Introduction
Storytelling in artificial intelligence has long grappled with a fundamental trade-off: the rigid logical consistency of automated planning versus the fluid creativity of deep learning. Traditional automated planners, often utilizing Planning Domain Definition Language (PDDL), excel at maintaining causal integrity—ensuring, for instance, that a character cannot use an item they have not yet acquired. However, these systems typically lack the ability to generate diverse or "surprising" narratives without extensive manual engineering. Conversely, large-scale neural models can produce fluent text but often struggle with long-term plot coherence, leading to narratives that wander or violate their own internal logic.
This research proposes a hybrid approach that treats narrative generation as a structural sequence modeling task. By training deep learning models on synthetic, logically valid plan sequences rather than raw text, the authors demonstrate that models can learn the underlying "grammar" of a genre and generate novel, consistent story plans.
Methodology: Learning from Structured Plans
The core of the proposed system is a self-supervised Encoder-Decoder (Seq2Seq) architecture designed to operate on symbolic action sequences.
The process involves three distinct stages:
- Synthetic Dataset Construction: Using established narrative planning domains (Aladdin, Medical, Red, and Western), the researchers employed random-search planners to generate thousands of unique, well-formed plans. These plans consist of symbolic actions (e.g.,
shout-at-family,formal-complaint) rather than natural language. - Neural Model Training: The architecture utilizes a 1D Convolutional layer over operator embeddings to capture local action patterns, followed by a Recurrent Neural Network (RNN) to map the entire sequence into a latent vector ($z$). A decoder RNN then attempts to reconstruct the original plot from this vector.
- Plan Generation: To generate new stories, the encoder is bypassed. The decoder is "primed" with random noise ($z_G$), producing a novel sequence of actions. A PDDL validator then checks these sequences for logical well-formedness.
The researchers evaluated three model variants: a deterministic Autoencoder (AE), a standard LSTM baseline, and a Variational Autoencoder (VAE) designed to encourage diversity through latent space regularization.
Experimental Results
The models were tested across four diverse narrative domains. A primary metric for success was the Generative Factor ($f_G$), defined as the ratio of novel, valid plans generated to the size of the training set.
Novelty and Validity
The results indicate that the models successfully synthesized new knowledge rather than merely memorizing training data. In the Aladdin domain, the AE model achieved an $f_G$ of 3.46, meaning it generated nearly 3.5 times more unique, valid plans than were present in its training set. Notably, the AE variant proved more robust than the VAE, particularly in maintaining long-term dependencies in complex domains like Western.
Narrative Quality and Diversity
Beyond logical validity, the generated plans maintained the dramatic pacing of their respective genres. By tracking "beats" (pre-defined dramatic turning points), the authors found that the neural-generated plots mirrored the pacing distributions found in human-authored or classically planned stories. Furthermore, Levenshtein distance measurements confirmed high diversity among generated outputs, suggesting the models avoided the repetitive "mode collapse" often seen in generative AI.
Discussion and Future Work
This approach offers a "media-neutral" framework for storytelling. Because the output is a structured plan of actions, it can serve as the backbone for various media, including 3D animation engines, quest systems in video games, or text generation pipelines.
However, several limitations remain. The models are currently domain-specific; a model trained on Medical dramas cannot generalize to Westerns. Additionally, while the system generates valid plans, it lacks an internal "critic" to evaluate the emotional resonance or artistic quality of the stories. Future research directions include the implementation of controllable latent codes, allowing users to influence narrative parameters such as mood or length during the generation process.
Summary updated