ArticlePDF Available

Evidence for Gender Specific Approaches to the Development of Emotionally Intelligent Learning Companions

Authors:

Abstract and Figures

A 2 x 2 experiment investigated the effect of elements of an affective learning companion's emotional intelligence on seventy-six 11-13 year-old participants during a challenging problem solving activity. The experiment contrasted use of an agent showing sensor-driven non-verbal mirroring with one showing pre- recorded non-verbal interactions and, separately, affective support vs. task support interventions. The effect of emotional intelligence, in terms of the presence of active listening, delivery of appropriate interventions, and type of non-verbal interactions on participant's experience, including: frustration, perseverance, intrinsic motivation, and meta affective skill were examined. Hypothesized effects of interacting with a more vs. less emotionally intelligent agent did not hold true at the group level, however significant gender differences were found. Discussing these, this paper contributes new evidence on the importance of appropriately coordinating the relationships between affect and task based intervention and non-verbal mirroring with respect to the affective state of girls and boys.
Content may be subject to copyright.
Evidence for Gender Specific Approaches to the Development of
Emotionally Intelligent Learning Companions
Winslow Burleson
Computer Science and Engineering/
Arts, Media and Engineering
Arizona State University
Winslow.Burleson@asu.edu
Rosalind Picard
MIT Media Lab
Picard@media.mit.edu
Abstract:
A 2 x 2 experiment investigated the effect of elements of an affective learning companion’s emotional
intelligence on seventy-six 11-13 year-old participants during a challenging problem solving activity. The
experiment contrasted use of an agent showing sensor-driven non-verbal mirroring with one showing pre-
recorded non-verbal interactions and, separately, affective support vs. task support interventions. The
effect of emotional intelligence, in terms of the presence of active listening, delivery of appropriate
interventions, and type of non-verbal interactions on participant’s experience, including: frustration,
perseverance, intrinsic motivation, and meta affective skill were examined. Hypothesized effects of
interacting with a more vs. less emotionally intelligent agent did not hold true at the group level, however
significant gender differences were found. Discussing these, this paper contributes new evidence on the
importance of appropriately coordinating the relationships between affect and task based intervention and
non-verbal mirroring with respect to the affective state of girls and boys.
Introduction:
Social bond between teachers and learners and affective support has been shown to have
considerable impact on learners’ performance and motivation. Wentzel has shown that caring
bonds between middle school children and their teachers are predictive of learners’ performance
[1]. Lester has shown that intelligent tutoring systems that employ agents elicit a social presence
or “persona effect” that increases learners’ engagement [2]. Beyond the persona effect, Bailenson
and Yee have shown that non-verbal mirroring in the form of behavioral mimicry can increase the
likeability and persuasive effect of a virtual agent [3]. Bickmore and Picard have developed
interaction and evaluation strategies to increase empathetic and caring relationships between
agents and participants [4]. Providing participants a choice, in terms of the ethnicity and gender
of an agent-tutor, has also been shown to have beneficial impacts on learners’ impressions of the
agent and on their own performance; similarly, matching learners’ gender and ethnicity also leads
to more positive impressions and performance [5].
One of the ways to develop a social bond with learners is to provide assistance. Systems that
provide affective support at times of user frustration have been shown to reduce frustration [6].
Emotional support is an important factor in learning activities. In fact, in a study of expert human
tutors’ interactions with their students it was found that up to half of these interactions are
focused on support of the learner’s affective state [7], yet most intelligent tutoring systems
provide predominantly task based support.
Dweck has shown that supporting learners by encouraging them to “think of the mind as being
like a muscle and believe that they can increase their intelligence through effort, even when
experiencing frustration” helps learners in their approach to and perseverance in challenging
learning activities [8]. Her message supports learners’ development of “meta-affective skill” –the
ability to coordinate meta-affective knowledge (knowing a strategy based on affect, such as
“when you feel frustrated it helps to think of the mind as a muscle”…) with meta-affective
experience (a conscious reflection on what an emotion, such as frustration, is doing to you, or
may do to you).
Inspired by these findings, and many others that point to the importance of supporting the
emotions of people, we undertook the design of an automated companion that could sense and
respond to certain aspects of human emotion in a learning context. Because no automated
system today can reliably sense all the emotions that occur during learning, and no system is
smart enough to know how to respond appropriately all the time to the affective information that
is sensed, this undertaking is an extremely ambitious one, and illuminates the challenges in
creating successful versions of such future technologies. The research described in this paper
implements just a few of the multitudinous possibilities for intelligently sensing and responding
to learner affect, but provides the first experiment that we know of to implement real-time
character responses to affective cues based on theory of how to support learners. The rest of this
paper describes the experiment, main hypotheses and findings, together with discussion and
recommendations about future experiments in this area.
Experiment:
A multi-modal real-time affective agent research platform [9] that incorporates a facial expression
camera, pressure mouse, skin conductance sensor, and posture chair to engage in sensor driven
non-verbal mirroring was built and used to begin to develop elements of an affective learning
companion’s emotional intelligence (Figure 1). This system collects data from the sensors that
relates to the users’ affective states. The data is both processed off-line with a classifier to
determine affective state [9] and processed in real-time via a system server to influence the
character’s interactions with the user. The system server coordinates the user interface, activity,
behavior engine and character interactions. The behavior engine processes the real-time data
from the sensors to determine non-verbal interactions that are in turn displayed by the character
engine. The character’s behaviors include speaking, nodding, smiling or fidgeting the mouth,
shifting its posture forward or backward, changing its color and fidgeting very slightly. These are
the main behaviors controlled in this experiment, even though the character is capable of much
more (e.g., turning its head to watch the actions of the learner, or walking around.)
An experiment was conducted with seventy six 11-13 year old girls and boys, who interacted with
the agent and sensing system in the context of a challenging problem solving activity, the Towers
of Hanoi activity with 7 disks [9]. The character followed one of two strategies for its non-verbal
movements: (1) Sensor driven non-verbal mirroring, in which the four sensors were used to
create a 4-second delayed behavioral mimicry of elements of facial expression, agitated swaying
proportional to mouse pressure, reddening skin tone relative to skin conductance values, and
leaning forward/back posture mirroring that of the learner; or (2) Pre-recorded interactions
generated from the recorded files of the “most average” pilot participant interactions (determined
using the standard deviation of each behavioral channel to categorize five naturally-occurring
pilot files as being most average across the behavioral channels; for each new participant one of
these five naturally-occurring files was randomly selected and used to present a pre-recorded
control condition that exhibited a similar range of non-verbal behavioral expressions to those
receiving sensor-driven non-verbal interactions.) Thus, both cases involved non-verbal
movements by the character, but in only the first case were these synchronized to the learner’s
current sensor outputs. In a series of pilot studies participants were found to be unaware of the
agent’s 4-second delayed mirroring. In addition to non-verbal interactions, the character also
practiced one of two interventions: (1) Affective support intervention included adaptive “active
listening” strategies [6] and support of meta-affective skill based on Dweck’s message (“the mind
is like a muscle and you can increase your intelligence, through mental exercise”) or (2) Task
support intervention (“another way to think about this is to think about the small disks that are in
the way. If you move these out of the way then you can move the disk that you want to move”).
Both affective and task based interventions concluded with similar phrases, encouraging the
learner to continue with the activity. The overall 2x2 design is summarized in Table 1.
Figure 1. Affective Agent Research Platform with sensors listed from right to left: off-line video
camera, facial expression camera, pressure mouse, skin conductance sensor, and posture chair.
