ArticlePDF Available


Narrative-centered learning environments offer significant promise for promoting interactive learning experiences that are both effective and engaging. Models of narrative generation and reasoning can balance the motivational and pedagogical aspects of narrative-centered learning interactions. Affect recognition and affect expression models are useful for shaping students' affective trajectories during narrative-centered learning. Conducting empirical evaluations are critical for determining what factors contribute to the potential pedagogical and motivational benefits of narrative-centered learning environments. This paper presents an overview of progress in these areas by summarizing work on CRYSTAL ISLAND, a narrative-centered learning environment for eighth grade microbiology.
CRYSTAL ISLAND: A Narrative-Centered
Learning Environment for Eighth Grade
Jonathan P. ROWE
, Bradford W. MOTT
Jennifer L. ROBISON
, Sunyoung LEE
, and James C. LESTER
Department of Computer Science, North Carolina State University, Raleigh NC 27695
Education Practice, SAS Institute, Inc., Cary NC 27513
Abstract. Narrative-centered learning environments offer significant promise for
promoting interactive learning experiences that are both effective and engaging.
Models of narrative generation and reasoning can balance the motivational and
pedagogical aspects of narrative-centered learning interactions. Affect recognition
and affect expression models are useful for shaping students’ affective trajectories
during narrative-centered learning. Conducting empirical evaluations are critical
for determining what factors contribute to the potential pedagogical and
motivational benefits of narrative-centered learning environments. This paper
presents an overview of progress in these areas by summarizing work on CRYSTAL
ISLAND, a narrative-centered learning environment for eighth grade microbiology.
Keywords. Narrative-centered learning environments, game-based learning
environments, affect recognition, affect expression, evaluation.
1. Introduction
Narrative-centered learning environments (NLEs) afford significant opportunities for
students to participate in motivating story-based educational experiences. By
combining commercial game technologies, intelligent tutoring systems, and rich
narrative structures, NLEs seek to provide effective, engaging learning experiences that
are tailored to individual students. NLEs show promise for encouraging problem
solving, strategic and analytical thinking, decision-making, and other 21
century skills
[1]. They also serve as a natural platform for adapting learning experiences to
individual students. Recent work on narrative-centered learning environments has
investigated pedagogical agents with rich models of dialogue [2, 3], affect [4, 5, 6] and
social behavior [7], director agents that can manipulate the pedagogical and narrative
directions of learning experiences [8], models for detecting students emotional and
motivational states [9], and predictive models of students’ goals [10] and other in-game
behaviors. NLEs are currently under investigation in a range of domains, including
language learning [2, 11], anti-bullying education [5], and science learning [12].
This paper presents work on several experimental versions of CRYSTAL ISLAND, a
narrative-centered learning environment developed for middle school students in the
Corresponding Author: Jonathan P. Rowe, Department of Computer Science, North Carolina State
University, Raleigh, NC 27695; E-mail:
domain of eighth-grade microbiology. Specifically, it summarizes several lines of
investigation that have explored how computational models of narrative and affect can
be leveraged in NLEs to create effective, engaging learning experiences. The structure
of the paper is as follows. Section 2 provides background and related work on
narrative-centered learning environments. Section 3 provides a detailed description of
the CRYSTAL ISLAND virtual environment. Following in Section 4 is a summary of
work on narrative generation and reasoning, affective recognition and expression, and
evaluation in CRYSTAL ISLAND. Section 5 provides concluding remarks and a brief
discussion of future work.
2. Background and Related Work
Narrative-centered learning environments offer significant potential for enhancing
students’ learning experiences. Stories draw audiences into plots and settings, thereby
opening perceptual, emotional, and motivational opportunities for learning.
Establishing concrete connections between narrative context and pedagogical subject
matter has been said to support the assimilation of new ideas in young learners [13].
Narratives can also facilitate students’ semantic encoding of new information and
making commitments to long-term memory in the form of episodic memories [14].
Furthermore, fantasy contexts in educational games have been shown to provide
motivational benefits for learning [15]. Although it is important to remain mindful of
potential disadvantages such as seductive details [16], a dynamically generated
narrative that draws students into the evolving plot has the potential to be
pedagogically compelling.
Narrative-centered learning environments leverage a range of techniques for
providing effective, engaging learning experiences. Multi-user virtual environments
such as Quest Atlantis [17] and River City [18] use rich narrative settings to
contextualize inquiry-based science learning scenarios with strong social and ethical
dimensions. Although the systems do not use artificial intelligence to provide tailored
narrative or learning experiences, several classroom studies have yielded promising
learning results. Other work on narrative-centered learning environments has applied a
range of techniques to generating engaging interactive narrative experiences that are
pedagogically effective and tailored to individual students. FearNot! uses affectively-
driven autonomous agents to generate dramatic, educational vignettes about bullying
[5]. In between the non-interactive vignettes, the virtual agent consults the student for
advice about prior bullying scenarios, and then uses this feedback to inform its
behavior in subsequent vignettes. The Thespian architecture takes a decision-theoretic,
multi-agent approach to controlling virtual characters in the Tactical Language and
Culture Training System’s narrative scenarios [11, 2]. The agents’ goal-driven
behaviors are trained using a corpus of linear, pre-authored scripts, providing the
agents with believable behavior models for conversing with students during language
and culture training scenarios [2]. The SASO (Stability and Support Operations)
narrative-centered learning environment uses robust, socially intelligent virtual humans
as actors in military training scenarios. SASO’s virtual humans implement models of
multimodal conversational behavior [19, 3], affective reasoning [4], and social
behavior [7].
