2)

(adapt (see-in-2 inhabit-2) adapt-2)

The overall inference is that coelacanths inhabit deep water because they are like the centroscysmus and the ruvettus sources in having rod pigments that adapt them to life in deep water. Notice that these are deep, systematic analogies, because the theory of natural selection suggests that the two source fishes have the rod pigments because the pigments adapt them to life in deep ocean water environments. When HOTCO maps the ruvettus source to the coelecanth target after mapping the centroscysmus source, it creates excitatory links from the inferred proposition INHABIT-NEW with both INHABIT-1 in the first source and INHABIT-2 in the second source. Hence activation can flow from both these propositions to INHABIT-NEW, so that the inference is supported by multiple analogies. If another analog or other information suggests a contradictory inference, then INHABIT-NEW will be both excited and inhibited. Thus multiple analogies can contribute to the defeasible and holistic character of analogical inference.

The new links created between the target proposition and the source proposition can also make possible emotional transfer. The coelacanth example is emotionally neutral for most people, but if an emotional valence were attached to INHABIT-1 and INHABIT-2, then the excitatory links between them and INHABIT-NEW would make possible spread of that valence as well as spread of activation representing acceptance. In complex analogies, in which multiple new excitatory link are created between aspects of one or more sources and the target, valences can spread over all the created links, contributing to the general emotional reaction to the target. Section 5 below provides detailed examples of this kind of emotional analogical inference.

The Columbia Dictionary of Quotations (available electronically as part of the Microsoft Bookshelf) contains many metaphors and analogies concerning love and other emotions. For example, love is compared to religion, a master, a pilgrimage, an angel/bird, gluttony, war, disease, drunkenness, insanity, market exchange, light, ghosts, and smoke. It is not surprising that writers discuss emotions non-literally, because it is very difficult to describe the experience of emotions in words. In analogies about emotions, verbal sources help to illuminate the emotional target, which may be verbally described but which also has an elusive, non-verbal, phenomenological aspect. Analogies are also used about negative emotions: anger is like a volcano, jealousy is a green-eyed monster, and so on.

In order to handle the complexities of emotion, poets often resort to multiple analogies, as in the following examples:

(1) John Donne:

Love was as subtly catched, as a disease;

But being got it is a treasure sweet.

(2) Robert Burns:

O, my love is like a red, red rose,

That's newly sprung in June:

My love is like a melodie,

That's sweetly play'd in tune.

Love is a smoke made with the fume of sighs,

Being purged, a fire sparkling in lovers' eyes,

Being vexed, a sea nourished with lovers' tears.

What is it else? A madness most discreet,

A choking gall and a preserving sweet.

In each of these examples, the poet uses more than one analogy or metaphor to bring out different aspects of love. The use of multiple analogies is different from the scientific example described in the last section, in which the point of using two marine sources was to support the same conclusion about the depths inhabited by coelacanths. In these poetic examples, different source analogs bring out different aspects of the target emotion, love.

Analogies about emotions may be general, as in the above examples about love, or particular, used to describe the emotional state of an individual. For example, in the movie Marvin's Room, the character Lee played by Meryl Streep describes her reluctance to discuss her emotions by saying that her feelings are like fishhooks - you can't pick up just one. Just as it is hard to verbalize the general character of an emotion, it is often difficult to describe verbally one's own emotional state. Victims of post-traumatic stress disorder frequently use analogies and metaphors to describe their own situations (Meichenbaum (1994, pp. 112-113):

In these particular emotional analogies, the target to be understood is the emotional state of an individual, and the verbal source describes roughly what the person feels like.

The purpose of analogies about emotions is often explanatory, describing the nature of a general emotion or a particular person's emotional state. But analogy can also be used to help deal with emotions, as in the following quote from Nathaniel Hawthorne: "Happiness is a butterfly, which, when pursued, is always just beyond your grasp, but which, if you will sit down quietly, may alight upon you." People are also given advice on how to deal with negative emotions, being told for example to "vent" their anger, or to "put a lid on it."

In principle, analogies about emotions could be simulated by the standard models such as ACME and SME, with a verbal representation of the source being used to generate inferences about the emotional target. However, even in some of the above examples, the point of the analogy is not just to transfer verbal information, but also to transfer an emotional attitude. When someone says "I feel like I am caught up in a tornado," he or she may be saying something like "My feelings are like the feelings you would have if you were caught in a tornado." To handle the transfer of emotions, we need to go beyond verbal analogy.