Sensor-driven non-verbal mirroring Prerecorded non-verbal interaction
Affect support
intervention
Affect support, non-verbal mirroring
16 valid out of 20 assigned
8 girls valid out of 10
8 boys valid out of 10
Affect support, prerecorded non-verbal
interaction
14 valid out of 19 assigned
5 girls valid out of 8
9 boys valid out of 11
Task support
intervention
Task support, non-verbal mirroring
15 valid out of 18 assigned
7 girls valid out of 8
8 boys valid out of 10
Task support, prerecorded non-verbal
interaction
16 valid out of 19 assigned
9 valid out of 11
7 boys valid out of 8
Table 1. The 2 x 2 design contrasting intervention x mirroring conditions. Cells depict the number of
valid participants included in the analysis of the hypotheses (See Participants section).
In this experiment, the affective learning companion was coded as being more emotionally
intelligent when it engaged in sensor driven non-verbal mirroring and likewise when it provided
affective support interventions than when it provided neither. These interventions and the
mirroring condition were considered additive; an agent that provided both mirroring and affective
support would be considered more emotionally intelligent than one that provided either
separately, or neither.
Hypothesis
There were four specific areas to which this work had planned contributions:
First, it sought to extend Bailenson’s use of Transformed Social Interactions [3], where he
showed that when a participant wore an immersive head-mounted display that sensed head
motions, and a virtual agent mimicked the head motions, it made the agent more likable and
persuasive. Bailenson’s findings were with college-aged participants and the persuasive message
concerned security card usage. The extension of Bailenson’s approach in this experiment
includes four significant components: providing a new domain (a learning platform), including a
different kind of persuasive message (meta affective skill based), addressing a new audience (11-
13 year old learners), and using a new set of less invasive sensors to extend the mirroring beyond
head movements.
H1: the affective learning companion is expected to be more persuasive (as measured
by self report during introduction, perseverance measures, and intrinsic motivation
measures), and users will form a stronger social bond (as measured by bye.button
response and positive/negative impressions assessed with the Modified Working
Alliance Inventory) with the affective learning companion, when sensor-driven non-
verbal mirroring informs the affective learning companion’s interactions than when
pre-recorded non-verbal interactions are displayed.
* See the Methodology section and Table 3 for further explanation of the H1-H4
measures.
Second, this research sought to create new applications of Dweck’s strategies of intervention [8]
that facilitate learners’ metacognitive strategy and meta affective skill. Additionally, the
pedagogical benefits, of increased social bond [1] and persuasion [3], due to the approach taken
in H1 was expected to leverage Dweck’s message.
H2.A: A learner’s social bond (as measured by bye.button response and
positive/negative impressions assessed with the Modified Working Alliance
Inventory) with an affective learning companion will positively correlate with his or
her perseverance (time from character’s departure until participant clicks on a quit
button or until time limit) and self-theories adoption of internal beliefs that he or
she can increase his or her own intelligence and the adoption of mastery orientation
(as measured by Dweck’s Self Theories of Intelligence and Goal Master Orientation
survey instruments).
H2.B: The level of persuasion (as measured by self report during introduction,
perseverance measures, and intrinsic motivation measures) a learner
experiences from the affective learning companion’s metacognitive message
(presented during the introduction) will positively correlate with the social bond (as
measured by bye.button response and positive/negative impressions assessed with the
Modified Working Alliance Inventory), with perseverance (time from character’s
departure until participant clicks on a quit button or until time limit), and will
negatively correlate with frustration (self-reported at the time of intervention and in
the post-activity survey).
Third, this research sought to design interventions that would increase intrinsic motivation and
reduce frustration by taking into account strategies for empathetic and caring relationship
development [4] and “frustration handling” [6].
H3: An affective learning companion that exhibits emotional intelligence (active
listening provided during the affect support intervention, appropriate interventions --
see Table 5 and related discussion of the congruence measure, and sensor-driven
non-verbal mirroring rather than pre-recorded non-verbal interactions) will increase
learner’s intrinsic-motivation (as measured by voluntary re-engagement with the
activity) and reduce frustration (self-reported at the time of intervention and in the
post-activity survey).
Fourth, the research intended to evaluate the impact of this system on learners’ meta affective
skill development. Meta affective skill addresses a learner’s awareness of feelings during an
activity. An affectively aware Learning Companion might facilitate a learner’s awareness of their
feelings.
H4: Metacognitive skill will be exhibited at higher levels when learners interact with
emotionally intelligent agents (see H3) and will positively correlate with
perseverance (time from character’s departure until participant clicks on a quit
button or until time limit), willingness to continue (as measured by self-report at the
time of intervention), and intrinsic motivation (as measured by
reengaging in the
activity during the final 2 minutes of the protocol)
.
Methodology
The methodology (Table 2) included a pre-test, administered to determine children’s self theories
of intelligence and their goal mastery orientation [8]. The learning companion presented itself
saying “Hi there. My name is Casey. I’m a digital character...”; its introduction was the same
(other than non-verbal interactions, determined by the mirroring vs. pre-recorded condition)
across the affective support vs. task support conditions. The character engaged in either non-
verbal mirroring or prerecorded non-verbal interactions throughout the time of its presence. The
learning companion presented a slide show, during which it asked the learner several questions.
The slide show was based on a script used by Dweck that has been shown to shift children’s
beliefs about their own intelligence toward incremental self theories [8]. The learning companion
then presented the Towers of Hanoi activity and explained that it may have to leave before the
learner completes the activity. The companion instructs the learner to, “Click on a disk to start
whenever you want, I’ll just watch and help if I can.” The learner is given four minutes to engage
with the activity before the character intervenes with either an affect support or task support
based intervention. (see [9], for the exact intervention dialogues). During the intervention, self-
report measures are obtained through the interaction with the companion when it asks face to face
questions of the learner, e.g. ”On a scale from 1 to 7, how frustrated are you feeling right now?”
Then the character says that it will need to leave and tells them, ”I have to go now. Thank you
for letting me watch you do this activity. Watching you has helped me learn too. Sorry that I
have to leave now.” Then the companion encourages them to continue. Finally it says, “If you
feel like you would like to stop there will be a few buttons in the upper right hand corner that you
can press. Bye bye.” Participants have the opportunity to respond by pressing one of three
bye.buttons: “Ok, bye”, “Ok, bye I was glad to have you here”, or “Ok glad you are finally
going”; presented in different random order for each participant, to control for presentation order
effects. After they select one of the three bye-responses, or after 20 seconds elapses (when the
bye-response choices disappear/time out so as not to distract the participant), the character
disappears and three quit buttons that the character previously discussed appear in the upper right
corner of the screen offering the opportunity for the learner to end the activity. The three buttons
appear with the labels: "I want to stop because I'm too frustrated to continue", "I've put in all the
effort that I can and want to stop", and "I want to stop for some other reason".
