In this article we discuss an approach to reasoning about emotion intensity within the context of the Affective Reasoner [Elliott1992] (hereafter, AR) a computer simulation that reasons about emotions in a multi-agent system. The AR is designed around the constraining hypothesis that there are twenty-four distinct categories of emotion, each based on a different set of eliciting conditions (c.f., [Ortony et al. 1988]). In this context we have postulated the existence of functions that map interpretations of simulated situations into scalar intensity ratings for emotion instances within those categories. Working abductively from descriptions of emotion-generating situations, we used emotion intensity variables to explain the ways in which changes in various aspects of the simulation might cause changes in emotion intensity. These variables, which represent both situations external to the agent, and dispositions and moods internal to the agent, together with some preliminary functions that illustrate their use in computing emotion intensities, are described in this paper.
Ortony, et al. [Ortony et al. 1988] discuss twenty-two emotion types based on valenced reactions to situations being construed as goal-relevant events, acts of accountable agents, or attractive or unattractive objects (including agents interpreted as objects). This theory has been extended to include the two additional emotion types of love and hate [Elliott1992]. A summary of these emotion types using groupings based on the associated eliciting conditions appears in figure 1.
Previous implementations of the AR allowed the mapping of situations into the twenty-four emotion types by reasoning about simple eliciting conditions, but they did not provide for the determination of emotion intensity. Using the work of Ortony, et al. [Ortony et al. 1988] as a guide, we analyzed a set of descriptions of such situations and created a set of emotion intensity variables to explain the causes of varying emotion intensity, within a coarse-grained simulation paradigm. We reduced the resulting set of variables to a computable formalism, and represented sample situations in the AR. We then isolated three areas of the simulation where variables in either the short-term state of an agent, the long-term disposition of an agent, or the emotion-eliciting situation itself helped to determine the intensity of the agent's subsequent affective state. These three areas, and the intensity-relevant variables embodied in them, are discussed in the next three sections.
We have segregated variables thought to affect the intensity of emotion into three groups. The first group, simulation-event variables, comprises variables whose values change independently of situation interpretation mechanisms. The second group, stable disposition variables, consists of variables that are involved in an agent's interpretation of situations, tend to be constant, and help to determine an agent's personality and role in the simulation. The last group, mood-relevant variables, contains those variables that contribute to an agent's mood state.
The values in the simulation-event variables group change independently of an agent's interpretation of them. In other words, although these variables are still subject to appraisal by the agent, the value changes themselves are external to the appraising agents. Another way to consider these variables is that a single change in the value of one of them may simultaneously, and differentially, affect several appraising agents. Such variables capture ``objective'' features of the world such as the loudness of a noise or the brightness of a light. Thus, for example, the objective degree of loudness of a piece of music might relate to its appealingness for one agent, but to its repulsiveness for another.
Nonetheless, the volume of the music, a variable within the emotion-eliciting situation, in both cases contributes to the calculation of the intensity for the resulting emotions.
By contrast, the stable disposition variables group contains variables that help to determine an agent's bias toward interpreting an emotion-eliciting situation one way or another. For example, it would be these which would allow us to specify that, for a given volume level, one agent experiences intense dislike for the music, whereas another experiences only mild dislike. That these variables are considered stable does not mean that they can never change but rather that such changes are rather slow, and tend to be unidirectional. For example, someone who finds rap music repulsive, may, over time, come to dislike it less, or even, ultimately, to find it appealing. This group also includes some relationship variables that help to define the strength of friendship or animosity between agents, and their emotional distance.
The last group, the mood-relevant variables, contains variables which (1) alter an agent's interpretation of situations, (2) are volatile, (3) are bipolar in nature, (4) are not dispositional, so that they naturally return to (agent-specific) default values over time, and (5) may be the result of prior affective experience. For example, if an agent is represented as ``feeling unwell or depressed,'' then hearing rap music might be simulated as being particularly uplifting, ``taking the agent's mind off her pain,'' or as particularly intolerable, ``driving her to distraction.''
