Abstract:

The United States Marine Corps (USMC)’ College of Distance Education and Training (CDET) has identified a curriculum-gap for an automated course-content-creation capability that can generate interactive-content such as HTML5-Packages (H5P) for eLearning-systems from legacy schoolhouse training-content such as student-handouts in Portable-Document-Format (PDF) or Word-format, and classroom-presentations in PowerPoint (PPTX)-format. This need stems from the fact that schoolhouses are resource-constrained in personnel having the requisite subject-matter-expertise (SME) who are busy teaching classes with limited time to convert the static course-curricula into dynamic format within a short duration. Recent advances in generative-Artificial-Intelligence (genAI) has the potential to fill this need by automatically performing knowledge-extraction from USMC training-content and producing interactive-content with knowledge-checks. However, cloud-based genAI-models such as ChatGPT require uploading classified USMC content over the internet-connection, thus violating USMC security-policies. Consequently, there is a need for a non-cloud, non-subscription genAI capability to convert legacy-content to web-ready-content for learning-management-servers (LMSs) such as Moodle to enable distance-learning.

In this paper, the research-team presents a genAI interactive course-package-creation pipeline that uses the power of local computing-devices such as laptops/desktops and offline ChatGPT-like Large-Language-Models (LLMs) for auto-generating Moodle-compliant H5P interactive-learning-content from schoolhouse course-material within a short period of time. A web-based user-interface allows instructors to upload PPTX and Word/PDF material as contextual inputs, choose tuning-presets for creation of interactive-content such as Quizzes, Dialog-Cards, Fill-in-the-Blanks, Flashcards, etc., and review/modify the LLM responses in human-in-the-loop interface before sending it to an automated H5P-generator. The process is modular and makes provisions for swapping underlying LLMs without the need to redeploy software. It is also secure since it is laptop-based and uses on-premise genAI without requiring cloud-access.

This paper documents the results of scalability-tests on a standard benchmark-conversion from course-curriculum to H5P-activity, comparing performance from various open-source LLMs on end-user desktops versus high-end GPU-enabled hardware. The paper concludes with best-practices and lessons learned from these content-conversion-experiments.

Keywords: ADVANCED DISTRIBUTED LEARNING;AI;AUTHORING TOOLS;CONTENT GENERATION;ELEARNING;EMERGING TECHNOLOGIES;TRAINING

Menu

Jordan Thurston