While simulated story worlds have proven to be fertile ground for emergent storytelling in research and in commercial game genres such as life simulation, 4X, and roguelikes, they have also proven challenging to create. Building these simulations can involve hand-authoring large amounts of content that provide context for character decision-making, ensure variation among generated narratives, and align with the simulation's narrative setting and themes. This short paper discusses preliminary work on leveraging ChatGPT to generate theme-relevant content for Neighborly, a character-driven settlement simulation framework designed for narrative generation. Given a textual description of the narrative setting, we demonstrate how ChatGPT can be used to generate Python code for characters' businesses, occupations, and traits. We discuss challenges with development and future work.
Simulated story worlds like those seen in commercial Workshop Proce dings htp:/ceur-ws.org ISN1613-073
CEUR area has explored level generation[4], automated game design[5], and interactive storytelling[6, 7].
GPT have recently gained popularity for their versatile and unprecedented performance at various tasks, includsummarization, and image generation. Companies such as NVIDIA and Unity have announced new LLM-enabled technology stacks that enable game developers to create a wide array of content ranging from animations to code to autonomous virtual characters [8, 9].
This short paper discusses preliminary work on leveraging ChatGPT to generate narrative theme-relevant content for Neighborly, a character-driven settlement simulation framework for narrative generation[10]. Given a textual description of the narrative setting, we show how ChatGPT can be used to generate Python code for is to eventually enable users to create simulations from natural language prompts.
with development and future work.
LLM technology became popular for interactive storytelling following the release of AI Dungeon in 2019. The creators trained a GPT-3 model on choose-your-ownadventure stories and used it to generate interactive fiction experiences[11]. Since then, we have seen LLMs applied to other storytelling tasks like mixed-initiative
Mixed-initiative content generation is a promising so- characters’ businesses, occupations, and traits. Our goal to create content that neither could easily create inde- stages of our generation pipeline and discuss challenges CEUR
ceur-ws.org © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License storytelling [12], AI-enabled editors for story writing[13], Attribution 4.0 International (CC BY 4.0).
Companies within the commercial games industry are also using LLMs to power new generative AI tools aimed at game developers. Unity recently released Unity Muse, a This section describes our preliminary work using chattool to improve developer productivity by allowing users GPT to generate content for simulated story worlds. We to generate assets (textures, models, animations, etc.) want to explore how LLMs, like ChatGPT, can support using a natural language interface[8]. Another example creators during the rapid prototyping and iteration phase is NVIDIA ACE for Games, a suite of tools that leverage of simulation development. When brainstorming conLLMs and other AI technologies to help game developers tent, designers might spend time consulting multiple refcreate hyperrealistic virtual characters in their games[9]. erences for inspiration. Generating initial content gives designers a starting point to jumpstart the rest of their design process. Our goals with this project are to: 3. Generating simulation content with ChatGPT • Generate content that fits the narrative setting of
the story world. • Output editable source code that integrates with
the existing development ecosystem. • Evaluate if this improves the prototyping worklfow for simulation designers. interactive story generation[14], and dialog generation for NPCs in table-top role-playing games [15]. be handled by simulated systems implemented using a general-purpose programming language.
ChatGPT is an interactive LLM application within the family of generative pre-trained transformer models released by OpenAI. Its main user interface resembles a messaging application, and users can send natural language prompts to the model. It was trained using reinforcement learning techniques [16], allowing it to respond fluently to input from the user. One of its greatest strengths is its ability to reason about semantic relationships.