EMOTION CATEGORIES
|
GROUP |
SPECIFICATION |
CATEGORY LABEL AND EMOTION TYPE |
| Well-Being | appraisal of a situation as an event | joy: pleased about an event distress: displeased about an event |
| Fortunes-of-Others | presumed value of a situation as an event affecting another |
happy-for: pleased about an event desirable for another
gloating: pleased about an event undesirable for another resentment: displeased about an event desirable for another sorry-for: displeased about an event undesirable for another |
| Prospect-based | appraisal of a situation as a prospective event |
hope: pleased about a prospective desirable event fear: displeased about a prospective undesirable event |
| Confirmation | appraisal of a situation as confirming or disconfirming an expectation |
satisfaction: pleased about a confirmed desirable event relief: pleased about a disconfirmed undesirable event fears-confirmed: displeased about a confirmed undesirable event disappointment: displeased about a disconfirmed desirable event |
| Attribution | appraisal of a situation as an accountable act of some agent |
pride: approving of one's own act admiration: approving of another's act shame: disapproving of one's own act reproach: disapproving of another's act |
| Attraction | appraisal of a situation as containing an attractive or unattractive object |
liking: finding an object appealing disliking: finding an object unappealing |
| Well-being/ Attribution |
compound emotions | gratitude: admiration+joy anger: reproach+distress gratification: pride+joy remorse: shame+distress |
| Attraction/ Attribution |
compound emotion extensions | love:admiration+liking hate:reproach+disliking |
Our current simulation work is based on the proposals of Ortony, et al. [Ortony et al. 1988], and Frijda et al. [Frijda et al. 1992], and most of our intensity variables derive in one form or anther from their work. The main question that arises in trying to implement these and other variables in a computer program that generates simulated emotions is, what features are present (or might be present) in the simulation events, and how can they be represented in ways that lead to human-like emotions on the parts of our automated agents?
The three central variables described by Ortony, et al. [Ortony et al. 1988] (i.e., the desirability of an event, the blameworthiness or praiseworthiness of an act, or the attractiveness of an object, or agent construed as an object) determine the valence of the resulting emotion. In addition, these variables, each given here as a simulation event variable / stable disposition variable pair representing the actual feature values within the simulation, and their importance to the agent, are also the primary determinants of emotional intensity. Secondary to these are variables, such as that representing physiological arousal, which modify the base intensity levels generated by the central variables. Some variables only admit values that can reduce or maintain the intensity of an emotion. For example, events are ordinarily assumed to be perceived as real. However, if explicitly represented otherwise, the intensity of an emotion resulting from the (at least partially perceived as unreal) situation will be lessened. Lastly, some variables are represented as pairs containing a bias factor and a strength, and will act differently on the intensities of differently valenced emotions.
Moreover, the ranges we have specified for variables are not important to the design of our intensity architecture. Instead we have merely provided arbitrary first designations for these ranges, based loosely on introspective analysis. These ranges, default values for variables, and the intensity functions that use them, have been isolated in the implementation, and are treated essentially as data, so that they may be easily altered to reflect the current cognitive theory.
In the basic intensity functions given at the end of this article we use values for the intensity variables as factors: those values below 1 reduce the strength of the emotion, those above 1 increase it. In the current model, primary determinants of emotion intensity range from zero to 10, with a default of 3. Factors which are treated as weaker modifiers of intensity range from (approximately) zero to 3. Modifiers which, within our model, can only reduce intensity, range from zero to 1. Variables whose effects on intensity calculations are determined by the valence of an emotion (such as a variable which heightens the intensity of negatively valenced emotions but lessens the intensity of positively valenced emotions) are given both a bias value, and a strength ranging from 1 to 3.
It is not desirable to have to specify a value for each of the
intensity variables in each emotion eliciting situation that arises in the
simulation. In addition, some variables (such as that representing the
concept of surprise), which we might desire to represent in a particular
situation, are sometimes difficult, if not impossible, to
calculate.
Intensity functions that are designed to use the missing values, must
therefore either use an alternate calculation method, or must use defaults,
as is done in the current implementation. Because in this implementation
intensities are calculated by multiplying the intensities of component
factors together, we use a default value of 1 for most variables since this
then gives them the convenient property of affecting the intensity
calculation only when otherwise specified.
Example: ``How big of a tip the waiter got for his services.'' Range: -10 to
10, where -10 represents complete blockage of a goal, and 10 complete
achievement. No default, this must be specified by calculation from some
feature value(s).
Range: zero to 10. No default.
The specification stable refers to the position these variables have in determining the relatively stable personalities and roles of the automated agents in the simulation. This is not to say that these values cannot change over time, but rather that such changes will be considered moderately permanent, with no tendency to return to the original state. Although we include the friendship and animosity variables in this category, this is an arbitrary decision since the relationships they represent, as we define them, can be affected by mood.
These variables, based on the central intensity variables of Ortony, et al. [Ortony et al. 1988], appear as part of the agents' appraisal mechanisms and represent the degrees to which various events, actions of agents, and objects are important to them. For example, two agents may like music, but for one it is very important and has the potential for generating ecstasy, whereas for the other it may lead, at most, to a moment of mild pleasure.
The importance of situations for which goal-based and standard-based appraisals are made may be assessed differently depending on whether a goal is achieved (standard upheld) or blocked (standard violated). This is because some concerns can only lead to negative or neutral outcomes, whereas others can only lead to neutral or positive outcomes. For example, one is not ordinarily happy about ``not getting cancer,'' but might be very distressed if such a preservation health goal were blocked. Similarly, serendipitously winning in the lottery is an important event, whereas losing is not. The same is true of standards, where, for example, it is not normally considered praiseworthy to refrain from committing crimes, but it is considered blameworthy to commit them. Preferences and non-preferences are represented separately, and so in each instance a single variable suffices. For discussion see also [Schank and Abelson1977], and [Elliott1992]).
