📝AI-driven Interactive Hierarchical Concept Maps
Traditional concept maps lack interactivity and scalability, necessitating enhancements with AI capabilities and interactive features in digital learning environments.
The proposed system leverages artificial intelligence to generate initial concept maps, refined through human-in-the-loop input for pedagogical accuracy and relevance.
A custom-designed interface enables learners to explore concepts at varying levels of depth, promoting active engagement and quick navigation.
Numerous studies have highlighted the pivotal role of visualization in enhancing the comprehension and retention of educational content.
Early work by Naps et al. emphasized visual engagement, particularly in computer science education, contributes significantly to student understanding and motivation.
The CoMPASS project and ActiveMath platform introduced navigable and adaptive visual structures to improve structural understanding and navigation efficiency.
Hollingsworth and Narayanan argue that interactive features like concept maps should be standard components of digital textbooks.
Barria-Pineda et al. developed visualization tools supporting self-regulated learning via concept-level mapping.
In ontology-oriented learning systems, concept maps serve as structured, navigable knowledge representations, exemplified by TM4L.
Many studies have shown that concept maps improve learning outcomes, particularly for complex material.
A meta-analysis by Schroeder et al. reviewed 142 studies and found concept maps significantly more effective than reading texts (g = .39) or reviewing outlines (g = .28).
Bolatli and Bolatli reported higher post-test scores and lower cognitive load in anatomy students using predefined concept maps.
Elgendi and Shaffer demonstrated interactive glossary maps increased student engagement and repeated glossary use in a computer science e-textbook.
Schwab et al. introduced booc.io, a system with drill-down hierarchical concept maps supporting adaptive navigation and targeted feedback.
Ma and Chen proposed an LLM framework for automated concept map construction from e-books, including segmentation, key concept extraction, and relationship identification.
Their evaluation of GPT-4o demonstrated strong performance, effectively extracting key concepts and accurately identifying hierarchical and cross-topic connections.
LLMs could generate concept maps that differ from textbook structure, reflecting a more logical and content-based organization.
Kluga et al. integrated causal concept maps into an intelligent anatomy textbook for personalized navigation, quiz adaptation, and content feedback.
Wehnert et al. presented a dynamic visualization system for exploring concept hierarchies extracted from legal textbooks, supporting various navigation modes.
This project explores the implementation and evaluation of an interactive, hierarchical concept map platform designed to enhance digital learning experiences.
The interface introduces a layered, LLM-augmented design supporting on-demand concept generation, infinite drill-down navigation, and in-context information display.
Our system, similar to Ma and Chen, uses LLMs for concept map generation but addresses interactive UI/UX design and cognitive overload.
| Aspect | Ma and Chen System | Proposed System |
|---|---|---|
| LLM Use | Automates extraction of concepts and relationships | Employs LLMs to generate concept maps |
| Real Educational Context | Does not address use in real educational contexts | Addresses use with validation based on real student feedback |
| Interactive UI/UX Design | Lacks production-ready interactive UI/UX design | Focuses on production-ready interactive UI/UX design |
| Cognitive Overload | Overlooks cognitive overload from large, flat maps | Addresses via hierarchical drill-down navigation and embedded pedagogical content |
| Content Alignment | Uses full course materials | Plans to incorporate RAG in future iterations |
The system was applied in Object-Oriented Programming and Data Structures, and UI Design and AI-Assisted Frontend Development university courses.
The Main Map of the Course functioned as a central, interactive overview of the entire curriculum, visually organizing key topics and their relationships.
Each concept node could be expanded to explore subtopics, allowing students to see the “big picture” and access detailed content.
The Information Panel provided contextual details for each concept selected within the map, displaying concise explanations, examples, or supporting materials.
This feature allowed learners to engage with the content without leaving the map interface, maintaining flow and minimizing distractions.
A Child Concept Map is a secondary, more focused map that expands upon a specific node from the main concept map.
Child maps support hierarchical learning by allowing users to progressively explore concepts at increasing levels of depth without overwhelming them.
