PART
I
PERSON PERCEPTION AND IMPRESSION FORMATION
CHAPTER
1
MAKING SENSE OF PEOPLE: COHERENCE MECHANISMS1
Paul Thagard
Ziva Kunda
University of Waterloo
THREE WAYS OF MAKING SENSE1
When trying to make sense of other people and ourselves, we may rely on several different kinds of cognitive processes. First, we form impressions of other people by integrating information contained in concepts that represent their traits, their behaviors, our stereotypes of the social groups they belong to, and any other information about them that seems relevant. For example, your impression of an acquaintance may be a composite of personality traits (e.g., friendly, independent), behaviors (e.g., told a joke, donated money to the food bank), and social stereotypes (e.g., woman, doctor, Chinese). Second, we understand other people by means of causal attributions in which we form and evaluate hypotheses that explain their behavior. To explain why someone is abrupt on one occasion, you may hypothesize that this person is impatient or that he or she is under pressure from a work deadline. You believe the hypothesis that provides the best available explanation of the personās behavior. A third means of making sense of people is analogy: You can understand people through their similarity to other people or to yourself. For example, you may understand the stresses that your friend is experiencing by remembering an occasion when you yourself experienced similar stresses. This will allow you to predict your friendās likely feelings and behavior.
All three of these ways of understanding people can be applied to oneself as well as to others. I may gain insight into myself by applying new concepts to myself (e.g., realizing I am impatient), forming new hypotheses about myself (e.g., conjecturing that I may be more upset by a setback than I realized), and by seeing myself as similar to others (e.g., noticing that I am acting just like my father did).
We propose that making sense of people through conceptual integration, explanation, and analogy can all be understood in terms of cognitive mechanisms for maximizing coherence. When integrating information about a person, we attempt to achieve coherence among concepts by reconciling conflicts among the different pieces of information that we have about an individual (Kunda & Thagard, 1996). Knowing that someone who is a lawyer responded meekly to an insult requires us to balance conflicting expectations generated by the stereotype that lawyers are aggressive and the unaggressive behavior (Kunda, Sinclair, & Griffin, 1997). Similarly, in causal attribution, we need to reconcile different explanations for an individualās behavior, choosing, for example, between explanations in terms of personality traits and explanations in terms of situational factors. Such choices require us to maximize explanatory coherence, accepting those explanations that fit best with the rest of our beliefs (Read & Miller, 1993a, 1993b; Thagard, 1989). Finally, making sense of people in terms of other similar people requires us to assess analogical coherence, finding a good fit between the complex of attributes of one person and the complex of attributes of another (Holyoak & Thagard, 1995).
In the next section, we first outline a general characterization of coherence that provides a uniform vocabulary for understanding a wide variety of cognitive processes in terms of parallel constraint satisfaction. We then show in more detail how conceptual integration, causal attribution, and analogical understanding (including empathy) can be understood as different kinds of coherence. We argue that connectionist models of conceptual, explanatory, and analogical coherence provide computationally powerful and psychologically plausible explanations of diverse ways in which people think about other people and themselves.
Coherence as Constraint Satisfaction
To make sense of people, we need to represent different kinds of information about them. These include concepts such as traits and stereotypes that apply to individuals as well as propositions such as Mary loves John that describe relations between people. Some representations fit together, but others conflict. For example, describing someone as loving fits with describing that person as kind, but conflicts with describing that person as hateful. The proposition that Mary loves John fits with the proposition that Mary is nice to John, but conflicts with the proposition that Mary hates John. When two representations fit together, there is a positive constraint between them: If you apply one of the representations to someone, then you will tend to apply the other representation as well. If two representations conflict, there is a negative constraint between them: If you apply one of the representations to someone, then you will tend not to apply the other representation. Coming up with a coherent interpretation of people is a matter of applying some representations to them and not applying others. Generally, coherence is a matter of accepting some representations and rejecting others in a way that maximizes compliance with positive and negative constraints.