As already mentioned, not all analogies are verbal: some involve transfer of visual representations (Holyoak and Thagard, 1995). In addition, analogies can involve transfer of emotions from a source to a target. There are at least three such kinds of emotional transfer, involved in persuasion, empathy, and reverse empathy. In persuasion, I may use an analogy to convince you to adopt an emotional attitude. In empathy, I try to understand your emotional reaction to a situation by transferring to you my emotional reaction to a similar situation. In reverse empathy, I try to get you to understand my emotion by comparing my situation and emotional response to it with situations and responses familiar to you.

The purpose of many persuasive analogies is to produce an emotional attitude, for example when an attempt is made to convince someone that abortion is abominable or that capital punishment is highly desirable. If I want to get someone to adopt positive emotions toward something, I can compare it to something else toward which he or she already has a positive attitude. Conversely, I can try to produce a negative attitude by comparison with something already viewed negatively. The structure of persuasive emotional analogies is:

You have an emotional appraisal of the source S.

The target T is like S in relevant respects.

So you should have a similar emotional appraisal of T.

Of course, the emotional appraisal could be represented verbally by terms such as "wonderful," "awful," and so on, but for persuasive purposes it is much more effective if the particular gut feeling that is attached to something can itself be transferred over to the target. For example, emotionally intense subjects such as the Holocaust or infanticide are commonly used to transfer negative emotions.

Blanchette and Dunbar (1997) thoroughly documented the use of persuasive analogies in a political context, the 1995 referendum in which the people of Quebec voted whether to separate from Canada. In three Montreal newspapers, they found a total of 234 different analogies, drawn from many diverse source domains: politics, sports, business, and so on. Many of these analogies were emotional: 66 were coded by Blanchette and Dunbar as emotionally negative, and 75 were judged to be emotionally positive. Thus more than half of the analogies used in the referendum had an identifiable emotional dimension. For example, the side opposed to Quebec separation said "It's like parents getting a divorce, and maybe the parent you don't like getting custody." Here the negative emotional connotation of divorce is transferred over to Quebec separation. In contrast, the yes side used positive emotional analogs for separation: "A win from the YES side would be like a magic wand for the economy."

HOTCO can naturally model the use of emotional persuasive analogies. The separation-divorce analogy can be represented as follows:

Source : divorce

(married (spouse-1 spouse-2) married-1)

(have (spouse-1 spouse-2 child) have-1)

(divorce (spouse-1 spouse-2) divorce-1) negative valence

(get-custody (spouse-1) get-custody-1)

(not-liked (spouse-1) not-liked-1) negative valence

Target: separation

(part-of (Quebec Canada) part-of-2)

(govern (Quebec Canada people-of-Quebec) govern-2)

(separate-from (Quebec Canada) separate-from--2)

(control (Quebec people-of-Quebec) control-2)

When HOTCO performs a broad inference on this example, it not only computes the analogical mapping from the source to the target and completes the target using copying with substitution and generation, but also transfers the negative valence attached to the proposition DIVORCE-1 to SEPARATE-FROM-2.

Persuasive analogies have been rampant in the recent debates about whether Microsoft has been engaging in monopolistic practices by including its World Wide Web browser in its operating system, Windows 98. In response to the suggestion that Microsoft also be required to include the rival browser produced by its competitor, Netscape, Microsoft's chairman Bill Gates complained that this would be "like requiring Coca-Cola to include three cans of Pepsi in every six-pack it sells," or like "ordering Ford to sell autos fitted with Chrysler engines." These analogies are in part emotional, since they are intended to transfer the emotional response to coercing Coca-Cola and Ford - assumed to be ridiculous - over to the coercion of Microsoft. On the other hand, critics of Microsoft's near-monopoly on personal computer operating systems have been comparing Gates to John D. Rockefeller, whose predatory Standard Oil monopoly on petroleum products was broken up by the U.S. government in 1911.

Persuasive analogies suggest a possible extension to the multiconstraint theory of analogical reasoning developed by Holyoak and Thagard (1995). In that theory, similarity is one of the constraints that influence how two analogs are mapped to each other, including both the semantic similarity of predicates and the visual similarity of objects. We conjecture that emotional similarity may also influence analogical mapping and predict that people will be more likely to map elements that have the same positive or negative valence. For example, if you have a positive feeling about Bill Gates and a negative feeling about John D. Rockefeller, you will be less likely to see them as analogs, impeding both retrieval and mapping. A recent advertisement for a book on cancer compared the genetically mutated cell that initiates a tumor growth to Jesus initiating Christianity. Regardless of how structurally informative this analogy might be, the correspondences between cancer cells and Jesus and between tumors and Christianity produce for many people an emotional mismatch that renders the analogy ineffective. During the Kosovo war in 1999, comparisons were frequently made between the Serbian leader Slobodan Milosovic and Adolf Hitler; these comparisons were emotionally congruent for most people, but not for Milosovic's supporters.