The Infinite Drill-Down AI-Based Domain Exploration feature enables learners to move beyond predefined content by dynamically generating new sub-maps using AI.
When a student reaches a terminal node and seeks further explanation, the system can generate an extended concept map based on the topic's semantic description.
This feature transforms the concept map into an adaptive learning environment capable of supporting personalized, self-directed exploration across limitless depth within a subject domain.
The system generates interactive concept maps by querying LLMs, such as GPT-4o-mini, based on provided course metadata and prompt-tuning commands.
The end-to-end concept map development process includes eight distinct steps, from course creation and initial generation to refinement and nested map creation.
The instructor defines a new course by providing a course title, a free-form course description, and optional prompt-tuning instructions.
The system queries the LLM to produce a draft map of key concepts and their relationships.
If the initial map is unsatisfactory, instructors can clear it and regenerate a new one from scratch.
Users can ask a secondary LLM (e.g., ChatGPT) to rephrase or restructure parts of the map for better clarity or alignment.
Clicking a concept opens an interactive panel that fetches descriptive content via the LLM.
Instructors can discard and regenerate concept information if needed.
Generated descriptions can be edited or enhanced using a separate LLM interface (e.g., ChatGPT).
Instructors can expand individual concepts by generating subordinate maps, enabling drill-down exploration.
The refinement process requires manual review of each map level and concept description to validate pedagogical relevance and accuracy.
For highly customized courses, the refinement process can be time-consuming; however, AI-assisted generation significantly reduces the baseline effort.
Several limitations emerge from relying solely on large language models (LLMs) without incorporating course-specific materials.
Initial concept maps are generated based only on a course title and concise metadata, risking shallow or misaligned content in specialized domains.
Future work will explore integrating retrieval-augmented generation (RAG) techniques to enable concept map generation based directly on actual course content.
Although infinite drill-down capability is innovative, it also poses challenges like disorientation and topic drift.
As users delve deeper into nested submaps, they may lose awareness of their location within the overall concept structure.
In the absence of clear semantic boundaries, LLMs may generate tangential or unrelated subtopics, resulting in conceptual divergence.
Future work will explore introducing depth limits and developing UX strategies for clearer visual cues to help users maintain orientation.
Because submaps are generated independently, a concept node on a higher-level map may not reproduce its child nodes when opened in a dedicated submap.
Scalable resolution for submap consistency will require algorithmic support and improved tooling across the concept hierarchy.
The system features a lightweight, modular architecture that combines AI-driven content generation with interactive frontend visualization.
The Python backend handles HTTP requests, interacts with the OpenAI GPT-4o-mini API, and stores course content in a file-based structure.
The frontend uses Vis.js to render dynamic, clickable concept maps, enabling content display and drill-down navigation.
Concept maps are generated by dynamically constructing and sending prompts to an LLM, such as GPT-4o-mini, driven by frontend input and course metadata.
The prompt instructs the LLM to generate at least 15 nodes, define a single root concept, and create a tree-like hierarchical structure without cycles or dangling nodes.
The prompt also requires formatting output as Vis.js code, labeling all edges with clear semantic relationships, ensuring domain-specific accuracy, and including a concrete output example.
The LLM response is cached as a plain text file, serving subsequent requests directly from the local file system to avoid redundant OpenAI API calls.
The system generates individual concept pages as HTML fragments through targeted prompt engineering when a user selects a concept node.
The prompt instructs the model to generate valid HTML fragments using semantic tags, format code snippets with pre/code, and mathematical content with LaTeX/MathJax.
The Content Management Panel is a core instructor-facing tool enabling efficient oversight of all AI-generated concept map content.
Instructors can create/manage courses, regenerate/refine maps, and customize individual concept pages using third-party LLMs for better alignment.
This section presents the design and results of a student survey evaluating the perceived effectiveness of AI-powered concept map interfaces.
Hierarchical concept maps were developed for two courses at American University Kyiv, and a survey was administered to students.
Concept maps were developed for Object-Oriented Programming and Data Structures for first-year Bachelor of Software Engineering/AI and Data Science students.
