Intelligent Tutoring Systems

Combination of Virtual Reality and Artificial Intelligence technologies offer very interesting possibilities for educational purposes, allowing to design creative, intelligent and dynamic 3D virtual learning environments. However, nowadays there are few programming environments and tools that support Artificial Intelligence and agent programming techniques to control virtual 3D avatars. Aiming to help in this question, this work introduces a logical programming environment, which extends Prolog with BDI and multi-agent programming concepts and is fully integrated with Virtual Reality technology. The paper shows how this programming environment was used to create an interactive, animated and intelligent virtual world, focused on teaching the beginnings of Industrial Evolution. This educational virtual world was positively evaluated through experiments carried out with simulated classes of History.

Despite the importance of introductory programming disciplines, it is quite common to find problems related to academic students performance. In such environments, we easily find unmotivated students with some doubts and that do not understand basic programming concepts. Monitoring each of the students is not trivial because the number of students is high and, to do so, it would be necessary to observe many characteristics of each code submitted for practical activities. The teacher, even when helped by TAs (Teacher Assistants), is not able to perform the reviews quickly, for this activity requires a huge amount of time. Fast feedback is extremely important to enable the learning of any concept. In this research, we investigate an adaptive approach to cluster codes in order to minimize the effort of evaluation. The results vary from reasonable to perfect concordances, considering the semiautomatic evaluations obtained with the clustering and the expert evaluations.

Social tagging activities allow the wide set of web users, especially learners, to add free annotations on educational resources to express their interests and automatically generate folksonomies. Folksonomies have been involved in a lot of recommendations approaches. Recently, supported by semantic web technologies, the Linked Open Data (LOD) allow to set up links between entities in the web to join information in a single global data space. This paper demonstrates how structured content accessible via LOD can be leveraged to support educational resources recommender in folksonomies and overcome the limited capabilities to analyze resources information. Another limitation of resources recommendation is the content overspecialization conducting in the incapacity to recommend relevant resources diverse from the ones that learner previously knows. To address these issues, we proposed to take advantage of the richness of the open and linked data graph of DBpedia and Ant Colony Optimization (ACO) to learn users’ behavior. The basic idea is to iteratively explore the RDF data graph to produce relevant and diverse recommendations as an alternative of going through the tedious phase of calculating similarity to attain the same goal. Using ant colony optimization, our system performs a search for the appropriate paths in the LOD graph and selects the best neighbors of an active learner to provide improved recommendations. In this paper, we show that ACO also in the problem of recommendation of novel diverse educational resources by exploring LOD is able to deliver good solutions.

Adapting ITSs that promote the use of metacognitive strategies can sometimes lead to intense prompting, at least initially, to the point that there is a risk of it feeling counterproductive. In this paper, we examine the impact of different prompting strategies on self-reported agent-directed emotions in an ITS that scaffolds students’ use of self-regulated learning (SRL) strategies, taking into account students’ prior knowledge. Results indicate that more intense initial prompting can indeed lead to increased frustration, and sometimes boredom even toward pedagogical agents that are perceived as competent. When considering prior knowledge, results also show that this strategy induces a significantly different higher level of confusion in low prior knowledge students when compared to high prior knowledge students. This result is consistent with the fact that higher prior knowledge students tend to be better at self-regulating their learning, and it could also indicate that some low prior knowledge students may be on their path to a better understanding of the value of SRL.

Cognitive, affective, metacognitive, and motivational (CAMM) processes are critical components of self-regulated learning (SRL) essential for learning and problem solving. Currently, ITSs are designed to foster cognitive, affective, and metacognitive (CAM) strategies and processes, presenting major gaps in the research since motivation is a key component of SRL and influences the remaining CAM processes. In our study, students interacted with MetaTutor, a hypermedia-based ITS, to investigate how 190 undergraduate students’ proportional learning gain (PLG) related to sub-goals set, cognitive strategy use and metacognitive processes differed based on self-reported achievement goal orientation. Results indicated differences between approach, avoidance, and students who adopted both approach and avoidance goal orientations, but no differences between mastery, performance and students who adopted both mastery and performance goal orientations on PLG for content related to sub-goal 1. Conversely, no differences were found between goal orientation groups on PLG for sub-goal 2, revealing possible changes in goal orientation following sub-goal 1. Analyses indicated no differences between goal orientation groups on metacognitive processes and cognitive strategy use. Thus, we suggest turning away from self-report data, where future studies aim to incorporate multi-channel data over durations of tasks as students interact with ITSs to measure motivation and its tendency to fluctuate in real-time. Implications for using multiple data channels to measure motivation could contribute to adaptive ITS design based on all CAMM processes.