Computational models of affect recognition and affect expression are also
important for effective narrative-centered learning environments. The AI in Education
community has seen the emergence of work on affective student modeling [20],
detecting frustration and stress [21, 22, 23], modeling agentsemotional states [24, 4],
devising affectively informed models of social interaction [25, 26, 27], detecting
student motivation [28], and diagnosing and adapting to student self-efficacy [29]. All
of this work seeks to increase the fidelity with which affective and motivational
processes are understood and utilized in intelligent tutoring systems in an effort to
increase the effectiveness of tutorial interactions and, ultimately, learning. Recent
work has also sought to characterize the affective experiences of learners interacting
with intelligent learning environments by considering student affective trajectories
during learning [30, 31, 32].
3. Crystal Island
CRYSTAL ISLAND (Figure 1) is a narrative-centered learning environment built on Valve
Software’s Source™ engine, the 3D game platform for Half-Life 2. CRYSTAL ISLAND
features a science mystery set on a recently discovered volcanic island. The curriculum
underlying CRYSTAL ISLAND’s science mystery is derived from the North Carolina state
standard course of study for eighth-grade microbiology. Students play the role of the
protagonist, Alyx, who is attempting to discover the identity and source of an infectious
disease plaguing a newly established research station. The story opens by introducing
the student to the island and members of the research team for which the protagonist’s
father serves as the lead scientist. Several of the team’s members have fallen gravely
ill, including Alyx’s father. Tensions have run high on the island, and prior to Alyx’s
arrival various team members began to accuse one another of having poisoned the sick
researchers. It is the student’s task to discover the outbreak’s cause and source, and
determine whether one of the team members is guilty of poisoning.
CRYSTAL ISLAND’s setting includes a beach area with docks, an outdoor field
laboratory, underground caves, and a research camp. Throughout the mystery, the
student is free to explore the world and interact with other characters while forming
questions, generating hypotheses, collecting data, and testing hypotheses. The student
can pick up and manipulate objects, take notes, view posters, operate lab equipment,
and talk with non-player characters to gather clues about the source of the disease.
During the course of solving the mystery, the student is minimally guided through a
five problem curriculum. The first two problems focus on pathogens, including viruses,
Figure 1. The CRYSTAL ISLAND narrative-centered learning environment.
bacteria, fungi, and parasites. The student gathers information by interacting with in-
game pathogen “experts” and viewing books and posters in the environment. In the
third problem, the student is asked to compare and contrast her knowledge of four types
of pathogens. In the fourth problem, the student is guided through an inquiry-based
hypothesis-test-and-retest problem. In this problem she must complete a “fact sheet”
with information pertaining to the disease afflicting members of the research team.
Once the “fact sheet” is completed and verified by the camp nurse, the student
completes the final problem concerning an appropriate treatment plan for the sickened
CRYSTAL ISLAND researchers.
To illustrate the behavior of CRYSTAL ISLAND, consider the following situation.
Suppose a student has been interacting with virtual agents in the storyworld and
learning about infectious diseases. In the course of having members of the research
team become ill, she has learned that an infectious disease is an illness that can be
transmitted from one organism to another. As she concludes her introduction to
infectious diseases, she learns from the camp nurse that the mystery illness seems to be
coming from food items the sick members recently ate. Some of the island’s characters
are able to help identify food items and symptoms that are relevant to the scenario,
while others are able to provide helpful microbiology information. The student
discovers through a series of tests that a container of unpasteurized milk in the dining
hall is contaminated with bacteria. By combining this information with her knowledge
about the characters’ symptoms, the student deduces that the disease is E. coli. The
student reports her findings back to the camp nurse, and they discuss a plan for
treatment. The E. coli diagnosis ultimately exonerates the members accused of
poisoning, and the sick researchers make a speedy recovery.
4. Current Progress
Narrative-centered learning environments are inherently complex systems, and multiple
lines of investigation are necessary to develop suites of technologies that can produce
adaptive, engaging learning experiences. To date three principal areas of research have
been conducted with the CRYSTAL ISLAND virtual environment: narrative generation
and reasoning, affect recognition and expression, and empirical evaluation of NLEs.
4.1. Narrative Generation and Reasoning
Because of their interactive nature, narrative-centered learning environments must cope
with a wide range of student actions that can be performed in a virtual environment.
Providing students with a strong sense of control and agency is important for
supporting motivation [33], but agency also introduces opportunities for students to
violate or ignore important aspects of an intended narrative experience. This presents a
major challenge for narrative-centered learning environments: to simultaneously
maintain the coherence and pedagogical effectiveness of a learning experience, but
permit significant user agency in the environment. Developing computational models
that can reason about students’ actions within the narrative, adapt and re-plan narrative
events in response to student actions, and promote robust, believable interactions with
virtual characters is critical.
Models of narrative generation for CRYSTAL ISLAND have taken a dual planning
space approach, with one planning space allocated to tutorial planning and a second
allocated to narrative planning [34]. A tutorial planner supports the requirements of
inquiry-based learning by formulating tutorial strategies that encourage question
formation, hypothesis generation, data collection, and hypothesis testing. A narrative
planner is responsible for generating plot elements, sequencing plot elements into
coherent and engaging stories, and directing characters’ actions and storyworld events
to achieve tutorial and narrative goals. However, for the two planners to work in
concert, they must effectively coordinate their actions, resulting in a single stream of
events occurring in the virtual storyworld. To this end, the tutorial planner posts goals
in the tutorial planning space that are achieved by operators in the narrative planning
space. This tutorial-driven model seeks to balance plot advancement and tutorial goal
achievement seamlessly by the built-in coordination of the two planning spaces via
lower-level tutorial constraints and the upper-level narrative goals [34].