Another, more personal, kind of persuasive emotional analogy is identification, in which you identify with someone and then transfer positive emotional attitudes about yourself to them. According to Fenno (1978, p. 58), members of the U.S. congress try to convey a sense of identification to their constituents. The message is "You know me, and I'm like you, so you can trust me." The structure of this kind of identification is:

You have a positive emotional appraisal of yourself (source).

I (the target) am similar to you.

So you should have a positive emotional appraisal of me.

Identification is a kind of persuasive analogy, but differs from the general case in that the source and target are the people involved. A full representation of the similarity involved in identification and other analogies requires specification of the causal and other higher-order relations that capture deep, highly relevant similarities between the source and target.

Empathy also involves transfer of emotional states between people; see Barnes and Thagard (1997) for a full discussion. It differs from persuasion in that the goal of the analogy is to understand rather than to convince someone. Summarizing, the basic structure is:

You are in situation T (target).

When I was in a similar situation S, I felt emotion E (source).

So maybe you are feeling an emotion similar to E.

As with persuasion and identification, such analogizing could be done purely verbally, but it is much more effective to actually feel something like what the target person is feeling. For example, if Thagard wants to understand the emotional state of a new graduate student just arrived from a foreign country, he can recall his emotional state of anxiety and confusion when he went to study in England. Here is a more detailed example of empathy involving someone trying to understand the distress of Shakespeare's Hamlet at losing his father by comparing it to his or her own loss of a job (from Barnes and Thagard, 1997):

Source: you Target: Hamlet

(fire (boss, you) s1-fire) (kill (uncle, father) t1-kill)

(lose (you, job) s2-lose) (lose (Hamlet, father) t2-lose)

(marry (uncle, mother) t2a-marry)

(cause (s1-fire, s2-lose) s3) (cause (t1-kill, t2-lose) t3)

(angry (you): s4-angry) (angry (Hamlet) t4-angry)

(depressed (you) s5-depressed) (depressed (Hamlet) t5-depressed)

(cause (s2-lose, s4-angry) s6) (cause (t2-lose, t4-angry) t6)

(cause (s2-lose, s5-depressed) s7) (cause (t2-lose, t5-depressed) t7)

(indecisive (you) s8-indecisive)

(cause (s5-depressed, s8-indecisive) s9)

The purpose of this analogy is not simply to draw the obvious correspondences between the source and the target, but to transfer over your remembered image of depression to Hamlet.


Unlike persuasive analogies, whose main function is to transfer positive or negative valence, empathy requires transfer of the full range of emotional responses. Depending on his or her situation, I need to imagine someone being angry, fearful, disdainful, ecstatic, enraptured and so on. As currently implemented, HOTCO transfers only positive or negative valences associated with a proposition or object, but it can easily be expanded so that transfer involves an emotional vector which represents a pattern of activation of numerous units, each of whose activation represents different components of emotion. This expanded representation would also make possible the transfer of "mixed" emotions.

Empathy is only one kind of explanatory emotional analogy. In section 4, we already saw examples of analogies whose function is to explain one's own emotional state to another, a kind of reverse empathy in that it enables others to have an empathic understanding of oneself. Here is the structure of reverse empathy:

Here is a final example of analogical transfer of emotion: "Psychologists would rather use each other's toothbrushes than each other's terminology." This analogy is complex, because at one level it is projecting the emotional reaction of disgust from use of toothbrushes to use of terminology, but it is also generating amusement. A similar dual role is also found in the following remark in The Globe and Mail: "Starbuck's coffee shops are spreading through Toronto faster than head lice through a kindergarten class." Both these examples convey an attitude as does the remark of country music star Garth Brooks: "My job is like what people say about pizza and sex: When it's good, it's great; and even when it's bad, it's still pretty good." Note that this is a multiple analogy, for what its worth. The writer Flaubert also used an analogy to convey his attitude toward his work: "I love my work with a love that is frantic and perverted, as an ascetic loves the hairshirt that scratches his belly." Let us now consider analogies that go beyond analogical transfer of emotions and actually generate new emotions.