Thagard and Verbeurgt (1998) provide a general definition of coherence problems. A coherence problem arises when one encounters a set of elements that mutually constrain each other, and wishes to accept some of these elements and reject the remaining ones. For example, one needs to decide which of a set of interrelated traits are characteristic of John (the accepted elements) and which are not (the rejected elements). The constraints among the elements may be positive or negative. A positive constraint among two elements means that the two should go togetherāthey should both be accepted or both be rejected. For example, if John is loving, he should be kind as well, and if he is not loving, he should not be kind either. A negative constraint means that the two elements should not go togetherāif one is accepted, the other should be rejected. For example, if John is loving, he should not be hateful; if he is hateful, he should not be loving. Each of the constraints carries a weight that reflects its importance.
When partitioning the elements into the accepted set and the rejected set, it is often not possible to satisfy all of the constraints because they may conflict with each other. For example, a person who is manipulative should also be interpersonally skilled. But a person who is interpersonally skilled should also be loving, whereas a person who is manipulative should not. It will be impossible to satisfy all of these constraints simultaneously. The coherence problem is to satisfy as many of the constraints as possible, while giving preference to the more important ones. More technically, the aim is to partition the elements into an accepted and rejected set in a way that maximizes the weight of the satisfied constraints. For a more precise definition, see the appendix.
In later sections, we show how conceptual, explanatory, and analogical coherence can all be understood as special cases of this general characterization of coherence. Each kind of coherence involves different sorts of elements and constraints.
Maximizing coherence is a difficult computational problem: Thagard and Verbeurgt (1998) prove that it belongs to a class of problems generally considered to be computationally intractable, so that no algorithms are available that are both efficient and guaranteed to be correct. Nevertheless, good approximation algorithms are available, in particular connectionist algorithms from which the above characterization of coherence was originally abstracted.
Here is how to translate a coherence problem into a problem that can be solved in a connectionist network:
1. Each element is represented as a unit (node) in a network of units. These units are very roughly analogous to neurons or groups of neurons in the brain.
2. A positive constraint between two elements is represented as an excitatory link between the corresponding units. Each link has a weight representing the strength of the constraint, as determined, for exam ple, by the strength of association between two concepts.
3. A negative constraint between two elements is represented as an inhibitory link between the corresponding units.
4. Each unit is assigned an equal initial activation, say .01. The activation of all the units is then updated in parallel. The updated activation of a unit is calculated on the basis of its current activation, the activation of the units to which it is linked, and the weights of these links. The activation of a given unit is increased with the activation of units to which it has excitatory links, and decreased with the activation of units to which it has inhibitory links. A number of equations are available for specifying how this updating is done (McClelland & Rumelhart, 1989). Typically, activation is constrained to remain between a minimum (e.g., ā1) and a maximum (e.g., +1).
5. The network goes through many cycles in which the activation of all units is updated. Updating is repeated until all units have settled, that is, achieved stable activation values that change only minimally from one cycle to another.
6. If a unitās final activation exceeds a specified threshold (e.g., 0), then the element represented by that unit is deemed to be accepted. Otherwise, that element is rejected.
This process results in a partition of elements into accepted and rejected sets by virtue of the network settling in such a way that some units end up with activation levels that are above the critical threshold for acceptance, and others do not. The final levels of activation can also be taken to represent degrees of acceptance and rejection.
Intuitively, this solution is a natural one for coherence problems. Just as we want two coherent elements to be accepted or rejected together, so two units connected by an excitatory link will be activated or deactivated together. Just as we want the outcome for two incoherent elements to be such that one is accepted and the other is rejected, so two units connected by an inhibitory link will tend to suppress each otherās activation, with one activated and the other deactivated. A solution that enforces positive and negative constraints on maximizing coherence is provided by the parallel update algorithm that adjusts the activation of all units at once based on their links and previous activation values. Table 1.1 summarizes the correspondences between coherence problems and connectionist networks.