Protocol Events for Subjects in all Four Conditions Duration in minutes
Assent and consent forms ~3
Initial Survey Questions and Pre-Test (including Self Theories of
Intelligence and Goal Mastery Orientation)
~10
Character arrives, introduces itself, the activity, and shows a Slide Show
(based on Dweck’s message [8])
~10
Participant engages in Towers of Hanoi activity 4
Character provides affect support or task support intervention
Obtains self-report measures
Introduces quit buttons
Then says goodbye (offers bye.button response)
~2
Participants persist in Towers of Hanoi task with three “quit” buttons
present
up to 15 minutes from the start
of the activity
Post-activity survey of experience ~3
Neutral affect inducement video 1.5
Post-Test (including Self Theories of Intelligence and Goal Mastery
Orientation)
~10
Modified Working Alliance Inventory ~2
Opportunity to reengage with Towers of Hanoi 2
Table 2. Experiment protocol with durations in minutes; the approximate values indicate that these
events have participant interactions and therefore some variation in duration
At the time the learner clicks one of the three quit buttons, or 15 minutes after the start of the
activity, whichever happens first, the learner is presented with post-activity questions about the
experience e.g. “How many minutes would you say this activity took from the time you first
moved a disk until now?”, "Mark how much of the time you were frustrated”, and others, (see [9]
for complete list). After these questions the learner is presented with a 1.5 minute video clip of a
seascape, as a neutral affect inducement [10] to help alleviate frustration that may bias answers to
subsequent questions. The learner is then presented with post-test questions on self theories of
intelligence and goal mastery orientation, followed by a modified Working Alliance Inventory [4]
to gauge his or her impression of the character. Finally the learner is again presented with the
Towers of Hanoi activity, still in its previous state, along with instructions indicating that, “It will
be a couple of minutes before the next activity is ready. You can do whatever you want now, just
stay seated here please.” After two minutes another message appears and says: “Thanks for
waiting.” The experimenter informs them they are done, and conducts a debriefing. This final
two minute period allows a learner to reengage in the activity, as an indication of intrinsic
motivation.
Measures
The measures used in this study use self-report surveys, learner’s responses during dialogue with
the learning companion, and learner’s behavioral activities e.g. duration of engagement and re-
engagement in the activity. An abbreviated set of the measures is presented in Table 3 (see
Burleson 2006 for a complete list and discussion of their precise implementations). The learners’
positive/negative impressions of the character were obtained using a Modified Working Alliance
Inventory self-report survey that was based on Bickmore’s research on users’ social bond with
agents [4]. The self theories of intelligences and goal mastery orientation instruments were
developed by Dweck [8] and have been used extensively in her research on learners’ approach to
and perseverance in challenging learning activities. The Flow/Stuck measure was a composite
developed from a self-report survey based on the theory of Stuck, a state of non-optimal
experience and Flow, a state of optimal experience [9]. Frustration was measured through
dialogue based self-report at the time of intervention and in a post-activity survey. We did not
employ specific learning measures in this study; instead we focus on affect and affective learning
(use of affective strategies and related behaviors). While the Tower of Hanoi is so well studied
that at times it is considered a “toy problem” with respect to traditional learning measures, the
focus of this research on affective learning and strategies during frustration deals with the Towers
of Hanoi activity as a very real (and frustrating) experience for learners who have not
encountered it before. Based on our findings in this study we now have plans to conduct studies
of affective learning companions in conjunction with intelligent tutoring systems that have
explicit learning measures in Science, Technology, Engineering, and Math (STEM) topics.
Measures Method
Self report measures during introduction:
Persuasion
Self report measures at time of intervention:
frustration
effort
willing to stick with it
able to use the strategies (presented in the
intervention, affective or task strategies)
Self reported as part of dialogue with the character
(Character asks a question, student selects a dialogue
response or Likert response that indicates their answer)
Perseverance Measure of time from character’s departure until
participant clicks on a quit button or until time limit
Social bond measures:
bye.button response
positive/negative impressions
Measured by the pressing of one of three bye buttons
Modified Working Alliance Inventory (bond dimension)
Post-activity frustration Composite scale from post-activity self report survey
Meta-affective skill Composite scale from post-activity self report survey
More Flow/less Stuck Composite scale from post-activity self report survey
Flow is a state of optimal experience
Stuck is a state of non-optimal experience
Dweck pre/post test measures [8]
goal mastery orientation
self theories of intelligence
Composite scale from pre/post-activity self report survey
Intrinsic-motivation Measure of whether or not learners reengaged in the
activity during the final 2 minutes of the protocol
Table 3. An abbreviated list of the measures used in this experiment (see [9] for comprehensive list)
Participants
The participants were seventy-six 11-13 year old children from three semi-rural schools in
western Massachusetts, randomly assigned to the conditions as shown in Table 1. Attrition
eliminated 10 students due to a variety of factors: some participants needed to leave prior to
completing the activity, due to unexpected all-school meetings or changes in transportation
schedules; power failures due to storms and some equipment failures occurred; a few participants
were unresponsive to the character interactions (e.g. did not answer several questions even after
the experimenter instructed them to do so, so the timing of the introduction and the intervention
were inconsistent with respect to other participants); and one participant was identified by her
teacher as a student with special needs (her rapid response times to the self report questions also
indicated that she did not take time to read the questions.) Five additional students were excluded
from the analysis due to their prior knowledge of the Towers of Hanoi learning activity and/or
because they completed the activity. The analysis below was then conducted on the remaining
sixty-one participants.
Results of the Investigation of the Hypotheses across Gender:
On the whole the results from the investigation of H1-H4 across gender did not support the
hypotheses (see chapters 6.1-6.4 in [9] for a detailed explanation and discussion of these
findings). However, there is some indication that the lack of significant findings may, in some
cases, be because of the boys and girls behaving in opposite ways with respect to the conditions.
Initial investigations with respect to gender, described below, indicated that there were several
differences that showed significant interactions. These interactions may be due to developmental
differences with respect to boys’ and girls’ emotional intelligence, which are particular to the 11-
13 year old age group, which were not adequately understood and incorporated into the initial
design of the experiment. These differences may have contributed to unanticipated variance that
could interfere with support for the primary hypotheses. Since these differences are likely to
generalize to a broader population, they are important to consider in future evaluations of
affective technologies with boys and girls aged 11-13. Thus, while the comparisons below were
unplanned, and do not carry the same weight as the planned comparisons, we think researchers
will find these data of interest for current and future efforts to build emotionally intelligent
learning companions for this age group.
Results of the Investigation of the Hypothesis with regard to Gender:
With an interest in explaining the general lack of support for the primary hypotheses H1-H4 and
to further explore the initial gender findings, this section presents exploratory analysis with
respect to gender differences and gender effects. Here, too, H1– H4 were not generally supported
for either gender separately. However, there were significant (p<0.05) exploratory findings that
have gender specific implications for the development of emotionally intelligent learning
companions (Table 4).
Exploratory Test Significance Girls Mean
and Range
Boys Mean
and Range
ANOVA found that girls were more likely than boys to
“think it will help to know that your mind is like a muscle
and that you can increase your learning through effort”
p=0.03
F=4.9
3.7
range = 1-4
3.3
range = 1-4
ANOVA found that girls felt they would be better able to
use strategies presented in the intervention than boys
p=0.003
F= 9.4
6.1
range = 1-7
4.8
range = 1-7
ANOVA found that girls persevere more than boys
p=0.016
F = 6
14.3 min
range = 0-15min
12.2 min
range = 0-15min
ANOVA found that boys that received task support
responded more positively (bye.button) and had more
positive impressions of the character (Modified Working
Alliance) than boys that received affect support
p=0.02
F=6.1
p=0.036
F=4.8
1.1 affect resp.
1.7 task resp.
range = 0-2
28.0 affect imp.
32.6 task imp.
range = 6-42
ANOVA found that girls that received affect support
responded more positively and had a trend (p = 0.09)
toward more positive impressions of the character
(Modified Working Alliance) than girls that received task
support
p=0.04
F=4.8
p=0.09 (* trend)
F=3
2.0 affect resp.
1.4 task resp.
range = 0-2
33.2 affect imp.
29.2 task imp.
range = 6-42
ANOVA found that post-activity frustration had a
congruence x gender interaction
p=0.043, F=4.3
6.9 high cong.
9.4 low cong.
range = 2-28
11.5 high cong.
9.4 low cong.
range = 2-28
Table 4. Summary of selected exploratory tests contrasting measures between girls and boys.
