Automated Storytelling via Causal, Commonsense Plot Ordering

Prithviraj Ammanabrolu, Wesley Cheung, William Broniec, Mark O. Riedl

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 field of automated story generation has long struggled with a trade-off between logical coherence and creative flexibility. Traditional symbolic planning systems guarantee causality through strict logical rules but suffer from high manual engineering costs and narrow domain applicability. Conversely, modern neural language models excel at generating fluent prose but often lack a coherent logical thread, leading to "rambling" narratives where events occur without clear cause or enablement.

This paper introduces C2PO (Commonsense, Causal Plot Ordering), a framework that bridges this gap by utilizing soft causal relations. Unlike hard logical preconditions, soft causality leverages everyday commonsense expectations—what a character likely wants to do or needs to have happened—to ensure narrative progression feels natural and motivated to a human reader.

The C2PO Framework

C2PO treats story generation as a plot-infilling task. The system first extracts a high-level plot skeleton from existing text to serve as the basis for generation.

The C2PO pipeline extracts plot points and aligns them with coreference mentions to create a structured plot line.

The extraction process involves three primary stages:

  1. Neural Coreference Resolution: Identifying all mentions of a specific character (e.g., mapping "husband", "he", and "him" to the same entity).
  2. Open Information Extraction (OpenIE): Extracting subject-relation-object tuples (e.g., "He scales the wall") to define discrete events.
  3. Alignment: Mapping these tuples back to the story text to create a chronologically ordered sequence of plot points.

Soft Causal Reasoning and Pathfinding

To bridge the narrative gap between two adjacent plot points (A and B), C2PO constructs a causal bridge using the COMET model, a transformer-based commonsense reasoner trained on the ATOMIC dataset.

C2PO uses COMET to generate forward (wants) and backward (needs) graphs to bridge gaps between plot events.

The system generates two directed acyclic graphs to find a path between events:

  • Forward Branching: Starting from plot point A, it queries COMET for the character's likely subsequent desires (wants).
  • Backward Branching: Starting from plot point B, it queries COMET for the necessary prerequisites (needs) for that event to occur.

C2PO then identifies the optimal path that connects these two graphs, using probability weights from COMET to ensure the resulting sequence of events is both motivated by the past and enables the future.

Evaluation and Results

The authors evaluated C2PO using a human-participant study involving 351 participants across two distinct genres: Mystery (which follows real-world logic) and Fairy Tales (which often subverts it). C2PO was compared against BERT+infill (a language model baseline) and Hierarchical Fusion (a storytelling-specific model).

Key findings include:

  • Superior Coherence: C2PO significantly outperformed BERT+infill across all metrics, including plausible ordering, local sense, and overall quality in both genres.
  • Causality vs. Style: Compared to Hierarchical Fusion, C2PO demonstrated better logical causality and plot maintenance. However, the two models tied in "enjoyability" and "genre resemblance," suggesting that while C2PO provides better structure, Hierarchical Fusion is effective at mimicking narrative prose style.
  • Genre Sensitivity: The effectiveness of everyday commonsense reasoning varies by genre. C2PO's improvements were more pronounced in Mystery stories than in Fairy Tales, where the subversion of norms (e.g., magic) can conflict with standard commonsense logic.

Conclusion

C2PO demonstrates that explicitly modeling soft causal relations through commonsense reasoning significantly enhances the coherence of automated stories. While the system currently relies on general-purpose commonsense, future work aims to incorporate genre-specific or thematic logic, allowing models to adapt their causal reasoning to the unique rules of fictional worlds like science fiction or fantasy.

Summary updated

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.
BibTeX
@inproceedings{ammanabrolu2021automated,
  author = {Prithviraj Ammanabrolu and Wesley Cheung and William Broniec and Mark O. Riedl},
  title = {{Automated Storytelling via Causal, Commonsense Plot Ordering}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=Rfvw6iTD34v},
}
Figure