In this paper, we present a solution for computer assisted emotional analysis of game session. The proposed approach combines eye movements and facial expressions to annotate the perceived game objects with the expressed dominate emotions. Moreover, our system EMOGRAPH (Emotional Graph) gives easy access to information about user experience and predicts player’s emotions. The prediction mainly uses both subjective measures through questionnaire and objective measures through brain wave activity (electroencephalography - EEG) combined with eye tracking data. EMOGRAPH’s method was experimented on 21 participants playing horror game “Outlast”. Our results show the effectiveness of our method in the identification of the emotions and their triggers. We also present our emotion prediction approach using game scene’s design goal (defined by OCC variables from the model of emotions’ cognitive evaluation of Ortony, Clore and Collins [1]) to annotate the player’s situation in a scene and machine learning algorithms. The prediction results are promising and would widen possibilities in game design.

The rote learning problem has plagued the education systems of developing world since long. To name a few, improperly designed assessments, teachers’ authority, rewarding verbatim answers, sheer class sizes, and individual learner differences are amongst the most notable mediators. The authors report on the design and development of an adaptive educational hypermedia, which disrupts the rote learning loop by hitting a few of the aforementioned reasons. The reported system provides a personalized learning experience to each learner, adapting on the basis of cognitive and learning styles. Further, the assessments are designed in a way that they loop each failed learning via variated paths, hence eliminating chances of rote learning. Moreover, the failed perturbations are traced back to the problematic domain segment for further knowledge acquisition. In-situ evaluations of the system with end-users (real students of Bachelor of Science in Computer Science) reveal a difference between control and experimental groups. The effect size is however moderate.

Learning by doing, such as when learners give explanations to peer learners in collaborative learning, is known to be an effective strategy for gaining knowledge. This study used two types of facilitation technology in a simple explanation task to experimentally investigate those influence on the performance of understanding self’s concept during collaborative explanation activity. Dyads were given a topic about cognitive psychology and were required to use two different theoretical concepts, each of which was provided separately to one or the other of them, and explain the topic to each other. Two types of facilitation were examined: (1) use of a pedagogical conversational agent (PCA) and (2) visual gaze feedback using eye-track sensing. The PCA was expected to enable greater support of task-based activity (task-work) and visual gaze feedback to support learner coordination within the dyads (team-work). Results show that gaze feedback was effective when there was no PCA, and the PCA was effective when there was no gaze feedback on explaining self’s concept. This work provides preliminary implications on designing collaborative learning technologies using tutoring agents and sensing technology.

Motivation is a key factor for learning and retention. Motivation in learning, which refers to an individual’s desire to learn, is influenced by a number of factors (e.g., interest, self-regulation abilities, self-efficacy, personality) and is further complicated by an individual’s sensitivity to those factors. Thus, identifying a learner’s general and fine-grained motivation factors is essential to designing individualized adaptations or interventions for implementation in an Intelligent Tutoring System (ITS). The present study addressed the development and validation of the Motivational Assessment Tool to identify correlations between motivation variables and factors from education and psychology. The results indicate an overlap between the scales, which implies a higher-order dimension structure not captured by existing instruments, enabling instructional designers to use the MAT to evaluate the motivation support provided by an ITS overall and identify motivation needs for individual learners.

