Reducing Cognitive Load in Digital Reading: An LLM-Powered Approach for Universal Reading Comprehension Öffentlichkeit
Han, Junzhi (Spring 2025)
Abstract
This research investigates how pre-trained large language models (LLMs) can generate concept maps to enhance digital reading comprehension in higher education. While particularly focused on supporting neurodivergent students with their distinct information processing patterns, this approach benefits all learners facing the cognitive challenges of digital text. The study employs GPT-4o-mini to extract concepts and relationships from educational texts across ten diverse disciplines using open-domain prompts without predefined concept categories or relation types, enabling truly discipline-agnostic extraction applicable to all educational domains. Evaluation of three text processing approaches against a manually annotated gold dataset reveals that for concept extraction, section-level processing achieves the highest precision (83.62%), while paragraph-level processing demonstrates superior recall (74.51%). For relation extraction, similar patterns emerge with section-level processing showing the highest precision (78.61%) and paragraph-level processing yielding better recall (69.08%). Disciplinary variations are observed in both extraction tasks, with biology showing the strongest concept (F1=77.52%) and relation (F1=73.65%) extraction performance while humanities disciplines have comparatively lower performance. An interactive web-based visualization tool was developed that transforms extracted concepts into navigable concept maps using D3.js force-directed layouts, accessible at https://simplified-cognitext.streamlit.app/. User evaluation (n=14) revealed that while participants spent more time engaging with concept maps (22.6% increase), they experienced substantially reduced cognitive load (31.5% decrease in perceived mental effort) and completed comprehension assessments more efficiently (14.1% faster) with marginal improvements in accuracy. Qualitative feedback (mean rating: 4.21/5) highlighted the tool's effectiveness in visualizing conceptual relationships, though initial adaptation challenges were noted. This work contributes to educational technology by establishing a framework for LLM-based concept extraction, providing evidence on processing granularity effects, developing a concept categorization system for educational mapping, and creating a visualization tool with demonstrated learning benefits.
Table of Contents
1 Introduction
1.1 The challenge of digital reading and concept maps as a potential solution
1.2 Current approaches and limitations
1.3 Research questions and hypothesis
2 Background
2.1 Cognitive load in digital reading
2.2 Reading challenges for neurodivergent students
2.3 Natural language processing for concept extraction
2.4 Relation extraction approaches
2.5 Visualization techniques for knowledge representation
3 Approach
3.1 Overview of methodology
3.2 Dataset selection and preparation
3.2.1 Preprocessing steps
3.3 Text processing modes
3.4 Model selection
3.4.1 GPT-4o-mini
3.4.2 all-MiniLM-L6-v2 sentence transformer
3.5 Evaluation framework
3.5.1 Quantitative assessment metrics
3.5.2 Matching criteria implementation
3.6 Knowledge extraction process
3.6.1 Gold standard dataset construction
3.6.2 Concept extraction
3.6.3 Post concept extraction analysis
3.6.4 Relation extraction
3.6.5 Post relation extraction analysis
3.7 Full automated extraction workflow
3.8 Concept map visualization
3.8.1 Concept map construction
3.8.2 Technical implementation
3.8.3 User reading comprehension assessment
4 Results
4.1 Data section
4.1.1 Data characteristics
4.2 Concept extraction performance by discipline
4.2.1 Concept categorization results
4.2.2 Concept extraction performance results
4.3 Relation extraction performance by discipline
4.3.1 Relation categorization results
4.3.2 Relation extraction performance results
4.4 Concept map visualization
4.4.1 Web application implementation
4.4.2 Reading comprehension assessments
4.4.3 User feedback
5 Analysis
5.1 Concept
5.1.1 Concept inter-annotator agreement
5.1.2 Concept categorization analysis
5.1.3 Concept extraction performance analysis
5.1.4 Error analysis
5.2 Relation
5.2.1 Relation inter-annotator agreement
5.2.2 Relation categorization analysis
5.2.3 Relation extraction performance analysis
5.2.4 Error analysis
5.3 Concept map visualization evaluation
5.3.1 Cognitive processing and comprehension performance
5.3.2 User experience and feedback
5.3.3 Technical implementation challenges
5.3.4 Information fidelity in concept map simplification
5.4 Broader discussion
5.4.1 Limitations
5.4.2 Theoretical implications
5.4.3 Future research directions
6 Conclusion
6.1 Summary of findings
6.2 Contributions
6.3 Implications for learning
6.4 Limitations and future directions
6.5 Concluding remarks
A Appendix
A.1 Concept and relation data
A.1.1 Concept categorization by discipline
A.1.2 Relation categorization by discipline
A.2 Extraction methodology
A.2.1 Concept extraction prompt
A.2.2 Concept linking prompt
A.2.3 Local relation extraction prompt
A.2.4 Global relation extraction prompt
Bibliography
About this Honors Thesis
| School | |
|---|---|
| Department | |
| Degree | |
| Submission | |
| Language |
|
| Research Field | |
| Stichwort | |
| Committee Chair / Thesis Advisor | |
| Committee Members |
Primary PDF
| Thumbnail | Title | Date Uploaded | Actions |
|---|---|---|---|
|
|
Reducing Cognitive Load in Digital Reading: An LLM-Powered Approach for Universal Reading Comprehension () | 2025-04-08 15:18:54 -0400 |
|
Supplemental Files
| Thumbnail | Title | Date Uploaded | Actions |
|---|