TL;DR
Large Language Models (LLMs) are great at helping writers overcome "writer's block" by suggesting continuations, but they fall short in supporting crucial aspects of creative writing like maintaining narrative consistency, developing plot structure, managing reader experience, and refining expressive intent. This paper surveys existing computational approaches, primarily from the AI and digital games research communities, that show promise in addressing these unmet needs, advocating for a hybrid approach that bridges LLMs with world-modeling and symbolic reasoning techniques.
What Problem Does This Paper Address?
The advent of powerful Large Language Models (LLMs) like GPT-3 has opened up exciting possibilities for computational support in creative writing. Writers are increasingly exploring these tools to help them "get unstuck" by providing plausible text continuations. However, the paper argues that current LLM-based tools primarily address only this one aspect of creative writing. The authors identify several other critical, yet largely unaddressed, needs:
- Maintaining Consistency: Ensuring that characters, settings, and plot points remain coherent throughout a long narrative.
- Developing Plot Structure: Architecting a satisfying overall story arc, including a clear beginning, rising action, climax, and resolution.
- Managing Reader Experience: Intentionally shaping how a reader perceives the story through elements like pacing, tension, and surprise.
- Refining Expressive Intent: Helping writers clarify and iterate on their high-level goals and the underlying message they wish to convey.
The paper highlights that LLMs, due to their focus on local coherence and often divergent suggestions, can even hinder these higher-level creative processes. The core problem is the gap between the capabilities of current LLM-centric tools and the multifaceted needs of creative writers for robust, long-term support.
Methods
This paper is primarily a survey and conceptual piece, rather than a presentation of new algorithms or experimental results. The authors:
- Define and Ground Needs: They articulate the five creative writing support needs by drawing on cognitive and theoretical literature from psychology, narratology, and human-computer interaction.
- Survey Existing Research: They review computational approaches from various research communities (NLP, AI, digital games) that have the potential to address the identified needs. This includes:
- World Models: Computational representations of the story world, its entities, and their relationships, used for consistency checking (e.g., "stories as plans" and "stories as proofs" approaches).
- Symbolic Representation Techniques: Methods like context-free grammars and logic programming for modeling plot structures (e.g., Propp's functions, Freytag's pyramid).
- Cognitive Modeling: Approaches to simulating reader understanding and experience (e.g., modeling pacing, tension, surprise).
- Goal-Oriented Systems: Tools that allow users to specify high-level author goals and use them to guide generation or provide suggestions.
- Advocate for Hybrid Approaches: The paper implicitly and explicitly suggests that future systems should integrate LLMs with these more symbolic and world-modeling techniques to achieve comprehensive creative support.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- LLMs are a Double-Edged Sword: While excellent for overcoming immediate writer's block, LLMs' tendency for divergent suggestions can actively work against the construction of coherent, long-term plot structures.
- The Importance of World Models: To maintain consistency, computational tools need to go beyond text generation and build explicit, structured representations of the story world—its characters, settings, and their interrelationships.
- Plot Structure is Global, Not Local: Unlike LLMs that excel at local text coherence, plot structure involves higher-level, global narrative arcs. Symbolic methods are better suited for representing and manipulating these structures.
- Reader Experience is an Active Process: Understanding how readers process narratives (as problem-solvers) is key to designing tools that can help authors intentionally shape pacing, tension, and surprise.
- Intent Refinement is Crucial: Creative intent isn't static; it evolves. Tools that can help writers articulate, refine, and maintain consistency with their high-level goals are vital.
- Bridging Communities is Key: The paper advocates for a synthesis of ideas from NLP (LLMs) and AI/digital games research (world models, planning, interactive drama) to create more effective creative writing support systems.
Key Results
As this is a survey and conceptual paper, there are no quantitative experimental results presented. The "results" are the identification and articulation of the unmet needs and the cataloging of promising existing technologies.
Limitations and Future Work
- Maturity of Technologies: The authors note that many of the surveyed technologies are at varying levels of maturity and have primarily been tested in automated or generative modes, not yet in mixed-initiative, co-creative interactions with human writers.
- Evaluation in Co-Creative Settings: A significant line of future work is to evaluate these technologies in mixed-initiative, co-creative modes to understand how they can best amplify human creativity.
- Interface Design: Developing effective interfaces and interaction protocols for these hybrid systems is crucial.
- Hybrid and Unified Models: The paper strongly suggests that future research should focus on developing hybrid and unified models that combine the strengths of LLMs with world-modeling and symbolic reasoning approaches.
Summary updated