Computer-assisted assessment environments, such as intelligent tutoring systems, simulations, and virtual environments are now being designed to measure students’ science inquiry practices. Some assessment environments not only evaluate students’ inquiry practice competencies, but also provide real-time scaffolding in order to help students learn. The present study aims to examine the impact of real-time scaffolding from an animated, pedagogical agent on students’ inquiry performance across a number of practices. Participants were randomly assigned to one of two conditions: receiving scaffolding or no scaffolding. All participants completed three virtual labs: Flower (a general pretest), Phase Change, and Density. Results showed that students who received immediate feedback during assessment performed better on subsequent inquiry tasks. These findings have implications for designers and researchers regarding the benefits of including real-time scaffolding within intelligent assessment systems.

Understanding how students allocate their time to different learning behaviors, especially those that distinguish students’ performances, can yield significant implications for the design of intelligent tutoring systems (ITS). Time on task is a typical indicator of students’ self-regulated learning (SRL) and student engagement. In this paper, we analyze log file data to identify patterns in the behavior durations of 62 medical students in BioWorld, an ITS that supports them in regulating their diagnostic reasoning while solving complex patient cases. Results demonstrated that task complexity mediated the relationship between students’ allocation of time and diagnostic performance outcomes. The high-performing students showed different patterns of time management with low-performing students when solving both simple and complex cases. Moreover, the durations of behaviors predicted students’ performance in clinical reasoning.

The cognitive load theory has been mainly investigated in declarative knowledge learning, typically learning with hyper-media material. In this study, the preceding findings are examined in problem solving domain with a different type of experimental task such as Reversi game. The experimental results were consistent with preceding studies, showing that extraneous cognitive load is harmful to the learning process, but the effects of intrinsic load are subject to debate. Additionally, the participants correctly evaluated each cognitive load, using a questionnaire. In addition, it was confirmed that the subjective evaluation predicted learning outcomes.

While it is important to individualize instruction, identifying and implementing the right intervention for individual students is too time-consuming for instructors to do manually in large classes. One approach to addressing this challenge is to identify groups of students who would benefit from the same intervention. As such, this work attempts to identify groups of students with similar academic and behavior characteristics who can benefit from the same intervention. In this paper, we study a group of 700 students who have been using ALEKS, a Web-based, adaptive assessment and learning system. We group these students into a set of clusters using six key characteristics, using their data from the first half of the semester, including their prior knowledge, number of assessments, average days and score increase between assessments, and how long after the start of the class the student begins to use ALEKS. We used mean-shift clustering to select a number of clusters, and k-mean clustering to identify distinct student profiles. Using this approach, we identified five distinct profiles within these students. We then analyze whether these profiles differ in terms of students’ eventual degree of content mastery. These profiles have the potential to enable institutions and instructors using ALEKS to identify students in need and devise and implement appropriate interventions for groups of students with similar characteristics and needs.

Students need to accurately monitor and judge the difficulty of learning materials to effectively self-regulate their learning with advanced learning technologies such as intelligent tutoring systems (ITSs), including MetaTutorIVH. However, there is a paucity of research examining how metacognitive monitoring processes such as ease of learning (EOLs) judgments can be used to provide adaptive scaffolding and predict student performance during learning ITSs. In this paper, we report on a study investigating how students’ EOL judgments can influence their performance and significantly predict their learning outcomes during learning with MetaTutorIVH, an ITS for human physiology. The results have important design implications for incorporating different types of metacognitive judgements in student models to support metacognition and foster learning of complex ITSs.

Massive Open Online Courses (MOOCs) are a promising approach for scalable knowledge dissemination. However, they also face major challenges such as low engagement, low retention rate, and lack of personalization. We propose AttentiveLearner2, a multimodal intelligent tutor running on unmodified smartphones, to supplement today’s clickstream-based learning analytics for MOOCs. AttentiveLearner2 uses both the front and back cameras of a smartphone as two complementary and fine-grained feedback channels in real time: the back camera monitors learners’ photoplethysmography (PPG) signals and the front camera tracks their facial expressions during MOOC learning. AttentiveLearner2 implicitly infers learners’ affective and cognitive states during learning from their PPG signals and facial expressions. Through a 26-participant user study, we found that: (1) AttentiveLearner2 can detect 6 emotions in mobile MOOC learning reliably with high accuracy (average accuracy = 84.4%); (2) the detected emotions can predict learning outcomes (best R2 = 50.6%); and (3) it is feasible to track both PPG signals and facial expressions in real time in a scalable manner on today’s unmodified smartphones.