H1: As mentioned above, conducting an exploratory analysis of H1 for the separate gender
groups did not show general support for H1 for either group. Here is a summary of the
statistically significant gender differences found between girls and boys. Boys were less likely to
“think it will help to know that your mind is like a muscle and that you can increase your learning
through effort” (p = 0.03, F = 4.9). Girls indicated they would be able to use the strategies
presented in the intervention to a greater extent than boys (p=0.003, F =9.4). As others have also
found, we found that girls persevere longer than boys (p=0.016, F = 6).
The intervention had opposing effects for boys and girls with respect to the bye.button response,
with boys responding more positively in the task support condition than boys in the affect support
condition and girls having the opposite relationship with respect to these two conditions. Boys
also had more positive impressions of the character that provided task support than the character
that provided affective support, while girls showed a trend toward the opposite response. These
are the types of differential findings that may explain the lack of significant results in the analysis
across genders.
H2: Conducting an exploratory analysis of H2 for the separate gender groups did not show
support for H2 for either group, although boys self-report using more effort than girls.
H3: Analysis did not find H3 to be generally supported for either girls or boys; however several
interesting findings were made. The investigation of H3 focuses on three components of
emotional intelligence (active listening, appropriate interventions, and sensor-driven non-verbal
mirroring). Active listening is present in the affective support intervention and not present in the
task support intervention. The appropriateness of the affect-support or task-support intervention
provided with respect to a participant’s level of frustration was encoded in the congruence
measure presented in Table 5.
Two caveats should be noted with respect to the congruence measure. First, when considering
those individuals experiencing Low levels of frustration, Low frustration could have meant a
Flow state, in which case an intervention would probably be unwelcome. However, it could have
alternatively meant boredom, in which case any intervention may have been welcome. Thus,
referring to this condition as low or high congruence (see * in Table 5) is a rough approximation
based on limited information. More complex affective state recognition is an open challenge.
Second, because the affective support intervention is adaptive, it may be more appropriate for
individuals experiencing low levels of frustration than the task support intervention for these
same individuals (see ** in Table 5).
While there are a few frustration x intervention conditions,
acknowledged in the table, that might be more appropriate or less appropriate than the
congruence measure indicates this measure was still found to be productive in assessing the
impact of these elements of an agent’s emotional intelligence.
Level of frustration Type of intervention Congruence measure
High Affective intervention High congruence
High Task based intervention Low congruence
Low Affective intervention ** Low congruence *
Low Task based intervention High congruence *
Table 5. Congruence is a function of frustration and intervention that encodes the appropriateness of
the intervention provided with respect to a participant’s level of frustration (*, ** see explanation in
the preceding paragraph).
In the assessment of the relationship between the character’s emotional intelligence (intervention,
congruence, and mirroring) and girl participants’ intrinsic motivation the only significant finding
was an interaction between intervention x congruence (p =0.02, F = 6.288). Girls who received
affect support and had lower levels of congruence (i.e. girls that received affect support and were
less frustrated) did not have as much intrinsic-motivation as those who had higher levels of
congruence (i.e. girls who received affect support and were more frustrated). On the other hand,
girls who received task support and had lower levels of congruence (i.e. girls that received task
support and were more frustrated) had more intrinsic-motivation than those that had higher levels
of congruence (i.e. girls who received task support and were less frustrated). For girls that were
frustrated either intervention increased their intrinsic motivation over those that were less
frustrated (at the time of intervention) (this will be specifically discussed in the next section). For
boys the only measure of motivation that was effected by the character’s emotional intelligence
(intervention, congruence, and mirroring) was their willingness to stick with it which showed a
trend toward significance (p=.065, F=3.8) suggesting that boys may be more willing to stick with
it when they received sensor driven non-verbal interactions than when they received pre-
recorded non-verbal interactions.
The relationship between the character’s emotional intelligence (intervention, congruence, and
mirroring) and participants’ frustration was also assessed separately for girls and boys, using
covariates of age, school, pre-test self theories, and frustration response at the time of
intervention. Girls were reported to be less frustrated than boys at the end of the activity. There
was also an interaction of congruence x gender (p= 0.043, F=4.327.) Boys and girls that received
interventions with lower congruence had similar levels of post-activity frustration mean = 9.4;
boys that received interventions that had higher congruence had a mean = 11.5, indicating more
frustration, while girls that received interventions that had higher congruence had a substantially
lower mean = 6.9, indicating less frustration.
For girls an interaction between intervention x mirroring (p = 0.001, F=16.3) was highly
significant indicating that girls who received non-verbal mirroring and affect support had lower
post-activity frustration than girls who received non-verbal mirroring with task support. Girls
without non-verbal mirroring had the opposite relationship with the interventions girls who
received affect support had higher levels of post-activity frustration than girls with out non-
verbal mirroring who received task support. Girls showed no main effect differences in post-
activity frustration with respect to the type of intervention (affect support vs. task support) that
they received.
For boys, there were significant (p= .009, F= 8.4) differences when grouped by intervention. Boys
showed twice as much post post-activity frustration if they received the affective support than if
they received task support. The higher levels of congruence were also detrimental for boys; they
showed almost twice as much frustration for higher levels of congruence when compared to boys
with lower levels of congruence. There was also a trend toward significance (p= .061, F= 4.0)
for mirroring: boys that received non-verbal mirroring reported a third less frustration than boys
that did not receive mirroring.
There was a significant interaction between congruence x non-verbal mirroring (p= .047, F= 4.6).
Boys without non-verbal mirroring that had more congruent interventions reported levels of
frustration approximately twice the level of frustration experienced by boys in other congruence x
mirroring conditions (i.e. boys that had low levels of congruence, with and without mirroring,
and boys that received high levels of congruence with mirroring).
H4: Here, too, further analysis did not find H4 to be generally supported for either girls or boys;
however several additional interesting findings were made. For girls the affective support
intervention was positively correlated to the meta-affective skill (p=.040, r =.37) and (more
Flow/less Stuck) (p=.006, r=.52). Neither of these correlations was significant for boys.
In contrast to the result that no significant correlation between meta-affective skill and Flow/Stuck
was present when assessed across both genders, the assessment with only girls shows significant
correlation between meta-affective skill and Flow/Stuck (p = .010, r=.49). The assessment of only
boys, for these same measures also shows a significant correlation (p=.021 r=-.40), but for boys
this is a negative correlation. This is another clear instance where the grouping of the genders
mixes different gender effects, yielding no significance when assessed together.
For boys controlling for age, school, self theories, mirroring and intervention, through partial
correlation shows that there is significance (p = .048, r=.34) for meta-affective skill correlating
with perseverance, while there is no significance for girls. With these covariates, neither gender
shows significance for Flow/Stuck correlating with perseverance.
Meta-affective skill and Flow/Stuck were investigated with measures of motivation (stick with it,
strategies, post test goal mastery orientation, “I would like to try this activity again”, and
intrinsic-motivation) using the same covariates. For girls there was no significance for stick with
it; significance was found for both measures with respect to strategies, (p= .027, r=.4766) for
meta-affective skill, and (p= .009 r=.5691) for Flow/Stuck. For girls there was also significant
correlation to changes in goal mastery orientation for the meta-affective skill measure (p=.008,
r=.57) indicating that girls that report higher levels of meta-affective skill also report higher levels
of mastery orientation.
Controlling for the same variables (age, school, self theories, mirroring and intervention) there
was no significant difference in meta-affective skill when the girls that showed intrinsic-
motivation (measured by their reengagement in the task after the post-test surveys) were
compared with those that did not reengage. There was a trend toward significant differences
between these two groups of girls in terms of their Flow/Stuck (p = .067, F=3.8). Those that
reengaged also had slightly higher levels of Flow/slightly lower levels of Stuck; both groups were
fairly high on this measure, so there may also have been a ceiling effect – i.e. the differences may
have been greater. In similar tests boys showed no significant differences across these groups and
measures.
