Emotions as metarepresentational states of mind11.pdf

Available online at www.sciencedirect.com

Cognitive Systems Research 10 (2009) 6–20

www.elsevier.com/locate/cogsys

Emotions as metarepresentational states of mind: Naturalizing

the belief–desire theory of emotion

Action Editor: Jonathan Gratch

Rainer Reisenzein

Institute of Psychology, University of Greifswald, Franz-Mehring-Straße 47, 17487 Greifswald, Germany

Received 14 September 2007; accepted 31 March 2008

Available online 18 June 2008

Abstract

Describes the outlines of a computational explication of the belief–desire theory of emotion, a variant of cognitive emotion theory.

According to the proposed explication, a core subset of emotions including surprise are nonconceptual products of hardwired mecha-

nisms whose primary function is to subserve the monitoring and updating of the central representational system of humans, the

belief–desire system. The posited emotion-producing mechanisms are analogous to sensory transducers; however, instead of sensing

the world, they sense the state of the belief–desire system and signal important changes in this system, in particular the fulﬁllment

and frustration of desires and the conﬁrmation and disconﬁrmation of beliefs. Because emotions represent this information about the

state of the representational system in a nonconceptual format, emotions are nonconceptual metarepresentations. It is argued that this

theory of emotions provides for a deepened understanding of the role of emotions in cognitive systems and solves several problems of

psychological emotion theory.

Keywords: Emotion; Belief–desire theory; Metacognition; Aﬀective computing; BDI

What are emotions, and what is their function in the

economy of the mind? I propose that at least for a core sub-

set of emotions including surprise, these questions can be

answered as follows: Emotions are nonconceptual outputs

of hardwired mechanisms whose primary function is to

subserve the monitoring and updating of the central repre-

sentational system of humans, the belief–desire system.

This theory of emotions, which closely connects emotions

to the monitoring and updating of representations and

assigns them important epistemic functions—in holding

that they convey to agents important information about

their own representational system—may at ﬁrst appear

unusual. In fact, however, it can be argued that this view

of emotions is already implicit in the currently dominant

theories of emotion, the cognitive emotion theories (for

an overview, see Scherer, Schorr, & Johnstone, 2001 ).

For these theories assume implicitly that emotions are clo-

sely tied to changes in beliefs and desires; and at least some

of them explicitly attribute to emotions an informational

function (e.g., Ortony, Clore, & Collins, 1988 ). In fact,

the computational model of emotions sketched in this arti-

cle is an attempt to ‘‘naturalize” (to integrate into the

scientiﬁc picture; Dretske, 1995 ) a particular version of

cognitive emotion theory, the belief–desire theory of emo-

tion (BDTE).

The motivation for this attempt was the observation

that many important issues are still controversial among

cognitive emotion theorists (see also, Reisenzein, 2001,

2006a ) and the belief that, to resolve these issues, it is

indispensable to consider the cognitive architecture that

underlies emotions. Although cognitive emotion theorists

are agreed that (most) emotions presuppose cognitions,

they diﬀer on exactly which cognitions are necessary for

E-mail address: rainer.reisenzein@uni-greifswald.de

doi:10.1016/j.cogsys.2008.03.001

R. Reisenzein / Cognitive Systems Research 10 (2009) 6–20

emotions (e.g., factual beliefs, evaluative beliefs, or both)

and in which sense they are necessary (e.g., are they

required as causes or components of emotions); whether

and in what sense emotions also presuppose desires; how

the cognitive processes that generate emotions look like

in detail; what the emotion itself is (i.e., how it is to be the-

oretically deﬁned) and which functions emotion have; how

emotions relate to emotional experiences; what accounts

for the distinctive quality and the intensity of emotional

experiences; and how the object-directedness of emotions

can be explained. The theory of emotion proposed in this

article gives or at least sketches answers to all these ques-

tion and thus oﬀers a coherent picture of the emotional

mind. At the same time, the theory seeks to preserve the

central insights of cognitive emotion theories but to avoid

objections that have been raised against them.