The default values for the variables in this section are meaningful only in that they highlight our assumption that the degree to which a goal, standard, or preference is important to an agent makes a relatively significant contribution to calculations of emotion intensity. However, to avoid such absurdities as having the breaking of a shoelace and the theft of one's car lead to similarly intense emotions, these values must be specified.
Mood representation in the AR is in the early stages of development. The variables given here represent a first attempt at accounting for some of the most salient cases in which prior mood contributes to emotion intensity. Discussion follows at the end of this section.
There are a number of psychological phenomena which may be
thought of as contributing to an agent's valence bias, notably
including mood disorders. Though such phenomena are not well
understood and rigorous models of their determination are scarce
[MacLeod and Mathews1989], there is a body of literature
pertaining to mood disorders and more common, less serious, analogs
of these disorders, experienced by non-disordered people in everyday
situations (e.g., mild, possibly transient depression).
We have therefore included variables to represent mild degrees
of two acknowledged mood-biasing phenomena, namely depression and
anxiety.
Rudimentary moods for agents in the AR can result from an emotion, or series of emotions. To make use of this mechanism, it has been necessary to specify several granularities of mood. General mood biasing factors (such as valence-bias) tend to be automatically generated by the simulation as a result of emotions, whereas more specific mood factors (such as depression-anxiety) tend to be hand-coded as a result of specific situations. When values for more specific mood-biasing variables are present, they are preferred over those for more general mood-biasing variables.
In our model, anxiety and depression interact to form distinct valence biases. For example a non-depressed, non-anxious, agent will react in a way which is biased towards positive anxiety-relevant traits (e.g., ``assurance'') while an anxious agent will have a bias towards negative anxiety-relevant traits (e.g., ``tension''), and a depressed agent will have the bias removed entirely. As such, depression may have the effect, in the model, of dulling an agent's usual anxious nervousness (e.g., a depression-induced representation of ``it just doesn't matter'').
To illustrate, let us suppose that in a simulation an agent is ``fired from her job.'' This agent's concerns lead to both sadness over the loss of comfortable circumstances and companions (stemming from the blockage of some career and social goals), and fear over the prospects of future financial hardship (stemming from the prospect of a financial goal being blocked). If we raise the prior value of her ``anxiety'' (by giving anxiety-invincibility a negative bias and increasing its value), the predominant emotion becomes fear, (``How am I going to pay the rent?''). Raising her prior level of ``depression'' (by giving depression-ecstasy a negative bias and increasing its value) mediates her fear but increases her sadness, (``Everything just seems to be going wrong!'').
Similarly, if a non-depressed, non-anxious, simulated agent is ``involved in a successful business deal,'' the agent might ordinarily be predisposed to attain an ``undue'' intensity of pride from the event (as non-depressives often have an inflated sense of self-involvement in positive causal interactions [Vasquez1987]), while depression would temper this effect, and anxiety would provoke the agent to fear the deal's consequences.
The intensity calculation functions that make use of these variables
within the simulation are still under consideration. Our initial goal is
simply to establish some reasonably monotonic relationship between the
feature values which we are able to represent within the simulator and the
intensity of the simulated emotions resulting from the specified eliciting
conditions. There are many such simple relationships that can be captured.
However, to model the subtle effects of some of the variables, including
their inter-dependence with one another, and their non-linear effects, more
complex functions are needed.
As with the determination of ranges and defaults, the particular intensity-calculation functions used are not important to the reasoning architecture. These can be changed to accommodate different kinds of psychologically motivated experiments. The only fixed link from the intensity functions to the rest of the design is that they must return the scalar values which are used to set the various thresholds for different intensities in the generation of emotion instances, and that they return values for the different categories of emotion that are consistent with one another. Also important to note is that values for variables must be used consistently in both the elicitation of emotions, and their manifestation, since the appraisal mechanism passes variable bindings to the emotion manifestation mechanism.
Lastly, one final implication of our approach is that an agent might appraise a situation many different ways such that all the appraisals lead to the same intensity level in the resulting emotion, even with respect to the same principles and goals. An agent might, for example, be particularly angry because she is already not feeling well, or alternatively because she is already annoyed with the one responsible for her anger. In our architecture, such differences are important because, as just noted, the bindings for the values of the different emotion-intensity variables are passed to the action generation component of the AR, and might influence the reasoner's choice of actions for the experiencing agent.
Normalization variables are included in each function so that the relative strengths of emotions will be correct. Normalization is important for two reasons. First, different emotion types have different numbers of variables, which would affect their overall intensity range. Second, the intensities for compound emotions (e.g., anger) are based on the intensities bindings of two construals rather than one, and this can only be done in a consistent manner if the constituent construals are first normalized.