TABLE 1.1
Comparison of Coherence Problems and Connectionist Networks
Coherence | Connectionist Network |
Element | Unit |
Positive Constraint | Excitatory Link |
Negative Constraint | Inhibitory Link |
Constraint Satisfaction | Parallel Updating of Activation |
Element Accepted | Unit Activated Above Threshold |
Element Rejected | Unit Activated Below Threshold |
Impression Formation as Coherence Among Concepts
Sometimes we form impressions of others by integrating their diverse characteristicsātheir behavior, their traits, the stereotypes of the groups they belong to, and any other kind of information deemed relevant. This process may be viewed as a coherence problem in which the elements are concepts representing the personās characteristics, and the positive and negative constraints are imposed by the positive and negative associations among these concepts and their associates.
Kunda and Thagard (1996) developed a parallel constraint-satisfaction theory of impression formation. This theory assumes that stereotypes, traits, and behaviors can be represented as interconnected nodes in a spreading activation network. The nodes can have positive, excitatory associations, or negative, inhibitory ones. To illustrate the model, consider the well-documented finding that stereotypes can affect the meaning of behavior. For example, when a Black person pushes someone, this is interpreted as violent push. But when a White person performs the identical behavior, this is interpreted as a jovial shove (Sagar & Schofield, 1980).
Figure 1.1 shows part of the network of concepts that would be used to make sense of the observation that a Black person or a White person pushed someone. The boxes depict the nodes representing the concepts. The lines connecting these nodes depict the associations among them. Bold lines indicate excitatory associations, and thin lines indicate inhibitory ones. Each of the concepts depicted also has many additional associates that are not portrayed in the figure. The observed information, in this case the behavior (pushed someone) and the stereotyped category (Black or White), is connected to a node termed observed to indicate its special status, and to distinguish it from inferred knowledgeāin this case, the traits associated with the stereotype and the possible interpretations of the behavior. Observed concepts receive strong activation from the observed node. Inferred knowledge becomes activated or deactivated through its positive or negative associations with the observed information. These associations are based on perceiversā prior beliefs about the interrelationships among characteristics.
FIG. 1.1. Stereotypes affect the meaning of behavior. The network on the left activates āviolent pushā and deactivates ājovial shove.ā The network on the right does the opposite. Reprinted from Kunda and Thagard (1996, p. 286).
When one observes that a person pushed someone, pushed someone activates both violent push and jovial shove. If one also observes that the pusher is Black then, at the same time, Black activates aggressive, which further activates violent push while deactivating jovial shove. If, on the other hand, one observes that the pusher is White, White does not activate aggressive. Therefore, both aggressive and violent push end up with less activation when the pusher is White than when the pusher is Black. In this manner, stereotypes color oneās understanding of a personās behavior and oneās impression of that person.
As this example illustrates, the identical behavior may be interpreted differently in different contexts (Kunda & Sherman-Williams, 1993; Sagar & Schofield, 1980; Wojciszke, 1994). Similar shifts in meaning from one context to another have been demonstrated for all the major ingredients of impression formation. These include traits (e.g., Asch, 1946; Asch & Zukier, 1984; Hamilton & Zanna, 1974; Kunda et al., 1997; Zanna & Hamilton, 1977), stereotypes (e.g., Deaux & Lewis, 1984; Kunda, Miller, & Claire, 1990), facial expressions (Trope, 1986), and self-conceptions (e.g., Sanitioso, Kunda, & Fong, 1990). Thus, there is broad support for the notion that the meaning of social constructs varies from one occasion to another.
Traditional models of representation cannot readily account for such shifts in meaning. Leading models of social cognition conceptualize representations as schemas, which are typically understood in terms of a filing cabinet or a storage bin metaphor (e.g., Wyer & Srull, 1986). In such models, each social construct has a fixed and discrete meaning that may be accessed independently, much like one might pull out a single file f...