The article has three parts. In the ﬁrst part, I give an

overview of BDTE. In the second part, I sketch a compu-

tational model of BDTE. In the third part, I present argu-

ments in favor of the proposed computational explication

of BDTE.

taken to be basic kinds of mental states—modes of relating

to objects—that cannot be reduced to one another. To

mention but one diﬀerence between the two: Beliefs, but

not desires, can be true or false; desires, but not beliefs,

can be satisﬁed or frustrated (e.g., Green, 1992 ). As dis-

cussed in the next section, the distinction between beliefs

and desires is ultimately based on the fundamentally diﬀer-

ent functional roles that these mental states play in the

economy of the mind.

BDTE theorists diﬀer, among other issues, on the ques-

tion of the precise sense in which emotions are ‘‘products

of” beliefs and desires. Whereas some regard beliefs and

desires as components of emotions (see Green, 1992 , for fur-

ther discussion), I follow Meinong (1906) in assuming that

beliefs and desires are the causes of the emotion, which is a

mental state of its own. Presupposing this ‘‘causalist” inter-

pretation of BDTE, its basic assumption can be stated

more precisely as follows: At least a core subset of the men-

tal states presystematically subsumed under the category

‘‘emotion” are reactions to the cognized actual (e.g., happi-

ness, unhappiness) or potential (e.g., hope, fear) fulﬁllment

or frustration of desires plus, in some cases (e.g., Surprise,

disappointment), conﬁrmations or disconﬁrmations of

beliefs.

To illustrate BDTE, consider the case of Mary, who

feels happy that Mr. Schroiber was elected chancellor.

(For the time being, I identify emotions with emotional

experiences. A distinction between the two is drawn later

in the computational explication of the theory.) According

to BDTE, Mary feels happy about Mr. Schroiber’s election

if Mary (a) comes to (ﬁrmly) believe that Schroiber was

elected and (b) desires this state of aﬀairs. In slightly more

detail, the process of emotion generation, here illustrated

for happiness, typically looks as shown in Fig. 1 . First,

the person comes to desire some state of aﬀairs or proposi-

tion p. 1 For example, Mary acquires the desire that Schro-

iber is elected chancellor. Some time later—as a result of

new information acquired through the senses, communica-

tion from others, or inference from existing beliefs—the

person acquires the belief that p obtains. For example,

when watching the news on TV, Mary comes to believe

that Schroiber was, indeed, elected chancellor. Thereupon,

the emotion occurs: Mary now feels happy or pleased that

Schroiber was elected. 2 Note that the belief and desire are

connected to the emotion not only as cause to eﬀect, but

also semantically: The belief that p, the desire for p and

happiness about p concern a common topic, they all refer

1. The belief–desire theory of emotion

The belief–desire theory of emotion belongs to the

broader class of cognitive emotion theories represented,

for example, by the theories of Arnold (1960), Frijda

(1986), Lazarus (1991), Oatley and Johnson-Laird (1987),

Ortony et al. (1988), and Scherer (2001) in psychology;

and those of Kenny (1963), Lyons (1980), Nussbaum

(2001), and Solomon (1976) in philosophy (for reviews,

see Ellsworth & Scherer, 2003; Goldie, 2007 ). Cognitive

emotion theories have come to dominate psychological

and philosophical theorizing on emotions during the past

two decades, and they also form the basis of most existing

computational models of emotion (e.g., Elliott, 1992;

Gratch & Marsella, 2004; Neal Reilly, 1996; Staller & Pet-

ta, 2001 ). As a distinct type of emotion theory within the

cognitive approach to emotions, BDTE has been primarily

promoted by philosophers (e.g., Davis, 1981; Green, 1992;

Marks, 1982; Searle, 1983 ; and for an early version, Mei-

nong, 1894, 1906 [summary in Reisenzein, 2006a ]). Propo-

nents of BDTE in psychology include Roseman (1979) and

Miceli and Castelfranchi (1997, 2007) .

1.1. Basic assumptions of BDTE

The conceptual framework of BDTE is the same as that

of the philosophical belief–desire theory of action (e.g.,

Bratman, 1987; Mele, 1992 ) which inspired the BDI

(belief–desire–intention) approach to artiﬁcial agents

( Bratman, Israel, & Pollack, 1988 ). Analogous to the

belief–desire theory of action, which holds that actions

are the product of cognitive or informational states

(beliefs) and motivational states (desires), BDTE posits

that emotions are the product of cognitions (beliefs) and

motives (desires). In this analysis, beliefs and desires are

1 This is the philosophical usage of ‘‘proposition”.Psychologists

typically use the term to denote a sentence in a language-like mental

representation system that represents states of aﬀairs (e.g., Kintsch, 1988;

Anderson & Lebiere, 1998 ).