Figure 2 shows the basic outline of a function that calculates an emotion intensity for simple goal-relevant appraisals. Note that the defaults for mood-variables (e.g., depression-ecstasy) are those for the function only, and may be superseded by a different default for a particular agent.
Figure 3 shows a portion of a similar function used to calculate the intensities of fortunes-of-others-emotions. In this case the variables represent the supposed values for the other agent. We add variables for friendship-animosity and emotional-interrelatedness. As these values increase so do the strength of the resulting fortunes-of-other emotions. In addition we add a calculation for the effect of perceived deservingness or undeservingness. For example, if one is resentful over the good fortune of an adversary, but the adversary is seen as deserving, then the intensity of the resentment is lowered according to the degree of deservingness.
Our task was to define a set of variables that would allow us to calculate coarse emotion intensities for all emotion episodes that the Affective Reasoner was capable of representing. We constrained our task by specifying three discrete levels of intensity for each of the twenty-four emotion classes, giving us seventy-two different emotion/intensity duples that were to be represented. To test our representation we analyzed a number of different emotion episodes, as well as complete scenes, for their affective content, and used the emotion-intensity variables to represent the perceived causes of the different intensities, within the context of our emotion theory. Episodes were analyzed with respect to the way in which the intensity variables allowed us to map simulation-event features into values for the eliciting conditions of the different emotion types (see figure 1). Complete scenes were additionally analyzed for the ways in which affective states could be mapped to mood changes for the automated agents, using the intensity variables, so that the intensities of subsequent emotions would be correct for those agents. The present set of variables was sufficient to represent all of these episodes and scenes, at the Affective Reasoner's level of granularity.
The intensity model has not been implemented for all emotion types. Unsolved representational problems exist for some of the variables, such as surprisingness and deservingness. The functions that use the emotion intensity variables are in the early stages of development. Moreover, the cognitive correlates of the variables we have discussed interact, and they have cognitive import beyond that of emotion intensity. For example, a disturbance in what we have termed sense of reality might not only lower the intensity of person's emotions but could also affect which emotions are generated, given the same eliciting conditions (e.g., when audience members laugh at gruesome scenes in a horror movie). Because we represent intensity variables and emotion intensities as scalars our model is not able to address these phenomena. A more cognitively faithful representation would allow multidimensional intensity variables to affect each other in context dependent ways, and would use such variables in not only the determination of emotion intensity but also in the basic construal process responsible for the selection of the emotion types themselves. We consider our work to be only a preliminary effort.
(defun goal-relevant-intensity
(achieved-blocked
&key importance-achieving importance-not-blocking certainty
sense-of-reality temporal-proximity surprisingness
effort arousal physical-well-being importance-gsps
valence-bias valence-bias-strength
anxiety-invincibility anxiety-invincibility-strength
depression-ecstasy depression-ecstasy-strength)
(let ((emo-valence (if (> achieved-blocked 0) 'positive 'negative)))
(*
*goal-relevant-normalization*
achieved-blocked
(if (eql emo-valence 'positive)
(or importance-achieving *ia-default*)
(or importance-not-blocking *inb-default*))
(or certainty *certainty-default*)
(or sense-of-reality *sense-of-reality-default*)
(or temporal-proximity *temporal-proximity-default*)
(or surprisingness *surprisingness-default*)
(or effort *effort-default*)
(or arousal *arousal-default*)
(or physical-well-being *physical-well-being-default*)
(or importance-gsps *importance-gsps-default*)
(if valence-bias
(if (eql valence-bias emo-valence)
valence-bias-strength
(protect-divide 1 valence-bias-strength))
*valence-bias-default*)
(if anxiety-invincibility
(if (eql anxiety-invincibility emo-valence)
anxiety-invincibility-strength
(protect-divide 1 anxiety-invincibility-strength))
*anxiety-invincibility-default*)
(if depression-ecstasy
(if (eql depression-ecstasy emo-valence)
depression-ecstasy-strength
(protect-divide 1 depression-ecstasy-strength))
*depression-ecstasy-default*))))
Figure 2.
(defun fortunes-of-other-intensity
...
*fortunes-of-others-normalization*
friendship-animosity-strength
(or emotional-interrelatedness *emotional-interrelatedness-default*)
...
(if deservingness-undeservingness
(if (eql deservingness-undeservingness 'deserving)
(if (eql emo-valence 'positive)
deservingness-undeservingness-strength
(protect-divide 1 deservingness-undeservingness-strength))
(if (eql emo-valence 'negative) ;;; note: d/u = 'undeserving
deservingness-undeservingness-strength
(protect-divide 1 deservingness-undeservingness-strength))
)
*deservingness-undeservingness-default*)))
...
Figure 3.