Discussion of Results with regard to Gender:
The exploratory investigation yielded several interesting results that support strong and
potentially important recommendations for further study. This section will summarize the results
of the gender specific analysis presented in the previous section and argue for the importance of a
deeper understanding of the impact of mirroring and of affect and task support, as these relate to
the frustration, meta-affective skill and Flow/Stuck of the 11-13 year-old sampled population. In
particular this section argues for the need for better understanding of the gender differences in the
impact of the elements of a learning companion’s emotional intelligence and for the importance
of the appropriate “coordination” of these elements with each other, for both girls and boys.
As presented in the previous section there were a few differences in the pattern of the social bond
that girls and boys develop with the character, with respect to the type of intervention the
participants received. Boys responded more positively to the character and had more positive
impressions of the character that provided task support than the character that provided affect
support; girls had the opposite pattern. Differences in the social and emotional skill
developments of girls and boys at these ages (11-13 year olds), with girls typically maturing
earlier than boys, may have contributed to these differences.
1
Boys also self-report using more
effort than girls. This finding and the frustration finding for girls discussed later in this
paragraph, may have influenced different levels of interest in this activity, for girls and boys.
There were very few differences found in the motivation measures with respect to the different
elements of the character’s emotional intelligence for either girls or boys. It was found that the
girls that were more frustrated at the time of intervention also showed higher levels of intrinsic
motivation, regardless of intervention. A possible explanation for this may be related to how
much a participant cares about the activity. Girls that care more about doing this activity may
also find it more frustrating. Independent of the frustration and independent of the type of
intervention they receive, the caring may also lead to their increased intrinsic-motivation. In
contrast to the girls, boys showed a strong difference in their levels of frustration due to the type
of intervention, with much lower levels of frustration occurring in the task support conditions.
This is probably related to the social bond differences discussed above, in which boys responded
better to the character in the task support intervention. Likewise it is likely related to the finding
that boys and girls that received interventions that were had lower congruence had similar levels
of post activity frustration; while boys that received interventions that had higher congruence had
higher levels of post activity frustration and girls that received interventions with higher
congruence had substantially less frustration.
One of the biggest gender differences was found in the relationship between meta-affective skill
and Flow/Stuck. In contrast to the result that no significant correlation between meta-affective
skill and Flow/Stuck was present when assessed across both genders, the assessment with only
girls shows a strong correlation between meta-affective skill and more Flow/less Stuck, while for
boys, these measures show a strong correlation in the opposite direction. This is a clear instance
where the grouping of the genders clearly mixes different gender effects, yielding no significance
when assessed together. One possible hypothesis for the discrepancy in gender at this age is that
girls aged 11-13 may be better able to assess their own emotions than boys. If girls are better at
assessing their emotions then they may be better able to use their meta affective skill to lead
themselves to more Flow/ less Stuck. Boys on the other hand may report that they have meta
affective skill but may actually be less able to recognize their own emotions; thus, even though
they have some meta affective skill, they may not be as capable at applying it to their own
experiential benefit.
While girls showed no main effect difference in the level of frustration based on the type of
intervention, a further analysis indicated that this masked a more complex relationship that
showed highly significant differences due to the interaction of the type of intervention and the
presence of mirroring. These differences can be explained in terms of the “coordination” of the
different elements of the character’s emotional intelligence. Girls that experienced an affective
support intervention in conjunction with non-verbal mirroring (condition 1) had lower levels of
frustration than girls who received either affective support without non-verbal mirroring or girls
who received task support with non-verbal mirroring. Condition 1 is a condition in which the
mirroring and intervention are “coordinated” so that the character displays higher levels of
emotional intelligence (as defined in this experiment as the presence of intervention, congruence,
and mirroring) than in the other two conditions. One might argue that girls that received task
support without mirroring were also in a “coordinated” condition that presents a character with
1
The effects for both males and females who received affect support from machines in non-
learning environments have previously been positive, but all of those results were for participants
over age 18, and were delivered in different contexts.
higher levels emotional intelligence; they could also argue that in this condition girls experienced
similar low levels of frustration when compared to the girls in condition 1. Extending this
argument one might then argue that the existing capabilities of Intelligent Tutoring Systems, to
provide task support without mirroring have similar benefits to girls, and the effort to develop
affect support and mirroring are unwarranted. However the importance of affect support for girls
is bolstered by the exploratory analysis of H4 showing that girls that receive affective support
have higher levels of meta-affective skill and more Flow/less Stuck (these relationships were not
found for boys). Meta-affective skill correlated significantly with beneficial changes in goal-
master orientation and there was a trend toward significance in the positive relationship between
Flow/Stuck and intrinsic-motivation. The findings from H3 and H4, taken together, support an
argument not only for the further development of affective support and its benefits for girls, but
also for the appropriate “coordination” of the elements of the character’s emotional intelligence.
These findings indicate that there are important opportunities to increase girls’ meta-affective
skills, increase their experience of Flow and decrease their experience of Stuck, increase their
mastery orientation, and increase their intrinsic-motivation.
Data from the boys also support the argument for coordinating the elements of the character’s
emotional intelligence. The significant interaction between congruence x non-verbal mirroring
indicated that the boys that experience more congruent intervention without mirroring also
experienced twice as much post activity frustration as boys in the other three mirroring x
congruence conditions. This particular form of discordant emotional intelligence displayed by
the character (i.e. more congruent intervention without mirroring) seems to have had a negative
impact on these boys.
Conclusion:
There are four primary experimental contributions of this research. First, the experiment
demonstrated that the primary hypotheses were not supported for this age group when genders are
combined. Second, further analysis illuminated opposing reactions, based on gender, to help
explain the outcomes. Third, affective interventions were positively associated with girl’s meta-
affective abilities, higher levels of Flow, and lower levels of Stuck. Fourth, it was demonstrated
that the various elements of a character’s emotional intelligence should be presented in a
“coordinated” manner. Inconsistencies between the presence or absence of non-verbal social
mirroring and the presence or absence of other elements of emotional intelligence (congruence or
affective support intervention) were associated with both girl’s and boy’s frustration.
In the experiment conducted, the type of intervention (affect support or task support), the level of
congruence of the intervention with respect to a learner’s frustration, and the presence or absence
of social non-verbal mirroring played several important and different roles with respect to girl’s
and boy’s frustration, meta-affective abilities, increased Flow/reduced Stuck, and intrinsic
motivation. If these findings are confirmed by further studies and if they generalize to broader
populations than the participants used in this study, then as Intelligent Tutoring Systems, and
other systems that use relational agent strategies, advance to incorporate greater levels of
emotional intelligence, developers and researchers should be able to make considerable advances
to their systems and to learners’ experiences by incorporating these elements of emotional
intelligence. At the same time developers and researchers must be careful to appropriately
coordinate the diverse elements of emotional intelligence and be well aware of the differences in
the impact of these elements on boys and girls aged 11-13.
Acknowledgements:
We acknowledge contributors to the system development: Ken Perlin and Jon Lippincott, Ashish Kapoor,
John Rebula, David Merrill, Rob Reilly, Barry Kort, Danielle Martin, Devin Neal, Louis-Philippe
Morency, Carson Reynolds, Phil Davis, Karen Liu, Marc Strauss and Selene Mota. This material is based
upon work supported in part by the MIT Media Lab’s Things That Think consortium, in part by the
National Science Foundation under Grant No. ITR 0325428, and in part by Deutsche Telekom. Any
opinions, findings, conclusions, or recommendations expressed in this publication are those of the authors
and do not necessarily reflect the views of the National Science Foundation or the official policies, either
expressed or implied, of the sponsors or of the United States Government.