The resulting plans are incorporated into a decision-theoretic director agent
architecture that manages the narrative-centered learning experience [8]. Decision-
theoretic narrative planning offers a unified approach to dynamically guiding narratives
in a storytelling environment. The director agent has access to three principal
knowledge sources: narrative objectives, storyworld state, and student state. To cope
with the uncertainty in narrative planning, the three sets of knowledge sources are
integrated into a dynamic decision network (DDN) that the director agent evaluates
regularly to select the next narrative action. Once the director agent has fully updated
the decision network, it selects the director action that maximizes the expected
narrative utility, waits to see what action the student takes (if any) and updates its
beliefs as necessary [8].
Student goal recognition is also important for narrative generation in game-based
learning environments. Providing narrative planners with the ability to recognize
students’ goals could enable planners to monitor students to determine if their goals
were consistent with the plot, and determine whether sufficient plot progress has been
made. Two families of goal recognition models have been investigated in CRYSTAL
ISLAND: n-gram models (unigrams and bigrams) and Bayesian network models [10].
The models, which exploit knowledge of narrative structure as well as locational
information about students’ activities in the world, are induced from training data
acquired from traces of students’ actions in the story environment. Experimental
results suggest that probabilistic models can accurately predict students’ goals, and that
they converge on correct interpretations as observations of a student’s activities
become available over time [10].
An additional feature of narrative-centered learning environments is their natural
support for encouraging rich, believable interactions with embodied pedagogical agents.
Character dialogue behavior is crucial for defining and advancing plots, as well as
scaffolding learning experiences. Character dialogue generators for interactive
narrative environments must meet several requirements. They must generate dialogues
that are appropriate for characters’ traits, such as personalities, motivations, and
preferences; they must consider narrative context and history as they formulate
dialogue; and they must be able to robustly handle the large number of possible
character-character and character-player interactions that may result in dialogue.
CRYSTAL ISLAND takes advantage of a probabilistic unification-based dialogue
generation architecture that considers multiple sources of information (character
archetypes, narrative context, and communicative goals) to dynamically generate
character- and situation-appropriate dialogue [35]. The generated dialogues use
preference information encoded within character archetype representations and yield
character-specific variations in the dialogue that satisfy the major objectives for
conversational interactions during narrative-centered learning.
4.2. Affect Recognition and Expression
Emotion is critical for both narrative and learning. Creating narrative-centered learning
environments that are in tune with students’ affective experiences can provide support
for guiding pedagogical scaffolding and engaging students in virtual story worlds.
Affective interactions proceed through a three-stage process termed the affective loop:
affect recognition, affect understanding, and affect expression (adapted from [36]).
Affect recognition is the task of inferring a user’s affective state from a sequence of
observations of behavior. Affect expression is the task of determining how a system
should communicate emotion. Affect understanding is the process of interpreting
recognized user emotions, determining what it means for the user to feel the recognized
emotion, and then formulating adaptation strategies based on how the user feels.
Collectively, affect recognition, expression, and understanding have been the subject of
growing attention in the AI in Education community. Work on CRYSTAL ISLAND has
primarily focused on data-driven approaches to affect recognition and affect expression,
and has yielded promising initial results.
Data-driven models of affect recognition are trained and validated using a rich
corpus of student actions, locations, goals, physiological information, and temporal
information collected during student interactions with the CRYSTAL ISLAND
environment. After students problem-solving traces have been recorded, affect
recognition models are induced using supervised machine learning techniques such as
naïve Bayes, decision trees, and support vector machines. This methodology has
yielded affect recognition models that are both accurate and efficient, with some
models capable of correctly predicting over 95% of students’ emotion self-reports [9].
The same methodology has also been successfully applied to induce models of student
self-efficacy in the CRYSTAL ISLAND environment [37]. Related work has investigated
students’ affective transitions in the CRYSTAL ISLAND environment, differentiating
typical affective transitions and those stemming from pedagogical agents’ empathetic
responses to student affect [38]. Also important is the ability to make “early”
predictions of student affect. Early detection allows systems adequate time to prepare
for particular affective states, opening a window of opportunity for the learning
environment to take corrective action. Inductive approaches using a combination of n-
gram models and decision trees have yielded results with accuracy, precision, and
recall exceeding 88%, which was significantly better than baseline comparisons [23].
Computational models of agents empathetic behavior are an important area of
investigation in affect expression. Defined as “the cognitive awareness of another
person’s internal states, that is, his thoughts, feelings, perceptions, and intentions” [39],
empathy enables people to vicariously respond to one another via “psychological
processes that make a person have feelings that are more congruent with another’s
situation than with his own situation” [40]. Initial work on constructing empathetic
virtual agents for CRYSTAL ISLAND explored learning empirically grounded models of
empathy from observations of human-human social interactions [6]. In this approach,
training data is first generated as a by-product of trainers’ interactions in a virtual
environment, and models of empathy are induced from the resulting datasets. Critically,
the training data include only features that can be directly observed in the environment,
so that at runtime, the same features can be used by the empathy models to drive the
behavior of virtual agents interacting with students. Two complementary lines of
evaluation, one investigating predictive accuracy and one investigating perceived
accuracy, were conducted on an implemented empathy modeler, and yielded promising
results. Follow-up work investigated more detailed models of empathetic agent
behavior, namely the use of parallel vs. reactive empathy [41]. In the parallel case, the
empathizer mimics the affective state of the target. In the reactive case, empathizers
exhibit a higher cognitive awareness of the situation and react with empathetic
behaviors that do not necessarily match those of the target’s affective state. The results
indicated that models of empathy induced from knowledge of the student’s situation
and the student’s affective state can effectively determine which type of empathy is
most appropriate for interactions requiring empathetic expression.
