Assessing the validity of appraisal - Institute for Creative Technologies

Assessing the validity of appraisalbased models of emotion
Jonathan Gratch, Stacy Marsella, Ning Wang, Brooke Stankovic
Institute for Creative Technologies
University of Southern California
The projects or efforts depicted were or are sponsored by the U.S. Army Research,
Development, and Engineering Command (RDECOM). The content or information
presented does not necessarily reflect the position or the policy of the Government, and
no official endorsement should be inferred.
Computational models of human emotion

Goal: Build accurate models of cognitive antecedents and
consequences of emotion
– To enhance predictive power of human decision-making models
(Loewenstein & Lerner, 2003; Frank 1988; Busemeyer 2007)
– To simulating human interpersonal behavior
For training (Swartout et al; Aylett et al; Paiva et al)
– For user modeling (Conati)
– Methodological tools for improving theories of emotion (Sanders&Scherer)
2
Theoretical Framework: Appraisal Theory
(Arnold, Lazarus, Frijda, Scherer, Ortony et al.)
Desirability
Environment
Goals/Beliefs/
Intentions
Expectedness
Controlability
Causal Attribution
Action
Tendencies
Problem-Focused
(act on world)
3
Emotion
“Affect”
Coping
Strategy
Physiological
Response
Emotion-Focused
(act on beliefs)
Computational Appraisal Models
TABASCO
Appraisal
Theories
Frijda
OCC
Staller&Petta
ActAffAct
FLAME
ACRES
WILL
El Nasr
Swagerman
Moffat
EMILE
AR
EM
Gratch
Elliott
Neal Reilly
Lazarus
Rank
ParleE
Bui
THESPIAN
Si et al.
EMA
FearNot!
Gratch/Marsella
Dias
CBI
PEACTIDM
Marsella
Marinier
Scherer
4
ALMA
WASABI
Gebhard
Becker-Asano
Many models, which is best?

Few efforts have systematically evaluated model validity

No efforts have directly compared models
– Models typically tested in context of application
or
– Models appeal to empirical support of appraisal theory
BUT don’t assess design choices in realizing theory
FURTHER, Models make many conflicting design choices and
thus are difficult to directly compare
Our approach: break models into constituent design choices
and evaluate these separately
5
A component model view of appraisal models
Affect Derivation
Model
Appraisal Derivation
Model
Affect Intensity
Model
Personenvironment
Relationship
Appraisal
variables
Emotion/
Affect
Affect Consequent
Model
Behavioral
Cognitive

Question for today’s talk
– What is mathematical relationship between appraisal
and intensity of emotional response?
6
Alternative intensity models
Models make different predictions as events change over time
Winning Probability
Expected Utility: hope determined by amount of certainty (EMA, FearNot!)
Expectation Change: hope determined by change in certainty (EM, PEACTIDM)
∆Prob(T1,T2)
∆Prob(T0,T1)
T0
T1
T2
Expected Utility principle:
hope increases over time
Expectation change principle: hope decreases over time
7
Alternative Intensity Models
Additive
Threshold
Expectation Change
Winning Probability
Expected Utility
Expectation Change Model
T0
8
Model emotion
intensity as
proportional to
probability and
utility of goal
attainment
U x ∆Prob(T1,T2)
U x ∆Prob(T0,T1)
T1
T2
Emotion Intensity Hypotheses
Hope
Joy
Fear
Sadness
ΔExpect
Model
EM.
PEACTIDM
ParleE,
PEACTIDM
EM,
PEACTIDM
ParleE,
PEACTIDM
Expected
Utility
EMA,
Silverman,
FearNot!
EMA
Silverman
EMA,
EM
Threshold
Model
EMA,
EM
Additive
Model
Cathexis.
FLAME
Cathexis,
FLAME
Cathexis,
FLAME
Cathexis,
FLAME
Hybrid
Model
Price et al85
Price et al85
Silverman
Price et al85
Price et al85
Silverman
9
Empirical investigation desiderata

Assess behavioral fidelity of competing models
consistent with human emotional responses in naturalistic settings?
– Focus on appraisal variables of goal probability and utility
As these most commonly implicated
 But explore other common variables

– Generate data on appraisals and emotional intensity
– Identify paradigm where emotion arises from task

In contrast to mood induction studies
– Identify task where emotions unfold over time
As most models are intended to be dynamic
 But most empirical findings in psychology focus on non-dynamic tasks

10
Study
Competitive Turn-based strategy game

– Partial Observability
– Dynamic: situation shifts over time
OBJECTIVE: examine dynamics of appraisal & coping responses as
goal of WINNING facilitated or threatened

Q1: How do appraisals relate to intensity of emotional response over time

Q2: How do people cope with the emotions wining or losing gives rise to?

Q3: Do appraisals uniquely determine emotional response?

