Search for: All records

Education is poised for a transformative shift with the advent of neurosymbolic artificial intelligence (NAI), which will redefine how we support deeply adaptive and personalized learning experiences. The integration of Knowledge Graphs (KGs) with Large Language Models (LLMs), a significant and popular form of NAI, presents a promising avenue for advancing personalized instruction via neurosymbolic educational agents. By leveraging structured knowledge, these agents can provide individualized learning experiences that align with specific learner preferences and desired learning paths, while also mitigating biases inherent in traditional AI systems. NAI-powered education systems will be capable of interpreting complex human concepts and contexts while employing advanced problem-solving strategies, all grounded in established pedagogical frameworks. In this paper, we propose a system that leverages the unique affordances of KGs, LLMs, and pedagogical agents – embodied characters designed to enhance learning – as critical components of a hybrid NAI architecture. We discuss the rationale for our system design and the preliminary findings of our work. We conclude that education in the era of NAI will make learning more accessible, equitable, and aligned with real-world skills. This is an era that will explore a new depth of understanding in educational tools. 

Pedagogical agents (PAs) are increasingly being integrated into educational technologies. Although previous reviews have examined the impact of PAs on learning and learning-related outcomes, it still remains unclear what specific design features, social cues, and other contextual elements of PA implementation can optimize the learning process. These questions are even more prevalent with regards to the K-12 population, as most reviews to date have largely focused on post-secondary learners. To address this gap in the literature, we systematically review empirical studies around the design of PAs for K-12 learners. After reviewing 1374 studies for potential inclusion, we analyzed 44 studies that met our inclusion criteria using Heidig and Clarebout’s (2011) frameworks. Our findings showed that learners had preferences for specific types of PAs. While these preferences were not always associated with increased learning outcomes, there is a lack of research specifically investigating the intersection of perceptions and learning. Our results also showed that pedagogical strategies that are effective for human teachers were effective when used by PAs. We highlight what specific design features instructional designers can use to design PAs for K-12 learners and discuss promising research directions based on the extant work in the field. 

Systematic reviews are a time-consuming yet effective approach to understanding research trends. While researchers have investigated how to speed up the process of screening studies for potential inclusion, few have focused on to what extent we can use algorithms to extract data instead of human coders. In this study, we explore to what extent analyses and algorithms can produce results similar to human data extraction during a scoping review—a type of systematic review aimed at understanding the nature of the field rather than the efficacy of an intervention—in the context of a never before analyzed sample of studies that were intended for a scoping review. Specifically, we tested five approaches: bibliometric analysis with VOSviewer, latent Dirichlet allocation (LDA) with bag of words, k-means clustering with TF-IDF, Sentence-BERT, or SPECTER, hierarchical clustering with Sentence-BERT, and BERTopic. Our results showed that topic modeling approaches (LDA/BERTopic) and k-means clustering identified specific, but often narrow research areas, leaving a substantial portion of the sample unclassified or in unclear topics. Meanwhile, bibliometric analysis and hierarchical clustering with SBERT were more informative for our purposes, identifying key author networks and categorizing studies into distinct themes as well as reflecting the relationships between themes, respectively. Overall, we highlight the capabilities and limitations of each method and discuss how these techniques can complement traditional human data extraction methods. We conclude that the analyses tested here likely cannot fully replace human data extraction in scoping reviews but serve as valuable supplements. 

Over the past three decades the field of pedagogical agents (PAs) has seen significant growth, but no review has specifically focused on the design and use of PAs for K-12 students, despite the fact that an early meta-analysis showed that they receive the most benefits from learning from or with PAs. Our systematic search revealed 112 studies that met the inclusion criteria and were analyzed. Our findings revealed a plethora of studies investigating the use of PAs with K-12 populations and a considerable number of longitudinal studies, both of which the field has long stated did not exist in significant numbers. Our findings contrast long-held findings in the field, further support others, and highlight areas where further experimentation and research synthesis are needed. 