6. Analogies that Generate Emotions

A third class of emotional analogies involves ones that are not about emotions and do not transfer emotional states, but rather serve to generate new emotional states. There are at least four subclasses of emotion-generating analogies, involving humor, irony, discovery, and motivation.

One of the most enjoyable uses of analogy is to make people laugh, generating the emotional state of mirth or amusement. The University of Michigan recently ran an informational campaign to get people to guard their computer passwords more carefully. Posters warn students to treat their computer passwords like underwear: make them long and mysterious, don't leave them lying around, and change them often. The point of the analogy is not to persuade anyone based on the similarity between passwords and underwear, but rather to generate amusement that focuses attention on the problem of password security.

A major part of what makes an analogy funny is a surprising combination of congruity and incongruity. Passwords do not fit semantically with underwear, so it is surprising when a good relational fit is presented (change them often). Other emotions can also feed into making an analogy funny, for example when the analogy is directed against a person or group one dislikes:

Why do psychologists prefer lawyers to rats for their experiments?

1. There are now more lawyers than rats;

2. The psychologists found they were getting attached to the rats;

3. And there are some things that rats won't do.

This joke depends on a surprising analogical mapping between rats in psychological experiments and lawyers in their practices, and on negative emotions attached to lawyers. Further surprise comes from the addendum that psychologists have stopped using lawyers in their experiments because the results did not transfer to humans. Another humorous analogy is implicit in the joke: "How can a single woman get a cockroach out of her kitchen? Ask him for a commitment."

Some analogical jokes depend on visual representations, as in the following children's joke: "What did the 0 say to the 8? Nice belt." This joke requires a surprising visual mapping between numerals and human dress. A more risqué visual example is. "Did you hear about the man with five penises? His pants fit like a glove." Here are some more humorous analogies, all of which involve mappings that generate surprise and amusement:

(1) Safe eating is like safe sex: You may be eating whatever it was that what you're eating ate before you ate it.

(2) Changing a university has all the difficulties of moving a cemetery.

(3) Administering academics is like herding cats.

(4) An associate dean is a mouse training to be a rat.

(5) The juvenile sea squirt wanders through the sea searching for a suitable rock or hunk of coral to cling to and make its home for life. For this task, it has a rudimentary nervous system. When it finds its spot and takes root, it doesn't need its brain anymore, so it eats it! (It's rather like getting tenure.) (Dennett 1991, p. 177)

(6) Bill James on Tim McCarver's book on baseball: But just to read the book is nearly impossible; it's like canoeing across Lake Molasses.

(7) Red Smith: Telling a non-fan about baseball is like telling an 8-year-old about sex. No matter what you say, the response is "But why?"

(8) Melissa Franklin (Harvard physicist) on quarks: It's weird. You've got six quarks; five of them are really light, and the sixth is unbelievably heavy. It's as if you had Sleepy, Dopey, Grumpy, Happy, and Kierkegaard.

Note that Franklin only mentions five dwarfs, so the mapping is not one-one, a fact that does not undermine the analogy. Failure of one-one mapping can even be funny, as in a 1998 cartoon that showed a ship labeled "Bill Clinton" about to hit an iceberg labeled "Bill Clinton".

In the emotional coherence theory of Thagard (forthcoming), surprise is treated as a kind of metacoherence. When HOTCO shifts from coherent interpretation to another, with units that were previously activated being deactivated and vice versa, the units that underwent an activation shift activate a surprise node. In analogical jokes, the unusual mapping produces surprise because it connects together elements not previously mapped, but does so in a way that is still highly coherent. The combination of activation of the surprise node, the coherence node, and other emotions generates humorous amusement.


Analogies that are particularly deep and elegant can also generate an emotion similar to that produced by beauty. A beautiful analogy is one so accurate, rich, and suggestive that it has the emotional appeal of an excellent scientific theory or mathematical theorem. Holyoak and Thagard (1995, ch. 8), describe important scientific analogies such as the connection with Malthusian population growth that inspired Darwin's theory of natural selection. Thus scientific and other elegant analogies can generate positive emotions such as excitement and joy without being funny.

Not all analogies generate positive emotions, however. Ironies are sometimes based on analogy, and they are sometimes amusing, but they can also produce negative emotions such as despair:

HONG KONG (January 11, 1998 AF-P) - Staff of Hong Kong's ailing

Peregrine Investments Holdings will turn up for work Monday still

in the dark over the fate of the firm and their jobs.

Other Peregrine staff members at the brokerage were quoted as

saying Sunday they were pessimistic over the future of the firm,

saddled with an estimated 400 million dollars in debts.