4.3. Empirical Evaluation
The third principal line of investigation in CRYSTAL ISLAND has been empirical
evaluation of narrative-centered learning environments. NLEs offer myriad
opportunities for enhancing learning experiences and motivating students. However,
determining appropriate metrics, criteria, and methodologies that can be used to assess
NLEs poses a number of challenges. Determining the impact of individual elements of
narrative-centered learning environments (setting, plot, game play activities,
empathetic pedagogical agents) also introduces a wide range of practical and
theoretical challenges. Further, because many narrative-centered learning
environments permit activities that are not strictly pedagogical, assessment of players
action traces and problem-solving paths can have implications for environment design
as well as models of adaptive scaffolding. The complexity inherent in intelligent
narrative-centered learning environments calls for a sophisticated, multi-faceted
approach to evaluation. Empirical evaluations of CRYSTAL ISLAND have sought to take
initial steps toward this objective.
A controlled, human participant experiment with middle school students
investigated the impact of narrative on learning [12]. The study compared two versions
of CRYSTAL ISLAND against a more traditional instructional approach, a narrated
slideshow that conveyed the same curricular material. The two CRYSTAL ISLAND
conditions featured varying levels of narrative content supplementing the curriculum.
The results showed that students in the NLE conditions did exhibit learning gains, but
that those gains were less than those produced by traditional instructional approaches.
However, the motivational benefits of narrative-centered learning, particularly with
regard to self-efficacy, presence, interest, and perception of control, were substantial.
Students reported the highest levels of presence in the full-narrative condition, a
finding that bears important implications for motivation. A recently completed follow-
up study with an updated version of the CRYSTAL ISLAND learning environment again
found that students in the NLE condition exhibited learning gains, and further, with the
updated version, the gains were on par with those of the slideshow condition.
Numerically, the learning gains in the CRYSTAL ISLAND condition exceeded those in
the slideshow condition, and analyses are underway to investigate these findings.
A different series of studies examined the impact of virtual charactersempathetic
behavior on student presence in narrative-centered learning environments [42]. In a
study with middle school students comparing non-empathetic and empathetic
characters, it was found that empathetic characters in narrative-centered learning
environments had a significant effect on measurements of students’ overall presence
(total PQ), involvement and control, naturalism of the experience, and resolution.
When the study was replicated with high school students, the same effects were found.
In short, it appears that empathetic interactions with characters in narrative-centered
learning environments can contribute to increased student presence. The results are
encouraging for the motivational potential of narrative-centered learning environments,
and they extend other results illustrating the relationship between narrative and
presence. The work points to a need for continued examination of the impact of
narrative content in virtual environments, as well as social and emotional interactions
that take place during those experiences [42].
Empirical evaluations of CRYSTAL ISLAND have also focused on the nature of
student behavior in narrative-centered learning environments. Work examining
students’ note-taking is one example. CRYSTAL ISLAND provides students with a note-
taking feature so that students can document useful information encountered during
learning interactions. A corpus of student notes was collected from a study involving
116 middle school students [43]. A team of judges annotated the corpus by classifying
individual notes into one of several categories including narrative, curricular, and
hypothesis notes. An analysis of the tagged corpus revealed that students who took
hypothesis notes performed better on posttests, confirming inquiry-based learning
findings suggesting the importance of scaffolding students’ hypothesis generation
activities. Individual differences were also able to suggest which students are likely to
take notes. Results illustrated significant gender effects on note-taking, where females
took significantly more notes than males. Goal orientation and efficacy for self-
regulated learning also exhibited significant correlations with note-taking behavior in
narrative-centered learning environments.
5. Conclusions and Future Work
Narrative is the subject of increasing attention in the AI in Education community as a
powerful medium for contextualizing learning. Narrative-centered learning
environments present a range of opportunities for investigating how different
computational models can be leveraged to create effective, engaging learning
experiences. Work on the CRYSTAL ISLAND environment has begun to illustrate how
models of narrative generation and reasoning, as well as affect recognition and
expression, can bear on game-based learning environments. A series of empirical
evaluations has begun to demonstrate the motivational and pedagogical potential of
narrative-centered learning.
Results to date suggest several promising directions for future work on narrative
generation and reasoning, affect recognition and expression, and empirical evaluation
in narrative-centered learning environments. Currently under investigation are
adaptive models for scaffolding students’ narrative and pedagogical progress through a
learning environment. Devising models that can integrate knowledge about the state of
a virtual environment, sets of intended narrative objectives, and individual student
qualities, poses serious challenges. Additionally, exploring computational models of
affect understanding will become increasingly important for closing the “affective
loop that exists in interactive learning environments. Finally, a critical step in this
research agenda will be conducting extensive empirical investigations to explore the
relationships between narrative, affect, character behavior, and their collective impact
on learning gains, self-regulated learning, and motivation.