42 concept maps and 117 individual concept pages were generated and integrated into Course 1.
24 students participated in the survey for Course 1.
Concept maps were developed for UI Design and AI-Assisted Frontend Development for first-year Bachelor of Software Engineering and AI program students.
27 concept maps and 128 concept pages were generated and refined for Course 2.
11 students participated in the survey for Course 2.
A total of 35 surveys were received from both courses.
The survey was proposed at the final part of the course as a means of recalling course content and evaluating the concept map’s effectiveness for structured review.
Questions assessed overall satisfaction, engagement, effectiveness for review/learning, and usefulness of drill-down navigation.
The analysis shows high student satisfaction and perceived effectiveness, quantitatively confirming that the interactive and visual structure of concept maps was well-received.
Students reported an average satisfaction score of 8.91 out of 10 with their learning experience using the concept map app.
Students rated the app’s usefulness for learning new content at an average of 8.31 out of 10.
A large majority (91.4%) of students agreed that the app made learning more engaging or enjoyable.
100% of students confirmed that the ability to click on concept names and navigate through nested maps was helpful.
Students expressed varying preferences for reviewing course material using interactive concept maps versus traditional materials.
60% of students found both interactive concept maps and traditional materials equally effective for review and recall.
34.3% of students preferred interactive concept maps for reviewing course material.
Only 5.7% of students preferred text-based materials alone for reviewing course material.
Open-ended responses revealed clickability/depth navigation and visual structure/relationships between topics as major advantages.
11 students mentioned appreciating the ability to click, navigate, and explore deeper levels of topics.
9 students highlighted the clarity and usefulness of the visual concept map structure and topic relationships.
Students suggested improvements related to UI, current interactions, and additional functionalities.
7 students felt the app was already perfect or good enough, stating “everything is perfect” or “all is good”.
8 students recommended UI improvements, including design changes for a more pleasant visual side and a dark theme.
6 students recommended improvements in current interactions, citing issues with dragging, zooming, and scrolling the graph.
Various individual ideas were suggested, such as online collaboration, shared comments, quizzes, and the ability to ask questions.
High ratings across both courses suggest AI-generated maps are relevant for learning, and the hierarchical drill-down structure offers an effective human-computer interaction pattern.
This study presented the design and implementation of an AI-driven system for generating interactive hierarchical concept maps for digital learning environments and intelligent textbooks.
The system positions itself as a practical implementation of the intelligent textbook paradigm, augmenting content delivery and navigational structure for student-centered learning.
Evaluation results, based on 69 AI-generated concept maps and 245 unique concept descriptions across two undergraduate computer science courses, confirm the system's educational value.
Feedback from 35 student surveys underscored the potential of integrating AI-generated maps into intelligent textbooks and online courses to enhance comprehension, engagement, and structured review.
Future work will address current limitations identified in the system’s architecture and usage, focusing on several key directions.
One key direction is the integration of retrieval-augmented generation (RAG) techniques to allow the LLM to incorporate course-specific materials into the generation process.
To improve usability and mitigate disorientation, future enhancements will include research on depth-limiting mechanisms and UX improvements for visual context and navigational cues.
Algorithmic methods and instructor-facing tools will be developed to support structural validation and map coherence across different levels of hierarchy.
The refinement process will be more tightly integrated into the instructor-facing UI for seamless editing and regeneration of maps and concept pages within the same workflow.
The admin dashboard will be extended to provide broader control and monitoring of content quality.
Additional research will explore the system’s effectiveness in other domains such as mathematics, management, and postgraduate software engineering education.
The implementation of agentic AI workflows, where multiple autonomous agents coordinate multi-step concept map construction and refinement, will be explored.
The author thanks American University Kyiv instructors Roman Tymoshuk, Andrii Tsabanov, and Ivan Danilov for reviewing and integrating the concept maps into the learning process.
The author used ChatGPT-4 and Grammarly to check grammar, spelling, improve writing style, paraphrase, and reword, taking full responsibility for the content.