In this paper, we introduce an artificial tutoring systems inspired by Applied Behavioral Analysis, named ABA tutor.Applied Behavioral Analysis is the application branch of analysis of behavior that derives from Behaviorism in psychology and has relevant features that can be transferred in an effective tutoring systems: the techniques of ABA are reproduced in the ABA tutor.Moreover we describe the first implementation of ABA tutor and the application to olfactory learning, as a case-study. In more detail, the ABA tutor has been applied to SNIFF, an integrated software and hardware system conceived to train the sense of smell with a gamified approach, and has been tested on 84 people. Results indicate the effectiveness of ABA tutor in promoting olfactory learning, thus supporting that this tutor can be successfully introduced in learning environments.

Self-regulated learning (SRL) and emotion regulation have been studied as separate constructs which impact students’ learning with intelligent tutoring systems (ITSs). There is a general assumption that students who are proficient in enacting cognitive and metacognitive SRL processes during learning with ITSs are also proficient emotion regulators. In this paper, we investigated the relationship between metacognitive and cognitive SRL processes and emotion regulation by examining students’ self-perceived emotion regulation strategies and comparing the differences between their (1) mean self-reported negative emotions, (2) proportional learning gains (PLGs), and the frequency of (3) metacognitive and (4) cognitive strategy use as they interacted with MetaTutor, an ITS designed to teach students about the circulatory system. Students were classified into groups based on self-perceived emotion regulation strategies and results showed students who perceived themselves as using adaptive emotion regulation strategies reported less negative emotions. Although no significant differences were found between the groups’ learning outcomes, there were significant differences between the groups’ frequency use of cognitive and metacognitive processes throughout the task. Our results emphasize the need to better understand how real-time emotion regulation strategies relate to SRL processes during learning with ITSs and can be used to enhance learning outcomes by encouraging adaptive emotion regulation strategies as well as increased frequencies of metacognitive and cognitive SRL processes.

This article presents the results of an experiment in which we investigated how prior algebra knowledge and personality can influence the permanence time from the confusion state to frustration/boredom state in a computer learning environment. Our experimental results indicate that people with a neurotic personality and a low level of algebra knowledge can deal with confusion for less time and can easily feel frustrated/bored when there is no intervention. Our analysis also suggest that people with an extroversion personality and a low level of algebra knowledge are able to control confusion for longer, leading to later interventions. These findings support that it is possible to detect emotions in a less invasive way and without the need of physiological sensors or complex algorithms. Furthermore, obtained median times can be incorporated into computational regulation models (e.g. adaptive interfaces) to regulate students’ emotion during the teaching-learning process.

Summarization enhances comprehension and is considered an effective strategy to promote and enhance learning and deep understanding of texts. However, summarization is seldom implemented by teachers in classrooms because the manual evaluation requires a lot of effort and time. Although the need for automated support is stringent, there are only a few shallow systems available, most of which rely on basic word/n-gram overlaps. In this paper, we introduce a hybrid model that uses state-of-the-art recurrent neural networks and textual complexity indices to score summaries. Our best model achieves over 55% accuracy for a 3-way classification that measures the degree to which the main ideas from the original text are covered by the summary . Our experiments show that the writing style, represented by the textual complexity indices, together with the semantic content grasped within the summary are the best predictors, when combined. To the best of our knowledge, this is the first work of its kind that uses RNNs for scoring and evaluating summaries.