our three goals. The evaluation is saved for a future full-length publication. We are working on turning our 2.4. ChatGPT for simulated storyworlds generation procedure into a single cohesive software tool. A fully-featured version of this tool might operate like a ChatGPT has started to garner some attention as a po- virtual pair programmer that provides suggestions and tential tool for storytelling with simulated story worlds. feedback during the development process to help ensure Méndez and Gervás explored using ChatGPT for story user meet their authorial goals. sifting[ 17] – searching for interesting narrative material Thus far, our tool generates character spawn inforwithin a repository of simulated story world data [18]. mation, business types, occupation types, and character They found that ChatGPT returned proper prose and traits for Neighborly, a character-driven simulation frameperformed well at summarization tasks. However, it of- work for narrative generation [10]. Neighborly simulates ten embellished the summaries, and the authors found the lives and relationships of generations of characters it challenging to influence/bias ChatGPT toward desired in a generated settlement. Characters are born, grow stories and narrative themes. older, work jobs, form relationships, have families, and
Park et al. (
GPT. The final product is valid Python source code that can be edited in a text editor and packaged inside a Neigh- Listing 1: The JSON output provided by ChatGPT for a borly plugin for later use. Figure 3.3 shows the existing business that might exist in a cyber punk world. data generation pipeline. If users have conflicts with (See prompt listing 2). or want to adjust the content generated by the pipeline, [ they can refine their initial prompt or edit the produced { Python code. ” name ” : ” C y b e r n e t i c
We focused on the three major content areas for theme- A u g m e n t a t i o n C l i n i c building in Neighborly: businesses and occupations, char- ” o w n e r _ t y p e ” : ” C y b e r n e t i c acters, and character traits. Based on what Neighborly’s S u r g e o n ” , APIs provide, these were the most straightforward to ” e m p l o y e e _ t y p e s ” : { generate content for since they rely on static configura- ” C y b e r n e t i c T e c h n i c i a n ” : tion data. They are not directly responsible for narra- 4 , tive generation, but they afect the growth of character ” A d m i n i s t r a t i v e A s s i s t a n t relationships and support the setting of the simulated ” : 2 , story world. Generating life events for characters would ” R e c e p t i o n i s t ” : 1 , have had a more direct impact on narrative generation.
However, Neighborly does not have a declarative way to define event types and their preconditions and post efects. } , ” s e r v i c e s ” : [ ” C y b e r n e t i c i m p l a n t s ” , ” N e u r a l E n h a n c e m e n t s ” , ” A u g m e n t a t i o n c o n s u l t a t i o n s ” ] ,
]
First, we attempted to generate new business types for characters to own and work at. In Neighborly, these set the tone for which characters interact and provide additional context to storytelling. Generating new types was a two-step process as business types define occupations that may need to be generated. After ChatGPT provides business types in the format provided in the listing below, we scrape the owner and employee types and re-query ChatGPT configuration for those occupa
Next, we tried generating configurations for the types of characters that spawn into the simulation. We found that ChatGPT had trouble with this task. It likes to generate the typical fantasy races (Humans, Elves, Orcs, etc.) found in role-playing games, even if the setting is not entirely appropriate for those categories. Also, it has trouble diferentiating between character types and roles. For example, in response to the prompt, “List races of characters that exist within a cyberpunk futuristic story world?”, we received output that included humans, cyborgs, and androids, but also definitions for organizations and roles such as “street gangs”, “company executives”, and “Fixers”.
Modifying the phrasing also did not help as expected. Swapping “race” for character type resulted in a list of narrative-inspired roles like hero, villain, and mentor. Furthermore, swapping “race” for species resulted in ChatGPT generating entirely made-up fantasy-style species like “Ferbles”, “Aquamids”, and “Plantlings” which did not fit the theme.
One solution to this problem was changing ChatGPT’s task from generating character types to determining which types from a specified list are most appropriate for appearing in the given setting and relative spawn frequencies based on the other entries.
The last content type we experimented with was character traits. Crusader Kings extensively uses traits to drive character relationship development and other mechanics.
However, creating a list of traits and determining their efects relative to each other is time-consuming manual work. We ofloaded this to ChatGPT and had it provide JSON feedback that included trait names, descriptions, and social rules for modifying the platonic and romantic compatibility of two given characters (see Figure 3.3). making, arbitrary string values can easily become an authoring nightmare. Adding 10 new business types could yield 10 to 30+ new service types that they would need to author new simulation rules. If we constrain the 4. Discussion list of services to be selected from a fixed set, then we could create new business types that leverage existing Overall, ChatGPT performs well at generating business rules, further simplifying content authoring. and occupation definitions. Occasionally, it will misin- This work needs to be evaluated in a user study. Tenterpret a configuration parameter. However, this can be tatively, the plan is to present participants with two aueasily corrected by the human author. Also, validation thoring tasks. The first task would ask participants to safeguards could be put in place during content gener- hand-author a subset of simulation content, and the secation to notify the user of any errors before final code ond would ask them to generate it with the tool and generation. modify the results. We believe that users will find the
Even with the occasional error, generating content this ChatGPT-powered workflow takes significantly less time. way was still much faster than authoring all the definitions by hand. ChatGPT really shines when generating fuzzy values for configuration settings such as the rela- 5. Conclusion tive socioeconomic status of occupations, the lifespan of businesses, and the relative numbers of employee types.