The authors wish to thank members of the IntelliMedia lab for their assistance, Omer
Sturlovich and Pavel Turzo for use of their 3D model libraries, and Valve Software for
access to the Sourceengine and SDK. This research was supported by the National
Science Foundation under Grants REC-0632450, IIS-0757535, DRL-0822200 and IIS-
0812291. Portions of this material are based upon work supported under a National
Science Foundation Graduate Research Fellowship. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the authors and
do not necessarily reflect the views of the National Science Foundation.
[1] Partnership for 21st Century Skills. A Policymaker's Guide to 21st century Skills. Retrieved April 1,
2006, from
[2] M. Si, S. Marsella, D. Pynadath, Thespian: Modeling socially normative behavior in a decision-
theoretic framework, in Proceedings of the 6
International Conference on Intelligent Virtual Agents
(2006), 369382.
[3] D. Traum, S. Marsella, J. Gratch, J. Lee and A. Hartholt, Multi-party, multi-issue, multi-strategy
negotiation for multi-modal virtual agents, in Proceedings of the 8
International Conference on
Intelligent Virtual Agents (2008), 117–130.
[4] J. Gratch and S. Marsella, A domain-independent framework for modeling emotion, Journal of
Cognitive Systems Research, 5(4) (2004), 269-306.
[5] R. Aylett, S. Louchart, J. Dias, A. Paiva and M. Vala, FearNot! an experiment in emergent narrative, in
Proceedings of the 5
International Conference on Intelligent Virtual Agents (2005), 305-316.
[6] S. McQuiggan and J. Lester, Modeling and evaluating empathy in embodied companion agents,
International Journal of Human-Computer Studies 65(4) (2007), 348-360.
[7] J. Gratch, N. Wang, J. Gerten, E. Fast and R. Duffy, Creating rapport with virtual agents, in
Proceedings of the 7
International Conference on Intelligent Virtual Agents (2007), 223–236.
[8] B. Mott and J. Lester, U-Director: A Decision-theoretic narrative planning architecture for storytelling
environments, in Proceedings of the 5
International Conference on Autonomous Agents and
Multiagent Systems (2006), 977-984.
[9] S. Lee, S. McQuiggan and J. Lester, Inducing user affect recognition models for task-oriented
environments, in Proceedings of the 11
International Conference on User Modeling (2007), 380-384.
[10] B. Mott, S. Lee and J. Lester, Probabilistic goal recognition in interactive narrative environments, in
Proceedings of the 21
National Conference on Artificial Intelligence (2006), 187-192.
[11] W. L. Johnson, Serious use of a serious game for language learning, in Proceedings of the 13
International Conference on Artificial Intelligence in Education (2007), 67-74.
[12] S. McQuiggan, J. Rowe, S. Lee and J. Lester, Story-based learning: the impact of narrative on learning
experiences and outcomes, in Proceedings of the 9
International Conference on Intelligent Tutoring
Systems (2008), 530-539.
[13] C. Wells, The meaning makers: children learning language and using language to learn, Heinemann,
Portsmouth, NH, 1986.
[14] J. Ormrod, Human Learning. 4
Edition. Pearson Education, Upper Saddle River, NJ, 2004.
[15] L. Parker and M. Lepper, Effects of fantasy contexts on children’s learning and motivation: Making
learning more fun, Journal of Personality and Social Psychology 62(4) (1992), 625 – 633.
[16] S. Harp and R. Mayer, How seductive details do their damage: a theory of cognitive interest in science
learning, Journal of Educational Psychology 90(3) (1998), 414-434.
[17] S. Barab, T. Dodge, H. Tuzun, K. Job-Sluder, C. Jackson, A. Arici, L. Job-Sluder, R. Carteaux, J.
Gilbertson Jr. and C. Heiselt, The Quest Atlantis Project: A socially-responsive play space for learning,
in B. E. Shelton and D. Wiley (Eds.) The Educational Design and Use of Simulation Computer Games,
(pp. 159-186), Sense Publishers, Rotterdam, The Netherlands, 2007.
[18] D. J. Ketelhut, C. Dede, J. Clarke, B. Nelson and C. Bowman, Studying situated learning in a multi-user
virtual environment, in E. Baker, J. Dickieson, W. Wulfeck and H. O'Neil (Eds.) Assessment of problem
solving using simulations, (pp. 37-58), Lawrence Erlbaum, Mahwah, NJ, 2007.
[19] J. Lee and S. Marsella, Nonverbal behavior generator for embodied conversational agents, in
Proceedings of the 6
International Conference on Intelligent Virtual Agents (2006), 243-255.
[20] C. Conati, and H. Mclaren, Data-driven refinement of a probabilistic model of user affect, in
Proceedings of the 10th International Conference on User Modeling (2005), 40-49.
[21] H. Prendinger, and M. Ishizuka, The empathic companion: a character-based interface that addresses
users’ affective states, Applied Artificial Intelligence 19(2005), 267-285.
[22] W. Burleson, 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, PhD thesis, Massachusetts Institute of Technology (2006).
[23] S. McQuiggan, S. Lee, and J. Lester, Early prediction of student frustration, in Proceedings of the 2
International Conference on Affective Computing and Intelligent Interaction (2007), 380-384.
[24] E. André and M. Mueller, Learning affective behavior, in Proceedings of the 10
Conference on Human-Computer Interaction (2003), 512-516.
[25] W. L. Johnson and P. Rizzo, Politeness in tutoring dialogs: “run the factory, that’s what I’d do”, in
Proceedings of the 7th International Conference on Intelligent Tutoring Systems (2004), 67-76.
