MetaTutor is a multi-agent, adaptive hypermedia learning environment that trains and fosters high school and college students' use of self-regulatory processes in the context of learning about science topics such as human body systems. The purpose of the MetaTutor environment is to examine the effectiveness of pedagogical agents (PAs) as external regulatory agents used to detect, trace, model, and foster students' self-regulatory processes during science learning with multiple representations of information. The multi-agent system provides adaptive tutoring based on students' evolving conceptual understanding of the topic and their strategic use of cognitive and metacognitive processes. Each of the four agents is responsible for specific aspects of SRL, including task definition, planning, metacognitive processes, and learning strategies. Based on their specialized roles, each PA has been designed to detect a specific set of SRL processes. For example, Mary the Monitor is in charge of detecting when students deploy metacognitive processes and make metacognitive judgments such as expressing a judgment of learning (JOL; e.g., used in relation to judging one's understanding of the current content) and also in determining the valence (e.g., JOL - or JOL +) associated with the metacognitive judgment. The presupposition of an accurate detection method (by each agent) leads the agent to model the temporal dynamics associated with each SRL process, across all SRL processes, and how they relate to several learning outcomes such as declarative, procedural, and inferential knowledge and mental models of the science topic. This evolving model is then used to foster SRL and content understanding by providing several levels of scaffolding. Instructional scaffolding involves the coordination of other architectural modules of MetaTutor that coordinate and manage the dialogue system between agents and the learner. Instructional scaffolding is provided based on current research on human and computerized tutoring research (Chi et al., 2004; Graesser, D'Mello & Person, in press; VanLehn et al., 2007; Wolff, 2009) and recent studies comparing SRL with ERL (externally-regulated learning; Azevedo et al., 2007, 2008). The types of scaffolding range from having the learner vicariously watch as the agent models the SRL process to having the student use a specific SRL process while being provided with elaborate feedback regarding the effective use of the process (based on Zimmerman & Moylan, in press).

Learning about complex and challenging science topics with advanced learning technologies requires students to regulate their learning. The deployment of key cognitive and metacognitive regulatory processes is key to enhancing learning in open-ended learning environments such as hypermedia. In this paper, we propose a metaphor—Computers as MetaCognitive tools—to characterize the complex nature of the learning context, self- regulatory processes, task conditions, and features of advanced learning technologies. We briefly outline the theoretical and conceptual assumptions of self-regulated learning (SRL) underlying MetaTutor, a hypermedia environment designed to train and foster students’ SRL processes in biology. Lastly, we provide preliminary learning outcome and SRL process data on the deployment of SRL processes during learning with MetaTutor.

We report preliminary data of an initial laboratory study examining the effectiveness of self-regulated learning (SRL) training versus no training on learners' ability to deploy SRL processes and learn about the circulatory system with MetaTutor. MetaTutor is an intelligent tutoring system (ITS) designed to train and foster learners' SRL processes while learning about several complex human body systems. We used a mixed methodology approach and include the results of a subset of the participants (N=30) whose product and process data we have analyzed. Overall, the results indicate that the SRL training group significantly outperformed the control group.

This study investigated the navigation patterns of 56 learners within a hypermedia learning environment, MetaTutor. Using K-Means cluster analysis, four types of navigational profiles were created. One cluster including participants who navigated primarily linearly through the learning environment, one showed high levels of non-linear progression, one included participants who opened the images very frequently, and the last cluster had a balance of all the navigational variables included in the analysis. Data from the learning measures used indicated that the balanced cluster and the cluster which opened the image frequently had the highest learning outcomes. Implications for the design of adaptive hypermedia learning systems are discussed.

This study explored learners’ utilization of multiple representations of information within a hypermedia learning environment during self-regulated and externally-regulated learning episodes. 135 middle school and high school participants were randomly assigned to either a self-regulated learning (SRL) condition, in which the learners attempted to use a hypermedia environment to learn about the circulatory system alone, or an externally-regulated learning (ERL) condition, in which learners attempted the same learning task, with access to a human tutor who facilitated their learning by administering prompts to engage in several adaptive selfregulated learning processes. Results indicate that learners in the ERL condition spent significantly more time constructing their own external representations of information (i.e. taking notes or drawing), and significantly less time reading text only, reading content with text and diagram, and watching the animation included in the hypermedia environment. In addition, correlations between the learning measures and time spent on these different types of representations indicate the learners who spent more time constructing external representations gained more from pretest to posttest, and those who spent more time in the remaining three types of representations gained less.

Self-regulated learning (SRL) involves a complex set of interactions between cognitive, metacognitive, motivational and affective processes. The key to understanding the influence of these self-regulatory processes on learning with open-ended, non-linear learning computer-based environments involves detecting, capturing, identifying, and classifying these processes as they temporally unfold during learning. Understanding the complex nature of the processes is key to building intelligent learning environments that adapt to learners’ fluctuations in their SRL processes and emerging understanding of the topic of domain. The foci of this paper are to: (1) introduce the complexity of SRL with hypermedia, (2) briefly present an information processing theory (IPT) of SRL and using it to analyze the temporally, unfolding sequences of processes during learning, (3) present and describe sample data to illustrate the nature and complexity of these processes, and (4) present challenges for future research that combine several techniques and methods to design intelligent learning environments that trace, model, and foster SRL. Self-Regulated Learning Learning about conceptually-rich domains with openended computer-based learning environments (CBLEs) such as hypermedia involves a complex set of interactions between cognitive, metacognitive, motivational, and affective processes (Azevedo, 2005, 2007, in press; Graesser, McNamara, & VanLehn, 2005; Jacobson, 2008; Moos & Azevedo, in press a; Vollmeyer & Rheinberg, 2006; Zimmerman, 2008). Current research from several fields including cognitive and learning sciences provides evidence that learners of all ages struggle when learning about these conceptually-rich domains with hypermedia. This research indicates that learning about conceptuallyrich domains with hypermedia is particularly difficult because it requires students to regulate their learning. Regulating one’s learning involves analyzing the learning context, setting and managing meaningful learning goals, determining which learning strategies to use, assessing whether the strategies are effective in meeting the learning goals, evaluating emerging understanding of the topic, and determining whether there are aspects of the learning context which could be used to facilitate learning. During self-regulated learning, students need to deploy several metacognitive processes to determine whether they understand what they are learning, and perhaps modify their plans, goals, strategies, and effort in relation to dynamically changing contextual conditions. In addition, students must also monitor, modify, and adapt to fluctuations in their motivational and affective states, and determine how much social support (if any) may be needed to perform the task. Also, depending on the learning context, instructional goals, perceived task performance, and progress made towards achieving the learning goal(s), they may need to adaptively modify certain aspects of their cognition, metacognition, motivation, and affect. Despite the ubiquity of hypermedia environments for learning, the majority of the research has been criticized as atheoretical and lacking rigorous empirical evidence (see Azevedo & Jacobson, 2008; Dillon & Jobst, 2005; Dillon & Gabbard, 1998; Jacobson, 2008; Jacobson & Azevedo, 2008; Neiderhauser, 2008; Tergan, 1997a, 1997b). In order to advance the field and our understanding of the complex nature of learning with hypermedia environments, we need theoretically-guided, empirical evidence regarding how students regulate their learning with these environments. Despite the recent rise of quality research on learning with hypermedia, we also raise several critical issues related to learning with hypermedia environments which have not yet been addressed by cognitive and educational researchers. For example, there is the question of how (i.e., with what processes) a learner regulates his/her learning with a hypermedia environment. Most of the research have used the product(s) of learning (i.e., pretest-posttest learning gains) to infer the connection between individual differences (e.g., prior knowledge, reading ability), learner characteristics (e.g., developmental level), cognitive processes (e.g., learning strategies used during learning), and structure of the hypermedia environment or the inclusion (or exclusion) of certain system features. In our research, we have adopted self-regulated learning (SRL) because it allows us to directly investigate how task demands, learner characteristics, cognitive and metacognitive processes, and system structure interact during the cyclical and iterative phases of planning, monitoring, and control while learning with hypermedia environments. In this paper, we provide a synthesis of existing research on learning with hypermedia, illustrate the complexity of self-regulatory processes during hypermedia learning, and present a theoretical account of SRL based on Winne and colleagues’ (1995, 1998, 2001, 2008) information processing theory (IPT) of SRL. We also provide a synthesis of recent research on using SRL as a framework with which to capture and study self-regulatory processes, and provide theoretically-driven and empirically-based guidelines for supporting learners’ self-regulated learning with hypermedia. Self-Regulated Learning with Hypermedia Learning about complex and challenging science topics, such as the human circulatory system and natural ecological processes with multi-representational, nonlinear computer-based learning environments (CBLEs), requires learners to deploy key self-regulatory processes (Azevedo et al., 2004a, 2004b, 2005, 2007, 2008; Biswas et al., 2005; Graesser et al., 2005; Jacobson, 2008; McNamara & Shapiro, 2005; Neiderhauser, 2008). Recent cognitive research with middle-school, high-school, and college students has identified several key self-regulatory processes which are associated with learning, understanding, and problem solving with hypermediabased CBLEs. First, there are planning processes such as activating prior knowledge, setting and coordinating subgoals that pertain to accessing new information, and defining which problem solution steps to perform for accomplishing a complex task. In addition, there are also several monitoring processes that are deployed during task enactment including monitoring one’s understanding of the topic, managing the learning environment and other instructional resources necessary to accomplish the learning goals, and engaging in periodic self-assessment (i.e., checking for the correctness of solution steps while solving problems and using this information to direct one’s future learning activities). During task performance a learner must also use several effective learning strategies for accomplishing the task such as coordinating several informational sources (e.g., text, diagram, animations), generating hypotheses, extracting relevant information from the resources, re-reading, making inferences, summarizing, and re-representing the topic based on one’s emerging understanding by taking notes and drawing. Lastly, the learner must continuously adjust during learning by handling task difficulties and demands such as monitoring one’s progress towards goals, and modifying the amount of time and effort necessary to complete the learning task. As such, we (Witherspoon, Azevedo, & D’Mello, 2008) and other colleagues (e.g., Biswas et al., 2005; Hadwin, Nesbit, Jamieson-Noel, Code, & Winne, 2007) emphasize that understanding the real-time deployment of these processes in the context of learning and problem solving tasks is key to understanding the nature of adaptivity in self-regulated learning (SRL) among learners of all ages and their influence on learning. Therefore, we propose that an IPT theory of SRL will best accommodate the complex nature of learning with hypermedia environments. Theoretical Framework: Information-Processing Theory of SRL Self-regulated learning (SRL) involves actively constructing an understanding of a topic/domain by using strategies and goals, regulating and monitoring certain aspects of cognition, behavior, and motivation, and modifying behavior to achieve a desired goal (see Boekaerts et al., 2000; Pintrich, 2000; Zimmerman & Schunk, 2001). Though this definition of SRL is commonly used, the field of SRL consists of various theoretical perspectives that make different assumptions and focus on different constructs, processes, and phases (Dunlosky & Lipko, 2007; Metcalfe & Dunlosky, 2008; Pintrich et al., 2000; Schunk, 2008; Winne & Hadwin, 2008; Zimmerman, 2008). We further specify SRL as a concept superordinate to metacognition that incorporates both metacognitive monitoring (i.e., knowledge of cognition or metacognitive knowledge) and metacognitive control (involving the skills associated with the regulation of metacognition), as well as processes related to planning for future activities within a learning episode and manipulating contextual conditions as necessary. SRL is based on the assumption that learners exercise agency by consciously monitoring and intervening in their learning. While most of our research has focused on the cognitive and metacognitive aspects of SRL with hypermedia, we are currently considering the incorporation of other key processes such motivation and affect (Moos & Azevedo, 2008, in press a, in press b). Recent studies on SRL with open-ended learning environments such as hypermedia (e.g., see Azevedo, 2005, in press, for recent reviews) have drawn on Winne and colleagues’ (Butler & Winne, 1995; Winne, 2001; Winne & Hadwin, 1998, 2008) Information Processing Theory (IPT) of SRL. This IPT theory suggests a fourphase model of self-regulated learning. The goal of this section is to explicate the basics of the model so as to emphasize the linear, recursive, and adaptive nature of selfregulated learning and make the link to its implication for the use in studying SRL with hypermed

This paper tested the deep-level reasoning questions effect in the domains of computer literacy between eighth and tenth graders and Newtonian physics for ninth and eleventh graders. This effect claims that learning is facilitated when the materials are organized around questions that invite deep-reasoning. The literature indicates that vicarious learners in college student populations show greater pretest to posttest learning gains when presented with deep-level reasoning questions before each content sentence, than when deep-level questions are omitted, or when learners interact with an intelligent tutoring system. This effect holds for vicarious learners across grade levels and domains.

In an attempt to discover the facial action units for affective states that occur during complex learning, this study adopted an emote-aloud procedure in which participants were recorded as they verbalised their affective states while interacting with an intelligent tutoring system (AutoTutor). Participants' facial expressions were coded by two expert raters using Ekman's Facial Action Coding System and analysed using association rule mining techniques. The two expert raters received an overall kappa that ranged between .76 and .84. The association rule mining analysis uncovered facial actions associated with confusion, frustration, and boredom. We discuss these rules and the prospects of enhancing AutoTutor with non-intrusive affect-sensitive capabilities.

This paper examines the dynamics of college students' self-regulatory processes within self-regulated learning (SRL) and externally-regulated learning (ERL) episodes during hypermedia learning. We re-analyzed and extended the results from an original study recently conducted by Azevedo and colleagues [1] to address four questions related to adaptivity, based on the temporal and dynamic deployment of self-regulatory processes by learners and human tutors in fostering complex science learning with hypermedia. Our questions include: (1) How does access to a human tutor affect the deployment of various SRL processes during learning?; (2) Which transitions between self-regulatory processes are more likely to occur within SRL and ERL?; (3) Which transitions between SRL classes are more likely or less likely to occur with SRL and ERL learning episodes?; (4) Are there significant correlations between learners' observed likelihood of transitions (between SRL processes) and learning outcomes? Lastly, we discuss implications for the design of MetaTutor, an adaptive hypermedia learning environment.

This study examines how passage availability and reading comprehension question format (open-ended vs. multiple-choice) influence question answering. In two experiments, college undergraduates read an expository passage and answered open-ended and multiple-choice versions of text-based, local, and global bridging inference questions. Half the participants were allowed to refer to the passage when answering the questions and half were not. Participants' prior domain knowledge relating to the text contents was assessed using multiple-choice and open-ended questions. Correlation-based analyses in the two experiments indicated: (a) a decline in the relationship between prior domain knowledge and comprehension when the passage was available during question answering; and (b) a high correlation between multiple-choice and open-ended question answering performance when the passage was not available for reference. Overall the results indicate that the nature of the reading comprehension assessment is influenced by the specific task with which comprehension is assessed.

Eighty-two (N = 82) college students with little knowledge of the circulatory system were randomly assigned either to the control condition (SRL; self-regulated learning) or human tutoring (ERL; externally-regulated learning) condition. Learners in the SRL condition regulated their own learning, while learners in the ERL condition had access to a human tutor who facilitated their self-regulated learning. All learners were given 40 minutes to learn about the circulatory system. We collected several pretest and posttest learning measures and collected think-aloud protocols during learning. Generally, the learners in the ERL condition significantly outperformed the learners in SRL condition. In comparison to the SRL condition, results indicate that learners in the ERL condition deployed significantly more SRL processes related to planning, monitoring, and handling task difficulties. Each of the classes of SRL processes was predictive of learners' performance on different posttest measures.

This study empirically examines the temporal nature of students' self-regulatory behavior while learning about a complex science topic using hypermedia. The experiment involved randomly assigning 74 undergraduate students to one of two tutoring conditions: self-regulated learning (SRL) or externally-regulated learning (ERL). Participants in the self-regulated learning condition used hypermedia environment to learn about the circulatory system on their own, while participants in the externally-regulated learning condition also used the hypermedia environment, but were given prompts and feedback from a human tutor during the session to facilitate their self-regulatory behavior. Results from product data (mental model pretest-posttest shifts) indicate that the ERL condition leads to a greater likelihood of having a high-level mental model at posttest. In addition, results from the process (think-aloud) data indicate that having access to a human tutor during learning affects the deployment of certain SRL classes at different times within a learning session. Implications for the design of a computer-based learning environment which is intended to stimulate learners' effective use of self-regulated learning are discussed.

We explored the reliability of detecting a learner's affect from conversational features extracted from interactions with AutoTutor, an intelligent tutoring system (ITS) that helps students learn by holding a conversation in natural language. Training data were collected in a learning session with AutoTutor, after which the affective states of the learner were rated by the learner, a peer, and two trained judges. Inter-rater reliability scores indicated that the classifications of the trained judges were more reliable than the novice judges. Seven data sets that temporally integrated the affective judgments with the dialogue features of each learner were constructed. The first four datasets corresponded to the judgments of the learner, a peer, and two trained judges, while the remaining three data sets combined judgments of two or more raters. Multiple regression analyses confirmed the hypothesis that dialogue features could significantly predict the affective states of boredom, confusion, flow, and frustration. Machine learning experiments indicated that standard classifiers were moderately successful in discriminating the affective states of boredom, confusion, flow, frustration, and neutral, yielding a peak accuracy of 42% with neutral ( chance = 20%) and 54% without neutral ( chance = 25%). Individual detections of boredom, confusion, flow, and frustration, when contrasted with neutral affect, had maximum accuracies of 69, 68, 71, and 78%, respectively ( chance = 50%). The classifiers that operated on the emotion judgments of the trained judges and combined models outperformed those based on judgments of the novices (i.e., the self and peer). Follow-up classification analyses that assessed the degree to which machine-generated affect labels correlated with affect judgments provided by humans revealed that human-machine agreement was on par with novice judges (self and peer) but quantitatively lower than trained judges. We discuss the prospects of extending AutoTutor into an affect-sensing ITS.

We investigated the impact of dialogue and deep-level-reasoning questions on vicarious learning in 2 studies with undergraduates. In Experiment 1, participants learned material by interacting with AutoTutor or by viewing I of 4 vicarious learning conditions: a noninteractive recorded version of the AutoTutor dialogues, a dialogue with a deep-level-reasoning question preceding each sentence, a dialogue with a deep-level-reasoning question preceding half of the sentences, or a monologue. Learners in the condition where a deep-level-reasoning question preceded each sentence significantly outperformed those in the other 4 conditions. Experiment 2 included the same interactive and noninteractive recorded condition, along with 2 vicarious learning conditions involving deep-level-reasoning questions. Both deep-level-reasoning-question conditions significantly outperformed the other conditions. These findings provide evidence that deep-level-reasoning questions improve vicarious learning.

The relationship between emotions and learning was investigated by tracking the affective states that college students experienced while interacting with AutoTutor, an intelligent tutoring system with conversational dialogue. An emotionally responsive tutor would presumably facilitate learning, but this would only occur if learner emotions can be accurately identified. After a learning session with AutoTutor, the affective states of the learner were classified by the learner, a peer, and judges trained on Ekman's Facial Action Coding system. The classification of the trained judges was more reliable and matched the learners much better than the low scores of untrained peers. This result suggests that peer tutors may be limited in detecting the affective states of peer learners. Classification accuracy was poor at constant intervals of polling (every 20 seconds) but much higher when individuals declared that an affect state had been experienced.

Detection of Emotions during Learning with AutoTutor Art Graesser (a-graesser@memphis.edu) Amy Witherspoon (awthrspn@mail.psyc.memphis.edu) Department of Psychology, 202 Psychology Building Memphis. TN 38152 USA Department of Psychology, 202 Psychology Building Memphis. TN 38152 USA Bethany McDaniel (btmcdanl@memphis.edu) Department of Psychology, 202 Psychology Building Memphis. TN 38152 USA Sidney D’Mello (sdmello@memphis.edu) Institute for Intelligent Systems, 365 Innovation Drive Memphis, TN 38152 USA Patrick Chipman (pchipman@memphis.edu) Department of Psychology, 202 Psychology Building Memphis. TN 38152 USA Barry Gholson (b.gholson@memphis.edu) Department of Psychology, 202 Psychology Building Memphis. TN 38152 USA between cognition and emotions, they do not directly explain and predict the sort of emotions that occur during complex learning, such as attempts to master physics, biology, or computer literacy. Researchers in many different fields are familiar with Ekman’s work on the detection of emotions from facial expressions (Ekman & Friesen, 1978). However, the emotions that Ekman intensely investigated (e.g., sadness, happiness, anger, fear, disgust, surprise) have minimal relevance to learning per se (Kort et al., 2001). The pervasive affective states during complex learning include confusion, frustration, boredom, flow/engagement, interest, and being stuck (Craig, Graesser, Sullins, & Gholson, 2004; Csikszentmihalyi, 1990). There are a number of ways in which tutors (and other types of learning environments) might adaptively respond to the learner’s affective states in the course of enhancing learning (D’Mello, Craig, Sullins, & Graesser, in press; Graesser et al., 2005; Lepper & Woolverton, 2002). If the learner is frustrated, for example, the tutor can give hints to advance the learner in constructing knowledge or can make supportive empathetic comments to enhance motivation. If the learner is bored, the tutor needs to present more engaging or challenging problems for the learner to work on. The tutor would probably want to lay low and stay out the learner’s way when the learner is in a state of flow (Csikszentmihaly, 1990), i.e., when the learner is so deeply engaged in learning the material that time and fatigue disappear. The flow experience is believed to occur when the learning rate is high and the learner has achieved a high level of mastery at the region of proximal learning (Metcalfe & Kornell, 2005). The affective state of confusion is particularly interesting because it is believed to play an important role in learning (Graesser et al., 2005; Guhe et al., 2004) and has a large correlation with learning gains (Craig et al., 2004). Confusion is diagnostic of cognitive disequilibrium, a state that occurs when learners face obstacles to goals, contradictions, incongruities, anomalies, uncertainty, and salient contrasts (Festinger, 1957; Graesser, Lu, Olde, Pye- Cooper, & Whitten, 2005; Graesser & Olde, 2003; Piaget, Abstract The relationship between emotions and learning was investigated by tracking the affective states that college students experienced while interacting with AutoTutor, an intelligent tutoring system with conversational dialogue. An emotionally responsive tutor would presumably facilitate learning, but this would only occur if learner emotions can be accurately identified. After a learning session with AutoTutor, the affective states of the learner were classified by the learner, a peer, and judges trained on Ekman’s Facial Action Coding system. The classification of the trained judges was more reliable and matched the learners much better than the low scores of untrained peers. This result suggests that peer tutors may be limited in detecting the affective states of peer learners. Classification accuracy was poor at constant intervals of polling (every 20 seconds) but much higher when individuals declared that an affect state had been experienced. Keywords: Emotion; Instruction and teaching; Human computer interaction; AutoTutor; Affective states Introduction Connections between complex learning and emotions have received increasing attention in the fields of psychology (Carver, 2004; Deci & Ryan, 2002; Dweck, 2002), education (Lepper & Henderlong, 2000; Linnenbrink & Pintrich, 2004; Meyer & Turner, 2002), neuroscience (Damasio, 2003), and computer science (Kort, Reilly, & Picard, 2001; Picard, 1997). A deep understanding of such affect-learning connections is needed in order to design engaging educational artifacts that range from responsive intelligent tutoring systems on technical material (DeVicente & Pain, 2002; Graesser, Person, Lu, Jeon, & McDaniel, 2005; Guhe, Gray, Schoelles, & Ji, 2004; Litman & Silliman, 2004) to entertaining media and games (Conati, 2002; Gee, 2003; Vorderer, 2003). There have been several theories that link cognition and affect very generally (Bower, 1981; Mandler, 1984; Ortony, Clore, & Collins, 1988; Russell, 2003; Stein & Levine, 1991). While these theories convey general links