Abstract
Automated story plot generation is the task of generating a coherent sequence of plot events. Causal relations between plot events are believed to increase the perception of story and plot coherence. In this work, we introduce the concept of soft causal relations as causal relations inferred from commonsense reasoning. We demonstrate C2PO, an approach to narrative generation that operationalizes this concept through Causal, Commonsense Plot Ordering. Using human-participant protocols, we evaluate our system against baseline systems with different commonsense reasoning reasoning and inductive biases to determine the role of soft causal relations in perceived story quality. Through these studies we also probe the interplay of how changes in commonsense norms across storytelling genres affect perceptions of story quality.
Introduction: The Gap Between Logic and Language
Automated story generation has historically struggled between two extremes. Symbolic planning systems utilize strict logic (hard causality) to ensure coherence, but they are limited by manual engineering and narrow vocabularies. Conversely, modern neural language models (e.g., GPT, BERT) produce fluent text but often lack "causal progression"—the logical thread that explains why a character acts.
To bridge this gap, researchers at Georgia Tech have introduced soft causal relations: causal links derived from everyday human expectations and commonsense reasoning rather than formal logical proofs. Their system, C2PO (Commonsense, Causal Plot Ordering), leverages these relations to generate narratives that are both fluent and causally coherent.
Methodology: The C2PO Pipeline
C2PO approaches storytelling as a plot-infilling task, moving through three distinct stages:
1. Plot Extraction
The system begins by processing an existing story summary. It uses Open Information Extraction (OpenIE) and Coreference Resolution to distill the text into a sequence of core "plot points."
2. Plot Graph Generation
This is the core innovation of C2PO. To bridge the gaps between extracted plot points, the system uses COMET, a transformer-based model trained on the ATOMIC commonsense dataset. COMET allows the system to reason in two directions:
- Forward Reasoning ("Wants"): Predicting what a character might do next based on their desires (e.g., if a character wants to be free, they might try to escape).
- Backward Reasoning ("Needs"): Identifying prerequisites for an action (e.g., to scale a wall, a character needs to know how to climb).
By intersecting these paths, C2PO constructs a branching graph of plausible intermediate events.
3. Story Generation
Finally, the system identifies the optimal path through the generated graph to connect the original plot points and then performs a "random walk" to produce the final narrative.
Evaluation and Results
The researchers evaluated C2PO using two genres with different commonsense profiles: Mystery Stories (Sherlock Holmes), which follow standard real-world norms, and Children’s Fairy Tales (Hansel and Gretel), which often subvert them. C2PO was compared against two baselines: BERT+infill and Hierarchical Fusion.
A human participant study involving 351 individuals yielded the following insights:
- Superior Plausibility: C2PO was significantly preferred over BERT+infill in both genres for "Plausible Ordering" and "Making Sense."
- Global Coherence: C2PO excelled at maintaining a "Single Plot," whereas baselines frequently drifted into unrelated topics.
- Genre Sensitivity: In the mystery genre, C2PO dominated in "Enjoyability" and "Genre Resemblance." However, in fairy tales, the gap in enjoyability narrowed. This suggests that the inherent randomness of baseline models occasionally mimics the "magical" or unpredictable nature of the fairy tale genre.
Conclusion
C2PO demonstrates that incorporating soft causal relations via commonsense reasoning significantly improves the perceived quality and coherence of AI-generated stories. While "everyday" commonsense is highly effective for realistic genres like mystery, future work will explore training models on genre-specific norms to better capture the unique logic of fantasy and science fiction.
Summary updated