[26] K. Porayska-Pomsta and H. Pain, Providing cognitive and affective scaffolding through teaching
strategies, in Proceedings of the 7th International Conference on Intelligent Tutoring Systems (2004),
[27] A. Paiva, J. Dias, D. Sobral, R. Aylett, S. Woods, L. Hall and C. Zoll, Learning by feeling: evoking
empathy with synthetic characters, Applied Artificial Intelligence 19(2005), 235-266.
[28] A. de Vicente and H. Pain, Informing the detection of the students’ motivational state: an empirical
study, in Proceedings of the 6th International Conference on Intelligent Tutoring Systems (2002), 933-
[29] C. Beal and H. Lee, Creating a pedagogical model that uses student self reports of motivation and mood
to adapt ITS instruction, in Workshop on Motivation and Affect in Educational Software, in conjunction
with the 12
International Conference on Artificial Intelligence in Education (2005).
[30] S. Craig, A. Graesser, J. Sullins and B. Gholson, Affect and learning: an exploratory look into the role
of affect in learning with AutoTutor, Journal of Educational Media 29(3) (2004), 241-250.
[31] R. Baker, M. Rodrigo and U. Xolocotzin, The dynamics of affective transitions in simulation problem-
solving environments, in Proceedings the 2
International Conference on Affective Computing and
Intelligent Interactions (2007), 666-677.
[32] S. D’Mello, R. S. Taylor and A. Graesser, Monitoring affective trajectories during complex learning, in
Proceedings of the 29
Annual Meeting of the Cognitive Science Society (2007), 203-208.
[33] T. Malone and M. Lepper, Making learning fun: a taxonomy of intrinsic motivations for learning, in R.
Snow and M. Farr (Eds.) Aptitude, learning, and instruction: III. Conative and affective process
analyses, (pp. 223-253), Erlbaum, Hillsdale, NJ, 1987.
[34] B. Mott and J. Lester, Narrative-Centered Tutorial Planning for Inquiry-Based Learning Environments,
in Proceedings of the Eighth International Conference on Intelligent Tutoring Systems (2006), 675-684.
[35] J. Rowe, E. Ha and J. Lester, Archetype-Driven Character Dialogue Generation for Interactive
Narrative, in Proceedings of the Eighth International Conference on Intelligent Virtual Agents (2008),
[36] P. Sundstrom, A. Stahl and K. Hook, In situ informants exploring an emotional mobile messaging
system in their everyday practice, International Journal of Human-Computer Studies 65(4) (2007),
[37] S. McQuiggan, B. Mott and James Lester. Modeling Self-Efficacy in Intelligent Tutoring Systems: An
Inductive Approach. User Modeling and User-Adapted Interaction 18(1-2) (2008), 81-123.
[38] S. McQuiggan, J. Robison and J. Lester, Affective Transitions in Narrative-Centered Learning
Environments, in Proceedings of the Ninth International Conference on Intelligent Tutoring Systems
(2008), 490-499.
[39] W. Ickes, Empathic Accuracy, Guilford Press, New York, NY, 1997.
[40] M. Hoffman, Empathy and Moral Development: Implications for Caring and Justice, Cambridge
University Press, Cambridge, UK, 2000.
[41] S. McQuiggan, J. Robison, R. Phillips and J. Lester, Modeling Parallel and Reactive Empathy in
Virtual Agents: An Inductive Approach, in Proceedings of the Seventh International Joint Conference
on Autonomous Agents and Multi-Agent Systems (2008), 167-174.
[42] S. McQuiggan, J. Rowe and J. Lester, The effects of empathetic virtual characters on presence in
narrative-centered learning environments, in Proceedings of the ACM Conference on Human Factors in
Computing Systems (2008), 1511-1520.
[43] S. McQuiggan, J. Goth, E. Ha, J. Rowe and James Lester. Student Note-Taking in Narrative-Centered
Learning Environments: Individual Differences and Learning Outcomes, in Proceedings of the Ninth
International Conference on Intelligent Tutoring Systems (2008), 510-519.
... However, it is known that some students may not learn well from this relatively unstructured and self-directed form of interaction, because they lack the skills to assess their progression and success [30]. Research has shown that this challenge can be alleviated by AI-driven help functionalities that are adaptive to the student needs, i.e., detect and respond to the learners' difficulties [27,34,37,43], and there is increasing interest in investigating if/how Intelligent Pedagogical Agents (IPAs) can facilitate students' learning in OELEs [9,12,36,40], by offering this help in a more engaging and motivating manner. Designing such IPAs for OELEs has two main challenges. ...
... First, we investigate the use of IPAs in free-form GD, a learning activity which has gotten increasingly popular in teaching CT to younger students [1,5,15,26], but has not been examined before in IPA research. Second, our IPA is implemented in a real-world commercial OELE for remote learning, whereas previous work was limited to OELEs that are designed specifically for research purposes and evaluated in ad-hoc learning activities [9,12,36,40]. Thus, our work is a further step toward showing the value of IPA for OELEs that are actively used in real-world education remotely, a setting that has become increasingly widespread with the Covid pandemic. ...
... Moreno et al. [36] designed an IPA that provides adaptive support in the Design-A-Plant microworld meant to engage college students into science topics. Crystal Island [40], is a narrative-centered OELE that teaches microbiology concepts with animated IPAs as students freely navigate in the 3D environment. Our work extends these previous works by considering IPAs for a novel learning domain (free-form GD to foster CT). ...
... One game includes Crystal Island, a microbiology game-based learning environment where learners have to diagnose a disease infecting people on an island. The plot of the story is able to adapt to the student in response to plot goals and the student's current interactions and progression in the game (Rowe et al., 2009). For example, based on the information the student has received, such as clues or character dialogue, the AI can predict the student's goals and adapt the plot if needed (Dever et al., 2021). ...