"I'm going to see the Titanic movie...that will be quite ironic,

another big thing going down," the South China Morning Post

quoted one broker as saying.

Shelley (forthcoming) argues that irony is a matter of "bicoherence," with two situations being perceived as both coherent and incoherent with each other. The Peregrine Investments-Titanic analogy is partly a matter of transferring the emotion of despair from the Titanic situation to the company, but the irony generates an additional emotion of depressing appropriateness.

The final category of emotion-generating analogies we want to discuss is motivational ones, in which an analogy generates positive emotions involved in inspiration and self-confidence. Lockwood and Kunda (1997) have described how people use role models as analogs to themselves, in order to suggest new possibilities for what they can accomplish. For example, an athletic African American boy might see Michael Jordan as someone who used his athletic ability to achieve great success. By analogically comparing himself to Michael Jordan, the boy can feel good about his chances to accomplish his athletic goals. Adopting a role model in part involves transferring emotions, e.g. transferring the positive valence of the role model's success to one's own anticipated success, but it also generates new emotions accompanying the drive and inspiration to pursue the course of action that the analogy suggests. The general structure of the analogical inference is:


My role model accomplished the goal G by doing the action A.

I am like my role model in various respects.

So maybe I can do A to accomplish G.

The inference that I may have the ability to do A can generate great excitement about the prospect of such an accomplishment.

The examples so far discussed in this paper were collected haphazardly, and thus amount to anecdotes rather than data. To compile analogies more systematically, Cam Shelley wrote a program to use the Internet to search for stories about analogy. Candidate news articles were collected by a keyword search for the term "analogy" and other comparative terms through Internet-based search engines, from February 1997 through September 1998. Many candidate articles were rejected due to lack of clarity. The others were classified by Shelley according to their primary practical function, including:

Inferential analogies were further divided into two types: "hot" and "cold." Hot analogies serve to transfer emotional tags or attitudes from the source to target, whereas cold analogies serve simply to transfer structured information without producing any particular affect.

Hot analogies were further broken down into the three types of emotional analogies discussed above, namely empathy, reverse empathy, and persuasion. Very often the analogies conveyed in each news article were complex and appeared to serve more than one of the functions listed here. However, we counted each analogy only once on the basis of its most salient function. Figure 1 displays the results. Note that well over half of the analogies had emotional content and purpose. Perhaps this result is not surprising, since news reports, particularly those of the major wire services such as the Associated Press or Reuters, tend to be sensational to attract readers.

Many of the non-emotional analogies came from reports of scientific advances, such as medical research on the origin of the human immune system. For example, when Dr. Gary Litman expresses an enthusiastic interest in the analogy between human and shark immune systems, the comparison serves largely to inform his research rather than to create excitement among his colleagues (from "Scientist goes fishing", 11 February 1998, St. Petersburg Times).

Remarkably few of the emotional analogies seem to display the process of empathy. News reports were counted as examples of empathy where they showed people attempting to understand another person by analogies about their own emotional experiences. Given the interest that reporters have in recording people in the grip of highly emotional experiences, it appears surprising that so few examples of empathy should be present in the database. But there are at least two reasons to suppose that the lack of examples is illusory.


Figure 1. Classification of analogies found on Internet wire services.

First, we might expect people to empathize based on their own personal experiences. But this expectation is implicitly based on the idea that our personal experiences are necessarily the most immediate or salient experiences that we might recall. This idea appears to be untrue. The examples of empathy in the database show people empathizing with others based on conventional or stereotypical experiences. For example, Dr. William Catalona expresses empathy with middle-aged men who are clamoring for the potency drug Viagra by comparing it to the demand for wrinkle cream (from "Baby Boomers fight aging in a variety of ways", by P. B. Librach, 1 June 1998, Scripps Howard News Service): "It's like the Fountain of Youth," said Catalona. "Viagra is analogous to anti-wrinkle cream. This is something that will turn back the clock and make men the way they were when they were young."

In other words, Catalona empathizes with middle-aged men by comparing their desire for Viagra with what he imagines to be the desire of middle-aged women for wrinkle cream. This source analog, although it does not represent Catalona's own experience, was selected due to (a) the surface similarity between Viagra as a drug and wrinkle cream as a cosmetic and (b) the representativeness of the wrinkle-cream situation as a stereotype of the (foolish?) desire to feel young. Because people tend to pick out stereotyped analogs when empathizing, it is not always clear whether they are truly empathizing instead of simply persuading the reporter that two emotional situations are indeed similar. Empathy and persuasion can be difficult to tell apart as they are reported in news articles.

