Adaptive Content and Process Scaffolding With Network-Based Tutors: Implications for Promoting Self-Regulated Learning

• In Event: 43.013 - Scaffolding Self-Regulated Learning Across Different Contexts

Abstract

Objectives and perspectives
Research on scaffolding student self-regulated learning (SRL) processes is an important area of study, specifically research on the type and timing of metacognitive prompting. Most prompts are delivered from human or virtual pedagogical agents relying on written protocols, scripts, and timers. Static prompt delivery is pre-determined on the basis of sequential production rules and is criticized for failing to consider the cyclical and dynamic processes that unfold during SRL (Kramarski & Kohen 2017; Sonnenberg & Bannert, 2018). An alternate approach is to deliver prompts on the basis of the content, or the information transformed by learners while engaging in SRL operations. The assumption is that more dimensions (i.e., time, event, and content) are critical to design production rules that deliver prompts in a valid and reliable manner.
Methods
Network-based tutoring systems provide a method for developing dynamic prompts by using web mining and natural language processing techniques to build a computational representation of educational content in the form of a network. The network consists of nodes that represent hypermedia elements interrelated through links weighed on the basis of semantic relationships. This paper examines three components of this domain modeling method: (1) authoring and validating network-based domain models, (2) adapting content and process scaffolds, and (3) evaluating and improving the scaffold delivery mechanisms.
Data sources
We apply semi-supervised machine learning algorithms to author multi-dimensional networks in an automated manner by leveraging metadata embedded in hypermedia-based collections of learning materials and resources. To demonstrate its feasibility, our analysis focused on 660 linguistic features extracted from 28 HTML elements using a combination of text pre-processing and feature selection steps as well as K-Means clustering and K-Nearest Neighbour classification algorithms.
Results
The analysis outcomes are shown in Figure 1. The conceptual similarity between any two nodes is computed as the angle of pairs of term vectors, while the quality of the latent dimensions is measured through both the distance to the centroid of each cluster and the error rate in cluster assignment based on the nearest neighbour. As the amount of predetermined dimensions increase, the nodes of the network map onto increasingly similar topics. One drawback is the amount of errors made in the assignment of new nodes to each of the dimensions also increases.
 
Significance
We will discuss the broader implications of the proposed approach to domain modeling for the design of content-driven prompts in terms of the contemporary research on process-driven prompt delivery mechanisms. Namely, prompts that are designed and validated from similar behaviors observed in expert tutors (Azevedo, Cromley, & Seibert, 2004), self-created by students (Pieger & Bannert, 2018), and induced from learner behaviors (Bouchet, Harley, Trevors, & Azevedo, 2013). Taken together, these findings warrant more attention from the research community towards the use of content analytics, a form of learning analytics that pertains to the different forms of content related to learning (see Kovanovi et al., 2016).

Authors

• Eric G. Poitras, University of Utah

• Laurel Udy, University of Utah

• Lingyun Huang, McGill University

• Susanne P. Lajoie, McGill University