References:
1. Wentzel, K. (1997). “Student Motivation in Middle School: The Role of Perceived Pedagogical Caring,
Journal of Educational Psychology.” Volume 89, Issue 3, p. 411.
2. Lester, J. C., S. A. Converse, S. Kahler, S. T. Barlow, B. A. Stone, R. Bhogal, (1997). “The persona
effect: Affective impact of animated pedagogical agents.” CHI'97, New York, NY, ACM Press.
3. Bailenson, J. N. and N. Yee (2005). “Digital chameleons: Automatic assimilation of nonverbal gestures
in immersive virtual environments.” Psychological Science 16: 814-819.
4. Bickmore, T. and R. W. Picard (2004). "Establishing and Maintaining Long-Term Human-Computer
Relationships." Transactions on Computer-Human Interaction 12(2): 293 - 327.
5. Baylor, A. L., E. Shen, and Huang, X.(2003). “Which pedagogical agent do learners choose? The effects
of gender and ethnicity.” E-Learn World Conference on E-Learning in Corporate, Government, Healthcare,
& Higher Education. Phoenix, Arizona.
6. Klein, J., Y. Moon, R. Picard. (2002). "This Computer Responds to User Frustration: Theory, Design,
Results, and Implications." Interacting with Computers 14: 119-140.
7. Lepper, M. R., M. Woolverton, M., Mumme, D. L., and Gurtner, J. L. (1993). “Motivational techniques
of expert human tutors: Lessons for the design of computer-based tutors.” Computers as cognitive tools, S.
P. Lajoie and S. J. Derry. Hillsdale, NJ, Erlbaum: 75-105.
8. Dweck, C. S. (1999). Self-Theories: Their role in motivation, personality and development. Philadelphia,
PA, Psychology Press.
9. Burleson, W. (2006), "Affective Learning Companions: Strategies for Empathetic Agents with Real-
Time Multimodal Affective Sensing to Foster Meta-Cognitive and Meta-Affective Approaches to Learning,
Motivation, and Perseverance," MIT PhD Thesis, September 2006.
(http://affect.media.mit.edu/publications.php)
10. Rottenberg, J., Ray, R. D., and Gross, J. J. (in press). Emotion elicitation using films. To appear in J. A.
Coan & J. J. B. Allen (Eds.), The handbook of emotion elicitation and assessment. London: Oxford
University Press.
Author’s Biographical Sketch:
Winslow Burleson is an Assistant Professor at Arizona State University
in the Department of Computer Science and Engineering and the Arts,
Media, and Engineering graduate program. His technical interests focus
on human-computer interaction applied to creativity, innovation, well-
being, design engineering, and educational technology. He received his
PhD in Media Arts and Sciences from the MIT Media Lab, Affective
Computing Group. He is a member of the ACM.
Arizona State University, 699 South Mill Avenue Room 395,Tempe,
AZ 85281, Winslow.Burleson@asu.edu
(pls crop image as needed)
Rosalind W. Picard is Professor of Media Arts and Science, Director of Affective Computing
Research, and Co-Director of the Things That Think Consortium at the MIT Media Laboratory.
Her interests include the
development of new technologies and theories that advance basic
understanding of affect and its role in human experience.
She holds a Doctorate of Science
from MIT in Electrical Engineering and Computer Science, and is a Fellow of the IEEE and a
member of the ACM and AAAI.
MIT Media Laboratory, E15-448; Cambridge, MA 02142-1308, Picard@media.mit.edu
QuickTime™ and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
... If the PWD becomes "stuck" (a nonoptimal experience state, eg, the user becomes confused or loses interest or focus) [29], as determined by a combination of the skin conductance level and context of recent sequence of problems in dressing activity, motivational prompting is provided to reengage the PWD. ...
... Inside part of the pants I Any markers from low part of the left side of the pants (28,29) are visible for 2+ seconds. Left leg of the pants worn L Any markers from low part of the right side of the pants (25,26) are visible for 2+ seconds. ...
... Many different social and relational factors can increase rapport, trust, and engagement with virtual agents and robots. For example, using appropriate social cues (Desteno et al., 2012;Lee et al., 2013;Breazeal et al., 2016b), contingent backchanneling ( Park et al., 2017a), nonverbal mirroring ( Bailenson et al., 2005;Burleson and Picard, 2007;Lubold et al., 2018), responsiveness and proactivity (Kim et al., 2006), increased social presence ( Lester et al., 1997), and matching ethnic communication styles (Cassell et al., 2009) all have had positive effects. ...
Article
Full-text available
In positive human-human relationships, people frequently mirror or mimic each other's behavior. This mimicry, also called entrainment, is associated with rapport and smoother social interaction. Because rapport in learning scenarios has been shown to lead to improved learning outcomes, we examined whether enabling a social robotic learning companion to perform rapport-building behaviors could improve children's learning and engagement during a storytelling activity. We enabled the social robot to perform two specific rapport and relationship-building behaviors: speech entrainment and self-disclosure (shared personal information in the form of a backstory about the robot's poor speech and hearing abilities). We recruited 86 children aged 3–8 years to interact with the robot in a 2 × 2 between-subjects experimental study testing the effects of robot entrainment Entrainment vs. No entrainment and backstory about abilities Backstory vs. No Backstory. The robot engaged the children one-on-one in conversation, told a story embedded with key vocabulary words, and asked children to retell the story. We measured children's recall of the key words and their emotions during the interaction, examined their story retellings, and asked children questions about their relationship with the robot. We found that the robot's entrainment led children to show more positive emotions and fewer negative emotions. Children who heard the robot's backstory were more likely to accept the robot's poor hearing abilities. Entrainment paired with backstory led children to use more of the key words and match more of the robot's phrases in their story retells. Furthermore, these children were more likely to consider the robot more human-like and were more likely to comply with one of the robot's requests. These results suggest that the robot's speech entrainment and backstory increased children's engagement and enjoyment in the interaction, improved their perception of the relationship, and contributed to children's success at retelling the story.
... Thus, the system needs to ensure that it is maintained. There is an abundance of literature on modeling affect and motivation (Afzal and Robinson 2011;Burleson and Picard 2007;D'Mello and Graesser 2012). It has been argued that, whereas a precise estimation of a learner's specific state might not be possible or even required, an approximation of the state can be helpful in the sense that it can allow a system to be more empathetic, leading to higher levels of engagement (Suleman et al. 2016). ...
Article
Full-text available
This paper describes an embodied conversational agent enhanced with specific conversational strategies aiming to foster learners’ readiness towards communication in a second language (L2). Willingness to communicate (WTC) in a second language is believed to have a direct and sustained influence on learners’ actual usage frequency of the target language. To help overcome the lack of suitable environments for increasing L2 learners’ WTC, our approach is to build embodied conversational agents that can help learners surmount their apprehension towards communication in L2. Here, we focus on the dialogue management aspects of our approach and propose a model based on a set of communication strategies (CS) and affective backchannels (AB) to foster such agents’ ability to carry on natural and WTC-friendly conversations with learners. We examined learners’ expected WTC after interacting with one of the following versions of the system: an agent featuring both CS and AB; an agent featuring only CS; and an agent featuring only AB. The results suggested that combining CS and AB empowers the conversational agent and leads to higher expected WTC among L2 learners. We also found that even the AB-only version of the system had the potential to enhance WTC to some extent. These findings are evidence of the feasibility of enhancing L2 learners’ engagement towards communication using a computer-based environment coupled with appropriate conversational strategies.