2 The converse temporal sequence is possible also. Mary may ﬁrst learn

that Schroiber was elected chancellor, and come to desire that state of

aﬀairs only later, when she reads about Schroiber’s political program. In

this case, the desire is cognized as being fulﬁlled as soon as it is formed,

resulting again in Mary’s happiness about Schroiber’s election.

R. Reisenzein / Cognitive Systems Research 10 (2009) 6–20

Desire(p)

elected (i.e., believes with uncertainty that p obtains);

whereas Mary fears at t that Schroiber is elected if she

diswants him to be elected but is uncertain whether or

not he will be elected. Mary is surprised at t that Schroiber

is elected chancellor if she up to t believed he would not be

elected ( : p) and at t comes to ﬁrmly believe that he was

elected (p). Mary is disappointed at time t that Schroiber

is not elected chancellor if she desires him to be elected

and up to t believed that this was likely or at least possible,

but at t comes to ﬁrmly believe that he was not elected ( : p).

Finally, Mary is relieved that Schroiber is not elected chan-

cellor if she diswants him to be elected and up to t believed

that this was at least possible, but at t comes to ﬁrmly

believe that he was not elected.

Although the emotions listed in Table 1 still comprise

only a small subset of the emotions distinguished in ordin-

ary language, from the perspective of BDTE they are basic

forms. Happiness and unhappiness are the emotional reac-

tions to the cognized actual fulﬁllment versus frustration of

desires; hope and fear to their cognized possible fulﬁllment

versus frustration. Surprise is the general emotional reac-

tion to belief-disconﬁrmation; if it co-occurs with desire–

fulﬁllment, it is pleasant surprise; if it co-occurs with desire

frustration, it is unpleasant surprise. Disappointment and

relief are the emotional reactions to the unexpected

(belief-disconﬁrming) nonoccurrence of desire fulﬁllment

versus frustration, respectively.

The diversity and complexity of humans’ emotional life

gets into the picture primarily, if not exclusively, via the

objects or contents of the beliefs and desires that give rise

to emotions. In other words, the diversity and complexity

of human emotions is due to the fact that humans have

diverse and complex desires and beliefs. In particular, as

a highly social species (e.g., Richerson & Boyd, 1998 ),

humans have desires that concern not only their own wel-

fare, but also the welfare of others, as well as their own

and others’ compliance with social and moral norms.

Accordingly, other-regarding emotions, such as joy for

another, Schadenfreude, pity,andenvy can be analyzed as

forms of happiness or unhappiness about, respectively, a

desired or undesired state of aﬀairs p that concerns the

positive or negative fate of another person (e.g., Meinong,

1894; Ortony et al., 1988 ). The classical ‘‘moral emotions”,

such as guilt and indignation on the negative side, and

pride or moral elevation on the positive side, can be incor-

porated into the belief–desire framework by assuming that

the content of the desire (the desired proposition) in these

cases is the compliance of oneself or another person with

a social or moral norm (e.g., Ortony et al., 1988; Staller

& Petta, 2001 ). In addition, BDTE can be extended to

cover ‘‘fantasy emotions” (i.e., emotional reactions to

anticipated or counterfactually imagined states of aﬀairs),

by positing that these emotions are based on assumptions

rather than beliefs ( Meinong, 1910; Reisenzein, Meyer, &

Sch¨ tzwohl, 2003 ). In the computational model of BDTE

described later, this would mean to add a ‘‘pretense store”

to the cognitive architecture (see Nichols & Stich, 2000 ).

Happiness(p)

Belief(p)

New

information

Fig. 1. Basic belief–desire analysis of emotions.

to p ( Green, 1992 ; I will later argue, however, that the

object-directedness of the emotion is only apparent).

