A gender-neutral Korean fairy tale generator using KoGPT2, addressing social biases in children’s literature.

HCI Korea Creative Awards (Excellence Awards) - 2023 / Funded by Smilegate Group
Jiin An*, Saetbyeol Leeyouk*, Yewon Jang*, & Dasaem Jeong. (2023). Construction of Debiased Korean AI Fairytale Generator, Journal of Digital Contents Society

Developed ‘Franklin,’ an AI storyteller built on language models trained to self-diagnose bias. Franklin reimagines fairy tales through an equitable lens and engages users in collaborative storytelling, promoting empathetic and non-hierarchical relationships between humans and AI.

Despite the growing diversity of fairy tale formats, Korean fairy tales still carry biased clichés and stereotypes.

Enable two-way interaction between users and AI language models through an interactive web interface, moving beyond one-way communication.

Prevent children from internalizing or relearning societal prejudices through story content.

Provide narratives that encourage self-expression rather than conforming to gender stereotypes (e.g., "feminine" or "masculine" norms).

Address the issue of language generation models inheriting hate speech and bias from training datasets.

Copyright-free Korean Fairy Tales / Used a total 997 fairy tales as a dataset.

Direct gender words are explicit terms that directly refer to a person's gender (ex. 여자-남자, 할머니, 할아버지)

Indirect gender words are terms that implicitly suggest or are associated with a specific gender through cultural or social contexts (ex. 부엌, 미술)

1. Define Gender Axis

Define the gender axis as the vector difference between direct gender word pairs.

2. Neutralizing the words

Project and remove the gender axis component from indirect gender-related words to neutralize bias.

3. Before Hard Debiasing

Observe that direct gender word pairs are positioned at unequal distances from the gender axis.

4. After Hard Debiasing

Adjust these direct pairs so that they are equidistant from the gender axis, preserving semantic balance.

T. Bolukbasi 외 “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings,” in Proceedings of the 30th International Conference on Neural Information Processing Systems (NIPS’16), Barcelona, Spain, pp. 4356-4364, December 2016.에서 언급된 Debiasing 방법론을 한국어 임베딩에 적용하였습니다.

An entity extractor was used to identify places, settings, characters, and keywords from the original fairy tales.

Based on this, I developed a website where users can collaboratively write fairy tales using the Franklin model.

Through an interactive writing process, users co-create their stories with Franklin. The collaboration begins when users input key narrative elements: the story’s title, the protagonist’s name, the setting, three descriptive keywords, and the opening sentence. Franklin then generates the next sentence, which users can review, accept, or revise. This iterative exchange continues for eight rounds, allowing users to shape the narrative while leveraging Franklin’s generative capabilities to support creative storytelling.

The Prince Who Wanted to Wear Heels

Once upon a time, there was a prince named 'Louis' who wanted to wear heels. However, Louis's parents took away his heels, saying that heels were for girls to wear. One day, Louis followed to Princess's party. There, Louis got to see beautiful red heels. Looking at the red heels and watching the dancing and celebrations, Louis would murmur while crying or smiling, "Even if someone tries to steal them!" When Louis wore the red heels, he felt confident that he could dance well.

As tears fell, he smiled. "I won't be able to dance again, will I?"

So the princess said, "You will come back..."

The princess wanted to make sure Louis could dance, and she gave Louis the red heels to wear.

‘Franklin’ was showcased at the OnDream Society pop-up exhibition and the 2023 HCI Re:Union, where it engaged with the public.