SS-GEN: A Social Story Generation Framework with Large Language Models (AAAI 2025 Oral)

🧠 Introduction

SS-GEN is a novel framework for automated generation of Social Stories™ aimed at helping children with Autism Spectrum Disorder (ASD) better understand and navigate social situations. Social Stories have traditionally been written by psychologists under strict guidelines — however, this process is costly, time-consuming, and lacks diversity.

SS-GEN leverages Large Language Models (LLMs) and a custom-designed, constraint-driven strategy (STARSOW) to generate personalized, high-quality Social Stories at scale.

📝 Abstract

Children with Autism Spectrum Disorder (ASD) often struggle to interpret social cues and engage in daily routines. Social Stories™, designed to improve these abilities, are typically handcrafted by experts, limiting their scalability. To address this, we propose SS-GEN, a framework that prompts LLMs to generate constraint-compliant Social Stories using a novel strategy named STARSOW. We further curate a high-quality dataset via human filtering and propose a structured evaluation framework. Finally, we fine-tune lightweight open-source models on our dataset, achieving strong results with lower cost and easier deployment. SS-GEN represents a significant step in creating accessible, affordable, and automated tools to assist ASD communities.

🌳 Framework: STARSOW

The STARSOW pipeline includes:

1. Taking Root: Generating diverse chapter themes from seed examples.
2. Branching Out: Producing multiple Social Story titles under each chapter.
3. Bearing Star Fruits: Completing full stories from titles, guided by strict structural and narrative constraints.
4. Gardening Work: Rigorous filtering to ensure quality, relevance, and safety.

🛠️ Implementation: StarSOW Method

This repository provides a complete implementation of the StarSOW method for generating Social Stories datasets. The implementation is organized into two main components:

📁 Repository Structure

📂 SS-GEN/
├── seed_data_gen/              # Seed dataset generation
│   ├── seed_data/              # Generated seed data
│   ├── seed_ main.py           # Main script for seed data generation
│   ├── openai_access.py        # OpenAI API interface
│   ├── utils.py                # Utility functions
│   └── explain_chapter.py      # Chapter explanation generation
├── hierarchical_instruct/      # StarSOW pipeline implementation
│   ├── bootstrap_chapters_breadth.py      # Step 1: Chapter generation
│   ├── bootstrap_titles_breadth_depth.py  # Step 2: Title generation
│   ├── bootstrap_stories_depth.py         # Step 3: Story generation
│   ├── gpt3_api.py                        # GPT API wrapper
│   ├── prompt_complete_social_story.py    # Story generation prompts
│   ├── data/                              # Generated data and processing
│   ├── visualization/                     # Data visualization tools
│   └── command_generate_story.sh          # Batch generation script
├── SS-GEN Dataset/             # Final dataset
├── README.md
└── Technical Appendix.pdf      # Detailed prompt templates

🚀 Quick Start

Prerequisites

pip install openai tqdm numpy pandas rouge-score datasets

Step 1: Generate Seed Dataset

First, prepare your seed Social Stories data and generate chapter explanations:

cd seed_data_gen
python seed_ main.py

This will:

• Load your original Social Stories data (pure-ori.json)
• Generate chapter explanations using GPT
• Create structured seed data for the StarSOW pipeline

Step 2: Generate Chapters (Taking Root)

Expand the seed chapters to create a diverse set of chapter themes:

cd hierarchical_instruct
python bootstrap_chapters_breadth.py \
    --seed_stories_path ../seed_data_gen/seed_data/seed_chapters_explanations.jsonl \
    --batch_dir data/gpt4_test_generations \
    --num_chapters_to_generate 56 \
    --engine gpt-4o \
    --api_key YOUR_API_KEY \
    --base_url YOUR_BASE_URL

Step 3: Generate Titles (Branching Out)

For each chapter, generate multiple Social Story titles:

python bootstrap_titles_breadth_depth.py \
    --seed_stories_path ../seed_data_gen/seed_data/seed_chapter_title_list.jsonl \
    --input_file machine_generated_chapters_explanations.jsonl \
    --batch_dir data/gpt4_test_generations \
    --num_titles_to_generate 70 \
    --engine gpt-4o \
    --api_key YOUR_API_KEY \
    --base_url YOUR_BASE_URL