By amending and reﬁning the just-described ‘‘basic for-

mula” of the belief–desire analysis of emotions, it is possi-

ble to specify the cognitive and motivational preconditions

of many emotions distinguished in ordinary language (e.g.,

Davis, 1981; Meinong, 1894, 1906; Miceli & Castelfranchi,

2007; Searle, 1983 ). Indeed, there is reason to believe that

all emotions with propositional objects (all emotions directed

at states of aﬀairs) are amenable to a belief–desire analysis.

These emotions cover the vast majority of the emotions dis-

tinguished in ordinary language (see also, Ortony et al.,

1988 ).

1.2. Qualitative belief–desire analysis of emotions

A qualitative belief–desire analysis of some emotions is

shown in Table 1 . To illustrate Table 1 , Mary feels happy

at time t that Schroiber is elected chancellor (= proposition

p), if Mary desires that Schroiber is elected and comes to

ﬁrmly believe (i.e., is certain) at t that he was elected. Mary

feels unhappy at t that Schroiber is elected chancellor if she

is aversive to p or ‘‘diswants” p to happen (here analyzed

as: Mary desires not-p, that Schroiber is not elected) and

comes to ﬁrmly believe at t that p obtains. Mary hopes at

t that Schroiber is elected chancellor if she wants him to

be elected but is uncertain at t whether or not he will be

Table 1

Belief–desire theory of emotions, qualitative formulation

Emotion if Belief at t Desire at t Belief at t-1

happy (p, t) Certain(p, t) Des(p, t)

unhappy(p, t) Certain(p, t) Des( : p, t)

hopes(p, t) Uncertain(p, t) Des(p, t)

fears(p, t) Uncertain(p, t) Des( : p, t)

surprised(p, t) Certain(p, t) – (irrelevant) Bel( : p, t-1)

disappointed( : p, t) Certain( : p, t) Des(p, t) Bel(p, t-1)

relieved( : p, t) Certain( : p, t) Des( : p, t) Bel(p, t-1)

Notation: Bel(p, t)...believes p at time t; Certain(p, t)...ﬁrmly believes p at t

Uncertain(p, t) ¼ : Bel(p, t)& : Certain(p, t)& : Certain( : p, t). Des(p, t)...

desires p at t; Des( : p, t)...desires not-p at t ( is aversive against p at t).

R. Reisenzein / Cognitive Systems Research 10 (2009) 6–20

1.3. Quantitative formulation of BDTE

ments that have been adduced for cognitive emotion theo-

ries generally (e.g., Ellsworth & Scherer, 2003; Goldie,

2007; Reisenzein et al., 2003 ). Arguments and evidence

more speciﬁcally supportive of BDTE stem from two

sources.

First, philosophical proponents of BDTE have argued

that this theory avoids a number of objections raised

against other versions of cognitive emotion theory, espe-

cially cognitive–evaluative or appraisal theories of emotion

(e.g., Green, 1992; Marks, 1982 ; see also, Reisenzein,

2006a ). A central diﬀerence between BDTE and the cogni-

tive–evaluative theory of emotion is that the latter regards

emotions as products of factual and evaluative beliefs

rather than of beliefs and desires. To illustrate, whereas

BDTE posits that Mary is happy about p if she believes p

and desires p, the evaluative theory of emotion assumes

that Mary feels happy about p if she believes p and evalu-

ates p as good for herself (i. e., believes that p is good for

her; e.g., Arnold, 1960; Lazarus, 1991 ). Against this

assumption, it has been objected that—even in the presence

of the necessary factual beliefs—evaluative beliefs are not

sucient for emotions. For example, one usually values

being healthy highly; nonetheless, the thought that one is

healthy (factual belief), and that this is good for oneself

(evaluative belief), normally does not cause noticeable

joy. BDTE can easily explain this: At the time of the eval-

uative judgment, the desire to be healthy has long been ful-

ﬁlled. It has also been argued that evaluative beliefs are not

generally necessary for emotions. BDTE can handle this

objection, too: To be happy about p, one must desire p

and come to believe that p obtains; it is not necessary that

one in addition believes that p is good for oneself ( Arnold,

1960 ), congruent with one’s motives ( Lazarus, 1991 ), or

goal-conducive ( Scherer, 2001 ).