In recent years, large language models (LLMs) have seen rapid advancement and adoption, and are increasingly being used in educational contexts. In this perspective article, we explore the open challenge of leveraging LLMs to create personalized learning environments that support the “whole learner” by modeling and adapting to both cognitive and non-cognitive characteristics. We identify three key challenges toward this vision: (1) improving the interpretability of LLMs' representations of whole learners, (2) implementing adaptive technologies that can leverage such representations to provide tailored pedagogical support, and (3) authoring and evaluating LLM-based educational agents. For interpretability, we discuss approaches for explaining LLM behaviors in terms of their internal representations of learners; for adaptation, we examine how LLMs can be used to provide context-aware feedback and scaffold non-cognitive skills through natural language interactions; and for authoring, we highlight the opportunities and challenges involved in using natural language instructions to specify behaviors of educational agents. Addressing these challenges will enable personalized AI tutors that can enhance learning by accounting for each student's unique background, abilities, motivations, and socioemotional needs. 

Virtual humans are on-screen characters that are often embedded in learning technologies to deliver educational content. Little research has investigated how virtual humans can be used to correct common misconceptions. In this study, we explored how different types of narrative structures, refutation text and expository text, influence perceptions of trust, credibility, and learning outcomes. In addition, we conducted exploratory analyses examining how different measures of trust and credibility are related to each other and how these measures may mediate learning outcomes. Results showed that the type of narrative used did not influence any measure. However, the trust and credibility measures, while related to one another, were measurably distinct. In addition, only perceptions of message trust were significantly related to learning. Perceptions of message trust did not mediate learning outcomes, but were significantly predictive of learning at nearly the same effect as prior knowledge. 

The ability to monitor mental effort during a task using a wearable sensor may improve productivity for both work and study. The use of the electrodermal activity (EDA) signal for tracking mental effort is an emerging area of research. Through analysis of over 92 h of data collected with the Empatica E4 on a single participant across 91 different activities, we report on the efficacy of using EDA features getting at signal intensity, signal dispersion, and peak intensity for prediction of the participant’s self-reported mental effort. We implemented the logistic regression algorithm as an interpretable machine learning approach and found that features related to signal intensity and peak intensity were most useful for the prediction of whether the participant was in a self-reported high mental effort state; increased signal and peak intensity were indicative of high mental effort. When cross-validated by activity moderate predictive efficacy was achieved (AUC = 0.63, F1 = 0.63, precision = 0.64, recall = 0.63) which was significantly stronger than using the model bias alone. Predicting mental effort using physiological data is a complex problem, and our findings add to research from other contexts showing that EDA may be a promising physiological indicator to use for sensor-based self-monitoring of mental effort throughout the day. Integration of other physiological features related to heart rate, respiration, and circulation may be necessary to obtain more accurate predictions. 

Trackers for activity and physical fitness have become ubiquitous. Although recent work has demonstrated significant relationships between mental effort and physiological data such as skin temperature, heart rate, and electrodermal activity, we have yet to demonstrate their efficacy for the forecasting of mental effort such that a useful mental effort tracker can be developed. Given prior difficulty in extracting relationships between mental effort and physiological responses that are repeatable across individuals, we make the case that fusing self-report measures with physiological data within an internet or smartphone application may provide an effective method for training a useful mental effort tracking system. In this case study, we utilized over 90 h of data from a single participant over the course of a college semester. By fusing the participant’s self-reported mental effort in different activities over the course of the semester with concurrent physiological data collected with the Empatica E4 wearable sensor, we explored questions around how much data were needed to train such a device, and which types of machine-learning algorithms worked best. We concluded that although baseline models such as logistic regression and Markov models provided useful explanatory information on how the student’s physiology changed with mental effort, deep-learning algorithms were able to generate accurate predictions using the first 28 h of data for training. A system that combines long short-term memory and convolutional neural networks is recommended in order to generate smooth predictions while also being able to capture transitions in mental effort when they occur in the individual using the device. 

Automated tracking of physical fitness has sparked a health revolution by allowing individuals to track their own physical activity and health in real time. This concept is beginning to be applied to tracking of cognitive load. It is well known that activity in the brain can be measured through changes in the body’s physiology, but current real-time measures tend to be unimodal and invasive. We therefore propose the concept of a wearable educational fitness (EduFit) tracker. We use machine learning with physiological data to understand how to develop a wearable device that tracks cognitive load accurately in real time. In an initial study, we found that body temperature, skin conductance, and heart rate were able to distinguish between (i) a problem solving activity (high cognitive load), (ii) a leisure activity (moderate cognitive load), and (iii) daydreaming (low cognitive load) with high accuracy in the test dataset. In a second study, we found that these physiological features can be used to predict accurately user-reported mental focus in the test dataset, even when relatively small numbers of training data were used. We explain how these findings inform the development and implementation of a wearable device for temporal tracking and logging a user’s learning activities and cognitive load.