... If the PWD becomes "stuck" (a nonoptimal experience state, eg, the user becomes confused or loses interest or focus) [29], as determined by a combination of the skin conductance level and context of recent sequence of problems in dressing activity, motivational prompting is provided to reengage the PWD. ...
Article
Full-text available
Background: Individuals living with advancing stages of dementia (persons with dementia, PWDs) or other cognitive disorders do not have the luxury of remembering how to perform basic day-to-day activities, which in turn makes them increasingly dependent on the assistance of caregivers. Dressing is one of the most common and stressful activities provided by caregivers because of its complexity and privacy challenges posed during the process. Objective: In preparation for in-home trials with PWDs, the aim of this study was to develop and evaluate a prototype intelligent system, the DRESS prototype, to assess its ability to provide automated assistance with dressing that can afford independence and privacy to individual PWDs and potentially provide additional freedom to their caregivers (family members and professionals). Methods: This laboratory study evaluated the DRESS prototype's capacity to detect dressing events. These events were engaged in by 11 healthy participants simulating common correct and incorrect dressing scenarios. The events ranged from donning a shirt and pants inside out or backwards to partial dressing-typical issues that challenge a PWD and their caregivers. Results: A set of expected detections for correct dressing was prepared via video analysis of all participants' dressing behaviors. In the initial phases of donning either shirts or pants, the DRESS prototype missed only 4 out of 388 expected detections. The prototype's ability to recognize other missing detections varied across conditions. There were also some unexpected detections such as detection of the inside of a shirt as it was being put on. Throughout the study, detection of dressing events was adversely affected by the relatively smaller effective size of the markers at greater distances. Although the DRESS prototype incorrectly identified 10 of 22 cases for shirts, the prototype preformed significantly better for pants, incorrectly identifying only 5 of 22 cases. Further analyses identified opportunities to improve the DRESS prototype's reliability, including increasing the size of markers, minimizing garment folding or occlusions, and optimal positioning of participants with respect to the DRESS prototype. Conclusions: This study demonstrates the ability to detect clothing orientation and position and infer current state of dressing using a combination of sensors, intelligent software, and barcode tracking. With improvements identified by this study, the DRESS prototype has the potential to provide a viable option to provide automated dressing support to assist PWDs in maintaining their independence and privacy, while potentially providing their caregivers with the much-needed respite.
... Forbes-Riley and Litman 2011a;) to include, for example, empathetic statements designed to motivate the students. Interestingly, research has shown that there is a gender difference in how empathetic feedback is perceived by students-for example, Burleson and Picard (2007) show that female students respond more positively to empathetic feedback than male students; while Vail et al. (2015) showed that female students were more engaged and less frustrated when provided with affect-aware support than male students. Other research (e.g. ...
Article
Full-text available
The paper is available at http://rdcu.be/GqPq This paper describes the design and ecologically valid evaluation of a learner model that lies at the heart of an intelligent learning environment called iTalk2Learn. A core objective of the learner model is to adapt formative feedback based on students' affective states. Types of adaptation include what type of formative feedback should be provided and how it should be presented. Two Bayesian networks trained with data gathered in a series of Wizard-of-Oz studies are used for the adaptation process. This paper reports results from a quasi-experimental evaluation, in authentic classroom settings, which compared a version of iTalk2Learn that adapted feedback based on students' affective states as they were talking aloud with the system (the affect condition) with one that provided feedback based only on the students' performance (the non-affect condition). Our results suggest that affect-aware support contributes to reducing boredom and off-task behavior, and may have an effect on learning. We discuss the internal and ecological validity of the study, in light of pedagogical considerations that informed the design of the two conditions. Overall, the results of the study have implications both for the design of educational technology and for classroom approaches to teaching, because they highlight the important role that affect-aware modelling plays in the adaptive delivery of formative feedback to support learning.
... Hence the system needs to ensure that such a state is maintained. There is an abundance of literature on modeling affect and motivation with varied views (Afzal, S. and Robinson, P., 2011, Burleson, W. & Picard, R. 2007, Conati, C. & Maclaren, H. 2009, Sidney D'mello and Art Graesser. 2013, Woolf, B. et al., 2010. ...
Article
Full-text available
Negotiation mechanism using conversational agents (chatbots) has been used in Open Learner Models (OLM) to enhance learner model accuracy and provide opportunities for learner reflection. Using chatbots that allow for natural language discussions has shown positive learning gains in students. Traditional OLMs assume a learner to be able to manage their own learning and already in an appropriate affective/behavioral state that is conducive for learning. This paper proposes a new perspective of learning that advances the state of the art in fully-negotiated OLMs by exploiting learner’s affective & behavioral states to generate engaging natural language dialogues that train them to enhance their metacognitive skills. In order to achieve this, we have developed the NDLtutor that provides a natural language interface to learners. Our system generates context-aware dialogues automatically to enhance learner participation and reflection. This paper provides details on the design and implementation of the NDLtutor and discusses two evaluation studies. The first evaluation study focuses on the dialogue management capabilities of our system and demonstrates that our dialog system works satisfactorily to realize meaningful and natural interactions for negotiation. The second evaluation study investigates the effects of our system on the self-assessment and self-reflection of the learners. The results of the evaluations show that the NDLtutor is able to produce significant improvements in the self-assessment accuracy of the learners and also provides adequate support for prompting self-reflection in learners.
... This section describes how sensor data were used for affective state detection. Our sensors and algorithms were minor adaptations of those used by Burleson and his collaborators in earlier work [46], [47], [48]. ...
Article
Full-text available
This project aimed to improve students’ learning and task performance using a non-cognitive learning companion in the context of both a tutor and a meta-tutor. The tutor taught students how to construct models of dynamic systems and the meta-tutor taught students a learning strategy. The non-cognitive learning companion was designed to increase students’ effort and persistence in using the learning strategy. It decided when to intervene and what to say using both log data and affective state monitoring via a fa-cial expression camera and a posture sensor. Experiments with high school students showed that the non-cognitive learning compan-ion increased students’ learning and performance. However, it had no effect on performance during a transfer phase in which the learning companion, meta-tutor and tutor were all absent. The transfer phase null effect must be interpreted with caution due to low power, a possible floor effect and other issues.
... Menu systems, email, and other static interfaces that respond to user input by performing a limited, predetermined task are not EAs, whereas a virtual assistant that takes input from users and responds intelligently to solve a problem is an EA. EAs are increasingly used in various systems, such as in elderly care (Huang et al., 2015), learning (Burleson and Picard, 2007), commerce (Gong, 2007), consumer-assisted product design (Kohler et al., 2011), exercise coaching (Bickmore et al., 2013b), tour guides (Bickmore et al., 2013a), video games (Isbister, 2006), automated interviewing (Nunamaker Jr. et al., 2011), and military applications (Gratch and Marsella, 2004). ...
Article
Because userstreatembodiedagents(EAs)associalactors,usersholdexpectationsabouthuman-to-EA communication (HtEAC)similartothoseinhuman-to-humancommunication.Thisstudyextends Expectancy ViolationsTheory(EVT)toexaminehowdifferentformsofinterfacesthatconfirm orviolate user expectationsaffectthecommunicationprocess,socialjudgments,abilitytoinfluence, andaccuracy of recallassociatedwithHtEAC.PositiveviolationsofexpectancyareactsorcharacteristicsoftheEAthat are unexpectedbutevaluatedfavorablybythehumanpartner.ResultssuggestthatwhentheEAdeviates from expectations,effectsontheHtEACprocessandrelatedoutcomescanbemorepronounced.EAs evaluatedaspositiveviolationshadmorefavorableeffectsontaskattractivenessthanotherhumanorEA interactionpartners.AspredictedbyEVT,EAinteractionsthatwerepositivelyevaluatedelicitedmore perceivedconnectedness,feelingsofbeingunderstood/receptivity,anddependabilitythanthoseEA interactionsevaluatednegatively.However,negativeviolationsdidnotproduceworseoutcomesthan negativeconfirmations. EVToffersausefullensforexaminingthecommunicationeffectsofHtEACand points tobenefits ofcreatingEAsthatevokepositiveviolationsofexpectancy.