Second, BDTE is supported by research on the quan-

titative relations between beliefs, desires, and emotions

( Table 2 ). Although studies of this kind are still few in

number and exist only for a few emotions, their results

are generally in line with predictions. For example, Rei-

senzein and Junge (2006) tested several of the ‘‘emotion

laws” shown in Table 2 in a monetary lottery paradigm

(Mellers et al., 1997) . Participants experienced occur-

rences and nonoccurrences of monetary wins and losses

that diﬀered in magnitude and probability and rated

the intensity of experienced surprise, disappointment

and relief caused by each outcome. Support for the pre-

dictions of BDTE was obtained even on the level of indi-

vidual participants. Further supporting the self-reports of

emotions collected by Reisenzein and Junge (2006) ,a

recent fMRI study ( Yacubian et al., 2006 ) found that

unexpected outcomes of a monetary lottery caused

changes in activity proportional to the ‘‘prediction error”

(the diﬀerence between the expected and actual outcome

value) in brain regions commonly associated with posi-

tive and negative emotions (the ventral striatum and

the amygdala, respectively) (see also Abler, Walter, Erk,

Kammerer, & Spitzer, 2006 ).

Apart from being able to explain the qualitative (i.e., type)

diﬀerentiation of emotions in a parsimonious way, BDTE

also allows a straightforward and parsimonious explanation

of the intensity aspect of emotion (the fact that emotions vary

in intensity). To explain emotional intensity, the qualitative

version of BDTE is reﬁned to a quantitative theory. In fact,

already the ﬁrst modern formulations of BDTE contained

proposals for a quantiﬁcation (e.g., Davis, 1981; Day,

1970 ). Quantiﬁcation is achieved by (a) introducing quantita-

tive concepts of belief and desire and (b) proposing quantita-

tive laws, expressed by numerical functions, that connect

degrees of belief and desire with the intensity of various emo-

tions. Table 2 shows a quantitative belief–desire analysis of

the emotions listed in Table 1 (based mostly on previous pro-

posals; e.g., Davis, 1981; Gratch & Marsella, 2004; Macedo,

Reisenzein, &Cardoso, 2004 ). For example, happiness about

p is experienced when b(p) = 1 and d(p) > 0, and the intensity

of happiness about p is a monotonically increasing function

of d(p). Fear is experienced whenever 0 < b(p)<1 and

d(p) < 0, and the intensity of fear is a monotonically increas-

ing function of j d(p) b(p) j .

1.4. BDTE as a psychological background theory for

computational models of emotion

As a member of the class of cognitive emotion theories,

BDTE is supported by the theoretical and empirical argu-

for domain subset (else emotion

intensity = 0)

Happiness(p, t)=U ha [d(p, t)] b(p, t)=1&d(p, t)>0

Unhappiness(p, t)=U uh [d(p, t)] b(p, t)=1&d(p, t)<0

Hope(p, t)=U ho [b(p, t) d(p, t)] 0 < b(p, t)<1&d(p, t)>0

Fear(p, t)=U fe [b(p, t) d(p, t)]

0 < b(p, t)<1&d(p, t)<0

Surprise(p, t)=U su [b(p, t-1)]

b(p, t)=0&b(p, t-1) > 0

Disappointment( : p, t)=U di [b(p, t-1)

d(p, t-1)]

b(p, t)=0&b(p, t-1) > 0 &

d(p, t-1) > 0

b(p, t)=0&b(p, t-1) > 0 &

d(p, t-1) < 0

Notation: b: p 1 , p 2 , ...} {t 1 , t 2 , ...} ) [0, 1] and

d : f p 1 ; p 2 ; ... gf t 1 ; t 2 ; ... g) R are the belief and desire functions. b(p, t)

represents the strength of belief in p at time t, with 1 denoting certainty that

p, 0.5 maximal uncertainty, and 0 certainty that not-p. d(p, t) represents the

direction and strength of the desire for p at time t, with values > 0 denoting

positive desire, 0 indiﬀerence, and values < 0 aversion against p. Happi-

ness(p, t), Unhappiness(p, t) etc. are the emotion intensities, ranging from 0

(absence of the emotion) to some maximumnumber. For arguments outside

the domain subsets indicated in column 2, the value of the emotion func-

tions is 0. Note that the emotion intensity functions (Happiness, Unhappi-

ness etc.) are only partly speciﬁed in the table, with U ha , U uh etc. representing

their unspeciﬁed components. For example, Gratch andMarsella (2004) use

U = jj (the absolute value function) to be able to represent all emotion

intensities, including those of negative emotions, by positive numbers. If no

further speciﬁcations are made, U is the identity function. The functions for

disappointment and relief only capture the case of the simple nonoccurrence

of a desired and undesired event, respectively. For a possible treatment of

the general case see Mellers, Schwartz, Ho, and Ritov (1997) .

