@inproceedings{zhao-etal-2026-ccstories,
title = "{CCS}tories: Improving the Safety and Age-Appropriateness of {LLM}-Generated {C}hinese Children{'}s Stories",
author = {Zhao, Jiaxu and
Zhang, Qin and
Wang, Xiaoxing and
Ding, Liang and
Pei, Yulong and
Chen, Xiaojun and
Fang, Meng and
K{\"a}ser, Tanja and
Pechenizkiy, Mykola},
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1905/",
pages = "43410--43429",
ISBN = "979-8-89176-459-0",
abstract = "Children{'}s stories play a crucial role in early cognitive and language development, yet existing story generation models lack the safety and age-appropriateness required for educational applications, especially in non-English languages. We introduce \textbf{CCStories}, a curated dataset of 22{,}200 high-quality Chinese children{'}s stories designed specifically to support safe and appropriate AI-generated content for young readers. Unlike general-purpose datasets derived from adult-oriented platforms, CCStories targets three critical dimensions of child-appropriate text: readability, syntactic complexity, and toxicity. We further propose a multi-dimensional evaluation framework that jointly assesses appropriateness and generation quality. As a validation of the dataset{'}s utility, we fine-tune five representative backbones: CPM, PanGu-$\alpha$, GLM-4, Qwen, a Chinese-adapted LLaMA-3 variant (Gigi-Llama-3-8B-zh), and additionally benchmark GPT-5-mini as a zero-shot instruction-tuned reference. Fine-tuning reduces offensive-prompt toxicity by up to 26.7 percentage points and raises readability by up to 14.3 percentage points, narrowing much of the gap to the GPT-5-mini reference. A human evaluation confirms that appropriateness rises from 53{\%} to 86{\%} after fine-tuning, and human validation of the toxicity classifier shows that the model ranking it induces matches the human ranking. CCStories provides a publicly available resource for building and benchmarking safer, more age-appropriate Chinese generative models for children{'}s education."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhao-etal-2026-ccstories">
<titleInfo>
<title>CCStories: Improving the Safety and Age-Appropriateness of LLM-Generated Chinese Children’s Stories</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jiaxu</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qin</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoxing</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liang</namePart>
<namePart type="family">Ding</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yulong</namePart>
<namePart type="family">Pei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaojun</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meng</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tanja</namePart>
<namePart type="family">Käser</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mykola</namePart>
<namePart type="family">Pechenizkiy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Children’s stories play a crucial role in early cognitive and language development, yet existing story generation models lack the safety and age-appropriateness required for educational applications, especially in non-English languages. We introduce CCStories, a curated dataset of 22,200 high-quality Chinese children’s stories designed specifically to support safe and appropriate AI-generated content for young readers. Unlike general-purpose datasets derived from adult-oriented platforms, CCStories targets three critical dimensions of child-appropriate text: readability, syntactic complexity, and toxicity. We further propose a multi-dimensional evaluation framework that jointly assesses appropriateness and generation quality. As a validation of the dataset’s utility, we fine-tune five representative backbones: CPM, PanGu-α, GLM-4, Qwen, a Chinese-adapted LLaMA-3 variant (Gigi-Llama-3-8B-zh), and additionally benchmark GPT-5-mini as a zero-shot instruction-tuned reference. Fine-tuning reduces offensive-prompt toxicity by up to 26.7 percentage points and raises readability by up to 14.3 percentage points, narrowing much of the gap to the GPT-5-mini reference. A human evaluation confirms that appropriateness rises from 53% to 86% after fine-tuning, and human validation of the toxicity classifier shows that the model ranking it induces matches the human ranking. CCStories provides a publicly available resource for building and benchmarking safer, more age-appropriate Chinese generative models for children’s education.</abstract>
<identifier type="citekey">zhao-etal-2026-ccstories</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1905/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43410</start>
<end>43429</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CCStories: Improving the Safety and Age-Appropriateness of LLM-Generated Chinese Children’s Stories
%A Zhao, Jiaxu
%A Zhang, Qin
%A Wang, Xiaoxing
%A Ding, Liang
%A Pei, Yulong
%A Chen, Xiaojun
%A Fang, Meng
%A Käser, Tanja
%A Pechenizkiy, Mykola
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F zhao-etal-2026-ccstories
%X Children’s stories play a crucial role in early cognitive and language development, yet existing story generation models lack the safety and age-appropriateness required for educational applications, especially in non-English languages. We introduce CCStories, a curated dataset of 22,200 high-quality Chinese children’s stories designed specifically to support safe and appropriate AI-generated content for young readers. Unlike general-purpose datasets derived from adult-oriented platforms, CCStories targets three critical dimensions of child-appropriate text: readability, syntactic complexity, and toxicity. We further propose a multi-dimensional evaluation framework that jointly assesses appropriateness and generation quality. As a validation of the dataset’s utility, we fine-tune five representative backbones: CPM, PanGu-α, GLM-4, Qwen, a Chinese-adapted LLaMA-3 variant (Gigi-Llama-3-8B-zh), and additionally benchmark GPT-5-mini as a zero-shot instruction-tuned reference. Fine-tuning reduces offensive-prompt toxicity by up to 26.7 percentage points and raises readability by up to 14.3 percentage points, narrowing much of the gap to the GPT-5-mini reference. A human evaluation confirms that appropriateness rises from 53% to 86% after fine-tuning, and human validation of the toxicity classifier shows that the model ranking it induces matches the human ranking. CCStories provides a publicly available resource for building and benchmarking safer, more age-appropriate Chinese generative models for children’s education.
%U https://aclanthology.org/2026.emnlp-main.1905/
%P 43410-43429
Markdown (Informal)
[CCStories: Improving the Safety and Age-Appropriateness of LLM-Generated Chinese Children’s Stories](https://aclanthology.org/2026.emnlp-main.1905/) (Zhao et al., EMNLP 2026)
ACL
- Jiaxu Zhao, Qin Zhang, Xiaoxing Wang, Liang Ding, Yulong Pei, Xiaojun Chen, Meng Fang, Tanja Käser, and Mykola Pechenizkiy. 2026. CCStories: Improving the Safety and Age-Appropriateness of LLM-Generated Chinese Children’s Stories. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 43410–43429, Budapest, Hungary. Association for Computational Linguistics.