Full-text available
New advancements in machine learning and AI can be used to augment student learning and teacher capabilities. Examples of AI approaches in education include generating personalized student recommendations, autograding essays, and improving educational resources. AI programs intended to improve education can be categorized informally into three groups: Guidance, Learning, and Teacher. These categories are general and not necessarily mutually exclusive, but provide a framework for organization and further development. This paper intends to look at the past approaches of AI to improve education and categorize them to help guide new development of AI applications in education. The potential benefits of AI-powered education is noteworthy as the current economy is based on higher education. AI can be used to speed up labor-intensive tasks and help close the knowledge gap. Additionally, this paper also looks at potential drawbacks, such as ethics concerns of using student data to power AI. By analyzing the past use of AI in education, this paper seeks to provide a grouping framework to improve understanding of the field and facilitate future development. Framework for organization and further development. This paper intends to look at the past approaches of AI to improve education and categorize them to help guide new development of AI applications in education. The potential benefits of AI-powered education is noteworthy as the current economy is based on higher education. AI can be used to speed up labor-intensive tasks and help close the knowledge gap. Additionally, this paper also looks at potential drawbacks, such as ethics concerns of using student data to power AI. By analyzing the past use of AI in education, this paper seeks to provide a grouping framework to improve understanding of the field and facilitate future development.
... Goal recognition is the task of inferring users' intended goals from observed sequences of actions. It was used in Crystal island (Rowe et al., 2009), a game-based learning environment where learners' goals are inferred through their responses to questions that the system asked narratively. Yannakakis and Hallam (2009) proposed a real-time adaptation model whose objective was to optimize user satisfaction. ...
Full-text available
Machine learning models are biased toward data seen during the training steps. The models will tend to give good results in classes where there are many examples and poor results in those with few examples. This problem generally occurs when the classes to predict are imbalanced and this is frequent in educational data where for example, there are skills that are very difficult or very easy to master. There will be less data on students that correctly answered questions related to difficult skills and who incorrectly answered those related to skills easy to master. In this paper, we tackled this problem by proposing a hybrid architecture combining Deep Neural Network architectures-especially Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN)-with expert knowledge for user modeling. The proposed solution uses attention mechanism to infuse expert knowledge into the Deep Neural Network. It has been tested in two contexts: knowledge tracing in an intelligent tutoring system (ITS) called Logic-Muse and prediction of socio-moral reasoning in a serious game called MorALERT. The proposed solution is compared to state-of-the-art machine learning solutions and experiments show that the resulting model can accurately predict the current student's knowledge state (in Logic-Muse) and thus enable an accurate personalization of the learning process. Other experiments show that the model can also be used to predict the level of socio-moral reasoning skills (in MorALERT). Our findings suggest the need for hybrid neural networks that integrate prior expert knowledge (especially when it is necessary to compensate for the strong dependency-of deep learning methods-on data size or the possible unbalanced datasets). Many domains can benefit from such an approach to building models that allow generalization even when there are small training data.
... This approach often uses Narrative generation to dynamically generate stories that can be tailored to the user's taste and ability, providing almost an infinite outcomes and possibilities. A multitude of research work have applied this approach of NL in various domain, including microbiology (Rowe et al., 2009), mathematics (Rodrigues et al., 2017), negotiation training (Kim Hill et al., 2009), and language learning (Lewis, 2010). ...
Full-text available
Career and technical education play a significant role in reducing high school and college dropouts as well providing necessary skills and opportunities to make suitable career decisions. The recent technological advances have benefited the education sector tremendously with the introduction of exciting innovations including virtual and augmented reality. The benefits of NL and game-based learning are well-established in the literature. However, their implementation has been limited to the education sector. In this research, the design and implementation of a Narrative Integrated Career Exploration (NICE) platform is discussed. The platform contains four playable tracks allowing students to explore careers in artificial intelligence, cybersecurity, internet of things, and electronics. The tracks are carefully designed with narrative problem-solving reflecting contemporary real-world challenges. To evaluate the perceived usefulness of the platform, a case study involving university students was performed. The results clearly reflect students’ interest in narrative and game-based career exploration approaches.
Research has shown that free-form Game-Design (GD) environments can be very effective in fostering Computational Thinking (CT) skills at a young age. However, some students can still need some guidance during the learning process due to the highly open-ended nature of these environments. Intelligent Pedagogical Agents (IPAs) can be used to provide personalized assistance in real-time to alleviate this challenge. This paper presents our results in evaluating such an agent deployed in a real-word free-form GD learning environment to foster CT in the early K-12 education, Unity-CT. We focus on the effect of repetition by comparing student behaviors between no intervention, 1-shot, and repeated intervention groups for two different errors that are known to be challenging in the online lessons of Unity-CT. Our findings showed that the agent was perceived very positively by the students and the repeated intervention showed promising results in terms of helping students make less errors and more correct behaviors, albeit only for one of the two target errors. Building from these results, we provide insights on how to provide IPA interventions in free-form GD environments.
Computational support for learning in the domain of esports has seen a great deal of attention in recent years as an effective means of helping players learn and reap the benefits of play. However, previous work has not examined the tools from a learning theory perspective to assess if learning is prompted and supported in the right place and time. As a first step towards addressing this gap, this paper presents the results of two studies: a review of existing computational tools, and an online survey of esports' players' learning needs supplemented with qualitative interviews. Using Zimmerman's Cyclical Phase Model of Self-Regulated Learning as a lens, we identify patterns in the types of support offered by existing tools and players' support interests during different learning phases. We identify 11 opportunities for future research and development to better support self-regulated learning in esports.