Step 4: Generate Stories (Bearing Star Fruits)

Complete the Social Stories from titles using strict structural constraints:

python bootstrap_stories_depth.py \
    --seed_stories_path ../seed_data_gen/seed_data/seed_chapter_story_list.jsonl \
    --input_folder Titles generation \
    --batch_dir data/gpt4_test_generations \
    --output_seed_folder Stories generation/Seed_Titles \
    --output_expand_folder Stories generation/Generated_Titles_from_gpt4 \
    --engine gpt-4o \
    --api_key YOUR_API_KEY \
    --base_url YOUR_BASE_URL

Step 4.1: Batch Generation (Optional)

Use the provided shell script for automated batch processing:

bash command_generate_story.sh

🔧 Configuration

API Configuration

Update your API credentials in the respective scripts:

# In gpt3_api.py or individual scripts
api_key = "your-openai-api-key"
base_url = "your-api-base-url"  # Optional, for custom endpoints

Generation Parameters

Key parameters you can adjust:

• --num_chapters_to_generate: Number of chapters to generate (default: 56)
• --num_titles_to_generate: Titles per chapter (default: 70)
• --num_prompt_demonstrations: Few-shot examples in prompts (default: 4-10)
• --engine: LLM model (gpt-4o, gpt-3.5-turbo, etc.)

📊 Data Processing and Visualization

The hierarchical_instruct/data/ folder contains:

• Generated chapters, titles, and stories
• Processing scripts for data cleaning

The hierarchical_instruct/visualization/ folder includes:

• Data distribution analysis

🎯 Quality Control

The implementation includes several quality control mechanisms:

1. Rouge-based Similarity Filtering: Removes highly similar generated content
2. Length Constraints: Ensures stories meet minimum/maximum word counts
3. Structural Validation: Verifies proper story structure (introduction, main body, conclusion)
4. Content Safety: Filters inappropriate or unsafe content
5. Human Refined: Human-curated and further refined for the final SS-GEN dataset.

📈 Monitoring Progress

Each generation step provides progress bars and detailed logging:

# Example output
Loaded 14 human-written seed chapters and explanations.
Loaded 0 existing machine-generated chapters and explanations.
Outer loop for total chapters which need to be titles-expanded: 100%|██████████| 14/14 [02:15<00:00]

🔄 Resuming Generation

The system supports resuming interrupted generation:

• Automatically detects existing generated files
• Continues from the last completed step
• Maintains consistency with request_idx tracking

📊 Dataset

We construct a large-scale Social Story dataset:

Item	Description
Chapters	57 diverse themes
Titles in each chapter	>=70
Total stories	5,085
Avg. chapter length (in words)	2.46
Avg. title length (in words)	5.28
Avg. story content length (in words)	281.65
Structure	Title + Introduction + Body + Conclusion
Constraints	Structural Clarity, Descriptive Orientation, Situational Safety

🧪 Load Dataset via Hugging Face 🤗 (Recommend)

from datasets import load_dataset

# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("FMiMiY/SS-GEN")

🔗 View on Hugging Face

📈 Results

• We fine-tuned several 2B–8B models (e.g., Gemma, Mistral, LLaMA3).
• Fine-tuned models significantly outperformed zero-shot baselines across BLEU, ROUGE, and BERTScore.
• Human evaluation confirmed improvements in empathy, coherence, and narrative safety.

Citation

If you use SS-GEN or our dataset, please cite:

@inproceedings{feng2025ss,
  title={SS-GEN: A Social Story Generation Framework with Large Language Models},
  author={Feng, Yi and Song, Mingyang and Wang, Jiaqi and Chen, Zhuang and Bi, Guanqun and Huang, Minlie and Jing, Liping and Yu, Jian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={2},
  pages={1300--1308},
  year={2025}
}

🤝 Acknowledgments

This work is supported by Beijing Jiaotong University, Tsinghua University, and Tencent. Special thanks to psychologists, educators, and collaborators who helped shape and evaluate this project.

📬 Contact

If you have any questions, suggestions or feedback, feel free to submmit a issue or contact :

Yi Feng – yifeng@bjtu.edu.cn