Table 2

Belief-desire theory of emotions, quantitative formulation

Emotion Intensity = function

of d and b

Relief( : p, t)=U re [b(p, t-1)

d(p, t-1)]

R. Reisenzein / Cognitive Systems Research 10 (2009) 6–20

The theoretical and empirical arguments for BDTE sug-

gest that BDTE may be a good choice as a psychological

background theory for computational models of emotion.

Other, more pragmatic considerations support this conclu-

sion. First, BDTE may be more tractable computationally

than are appraisal theories of emotion, as it attributes most

of the multidimensionality and complexity of appraisals to

the objects of the appraisals. As a consequence, much of

the computational labor that appraisal theories of emotion

impose on emotion mechanisms can be relegated to ordin-

ary propositional inference processes ( Reisenzein, 2001 ),

whereas the emotion-generating mechanism proper can

be kept simple. For in contrast to the impression conveyed

by appraisal theories, BDTE suggests that, behind the

apparent complexity of emotions lies a comparatively sim-

ple generating mechanism. This suggestion is borne out by

the computational model of BDTE proposed in the follow-

ing section. Second, the conceptual closeness of BDTE to

the BDI framework should facilitate the integration of

emotions into the BDI architecture (e.g., Becker &

Wachsmuth, 2006; Rank & Petta, 2005; Staller & Petta,

2001 ) and into other computational architectures that have

anities to the BDI approach (e.g., Gratch & Marsella,

2004; Marinier & Laird, 2006 ). Finally, BDTE connects

up naturally with attempts to model emotions in decision

theory (e.g., Mellers, 2000; Zeelenberg, van Dijk,

Manstead, & van der Pligt, 2000 ). This suggests straight-

forward ways of incorporating some eﬀects of emotions

on action into computational models of emotion (see also

Reisenzein, 1996 ; for further discussion).

are controlled by a belief–desire representation system?

That is, in what ways do these agents beneﬁt from having

emotions in addition to beliefs and desires?

The naturalization of BDTE proceeds in three steps. The

ﬁrst step consists in the modelling of the belief–desire sys-

tem, for beliefs and desires are the causes of emotions in

BDTE. Taking recourse to a strategy proposed by Fodor

(1987) , the belief–desire system is modeled as a proposi-

tional representation system. In the second step, this repre-

sentation system is amended by proposing: (a) The belief–

desire system comes equipped with hardwired mechanisms

that monitor and, if necessary, update the system in

response to newly acquired information. (b) The monitor-

ing mechanisms comprise two submechanisms: One that

compares newly acquired beliefs to existing beliefs and

another that compares newly acquired beliefs to existing

desires. (c) The outputs of these information-processing

mechanisms are ‘‘nonpropositional signals” (in the sense

of Oatley & Johnson- Laird, 1987 ) that carry information

about the degree of match or mismatch of newly acquired

beliefs with existing beliefs and desires. In the third step

(whose description is interwoven with that of the second)

the comparator mechanisms are linked to emotions.

2.1. The representational system

To develop a computational model of BDTE requires to

embrace the assumption that mental and especially cogni-

tive processes are computations in an internal representa-

tion system. I adopt the by now classical representational

assumption of cognitive psychology, that the central repre-

sentation system of humans is symbolic and more precisely,

that it is a language-like, or propositional system of repre-

sentations, a language of thought (e.g., Anderson & Lebi-

ere, 1998; Fodor, 1987; Kintsch, 1988 ). A main reason

for this assumption is that, in contrast to other proposed

representation systems, a propositional system lends itself

naturally to a transparent and plausible computational

analysis of the intentional mental states posited in com-

mon-sense psychology, such as beliefs and desires; and

thereby, to the naturalization of cognitive emotion theories

such as BDTE ( Aydede, 2004; Fodor, 1987; Sterelny, 1991 ;

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: