ILM: An AI-Powered Storytelling Educational Tool

1 Introduction

• Digital storytelling of prophetic narratives can support character formation, motivation, and engagement when stories are well designed and appropriately mediated by adults.The Qur’an also uses narrative to convey guidance and lessons, highlighting its pedagogical significance.Source: §1 · p. 1
• ILM is a storytelling platform for children and young adults that combines reading, visualization, and interactive learning.Source: §1 · p. 1
• Its story map visualizes people, places, and objects and their relationships, while quizzes assess factual knowledge and story comprehension.Source: §1 · p. 2

Existing apps and AI systems

• Apps such as Qissah and Qalam offer 40+ narrated Qur’an and prophet stories with professional voiceover and multilingual support, while other platforms use animation, audio, or games [6] [5] [17] [16].Source: §2 · p. 2
• Conversational systems including ChatGPT and specialized Islamic education systems provide sourced answers grounded in the Qur’an and authenticated hadith.Source: §2 · p. 2

Knowledge graph and RAG systems

• Prior work uses knowledge graphs for Islamic-text search and question answering, narrative-centric story exploration, and retrieval-augmented generation for narrative QA [18].Source: §2 · p. 2
• The paper identifies limited Arabic support and educator control, and presents ILM as supporting Arabic- and English-based exploration, assessment, story-grounded feedback, and educator-controlled content.Source: §2 · p. 2

3.1 Knowledge Graph Constructor Engine

• The KG Engine extracts entities and relations from Arabic paragraphs, using a four-expert NER ensemble, story-scoped entity linking, and a two-stage relation pipeline with NLI-based semantic resolution [3] [12] [1] [10] [2].Entities and relations become structured graph records; paragraph edits trigger retraction and reprocessing of associated records.Source: §3.1 · pp. 2–3

3 System Architecture and Design

• The React portals connect to a FastAPI backend, PostgreSQL and pgvector, a KG Constructor Engine, and a multilingual retrieval pipeline (Figure 1).Source: §3 · p. 2, Figure 1 · p. 3

3.1 Knowledge Graph Constructor Engine

• The educational application uses nine selected predicates from the broader WOJOOD ontology, while a separate, manually source-verified Quran enrichment layer is excluded from deterministic quiz generation [2].Source: §3.1 · p. 3

3.2 Multilingual Retrieval and RAG Pipeline

• The independent retrieval pipeline indexes reviewed, translated narrative paragraphs as multilingual embeddings, then supplies the top three retrieved chunks as context for question generation [9] [11].Post-generation checks verify source support, detect duplicates, and flag potentially redundant questions for manual review.Source: §3.2 · pp. 3–4

3.3 Question Evaluation

• The platform combines KG-based questions with retrieval-based open-ended questions, which use different evaluation strategies because they rely on different knowledge representations.Source: §3.3 · p. 4

3.3.1 Knowledge-Graph-Based Evaluation

• KG questions link to source triples and expected answers, enabling deterministic matching against canonical graph values and registered aliases.Source: §3.3.1 · p. 4

3.3.2 Retrieval-Based Answer Evaluation

• For open-ended answers, Gemini evaluates the learner’s response against retrieved passages and a reference answer, allowing reasonable phrasing differences and applying a benefit-of-the-doubt policy when evidence is insufficient.Source: §3.3.2 · p. 4

4 Conclusion

• ILM combines structured knowledge representation and multilingual retrieval to support learning from Islamic narratives.Source: §4 · p. 4
• The platform connects narrative content with interactive exploration and comprehension assessment while maintaining links to source material.Source: §4 · p. 4
• Automatically extracted relations and generated questions may contain errors; the authors describe approved content, human review, and source-verified Quranic enrichment as mitigations.Source: §4 · p. 4

• The current evaluation is limited to one story, motivating a wider prophetic-narrative corpus and evaluation across different texts.Source: §5 · p. 4
• Formal studies with children and educators are proposed to measure effects on comprehension, engagement, and knowledge retention.Source: §5 · p. 4
• Proposed engagement and assessment extensions include multimedia and activities such as matching characters, ordering events, and completing story maps.Source: §5 · p. 4

A.1 Open-Ended Question Generation Prompt

• Open-ended questions must be answerable from the source alone and must not introduce details absent from it.Source: §A.1 · p. 7
• Each generated question should ask about the story and invite a one- or two-sentence answer in the child’s own words.Source: §A.1 · p. 7
• The generation prompt requests a short expected answer and JSON fields for the prompt, reference answer, and explanation.Source: §A.1 · p. 7

A.2 Open-Answer Judge Prompt

• The judge assesses a child’s answer using only supplied passages and marks it correct when those passages do not resolve correctness.Source: §A.2 · p. 7
• The judge accepts meaning-consistent paraphrases and minor errors; it marks answers incorrect for contradiction, incompatibility, or missing the question’s main point.Source: §A.2 · p. 7
• Feedback should be one short, encouraging sentence, with the judgment and feedback returned as JSON.Source: §A.2 · p. 7

A.3 Shared Block for Question Generation Prompt

• The shared block directs question generation away from details already covered in earlier questions about the same passage.Source: §A.3 · p. 7