The second reason why empathy appears to be rare is that it may serve as a first step in the process of persuasion or reverse empathy, both of which involve someone attempting to influence someone else's judgment of emotions. But, in order to affect people's judgments on such issues to a particular end, it is necessary to have some estimate of their emotional repertoire or their opinions on emotional situations. Getting the correct emotional response out of people requires some idea of their view of emotions and emotional topics. Empathy is an important means for getting such an idea about someone. (There are other means, of course, such as using a general set of rules about people's emotional responses.) For example, consider the analogy made by Israeli Prime Minister Benjamin Netanyahu who compared the Palestinian Authority to "a regime facilitating terror" (from "U.S. raps Arafat, Netanyahu for rhetoric", 7 August 1998, Reuters):

Here Netanyahu compares Arafat's Palestinian Authority to Gaddafi's regime in Libya in order to project to his audience, which includes prominent Americans, his feelings about Arafat. However, Netanyahu's choice of Libya as the source analog is suggested not merely by any similarities between it and the Palestinian Authority, but also by his knowledge that Libya is particularly reviled by Americans. In other words, Netanyahu empathizes with Americans enough to know that mention of Libya and Gaddafi is likely to arouse strong, negative emotions in them. Certainly, U.S. Secretary of State Madeline Albright found his choice of source analog to be inflammatory.

Table 1 indicates the frequency with which emotions of positive and negative valence occur in the analogy database. Valences were assigned on a conventional basis, with emotional states suggesting fear, anger, disgust, sadness, and surprise counted as negative, whereas emotional states suggesting happiness, pride, admiration, calm, and trust counted as positive. Of course, the emotional states described in the news articles were typically complex and could not be completely captured simply as instances of happiness or anger, for example. We used the so-called basic emotions (anger, fear, disgust, surprise, and happiness) to catalog each example of an emotional analogy. Where such a description was obviously impossible or inadequate, terms referring to more complex states were used, e.g., pride and admiration. Eleven emotion terms were employed for the purposes of basic categorization: fear, happy, admiration, calm, disgust, anger, pride, trust, sad, surprise, and discomfort. Other terms were used to supplement these terms where they failed to completely capture the emotional content of an analogy.



positive emotion


negative emotion








reverse empathy











Table 1. Frequency of positive and negative emotions in emotional analogies.

The low number of examples of empathy precludes drawing any conclusions about that process, but the figures for reverse empathy and persuasion are more interesting. Instances of reverse empathy were fairly evenly distributed between positive and negative valences. As we described in sections 4 and 5, reverse empathy involves using an analogy in order to get someone else to understand your own emotional state. Consider the example given by Norman Siegel, executive director of the Civil Liberties Union, who joined a 28 member task force created by New York Mayor Rudolph Giuliani to look into the problem of police brutality (from "Under fire, Guiliani forms panel on police", by Fred Kaplan, 20 August 1997, Boston Globe.). Although Siegel remained cautious about Guilani's committment to the task force, he emphasized that he is willing for now to take the mayor at his word, saying Giuliani is in a unique position "to, once and for all, confront the systemic problems of police brutality." The analogy is Nixon going to China, Siegel said in a telephone interview earlier yesterday. Just as President Nixon, who long was known as a foe of communism, could recognize China without facing accusations of being soft on communism, so Giuliani - a former US prosecutor who has supported the police at nearly every turn of his political career - can mount an attack on the police force without prompting charges of being soft on crime.

This analogy suggests that Guiliani is at least in a good position to undertake a review of police misbehavior. But it also conveys Siegel's emotional attitude towards the task force. Siegel explains why he trusts Guiliani (takes him at his word) in terms of the trust that people invested in Nixon due to his long-standing anticommunist activities when he recognized communist China. In both cases, the trust signals belief in a positive outcome of the activity concerned, although the choice of "Tricky Dick" as a source analog might also signal that Siegel's belief in Giuliani's sincerity remains qualified by doubt.

An excellent example of the projection of a positive emotion in reverse empathy is given by astronaut David Wolf's description of what it felt like when he began his mission on the Mir space station (from "Astronaut describes life on Mir in wry e-mail home", by Adam Tanner, 2 December 1997, Reuters):

Here, Wolf projects his feelings about being left behind on a space station by reference to his feelings about being left behind at summer camp as a youth. Besides conveying the structure of the event, Wolf's use of a source analog shared by many Americans helpfully conveys the positive thrill of anticipation that he felt during an event very few people have ever experienced themselves.