Traditional concept maps lack interactivity and scalability, necessitating enhancements with AI capabilities and interactive features in digital learning environments.
The proposed system leverages artificial intelligence to generate initial concept maps, refined through human-in-the-loop input for pedagogical accuracy and relevance.
A custom-designed interface enables learners to explore concepts at varying levels of depth, promoting active engagement and quick navigation.
Numerous studies have highlighted the pivotal role of visualization in enhancing the comprehension and retention of educational content.
Early work by Naps et al. emphasized visual engagement, particularly in computer science education, contributes significantly to student understanding and motivation.
The CoMPASS project and ActiveMath platform introduced navigable and adaptive visual structures to improve structural understanding and navigation efficiency.
Hollingsworth and Narayanan argue that interactive features like concept maps should be standard components of digital textbooks.
Barria-Pineda et al. developed visualization tools supporting self-regulated learning via concept-level mapping.
In ontology-oriented learning systems, concept maps serve as structured, navigable knowledge representations, exemplified by TM4L.
Many studies have shown that concept maps improve learning outcomes, particularly for complex material.
A meta-analysis by Schroeder et al. reviewed 142 studies and found concept maps significantly more effective than reading texts (g = .39) or reviewing outlines (g = .28).
Bolatli and Bolatli reported higher post-test scores and lower cognitive load in anatomy students using predefined concept maps.
Elgendi and Shaffer demonstrated interactive glossary maps increased student engagement and repeated glossary use in a computer science e-textbook.
Schwab et al. introduced booc.io, a system with drill-down hierarchical concept maps supporting adaptive navigation and targeted feedback.
Ma and Chen proposed an LLM framework for automated concept map construction from e-books, including segmentation, key concept extraction, and relationship identification.
Their evaluation of GPT-4o demonstrated strong performance, effectively extracting key concepts and accurately identifying hierarchical and cross-topic connections.
LLMs could generate concept maps that differ from textbook structure, reflecting a more logical and content-based organization.
Kluga et al. integrated causal concept maps into an intelligent anatomy textbook for personalized navigation, quiz adaptation, and content feedback.
Wehnert et al. presented a dynamic visualization system for exploring concept hierarchies extracted from legal textbooks, supporting various navigation modes.
This project explores the implementation and evaluation of an interactive, hierarchical concept map platform designed to enhance digital learning experiences.
The interface introduces a layered, LLM-augmented design supporting on-demand concept generation, infinite drill-down navigation, and in-context information display.
Our system, similar to Ma and Chen, uses LLMs for concept map generation but addresses interactive UI/UX design and cognitive overload.
| Aspect | Ma and Chen System | Proposed System |
|---|---|---|
| LLM Use | Automates extraction of concepts and relationships | Employs LLMs to generate concept maps |
| Real Educational Context | Does not address use in real educational contexts | Addresses use with validation based on real student feedback |
| Interactive UI/UX Design | Lacks production-ready interactive UI/UX design | Focuses on production-ready interactive UI/UX design |
| Cognitive Overload | Overlooks cognitive overload from large, flat maps | Addresses via hierarchical drill-down navigation and embedded pedagogical content |
| Content Alignment | Uses full course materials | Plans to incorporate RAG in future iterations |
The system was applied in Object-Oriented Programming and Data Structures, and UI Design and AI-Assisted Frontend Development university courses.
The Main Map of the Course functioned as a central, interactive overview of the entire curriculum, visually organizing key topics and their relationships.
Each concept node could be expanded to explore subtopics, allowing students to see the “big picture” and access detailed content.
The Information Panel provided contextual details for each concept selected within the map, displaying concise explanations, examples, or supporting materials.
This feature allowed learners to engage with the content without leaving the map interface, maintaining flow and minimizing distractions.
A Child Concept Map is a secondary, more focused map that expands upon a specific node from the main concept map.
Child maps support hierarchical learning by allowing users to progressively explore concepts at increasing levels of depth without overwhelming them.