As one of the most desired skills for contemporary education and career, problem-solving is fundamental and critical in game-based learning research. However, students' implicit and self-controlled learning processes in games make it difficult to understand their problem-solving behaviors. Observational and qualitative methods, such as interviews and exams, fail to capture students' in-process difficulties. By integrating data mining techniques, this study explored students' problem-solving processes in a puzzle-based game. First, we applied the Continuous Hidden Markov Model to identify students' problem-solving phases and the transition probabilities between these phases. Second, we employed sequence mining techniques to investigate problem-solving patterns and strategies facilitating students' problem-solving processes. The results suggested that most students were stuck in a certain phase, with only a few able to transfer to systematic phases by applying efficient strategies. At the beginning of the puzzle, the most popular strategy was testing one dimension of the solution at each attempt. In contrast, the other two strategies (remove or add untested dimensions one by one) played pivotal roles in promoting transitions to higher problem-solving phases. The findings of this study shed light on when, how, and why students advanced their effective problem-solving processes. Using the Continuous Hidden Markov Model and sequence mining techniques, we provide considerable promise for uncovering students' problem-solving processes, which helps trigger future scaffolds and interventions to support students’ personalized learning in game-based learning environments.
Virtual humans are virtual characters within multimedia learning environments designed to aid the learning process. While there is a large variety of research examining how to design the physical appearance of the character or the teaching strategies it should embody, there is comparatively little work around the design of the narrative the virtual human uses to communicate with the learner. In this study, we examine the use of four different types of text structures to structure the character’s narrative in an instructional video: expository, refutation text, deep reasoning questions, and refutation text and deep reasoning questions; and their effects on learning, perceptions, attitudes, and emotions about genetically modified foods. Our results largely indicated no significant differences between the different communication strategies. We hypothesize that the length, dosage, and pacing of the intervention could explain why we did not see benefits from the different narrative structures that have been found in other contexts.
Full-text available
Affect has been the subject of increasing attention in cognitive accounts of learning. Many intelligent tutoring systems now seek to adapt pedagogy to student affective and motivational processes in an effort to increase the effectiveness of tutorial interaction and improve learning outcomes. To this end, recent work has begun to investigate the emotions experienced during learning in a variety of environments. In this paper we extend this line of research by investigating the affective transitions that occur throughout narrative-centered learning experiences. Further analysis differentiates the likelihood of affective transitions stemming from pedagogical agent empathetic responses to student affect.
Full-text available
Our project focuses on the design, implementation and evaluation of a ITS pedagogical model that considers student motivation, mood and cognitive processes in making instructional decisions in the domain of secondary school mathematics. Students complete integrated self report assessments of motivation and mood. Cognitive skillls such as math fact knowledge, spatial cognition, and prior math achievement are also assessed. The pedagogical model adapts instruction (problem selection, problem difficulty, topic area, choice of activity, choice of help type, and availability of help) following a model of human tutoring expertise that balances motivational and cognitive goals.
Full-text available
The following three chapters describe situations in which gaming environments have been taken into innovative subject areas, and studied through a variety of complex, triangular means. The first offers insight into the Quest Atlantis project and the implementation of multi-participant environments to help teach children social awareness and responsibility... These chapters provide insight into how studying the teaching and learning that takes place naturally within simulated realms can inform the effective design of educational games. The lessons learned lead us to recommendations in how we can design proper support mechanisms for the learning that takes place within these realms.
Full-text available
The role that affective states play in learning was investigated from the perspective of a constructivist learning framework. We observed six different affect states (frustration, boredom, flow, confusion, eureka and neutral) that potentially occur during the process of learning introductory computer literacy with AutoTutor, an intelligent tutoring system with tutorial dialogue in natural language. Observational analyses revealed significant relationships between learning and the affective states of boredom, flow and confusion. The positive correlation between confusion and learning is consistent with a model that assumes that cognitive disequilibrium is one precursor to deep learning. The findings that learning correlates negatively with boredom and positively with flow are consistent with predictions from Csikszentmihalyi's analysis of flow experiences.
Conference Paper
Full-text available
Affect has been the subject of increasing attention in cognitive accounts of learning. Many intelligent tutoring systems now seek to adapt pedagogy to student affective and motivational processes in an effort to increase the effectiveness of tutorial interaction and improve learning outcomes. To this end, recent work has begun to investigate the emotions experienced during learning in a variety of environments. In this paper we extend this line of research by investigating the affective transitions that occur throughout narrative-centered learning experiences. Further analysis differentiates the likelihood of affective transitions stemming from pedagogical agent empathetic responses to student affect.
In 4 experiments, students who read expository passages with seductive details (i.e., interesting but irrelevant adjuncts) recalled significantly fewer main ideas and generated significantly fewer problem-solving transfer solutions than those who read passages without seductive details. In Experiments 1, 2, and 3, revising the passage to include either highlighting of the main ideas, a statement of learning objectives, or signaling, respectively, did not reduce the seductive details effect. In Experiment 4, presenting the seductive details at the beginning of the passage exacerbated the seductive details effect, whereas presenting the seductive details at the end of the passage reduced the seductive details effect. The results suggest that seductive details interfere with learning by priming inappropriate schemas around which readers organize the material, rather than by distracting the reader or by disrupting the coherence of the passage.