Persuasive emotional analogies usually serve not to communicate an emotion between people of varied experience, but to have one person transfer some emotional attitude from one concept to another. For example, in order to persuade people to feel calm about an apparently dangerous situation, you might remind them of analogous situations that they regard calmly. Such a case actually occurred in response to volcanic activity in the neighborhood of Mexico City (from an untitled article, 24 April 1997, Associated Press):

In this example, Quass attempts to persuade the public to feel calm regarding the volcanic activity of Popocatepetl by comparing it to a belch, an event no one presumably feels is too grave. Positive occurrences of persuasion tend to involve such humorous belittlement of some situation or concern.

Table 1 shows that there is a great asymmetry in the database between positive and negative valences in cases of persuasion. It appears that persuasive analogies are mostly used to condemn a person or situation. Consider the reaction of a Boston Globe reporter to a vegetarian burger being tested in McDonald's outlets in New York's Greenwich Village by its inventor Jim Lewis (from "Will McDonald's beef up menu with veggie burger?", by Alex Beam, 10 June 1998, Boston Globe):

Here, Beam expresses disgust, or at least distaste and derision, for the vegetarian burger patty by comparing the act of eating it with the act of having sex with an inflatable doll. Beam does not need to elaborate on his source analog; it and the emotional attitude attached to it are readily available (or imaginable) to the reading audience.

It is not clear why negative emotions should dominate persuasive analogies. One possibility is bias in story selection, i.e., it may be that news reporters are more interested in describing negative emotions. But this possibility does not gibe with the symmetry of positive and negative emotions in cases of reverse empathy. A second and more likely possibility is bias in the source itself, i.e., it may be that the use of analogies for the purpose of persuasion lends itself better to negative emotions than to positive ones. But the reason for this situation is unclear: Do people simply have more negative experiences stored up for others to remind them of? Are negative experiences more readily recalled than positive ones? Are negative experiences more readily attached to new situations than positive ones? Is it more socially acceptable to persuade people that something is blameworthy rather than praiseworthy? The database suggests no clear answer, but the nature of reverse empathy suggests that social factors could play an important role. Of the three kinds of emotional analogies discussed here, persuasion is the one by which a speaker imposes on his or her audience to the greatest degree; that is, the speaker is deliberately trying to change other people's minds. The indirection achieved by evaluating something analogous to the real topic of discourse may serve to mitigate the imposition and make the audience more likely to adopt the speaker's opinion.

Social factors such as this one also tend to obscure the distinction between reverse empathy and persuasive analogies in practice. As Grice (1989) observed, much human communication and interaction depends on people adopting a cooperative stance in dealing with each other. Thus, to return to a classic example, someone might request a salt shaker not by asking for it directly but by stating something like "this lasagna could use some salt". Similarly, someone could employ an indirect method of persuasion by using reverse empathy. Consider the following analogy stated by Microsoft's William Neukom, senior vice president for law and corporate affairs, during a court appeal concerning Microsoft's practice of forcing computer makers to bundle its Internet Explorer with every installation of the Windows 95 operating system (from "Microsoft unveils defiant strategy in appeal", by Kaitlin Quistgaard and Dan Brekke, 16 December 1997, Reuters): "The central point of our position is that when a computer manufacturer licenses Windows, it should install the entire product, just as Ford requires that all its vehicles be sold with Ford engines. This is the only way to guarantee customers a consistent Windows experience." In this case, Neukom is attempting to persuade the court judge that the suit against Microsoft is as unreasonable as a suit against Ford would be. In so doing, Neukom does not directly urge the court to adopt Microsoft's position but rather simply states what their position is. In other words, he just states Microsoft's official attitude for the court in such as way as to invite the judge to adopt it - an indirect use of persuasion. This indirect approach is only one example of how social factors such as the exhibition of cooperative behavior tends to obscure the distinctions between the use of emotional analogies in persuasion and reverse empathy.

We cannot assume that the analogies found in the Internet survey are characteristic of emotional analogies in general. Perhaps a different methodology would find more instances of empathy and positive emotions. But the survey has served to collect many interesting analogies and to illustrate further the emotional nature of much analogical inference.

In this paper, we have provided numerous examples of emotional analogies including: analogies about emotions, analogies that transfer emotions in persuasion, empathy, and reverse empathy; and analogies that generate emotions in humor, irony, discovery, and motivation. In order to understand the cognitive processes involved in emotional analogies, we have proposed an account of analogical inference as defeasible, holistic, multiple, and emotional. The HOTCO model of emotional coherence provides a computational account of the interaction of cognitive and emotional aspects of analogical inference.


