Authentic Folk-Tale Story Generator

ISEF Category: Technology Enhances the Arts

Subcategory: Games  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

AI can remix a story in seconds, but a bad remix can flatten a culture. That makes folktales a tough test case. You can ask whether a story generator keeps the spirit of a tradition, not just the plot. That turns a cool app idea into a real research question.

What Is It?

This project studies whether an AI story generator can create interactive folk-tale games that stay faithful to cultural traditions. Think of it like making a remix, but with strict rules. You are not only asking if the model can tell a story, you are asking if it can keep the right characters, motifs, and patterns from the source tales.

The key idea is retrieval-augmented generation. That means the model first pulls facts or story pieces from a curated library, then writes from that material. In this case, the library comes from public-domain folktales, grouped with the Aarne-Thompson index, which is a system folklorists use to classify tale types. You can also constrain motifs, which are recurring story parts like tricksters, magical helpers, or impossible tasks, so the output stays closer to the tradition you want to study.

Why This Is a Good Topic

This is a strong science fair topic because you can test it in a clear, measurable way. You can compare story outputs with and without retrieval, with and without motif constraints, and with different cultural source sets. The project connects to real problems in AI ethics, digital storytelling, and cultural representation, and you can measure the results with reader ratings, classification accuracy, or motif-matching scores.

Research Questions

• How does retrieval from a curated folktale corpus affect cultural-fidelity ratings from readers?
• What is the effect of motif constraints on the consistency of story structure across generated tales?
• Does adding the Aarne-Thompson tale type as context improve the match between generated and source-story themes?
• To what extent do native readers rate retrieval-augmented stories as more authentic than unconstrained stories?
• Which type of prompt setup produces the highest motif preservation without lowering story readability?
• How does the size of the retrieved folktale set change the balance between originality and cultural fidelity?

Basic Materials

• Laptop or desktop computer with internet access.
• Free large language model access or a local open-source model.
• Curated public-domain folktale corpus.
• Aarne-Thompson tale-type reference list.
• Spreadsheet software for logging story outputs and ratings.
• Survey form tool for reader feedback.
• Basic rubric for cultural-fidelity scoring.
• Text editor for prompt drafts and experiment notes.

Advanced Materials

• High-performance workstation or university computing access.
• Python environment with retrieval libraries.
• Vector database or embedding index for document retrieval.
• Open-source language model with prompt control.
• Annotation platform for tale-type and motif tagging.
• Statistical analysis software or Python packages for interrater reliability and significance testing.
• Fine-grained rubric for motif preservation, narrative coherence, and cultural fidelity.
• Human participant survey system with secure data storage.

Software & Tools

• Python: Organizes the retrieval pipeline, scoring scripts, and data analysis.
• Jupyter Notebook: Lets you test prompts, inspect outputs, and keep your workflow organized.
• Google Forms: Collects reader ratings on cultural fidelity, clarity, and story quality.
• ImageJ: Not needed for text generation, so skip it unless you also study visual story assets.
• R: Runs statistical tests, agreement scores, and comparison plots for your reader data.

Experiment Steps

1. Define one cultural question you want to test, such as whether retrieval improves authenticity more than prompting alone.
2. Build a small, curated corpus of public-domain folktales and tag each tale with type and motif labels.
3. Design at least two generation conditions, one with retrieval and motif constraints, and one baseline without them.
4. Create a rating rubric that asks readers to judge fidelity, coherence, and cultural fit separately.
5. Plan a comparison method that scores both human ratings and text features, such as motif overlap or tale-type match.
6. Decide how you will check whether your results hold across different cultures, tale types, or prompt styles.

Common Pitfalls

• Using a corpus with mixed or poorly labeled tale origins, which makes authenticity claims weak.
• Treating cultural fidelity as one vague score instead of separate ratings for motifs, voice, and structure.
• Asking only one or two readers for feedback, which makes the results too noisy to trust.
• Letting the model train on or copy modern retellings, which can blur the public-domain source material.
• Comparing stories without controlling prompt length, tale type, or retrieval size, which confounds the results.

What Makes This Competitive

A stronger project would test more than one culture or tale family, then compare how well the system preserves each one. You can add a careful scoring rubric, blind raters, and agreement checks so your data means more than a simple opinion poll. A very strong entry would also compare multiple retrieval strategies or motif constraints, then analyze where the model succeeds and where it distorts the source traditions.

Project Variations

• Test whether the same generator works better for trickster tales than for quest tales.
• Compare reader ratings for stories generated from one culture's folktales versus stories built from mixed-culture corpora.
• Measure whether explicit motif constraints improve cultural fidelity more than longer prompts alone.

Learn More

• PubMed: Search for review articles on human evaluation, narrative generation, and cultural bias in AI text systems.
• NIH PubMed Central: Find free full-text papers on retrieval-augmented generation and story evaluation methods.
• Aarne-Thompson-Uther Index: Look up tale-type classification references through university folklore collections and library guides.
• University OpenCourseWare: Search for courses on natural language processing, information retrieval, and computational linguistics.
• Project Gutenberg: Find public-domain folktales and fairy tales for building a curated corpus.
• ACL Anthology: Search for peer-reviewed papers on story generation, retrieval-augmented generation, and evaluation metrics.