Positive academic emotions are generally associated with positive learning experiences, while the opposite is true for negative emotions. This study examined changes in learners’ emotional profiles as they participated in MetaTutor, a computer-based learning environment designed to foster self-regulated learning via study of the human circulatory system. Latent transition analysis was employed to determine distinct, parsimonious emotional profiles over time. Learners are shown to move systematically among three profiles (positive, bored/frustrated, and moderate) in fairly predictable patterns. Of these, boredom is the most pressing concern given the relatively small chance of moving from boredom to a different emotional profile. Students’ learning gains were also significant predictors of emotional transitions. The findings suggest the need for timely intervention for learners who are on the verge of negative emotional trajectories, and the complex relationship between learning gains and emotions. In addition, latent transition analysis is demonstrated as a potentially useful technique for analyzing and utilizing multivariate panel data.

The recommendation of micro Open Education Resources (OERs) suffers from the new-item cold start problem because little is known about the continuously published micro OERs. This paper provides a heuristic approach to inserting newly published micro OERs into established learning paths, to enhance the possibilities of new items to be discovered and appear in the recommendation lists. It considers the accumulation and attenuation of user interests and conform with the demand of fast response in online computation. Performance of this approach has been proved by empirical studies.

We investigated how college students’ (n = 40) different levels of action unit 4 (AU4: brow lowerer), metacognitive monitoring process use and pre-test score were associated with metacognitive monitoring accuracy during learning with a hypermedia-based ITS. Results revealed that participants with high pre-test scores had the highest accuracy scores with low levels of AU4 and use of more metacognitive monitoring processes, whereas participants with low pre-test scores had higher accuracy scores with high levels of AU4 and use of more metacognitive monitoring processes. Implications include designing adaptive ITSs that provide different types of scaffolding based on levels of prior knowledge, use of metacognitive monitoring processes, and emotional expressivity keeping in mind that levels of emotions change over time, and therefore must be monitored to provide effective scaffolding during learning.

The goal of this study was to examine 38 undergraduate and graduate students’ note taking and summarizing, and the relationship between emotions, the accuracy of those notes and summaries, and proportional learning gain, during learning with MetaTutor, an ITS that fosters self-regulated learning while learning complex science topics. Results revealed that students expressed both positive (i.e., joy, surprise) and negative (i.e., confusion, frustration, anger, and contempt) emotions during note taking and summarizing, and that these emotions correlated with each other, as well as with proportional learning gain and accuracy of their notes and summaries. Specifically, contempt during note taking was positively correlated with proportional learning gain; note taking accuracy was negatively correlated with proportional learning gain; and confusion during summarizing was positively correlated with summary accuracy. These results reveal the importance of investigating specific self-regulated learning processes, such as taking notes or making summaries, with future research aimed at examining the differences and similarities between different cognitive and metacognitive processes and how they interact with different emotions similarly or differently during learning. Implications of these findings move us toward developing adaptive ITSs that foster self-regulated science learning, with specific scaffolding based on each individual student’s learning needs.

We developed a system for programming practice that provides adaptive feedback based on the presence of confusion on the student. The system provides two types of adaptive feedback. First, it can control the complexity of the exercises presented to the student. Second, it can offer guides for the exercises when needed. These feedback are based on the presence of confusion, which is detected based on the student’s compilations, typing activity, and facial expressions using a hidden Markov model trained on data collected from introductory programming course students. In this paper we discuss the system, the approach for detecting confusion, and the types of adaptive feedback displayed. We tested our system on Japanese university students and discuss the results and their feedback. This study can lay the foundation for the development of intelligent programming tutors that can generate personalized learning content based on the state of the individual learner.

Digital doppelgangers are virtual humans that highly resemble the real self but behave independently. An emerging computer animation technology makes the creation of digital doppelgangers an accessible reality. This allows researchers in pedagogical agents to explore previously unexplorable research questions, such as how does increasing the similarity in appearance between the agent and the student impact learning. This paper discusses the design and evaluation of a digital doppelganger as a virtual listener in a learning-by-explaining paradigm. Results offer insight into the promise and limitation of this novel technology.

Accurate classification of learner responses is a critical component of dialog based tutoring systems (DBT). Errors in identifying the intent and context of responses can have cascading effects on the ongoing interaction thereby affecting the learning experience and outcome. In this paper we attempt to quantify the impact of Tutor misclassifications on student behavior by analyzing differences across our hypothesized conditions namely, no-misclassification vs. misclassification using various dialog metrics. We find that not only are there significant changes in behavior across the two groups but that Tutor errors related to misunderstanding of Intent - although fewer in occurrence, appear to have a higher impact than a misclassification of a valid student answer. We also see some evidence of the effectiveness of scaffolds like FITBs in sustaining dialog thereby mitigating the effects of a Tutor error.

While prior research has typically treated student self-confidence as a static measure, confidence is not identical in all situations. We study the degree to which confidence varies over time using entropy, investigating whether high variation in confidence is more characteristic of highly confident or highly uncertain students, using data from 118,000 students working within 8 courses within the LearnSmart adaptive platform. We find that more confident students are also more consistent in their confidence. Confident students were more likely to answer correctly but also more likely to be overconfident, making unexpected mistakes. Finally, we develop interpretable clusters of students based on their confidence entropy, degree of over/underconfidence, and related variables.

Virtual Reality is gathering increasing popularity for Intelligent Tutoring Systems. We introduce an approach that improves the baseline VR experience for ITS by enabling access to open educational resources and more intelligent navigation with the support of multiple artificial intelligence algorithms. A preliminary user study result not only reveals the potential of the proposed method, but also helps to identify the clues to improve the current design.

The performance and reactions of an individual in urgent situations vary according to his emotional states, which are subject to sudden changes. In medical situation, when patient’s life is in danger, the doctor’s emotional states are provoked and could affect his decisions. Virtual reality environments represent immersive situations in which we can dynamically simulate medical cases and particularly emergency cases. Using virtual reality and EEG devices, we can analyze the doctor’s emotional state and behavior without risk. In this paper, we propose to generate medical cases that can induce frustration or stress which can be at the origin of mistakes from the student. For that, we created a neurofeedback system named “Hypocrates” composed of a virtual reality environment, a medical cases generator and an intelligent agent. An experimental study involving 15 students in medicine was conducted to evaluate our approach. Results show that, the mistakes generally increase the frustration of medical students and decrease their performance.

This paper explores the application of deep learning in automated essay scoring (AES). It uses the essay dataset #8 from the Automated Student Assessment Prize competition, hosted by the Kaggle platform, and a state-of-the-art Suite of Automatic Linguistic Analysis Tools (SALAT) to extract 1,463 writing features. A non-linear regressor deep neural network is trained to predict holistic scores on a scale of 10–60. This study shows that deep learning holds the promise to improve significantly the accuracy of AES systems, but that the current dataset and most essay datasets fall short of providing them with enough expertise (hand-graded essays) to exploit that potential. After the tuning of different sets of hyperparameters, the results show that the levels of agreement, as measured by the quadratic weighted kappa metric, obtained on the training, validation, and testing sets are 0.84, 0.63, and 0.58, respectively, while an ensemble (bagging) produced a kappa value of 0.80 on the testing set. Finally, this paper upholds that more than 1,000 hand-graded essays per writing construct would be necessary to adequately train the predictive student models on automated essay scoring, provided that all score categories are equally or fairly represented in the sample dataset.

Our Virtual Learning Environment aims at improving the abilities of experienced technicians to handle critical situations through appropriate use of non-technical skills (NTS), a high-stake matter in many domains as bad mobilization of these skills is the cause of many accidents. To do so, our environment dynamically generates critical situations designed to target these NTS. As the situations need to be adapted to the learner’s skill level, we designed a hybrid architecture able to diagnose NTS. This architecture combines symbolic knowledge about situations, a neural network to drive the learner’s performance evaluation process, and a Bayesian network to model the causality links between situation knowledge and performance to reach NTS diagnosis. A proof of concept is presented in a driving critical situation.

The learning early-warning is an effective way to optimize the teaching effect and teach students in accordance of their aptitude. At present, the learning early-warning faces low accuracy, high value of MSE and MAE. We propose a novel learning early-warning model: LEWM-RFA. The model divides students’ learning behaviors data into three dimensions: knowledge, behavior and attitude. Then the model uses random forest algorithm to extract features that can affect students’ grades, and then predicts students’ final exam scores. Students are divided into three warning levels according to their grades. Compared with the model based on the linear regression algorithm, the LEWM-RFA’s MSE decreases by 27.498% and the LEWM-RFA’s MAE decreases by 26.960%.

Researches in areas such as neuroscience and psychology indicate that emotions directly impact learning. So, adapting to the learners’ affective reactions became a requirement and also a challenge for building a new generation of affect aware computing learning environments. In this paper, we present a hybrid approach for inferring learning related emotion that combines cognitive and physical data, gathered using minimal or non intrusive methods. In an initial experiment with students in a real education environment it was possible to obtain promising results when comparing some usual performance metrics with correlated works. In this study we achieved accuracy rates and Cohen’s Kappa near to 65% and 0.55, respectively. Furthermore, considering the open and expansible nature of this proposal, we believe that this results could be improved in the future by adding new data or new sensors to the model, for example.

Personalised recommendations feature prominently in many aspects of our lives, from the movies we watch, to the news we read, and even the people we date. However, one area that is still relatively underdeveloped is the educational sector where recommender systems have the potential to help students to make informed choices about their learning pathways. We aim to improve the way students discover elective modules by using a hybrid recommender system that is specifically designed to help students to better explore available options. By combining notions of content-based similarity and diversity, based on structural information about the space of modules, we can improve the discoverability of long-tail options that may uniquely suit students’ preferences and aspirations.

Many activities, such as learning a craft, involve learning how to manipulate physical objects by following a step-by-step procedure. In this paper we present our ongoing work on development of TUMA: an intelligent tutoring system for manual-procedural activities. We first introduce the notion of manual-procedural activity and then argue about the opportunities for creating intelligent tutors for manual-procedural activities. Such an intelligent tutoring system can be used in domains like teaching crafts, that involve acquiring cognitive knowledge along with specific motor skills. TUMA unifies the research from three different communities: intelligent tutoring systems, human motion tracking, and assistance systems for manual assembly in manufacturing. We describe the vision and the requirements of TUMA and its functional architecture inspired by high-level components of intelligent tutoring systems. Finally we report on our research road map for implementing a proof-of-concept and evaluating its impact.

We propose a method using eye-gaze tracking technology and machine learning for the analysis of the reading section of the Scholastic Aptitude Test (SAT). An eye-gaze tracking device tracks where the reader is looking on the screen and provides the coordinates of the gaze. This collected data allows us to analyze the reading patterns of test takers and discover what features enable test takers to score higher. Using a machine learning approach, we found that the time spent on the passage at the beginning of the test (in minutes), number of times switching between the passage and the questions, and the total time spent doing the reading test (in minutes) have the greatest impact in distinguishing higher scores from lower scores.

Assessment of skills and process knowledge is difficult and quite different from assessing knowledge of content. Many assessment systems use either multiple choice questions or other frameworks that provide a significant amount of scaffolding and this can influence the results. One reason for this is that they are easy to administer and the answers can be automatically graded. This paper describes an assessment tool that does not provide scaffolding (and therefore hints) and yet is able to automatically grade the free form answers through the use of domain knowledge heuristics. The tool has been developed for a tutoring system in the domain of red black trees (a data structure in computer science) and has been evaluated on three semesters of students in a computer science course.

This paper examines the use of “pacing plots” to represent variations in student learning sequences within a digital curriculum. Pacing plots are an intuitive and flexible data visualizations that have a potential for revealing the diversity of blended classroom instructional models. By using curriculum pacing plots, we identified several common implementation patterns in real-world classrooms. After analyzing two years’ worth of data from over 150,000 students in a digital math curriculum, we found that a PCA and K-Means clustering approach was able to discover pedagogically relevant instructional practices.

Programmers use Integrated Development Environments (IDEs) to write and test software, and students use them while learning programming. We explore the approach of embedding a tutoring companion inside Eclipse, a popular IDE. The embedded tutoring companion aims to be comparable to having an actual teaching assistant present all the time with each student throughout a course. The embedded tutoring companion tracks student’s actions while solving a problem (coding, compiling, running) and collects metadata including the time spent, the correctness of the work, and the amount of copied or auto-generated code in the work. Then it can determine the practical understanding of the topics and concepts associated with the presented problem, it can assist the student by providing immediate feedback, and it can help instructors by reporting real-time information about students’ performance. Our companion, implemented as an Eclipse plug-in, was evaluated with undergraduate students enrolled in a Java programming course.

In communities of practice (CoP), learning occurs through constant interactions of their participants. The social aspect is fundamental for the construction of knowledge. This work uses semantic web technologies and ontologies to structure and represent the interactions of CoPs participants around a dynamic user profile. This user profile describes a set of dispersed properties and relationships in CoPs, allowing collaborative trajectories recovery in these learning environments.

In the present paper we present analysis of gaming actions with MathSpring, an established ITS for mathematics for high school students. Our findings indicate that both student and problem features were similarly predictive of gaming behaviors, as well as that gaming was associated with lower excitement and lower learning gains.

E-learning systems based on a conversational agent (CA) provide the basis of an intuitive and engaging interface for the student. The goal of this paper is to propose a method for detecting conversational interaction behaviors of learners and CAs, using an agent-based framework, for the purpose of improving the communication between students and CA-based intelligent tutoring systems. Our framework models both the student and the CA and uses agents to represent data sources for each. We show how the framework uses the detection of conversational behaviors to initiate interventions to improve student conversational engagement. The results of initial user testing are reported.

Educational quizzes are useful not only to evaluate or test the knowledge acquired by a learner, but also to help her/him to deepen knowledge about a specific domain or topic in an informal and entertaining way. The production of quizzes is a time-consuming task that can be automated by taking advantage of existing knowledge bases available on the Web of Linked Open Data (LOD). However, automatically extracting from the LOD a knowledge graph composed by the information of a set of resources which are relevant to a given specific domain or topic, is a crucial phase for the automatic generation of quizzes.To address this issue, we propose a heuristic that extracts from DBpedia a set of resources related to a given specific domain. Such heuristic has been implemented and used for the automatic generation of quizzes in the geography and privacy domains. We report a comparative user evaluation of it.

Teaching surgical decision making aims at enabling students to choose the most appropriate action relative to the patient’s situation and surgical objectives. This requires a deep understanding of causes and effects related to the surgical domain as well as being aware of key properties of the current situation. To develop an intelligent tutoring system (ITS) for teaching situated decision making in the domain of dental surgery, in this paper, we present a planning-based representation framework. This framework is capable of representing surgical procedural knowledge with respect to situation awareness and algorithms that utilize the representation to generate rich tutorial dialog. The design of the tutorial dialogs is based on an observational study of surgeons teaching in the operating room. An initial evaluation shows that generated interventions are as good as and sometimes better than those of experienced human instructors.

We present a unified learner modeling approach in a childhood vocabulary tutoring platform that enables learning continuum across multiple learning experiences. By decoupling experiences from learner modeling, multiple learning experiences can be developed independently making learning approaches scalable in an inherently diverse setting: each child, his/her family, environment, geography, cultural differences can all make each learner really unique. By understanding the information-theoretic equivalence of different assessment types, and mapping different play and learning activities to one or more of these types, we can enable a rapid convergence of the learner model to better represent a young learner’s knowledge. More interestingly, when normalized, different assessment types converge at different levels of the normalized score.

As a pathway for learning, remix has become one of the most important practices within the field of open educational resources. In this study, we investigate the searching and re-editing behavior of students in an online web environment. Participants were asked to search and remix the retrieved web information, which was based on the content of textbook. To explore learning process in the remixing environment, the relationships among function of thinking styles, the search behaviors, the edit behavior, pre-performance, and the final remixing performance are analyzed. Various behavior data are recorded, including web visiting log, interaction log and eye tracking data. The finding provides insight into how to understand the behaviors associated with underlying cognitive and learning performances.