Acknowledgments: Thagard's work on this project was supported by a grant from the Natural Sciences and Engineering Council of Canada and by a Canada Council Killam Research Fellowship. Shelley was supported by a doctoral fellowship from the Social Sciences and Humanities Research Council of Canada. We are grateful to Usha Goswami for comments on an earlier draft.



Barnes, A., & Thagard, P. (1997). Empathy and analogy. Dialogue: Canadian Philosophical Review, 36, 705-720.

Blanchette, I., & Dunbar, K. (1997). Constraints underlying analogy use in a real-world context: Politics. In M. G. Shafto & P. Langley (Eds.), Proceedings of the Nineteenth Annual Conference of the Cognitive Science Society (p. 867). Mahwah, NJ: Erlbaum.

Bosanquet, B. (1920). Implication and linear inference. London: Macmillan.

Dennett, D. (1991). Consciousness explained. Boston: Little, Brown.

Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989). The structure-mapping engine: Algorithms and examples. Artificial Intelligence, 41, 1-63.

Fenno, R. F. (1978). Home style: House members in their districts. Boston: Little, Brown.

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155-170.

Grice, H. P. (1989). Studies in the way of words. Cambridge, MA: Harvard University Press.

Harman, G. (1986). Change in view: Principles of reasoning. Cambridge, MA: MIT Press/Bradford Books.

Holyoak, K. J., Novick, L. R., & Melz, E. R. (1994). Component processes in analogical transfer: Mapping, pattern completion, and adaptation. In K. J. Holyoak & J. A. Barnden (Eds.), Advances in connectionist and neural computation theory, Vol. 2: Analogical connections. (pp. 113-180). Norwood, NJ: Ablex.

Holyoak, K. J., & Thagard, P. (1989). Analogical mapping by constraint satisfaction. Cognitive Science, 13, 295-355.

Holyoak, K. J., & Thagard, P. (1995). Mental leaps: Analogy in creative thought. Cambridge, MA: MIT Press/Bradford Books.

Kolodner, J. (1993). Case-based reasoning. San Mateo, CA: Morgan Kaufmann.

Kunda, Z., & Thagard, P. (1996). Forming impressions from stereotypes, traits, and behaviors: A-parallel constraint-satisfaction theory. Psychological Review, 103, 284-308.

Lockwood, P., & Kunda, Z. (1997). Superstars and me: Predicting the impact of role models on the self. Journal of Personality and Social Psychology, 73, 91-103..

Meichenbaum, D. (1994). A clinical handbook/practical therapist manual for assessing and treating adults with post-traumatic stress disorder (PTSD). Waterloo, Ontario: Institute Press.

Mill, J. S. (1970). A system of logic. (8 ed.). London: Longman.

Salmon, W. (1984). Logic. (3rd ed.). Englewood Cliffs, NJ: Prentice-Hall.

Shelley, C. P. (1999). Multiple analogies in evolutionary biology. Studies in History and Philosophy of Science, Part C, Studies in History and Philosophy of Biological and Biomedical Sciences, 30, 143-180.

Shelley, C. P. (forthcoming). A cognitive theory of irony. Unpublished manuscript, University of Michigan, Ann Arbor.

Thagard, P. (1989). Explanatory coherence. Behavioral and Brain Sciences, 12, 435-467.

Thagard, P. (1992). Conceptual revolutions. Princeton: Princeton University Press.

Thagard, P. (forthcoming). Coherence in thought and action. Cambridge, MA: MIT Press.

Thagard, P., Eliasmith, C., Rusnock, P., & Shelley, C. P. (forthcoming). Knowledge and coherence. In R. Elio (Ed.), Common sense, reasoning, and rationality (Vol. 11, ). New York: Oxford University Press.

Thagard, P., Gochfeld, D., & Hardy, S. (1992). Visual analogical mapping, Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society (pp. 522-527). Hillsdale, NJ: Erlbaum.

Thagard, P., & Millgram, E. (1995). Inference to the best plan: A coherence theory of decision. In A. Ram & D. B. Leake (Eds.), Goal-driven learning: (pp. 439-454). Cambridge, MA: MIT Press.

Thagard, P., & Verbeurgt, K. (1998). Coherence as constraint satisfaction. Cognitive Science, 22, 1-24.

To return to the main Recent Articles Page, click here.

To return to Paul Thagard's Biography Page, click here.

Back to Computational Epistemology Laboratory.