Do results corroborate EMA model predictions?
11
Modeling game play
Probability
Sad
Play Game
p=.5
Lose $10
p=.5
Win $10
Joy
Fear
Utility
Hope
Manipulate Incentives (Utility)
Kahneman, D., & Tversky, A. (1979).
Potential Loss
Sad
Play Game
Play Game
p=.5
Lose $10
p=.5
Win Nothing
p=.5
Lose Nothing
p=.5
Win $10
Joy
Potential Gain
Fear
Hope
Manipulate Outcomes (Probability)
Lose
Sad
Play Game
Play Game
p= 1
p=.5
Lose Payoff
p= 0
p=.5
Win Payoff
p= 0
p=.5
Lose Payoff
p= 1
p=.5
Win Payoff
Joy
Win
Manipulate Probability of Winning over TIME
Start  Losing  Lost
Play Game
Play Game
p=.5
Lose Payoff
p=.5
Win Payoff
p=.5
Lose Payoff
p=.5
Win Payoff
Start  Winning  Won
2 x 2 x 3 design
Outcome and Incentive manipulated between subjects
Time manipulated within-subjects
Incentive (Gain vs Loss)
Outcome (win vs. lose)
Win $
Don’t win $
Don’t lose $
Lose $
Human subjects study 100 participants
Time 1
Prior
Expectations
WINNING
WON GAME
Time 21
Time 3
LOSING
LOST GAME
Hidden
Camera
Prior
Expectations
Subject
Confederate
Coping
Questionnaire
Measures

Demographic/Dispositional (start of experiment)
– Age, Education, Game experience
– Social value orientation: measure of cooperative/competitive

Appraisals (repeated T1, T2, T3)
–
–
–
–
Subjective value of winning
Subjective probability of winning
Subjective control over winning/losing
Subjective effort (how hard am I trying)

Emotion intensities (repeated T1, T2, T3)
– Prospective emotions: Hope, Fear
– Retrospective emotions: Joy, Sadness

Presented as visual analog scales
18
Manipulation check

Successfully manipulated perceived winning/losing over time

Failed to manipulate value of winning/losing (incentive)
– Did elicit positive and negative self-reported emotion
– No significant differences in appraisals/emotions by incentive
– Collapse data across incentive
19
Raw Emotion Intensity Scores
Hope
Joy
20
Fear
Sadness
What are the significant changes in intensity
as a function of probability
Hope
0.06
*
*
**
Consistent with
Expected Utility
Model
0.07
*
Fear
*
**
Joy
*
Sad
Lost
21
Consistent with
Threshold
Model
Losing
Tie
Wining
Won
100
JOY
80
30
FEAR
25
Model
Fitting
20
60
15
40
10
20
5
0
0
0 10 20 30 40 50 60 70 80 90 100
Probability
100
HOPE
80
0
30
10 20 30 40 50 60 70 80 90 100
Probability
SADNESS
25
20
60
15
40
10
20
5
0
0
0 10 20 30 40 50 60 70 80 90 100
Probability
0
10 20 30 40 50 60 70 80 90 100
Probability
Quantitative Fit
Joy = 1.41  Utility0.83  Probability1.54 + 2.37
Sad = 0.60  Utility0.82  (1-Probability)3.06 + 2.32
Hope = 0.02  Utility1.45  Probability1.0 + 1.45 where Probability < 1.0
Fear = 0.79  Utility0.98  (1-Probability)1.21 + 30.38 where Probability > 0.0
22
(r2 = 0.80)
(r2 = 0.83)
(r2 = 0.93)
(r2 = 0.92)
Q1: Emotion Intensity Results (nonlinear regression)
Hope
Realization EM.
PEACTIDM
Model
Expected
Utility
Joy
Fear
Sadness
ParleE,
PEACTIDM
EM,
PEACTIDM
ParleE,
PEACTIDM
EMA,
Silverman,
FearNot!
EMA
Silverman
EMA,
EM
Threshold
Model
EMA,
EM
Additive
Model
Cathexis,
FLAME
Cathexis,
FLAME
Cathexis,
FLAME
Cathexis,
FLAME
Hybrid
Model
Price et al85
Price et al85
Silverman
Price et al85
Price et al85
Silverman
RESULT: Strong support EMA (and date can refine model)
23
Discussion
– No effect of incentive framing

Possibly did a poor job of framing as win/loss
– Subjects may not have understood the manipulation

Suggests people have other incentives than monetary reward
– Competition with other
– Fun of playing game
– Social interaction

Future studies will explicitly examine other goals
– E.g., Use Subjective Value Inventory (SVI, Curhan 2006)
24
Discussion
Granularity of representation

– Our analysis assumes situation is construed by subjects as a single goal
(win) and a single abstract action (play-game)
Play Game
p= 1
p=.5
Lose Payoff
p= 0
p=.5
Win Payoff
– Actually situation more complex

Subgoals: sink ships, plot next shot
– Would tend to skew some of the analysis
E.g., Joy when Winning could reflect the joy associated with obtaining subgoals
 Suggests Joy, Sadness might be closer to threshold model than suggested by
results

25
Discussion
Other appraisal factors

– Some models consider several other intensity modifiers
– Probability and utility explained most of the variance in intensity
– No evidence that control or effort explained variance in intensity
Dynamics

– Prior psychological studies show evidence for expectation change model
in one-shot decision tasks (e.g., wheel of fortune)
– These models define change of probability against some reference point
– But this point not well defined if probability changes continuously over time
– Expectation change did not well-explain our data
26
Open issues (just starting to scratch surface)

Alternative explanations

Decision dynamics
– Explored monotonically-evolving decisions (losing vs. wining)
– Should explore other trajectories


does early failure impact future perceptions when circumstances improve?
Individual differences
– Subjects with low motivation to win show very different
behavioral/coping patterns
– Other appraisal/dispositional factors seem to improve predictions




Social Value Orientation
Personality
Cultural factors?
Social factors
– Battleship is a competitive game (theory of mind factors)
27
Conclusion

Identified that different models use different intensity fns.

Constructed study to assess these against human data

Evidence shows
– Expected utility good model for prospective emotions (hope/fear)
– Retrospective emotions (Joy, Sadness) fall between an expected utility
and threshold model

Results call into question the behavioral fidelity of several
popular models and support some.

Results particularly support EMA (Gratch and Marsella)
28