The Infinite Drill-Down AI-Based Domain Exploration feature enables learners to move beyond predefined content by dynamically generating new sub-maps using AI.
When a student reaches a terminal node and seeks further explanation, the system can generate an extended concept map based on the topic's semantic description.
This feature transforms the concept map into an adaptive learning environment capable of supporting personalized, self-directed exploration across limitless depth within a subject domain.
The system generates interactive concept maps by querying LLMs, such as GPT-4o-mini, based on provided course metadata and prompt-tuning commands.
The end-to-end concept map development process includes eight distinct steps, from course creation and initial generation to refinement and nested map creation.
The instructor defines a new course by providing a course title, a free-form course description, and optional prompt-tuning instructions.
The system queries the LLM to produce a draft map of key concepts and their relationships.
If the initial map is unsatisfactory, instructors can clear it and regenerate a new one from scratch.
Users can ask a secondary LLM (e.g., ChatGPT) to rephrase or restructure parts of the map for better clarity or alignment.
Clicking a concept opens an interactive panel that fetches descriptive content via the LLM.
Instructors can discard and regenerate concept information if needed.
Generated descriptions can be edited or enhanced using a separate LLM interface (e.g., ChatGPT).
Instructors can expand individual concepts by generating subordinate maps, enabling drill-down exploration.
The refinement process requires manual review of each map level and concept description to validate pedagogical relevance and accuracy.
For highly customized courses, the refinement process can be time-consuming; however, AI-assisted generation significantly reduces the baseline effort.
Several limitations emerge from relying solely on large language models (LLMs) without incorporating course-specific materials.
Initial concept maps are generated based only on a course title and concise metadata, risking shallow or misaligned content in specialized domains.
Future work will explore integrating retrieval-augmented generation (RAG) techniques to enable concept map generation based directly on actual course content.
Although infinite drill-down capability is innovative, it also poses challenges like disorientation and topic drift.
As users delve deeper into nested submaps, they may lose awareness of their location within the overall concept structure.
In the absence of clear semantic boundaries, LLMs may generate tangential or unrelated subtopics, resulting in conceptual divergence.
Future work will explore introducing depth limits and developing UX strategies for clearer visual cues to help users maintain orientation.
Because submaps are generated independently, a concept node on a higher-level map may not reproduce its child nodes when opened in a dedicated submap.
Scalable resolution for submap consistency will require algorithmic support and improved tooling across the concept hierarchy.
The system features a lightweight, modular architecture that combines AI-driven content generation with interactive frontend visualization.
The Python backend handles HTTP requests, interacts with the OpenAI GPT-4o-mini API, and stores course content in a file-based structure.
The frontend uses Vis.js to render dynamic, clickable concept maps, enabling content display and drill-down navigation.
Concept maps are generated by dynamically constructing and sending prompts to an LLM, such as GPT-4o-mini, driven by frontend input and course metadata.
The prompt instructs the LLM to generate at least 15 nodes, define a single root concept, and create a tree-like hierarchical structure without cycles or dangling nodes.
The prompt also requires formatting output as Vis.js code, labeling all edges with clear semantic relationships, ensuring domain-specific accuracy, and including a concrete output example.
The LLM response is cached as a plain text file, serving subsequent requests directly from the local file system to avoid redundant OpenAI API calls.
The system generates individual concept pages as HTML fragments through targeted prompt engineering when a user selects a concept node.
The prompt instructs the model to generate valid HTML fragments using semantic tags, format code snippets with pre/code, and mathematical content with LaTeX/MathJax.
The Content Management Panel is a core instructor-facing tool enabling efficient oversight of all AI-generated concept map content.
Instructors can create/manage courses, regenerate/refine maps, and customize individual concept pages using third-party LLMs for better alignment.
This section presents the design and results of a student survey evaluating the perceived effectiveness of AI-powered concept map interfaces.
Hierarchical concept maps were developed for two courses at American University Kyiv, and a survey was administered to students.
Concept maps were developed for Object-Oriented Programming and Data Structures for first-year Bachelor of Software Engineering/AI and Data Science students.
