High Expressivity/ Low Empathic Accuracy

Comforting Agents
Daniel Schulman
Timothy Bickmore
Relational Agents Group
College of Computer and Information Science
Northeastern University
Relational Agents Group

Relational agents
are designed to
establish long-term
social-emotional
relationships with
people to facilitate
health care and
education
programs.
Relational Agents Group

Just-in-Time
information for
exercise adoption.
Relational Agents Group

Virtual patient
advocates for
hospital discharge.
Relational Agents Group

Medication adherence for adults w/
schizophrenia
General Research Issues


What behaviors are necessary to
establish and maintain a long-term
relationship?
How can an agent produce these
behaviors?
Comforting


Comforting is a key
relationshipbuilding behavior.
And an important
determiner of
outcomes in
healthcare.
Agents that Limit User
Expressivity

Multiple-choice
responses – eliminate
misunderstandings.
Many users complain of lack of expressivity:
“She would ask a question and I would have a
choice, one, two, three, four. But I could never
explain…
I don’t feel that it gets the true feeling, the true
experience of a person.”
Expressivity vs. Empathic
Accuracy
High Empathic Accuracy/
Low Expressivity

Users dislike reduced
expressivity.

But we can provide
accurate comforting.
High Expressivity/
Low Empathic Accuracy

High user expressivity.

Cannot accurately infer
their emotional state.

Provide only a generic
comforting message.
Research Question: Is high empathic accuracy
or high user expressivity more effective for
comforting?
Related Work

(Klein et. al., 99) Induced frustration and
responded via text dialogs.



Multiple-choice input w/ empathic response
better than free input (venting) w/ no response.
(Hone 05) Embodied agent is more effective
than text-only.
Both only tested empathic response versus
no response at all.
Experimental Design



Within-subjects, 2 conditions.
Subjects receive a mood induction, then an
agent intervention.
Conditions use different intervention scripts.
Rest (Baseline)
Anxiety
Mood Induction
Intervention
Relational Agent
Multiple-choice input
Free-speech input
(Low user expressivity)
(High user expressivity)
Wizard-of-Oz Control


Wizard controls agent by indicating what response
user made.
For free speech, wizard controls head-nodding.
Person 1
Person 1
Relational Agent Interventions
High Accuracy
High Expressivity
“How are you
feeling right now?”
“How are you
feeling right now?”
great
Happy facial
display
stressed
exhausted
etc.
“…how ___
are you?”
Close-up
concern display
comforting
free speech
input
“Really? That is
interesting to hear”
Measures



Heart rate and Galvanic Skin Response
Positive and Negative Affect Schedule
(PANAS)
Attitude towards agent.
 Satisfaction
 Desire to continue
 Liking
 Perceived caring
Results

Preliminary results (N=16) indicate that
high empathic accuracy is better



Significant on self-reported affect and
satisfaction with agent.
Near-significant on most other measures.
More complete results (N=28) coming
soon…
Future Work


Input modalities (touchscreen vs.
speech)
Content of messages:


Varying person-centeredness and
nonverbal immediacy.
Generating comforting dialogue given
a description of eliciting events.
Mood Induction


Timed arithmetic
Time is manipulated to
slightly less than
needed:



3 correct → 90% time
used.
3 wrong → 10% slower.
Failure manipulation: all
subjects fail to get the
“average” 60% correct.