The Role of Content-Driven Prompts in Facilitating Preservice Teacher Information-Seeking With Network-Based Tutors

• In Event: Learner Support in Digital Learning Environments

Abstract

Background
Existing research recognizes the critical role played by prompts as instructional scaffolds for preservice teachers to become proficient at self-regulated learning (Kramarski & Kohen, 2017; Moos & Ringdal, 2012). The assumption is that teachers should become proficient themselves prior to teaching these skills to their students. Humans or artificial agents often prompt teachers given a set of pre-determined rules while considering relevant task conditions, such as elapsed time in the learning session. A major problem is the time and cost necessary to determine not only the content of prompts, but also the rules to deliver them to learners.
Objective
Network-based tutors are a type of intelligent tutoring system that relies on metaheuristic algorithms, natural language processing, and semantic web technologies to optimize the content of prompts over time (Poitras, Mayne, Huang, Doleck, Udy, & Lajoie, 2018). This paper addresses the problem of converging prompt to the optimal content to facilitate information seeking. We hypothesized that learners with assistance from content-driven prompts that are optimized over time will demonstrate improvements in various facets related to information seeking, acquisition, and task performance.
Dataset
51 preservice teachers were instructed to revise an online lesson either with or without assistance from content-driven prompts that are optimized over time (CP vs. NoCP conditions). The prompts were delivered by a virtual pedagogical agent implemented in nBrowser, an intelligent web browser designed to support learners to design technology-infused lesson plans.
Methods
We extracted metrics from the log trace data to characterize the following: (1) the breakdown of content-driven prompt convergence and compliance; (2) the operations that mediate information seeking (i.e., elapsed time, amount of characters, Latent Dirichlet Allocation topic mixtures); and (3) their outcomes (i.e., quality score for the desirability of linguistic content).
Results
From the data it is apparent that the content of prompts converged towards information assimilated by learners in accordance with the network topology, but that this process was unstable. This finding may be attributed to how learners complied with prompts, as the exam of cross-distances in the similarity of the original and revised information from the lesson suggests that most learners chose to elaborate, rather than quote or paraphrase information.
A linear mixed-effects model was used to predict the quality of learner revisions based on various facets of information seeking behaviors. It is apparent from Table 1 that learner revisions related to content knowledge were associated with greater quality scores when learners requested a prompt in the CP condition, but not when learners failed to make a request.
Significance
Together these results provide important insights into the design of a two-pass spreading activation algorithm to converge network-based domain models. We discuss the possibility that the content of prompts should be optimized over time at first based on information highlighted by learners from an online lesson, then from the quality of the revision saved by a learner and detected by the system using natural language processing techniques.

Authors

• Laurel Udy, University of Utah

• Eric G. Poitras, Dalhousie University

• Lingyun Huang, The Education University of Hong Kong

• Susanne P. Lajoie, McGill University