It allows authors to quickly reach the point of having a running simulation, allowing them to tweak the values at a later time.
ChatGPT’s ability to create arbitrary service types for business was surprising and feels like it adds to building up the theme of a Cyberpunk world. However, since AI systems use the list of services for character decisionThis short paper presents preliminary work on using ChatGPT to generate theme-relevant content for an agent-based social simulation. We aimed to simplify the content prototyping process by leveraging ChatGPT’s semantic knowledge to generate Python code. We describe our pipeline for generating characters, businesses, and character traits from a short description of a designer’s narrative setting/theme. We are working on turning our generation pipeline into a single tool to facilitate user studies. As future work, we would like to explore using an LLM to generate the storylet-style events that Neighborly uses for narrative generation. L i s t t h e t y p e s o f b u s i n e s s e s t h a t may e x i s t i n a CyberPunk c i t y , what s e r v i c e s t h e y o f f e r ( a s a l i s t o f s i n g l e words ) , t h e j o b t i t l e o f t h e owner , t h e j o b t i t l e s o f e m p l o y e e s , and t h e number o f e m p l o y e e s . Use t h e f o l l o w i n g YAML t e m p l a t e : name : < b u s i n e s s t y p e > components :
Name :
v a l u e : < b u s i n e s s name > B u s i n e s s : o w n e r _ t y p e : < owner j o b t i t l e > e m p l o y e e _ t y p e s : < e m p l o y e e t i t l e > : < q u a n t i t y > < e m p l o y e e t i t l e > : < q u a n t i t y > . . .
S e r v i c e s : s e r v i c e s : < s e r v i c e , s e r v i c e , . . . > {
Listing 3: Prompt for eliciting occupation status levels.
L i s t the t y p e s o f o c c u p a t i o n s t h a t e x i s t i n a CyberPunk c i t y and t h e i r r e l a t i v e s o c i a l s t a t u s on a s c a l e from 1 t o 5 with 5 b e i n g the h i g h e s t .
Use the f o l l o w i n g t e m p l a t e : [ ” name ” : ” < o c c u p a t i o n name >” ” s o c i a l s t a t u s ” : < s o c i a l s t a t u s >
Generate a l i s t o f c h a r a c t e r t r a i t s f o r
NPCs i n a s i m u l a t i o n . The l i s t s h o u l d i n c l u d e the name o f the t r a i t , a s h o r t d e s c r i p t i o n , and a l i s t o f o t h e r t r a i t s t h a t i t i s i n c o m p a t i b l e with . Each e n t r y s h o u l d a l s o i n c l u d e a l i s t o f r e l a t i o n s h i p m o d i f i e r s t h a t
l i s t the t r a i t o f an i n t e r l o c u t o r and c o r r e s p o n d i n g r o m a n t i c and p l a t o n i c c o m p a t i b i l i t y m o d i f i e r s on a s c a l e [ − 5 , 5 ] . Return the o u t p u t as v a l i d JSON . ” name ” : ” F r i e n d l y ” , ” d e s c r i p t i o n ” : ” T h i s c h a r a c t e r n a t u r a l l y g e t s a l o n g with o t h e r s , ” i n c o m p a t i b l e _ w i t h ” : [ ” Mean ” ,
. . . ] , ” r e l a t i o n s h i p _ m o d i f i e r s ” : [ { ” h a s _ t r a i t ” : ” F r i e n d l y ” , ” r o m a n t i c _ c o m p a t i b i l i t y :
0 , ” p l a t o n i c _ c o m p a t i b i l i t y :
3 ,