42 concept maps and 117 individual concept pages were generated and integrated into Course 1.
24 students participated in the survey for Course 1.
Concept maps were developed for UI Design and AI-Assisted Frontend Development for first-year Bachelor of Software Engineering and AI program students.
27 concept maps and 128 concept pages were generated and refined for Course 2.
11 students participated in the survey for Course 2.
A total of 35 surveys were received from both courses.
The survey was proposed at the final part of the course as a means of recalling course content and evaluating the concept map’s effectiveness for structured review.
Questions assessed overall satisfaction, engagement, effectiveness for review/learning, and usefulness of drill-down navigation.
The analysis shows high student satisfaction and perceived effectiveness, quantitatively confirming that the interactive and visual structure of concept maps was well-received.
Students reported an average satisfaction score of 8.91 out of 10 with their learning experience using the concept map app.
Students rated the app’s usefulness for learning new content at an average of 8.31 out of 10.
A large majority (91.4%) of students agreed that the app made learning more engaging or enjoyable.
100% of students confirmed that the ability to click on concept names and navigate through nested maps was helpful.
Students expressed varying preferences for reviewing course material using interactive concept maps versus traditional materials.
60% of students found both interactive concept maps and traditional materials equally effective for review and recall.
34.3% of students preferred interactive concept maps for reviewing course material.
Only 5.7% of students preferred text-based materials alone for reviewing course material.
Open-ended responses revealed clickability/depth navigation and visual structure/relationships between topics as major advantages.
11 students mentioned appreciating the ability to click, navigate, and explore deeper levels of topics.
9 students highlighted the clarity and usefulness of the visual concept map structure and topic relationships.
Students suggested improvements related to UI, current interactions, and additional functionalities.
7 students felt the app was already perfect or good enough, stating “everything is perfect” or “all is good”.
8 students recommended UI improvements, including design changes for a more pleasant visual side and a dark theme.
6 students recommended improvements in current interactions, citing issues with dragging, zooming, and scrolling the graph.
Various individual ideas were suggested, such as online collaboration, shared comments, quizzes, and the ability to ask questions.
High ratings across both courses suggest AI-generated maps are relevant for learning, and the hierarchical drill-down structure offers an effective human-computer interaction pattern.
This study presented the design and implementation of an AI-driven system for generating interactive hierarchical concept maps for digital learning environments and intelligent textbooks.
The system positions itself as a practical implementation of the intelligent textbook paradigm, augmenting content delivery and navigational structure for student-centered learning.
Evaluation results, based on 69 AI-generated concept maps and 245 unique concept descriptions across two undergraduate computer science courses, confirm the system's educational value.
Feedback from 35 student surveys underscored the potential of integrating AI-generated maps into intelligent textbooks and online courses to enhance comprehension, engagement, and structured review.
Future work will address current limitations identified in the system’s architecture and usage, focusing on several key directions.
One key direction is the integration of retrieval-augmented generation (RAG) techniques to allow the LLM to incorporate course-specific materials into the generation process.
To improve usability and mitigate disorientation, future enhancements will include research on depth-limiting mechanisms and UX improvements for visual context and navigational cues.
Algorithmic methods and instructor-facing tools will be developed to support structural validation and map coherence across different levels of hierarchy.
The refinement process will be more tightly integrated into the instructor-facing UI for seamless editing and regeneration of maps and concept pages within the same workflow.
The admin dashboard will be extended to provide broader control and monitoring of content quality.
Additional research will explore the system’s effectiveness in other domains such as mathematics, management, and postgraduate software engineering education.
The implementation of agentic AI workflows, where multiple autonomous agents coordinate multi-step concept map construction and refinement, will be explored.
The author thanks American University Kyiv instructors Roman Tymoshuk, Andrii Tsabanov, and Ivan Danilov for reviewing and integrating the concept maps into the learning process.
The author used ChatGPT-4 and Grammarly to check grammar, spelling, improve writing style, paraphrase, and reword, taking full responsibility for the content.