MoPS: Modular Story Premise Synthesis for Open-Ended Automatic Story Generation

Yan Ma^1,3,4, Yu Qiao³, Pengfei Liu^2,3,4

¹Fudan University, ²Shanghai Jiao Tong University,

³Shanghai AI Laboratory, ⁴Generative AI Research Lab (GAIR)

✒️ Contents

• Contents
• Overview
• Getting started
• Usage
• Citation

👀 Overview

"If a story is going to fail, it will do so first at the premise level." – Anatomy of a Premise Line

A story premise succinctly defines a story’s main idea, foundation, and trajectory. It serves as the initial trigger in automatic story generation.

Existing sources of story premises are limited by a lack of diversity, uneven quality, and high costs that make them difficult to scale. In response, we introduce Modular Story Premise Synthesis (MoPS) which breaks down story premises into modules like background and persona for automated design and generation. MoPS consists of three phases: (1) Precollect a consistent set of candidates for each module to form a nested dictionary. (2) Extract a key path from the nested dictionary as the premise design. (3) Instruct an LLM to integrate the design into a coherent premise sentence.

👨‍💻 Getting started

Installation

git clone https://github.com/GAIR-NLP/MoPS
pip install -r requirements.txt
poetry install

We use OpenAI model as the language backend and manage environment variables in .env via python-dotenv.

Warning

After setting up, please add .env to .gitignore to prevent uploading sensitive information.

#!/usr/bin/env bash

OPENAI_API_KEY="your openai api key"

OPENAI_API_BASE="your openai api url"

OPENAI_MODEL="gpt-3.5-turbo-1106"

Resources

1. Modules

Module candidates in our paper: ./assets/modules

2. Premises

Mops & 5 baselines premises and corresponding evaluation in our paper: ./assets/premises

3. Premised-based Stories

Mops & 5 baselines stories extended from premises and the corresponding evaluation in our paper: ./assets/stories

Stories come in two genres: scripts and novels.

Scripts are generated using Dramatron, and novels are generated using RecurrentGPT.

4. Huggingface dataset

We created a huggingface dataset including three versions of the MoPS dataset: complete, moderate, and curated, with each entry containing a premise and the extented stories.

from datasets import load_dataset

dataset=load_dataset("ManTle/mops")

print(dataset)

🎯 Usage

Stage 1: Induce Module Candidates

>>> python mops/induce.py --help
usage: induce.py [-h] [OPTIONS]
╭─ options ───────────────────────────────────────────────╮
│ -h, --help              show this help message and exit │
│ --module-dir PATH       (required)                      │
│ --step STR              (required)                      │
│ --max-backgrounds-per-theme INT                         │
│                         (default: 30)                   │
│ --max-personas-per-background INT                       │
│                         (default: 9)                    │
│ --max-events-per-persona INT                            │
│                         (default: 2)                    │
│ --max-endings-per-event INT                             │
│                         (default: 1)                    │
│ --max-twists-per-ending INT                             │
│                         (default: 1)                    │
╰─────────────────────────────────────────────────────────╯

Step 0: Pre-collect a few themes in module_dir/theme.json.

See examples: ./data/modules/theme.json where module_dir=./data/modules

Step 1: Induce background

python mops/induce.py --module-dir ./data/modules --step background --max-backgrounds-per-theme 1

Step 2: Induce persona

python mops/induce.py --module-dir ./data/modules --step persona --max-personas-per-background 1

Step 3: Induce event

python mops/induce.py --module-dir ./data/modules --step event --max-endings-per-event 1

Step 4: Induce ending

python mops/induce.py --module-dir ./data/modules --step ending --max-endings-per-event 1

Step 5: Induce twist

python mops/induce.py --module-dir ./data/modules --step twist --max-twists-per-ending 1

Stage 2: Synthesize Premises

>>> python mops/synthesize.py --help
usage: synthesize.py [-h] [OPTIONS]

╭─ options ──────────────────────────────────────────────────────────────╮
│ -h, --help              show this help message and exit                │
│ --module-dir PATH       (required)                                     │
│ --premise-dir PATH      (required)                                     │
│ --enable-verify, --no-enable-verify                                    │
│                         (default: False)                               │
│ --masks {None}|{[{theme,background,persona,event,ending,twist} [...]]} │
│                         (default: None)                                │
╰────────────────────────────────────────────────────────────────────────╯

For example, you can run the following command after stage 1:

python mops/synthesize.py --module_dir ./data/modules --premise_dir ./data

You can flexibly control the required modules through masks flag , and the candidates for corresponding modules in the masks will be set as empty strings.

For example, if you want to remove ending and twist during synthesis, you can run the following:

python mops/synthesize.py --module_dir ./data/modules --premise_dir ./data --masks ending twist

Stage 3: Evaluate Premises

We use the semantic Breadth and Density metrics proposed in the paper to evaluate diversity, and evaluate quality based on LLM in the three dimensions of Fascination, Completeness, and Originality.

Please refer to files in notebooks for the implementation details of evaluation.

Stage 4 [Optional]: Use your favorite premise-based automatic story generation pipeline to create long stories

We use Dramatron and RecurrentGPT in our paper.

If you are looking for more story generation work, we recommend you refer to Awesome-Story-Generation

Citation

@misc{ma2024mops,
      title={MoPS: Modular Story Premise Synthesis for Open-Ended Automatic Story Generation}, 
      author={Yan Ma and Yu Qiao and Pengfei Liu},
      booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
      address={Bangkok, Thailand},
      publisher={Association for Computational Linguistics},
      year={2024},
      url={http://arxiv.org/abs/2406.05690}
}