Conference Paper
Full-text available
Facial expressions transmit a variety of social, grammatical, and affective signals. For technology to leverage this rich source of communication, tools that better model the breadth of information they convey are required. MiFace is a novel framework for creating expression lexicons that map signal values to parameterized facial muscle movements. In traditional mapping paradigms using posed photographs, naïve judges select from predetermined label sets and movements are inferred by trained experts. The set of generally accepted expressions established in this way is limited to six basic displays of affect. In contrast, our approach generatively simulates muscle movements on a 3D avatar. By applying natural language processing techniques to crowdsourced free-response labels for the resulting images, we efficiently converge on an expression's value across signal categories. Two studies returned 218 discriminable facial expressions with 51 unique labels. The six basic emotions are included, but we additionally define such nuanced expressions as embarrassed, curious, and hopeful.
Conference Paper
We are investigating if an agent’s contingent smile during an interaction with a human participant affects the impression of the virtual character’s social status. Psychological studies show lower status individuals are more likely to show contingent feedback [4]. A “Rock-Paper-Scissors” game is used as the scenario to provide an infrastructure for an interaction between a 3D agent and a human participant. During the interaction we are using electromyographic measurements to determine when the human participant is smiling. Immediately after a detected smiling the virtual character mimics the smile. More specifically, we are expecting that participants form the impression that the character has a low social status when the agent shows contingent smiling behavior. We are currently performing the experiment and a next step is to evaluate the system by analyzing the results.
Article
Full-text available
This research investigates the meaning of “human-computer relationship” and presents techniques for constructing, maintaining, and evaluating such relationships, based on research in social psychology, sociolinguistics, communication and other social sciences. Contexts in which relationships are particularly important are described, together with specific benefits (like trust) and task outcomes (like improved learning) known to be associated with relationship quality. We especially consider the problem of designing for long-term interaction, and define relational agents as computational artifacts designed to establish and maintain long-term social-emotional relationships with their users. We construct the first such agent, and evaluate it in a controlled experiment with 101 users who were asked to interact daily with an exercise adoption system for a month. Compared to an equivalent task-oriented agent without any deliberate social-emotional or relationship-building skills, the relational agent was respected more, liked more, and trusted more, even after four weeks of interaction. Additionally, users expressed a significantly greater desire to continue working with the relational agent after the termination of the study. We conclude by discussing future directions for this research together with ethical and other ramifications of this work for HCI designers.
Article
Full-text available
This thesis has developed an affective agent research platform that advances the architecture of relational agents and intelligent tutoring systems. The system realizes non-invasive multimodal real-time sensing of elements of user's affective state and couples this ability with an agent capable of supporting learners by engaging in real-time responsive expressivity. The agent mirrors several non-verbal behaviors believed to influence persuasion, liking, and social rapport, and responds to frustration with empathetic or task-support dialogue. Pilot studies involved 60 participants, ages 10-14 years-old, and led to an experiment involving 76 participants, ages 11-13 years-old, engaging in the Towers of Hanoi activity. The system (data collection, architecture, character interaction, and activity presentation) was iteratively tested and refined, and two "mirroring" conditions were developed: "sensor driven non-verbal interactions" and "pre-recorded non-verbal interactions". The development and training of the classifier algorithms showed the ability to predict frustration/help seeking behavior with 79% accuracy across a pilot group of 24 participants.
Chapter
Research on emotion has undergone explosive growth during the past few decades, marked by new theories (e.g., evolutionary analyses; Tooby & Cosmides, 1990), methods (e.g., anatomically-based systems for coding facial expressive behavior, see Cohn this volume), and findings (see Cacioppo & Gardner, 1999). Some of the research in this area has been correlational, focusing on factors that naturally co-vary with emotional processes, such as chronological age, physical health, or social status. However, experimental research also has flourished, focusing on emotional processes in the context of relatively well-controlled laboratory environments. Our chapter on the use of emotion-eliciting films, like many of the contributions to the Handbook of Emotion Elicitation and Assessment, lies squarely within this second, experimental tradition.
Conference Paper
This study examined how learners' gender and ethnicity influenced their choice of pedagogical agents how they perceived the persona of the chosen agents. 183 undergraduates from two southeast universities participated in the study and were provided eight agents to choose from, each differing by gender (male, female), ethnicity (African American, Caucasian), and realism (realistic, cartoon). The results showed that African-American learners were significantly more likely to choose an agent with the same ethnicity and also to have positive attitudes toward the chosen agent after learning from it. Overall, the perceived agent demeanor was the most cited reason for why learners chose a particular agent. Female learners were more likely to choose a cartoon-like (as opposed to realistic) agent than male learners and were also more likely than males to choose an agent based on their previous experiences with human instructors.
Article
This study examined adolescents' perceptions of pedagogical caring in relation to their motivation to achieve positive social and academic outcomes in middle school. A longitudinal sample of 248 students was followed from 6th to 8th grade. Perceived caring from teachers predicted motivational outcomes, even when students' current levels of psychological distress and beliefs about personal control, as well as previous (6th grade) motivation and performance, were taken into account. Eighth-grade students characterize supportive and caring teachers along dimensions suggested by N. Noddings (1992) and models of effective parenting (D. Baumrind, 1971). Teachers who care were described as demonstrating democratic interaction styles, developing expectations for student behavior in light of individual differences, modeling a "caring" attitude toward their own work, and providing constructive feedback. The implications for understanding links between teacher behavior and student achievement are discussed. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
Conference Paper
Animated pedagogical agents that inhabit interactive learn- ing environments can exhibit strikingly lifelike behaviors. In addition to providing problem-solving advice in response to students' activities in the learning environment, these agents may also be able to play a powerful motivational role. To design the most effective agent-based learning environment software, it is essential to understand how students perceive an animated pedagogical agent with regard to affective dimen- sions such as encouragement, utility, credibility, and clarity. This paper describes a study of the affective impact of ani- mated pedagogical agents on students' learning experiences. One hundred middle school students interacted with animated pedagogical agents to assess their perception of agents' af- fective characteristics. The study revealed the persona effect, which is that the presence of a lifelike character in an interac- tive learning environment—even one that is not expressive— can have a strong positive effect on student's perception of their learning experience. The study also demonstrates the interesting effect of multiple types of explanatory behaviors on both affective perception and learning performance.
Article
Use of technology often has unpleasant side effects, which may include strong, negative emotional states that arise during interaction with computers. Frustration, confusion, anger, anxiety and similar emotional states can affect not only the interaction itself, but also productivity, learning, social relationships, and overall well-being. This paper suggests a new solution to this problem: designing human-computer interaction systems to actively support users in their ability to manage and recover from negative emotional states. An interactive affect-support agent was designed and built to test the proposed solution in a situation where users were feeling frustration. The agent's text-only interaction used components of active listening, empathy, and sympathy, in an effort to support users in their ability to recover from frustration. The agent's effectiveness was evaluated against two control conditions, which were also text-based interactions: (1) users' emotions were ignored, and (...