Abstract
Traditional Large Language Models (LLMs) excel at generating fluent text but often produce monotonous narratives that lack dramatic tension. This paper introduces CNGCI (Conflict-driven Narrative Generation through Commonsense Inference), a framework designed to intentionally inject conflict into stories. By utilizing defeasible inference to identify "soft causal threats" to a character's goals, the system generates coherent yet engaging narratives where protagonists encounter meaningful obstacles.
Introduction
While modern AI can write grammatically perfect prose, it frequently fails to capture the "soul" of storytelling: conflict. Standard language modeling objectives prioritize high-probability sequences, which often lead to predictable, "happily-ever-after" outcomes. To create compelling plots, a system must understand character intentions and realistically thwart them without sacrificing logical consistency.
The researchers address this by defining conflict as a disruption of expected causal flows. Instead of a linear generation process, they propose a neuro-symbolic approach that first establishes a dramatic "anchor" and then builds the narrative around it.
Methodology: The CNGCI Framework
The CNGCI framework operates in two primary stages: Conflict Generation and Story Completion.
1. Conflict Generation
Conflict is modeled as a "soft causal threat." Given an initial Context (S1), the system uses the COMET model to infer the protagonist’s likely Goal. It then applies defeasible inference to generate an Obstacle (S3). This process follows a "Yes, but..." logic: if the context is "Tom went on a singles cruise," the inferred goal might be "Tom has a good time," and the generated obstacle (the "weakener") could be "The cruise was a disaster."
2. Story Completion and Coherence Scoring
With the (Context, Goal, Obstacle) triple established, the system fills in the remaining sentences (S2, S4, S5) using a fine-tuned GPT-2 model. To prevent the introduction of conflict from degrading story quality, the authors implemented a multi-faceted Coherence Scoring metric.
This evaluation pipeline checks:
- Subject Transitions: Ensuring the protagonist remains the central character.
- Semantic Similarity: Using Sentence-BERT to maintain thematic focus.
- Implication/Contradiction Rules: Leveraging Natural Language Inference (NLI) to ensure that subsequent sentences either logically follow the context or meaningfully interact with the established conflict.
Datasets and Training
The system leverages several knowledge-rich datasets:
- ATOMIC-2020 & ConceptNet: Provide the commonsense backbone for character motivations.
- defeasible-NLI: Used to train the model to identify "weakeners" that diminish the likelihood of a goal.
- ROCStories: Used to train the story completion model and establish statistical norms for narrative structure.
Results and Discussion
Human evaluations conducted via Amazon Mechanical Turk compared CNGCI against two baselines: a knowledge-enhanced GPT-2 and C2PO.
- Conflict and Engagement: CNGCI demonstrated superior performance, achieving an 84.2% win rate for conflict presence and a 73.7% win rate for interestingness against C2PO.
- The Coherence Trade-off: While CNGCI significantly improved dramatic tension, it faced a slight decrease in perceived "Logic Coherency" compared to some baselines (42.1% vs 47.4%). This confirms the authors' hypothesis: intentional conflict is inherently disruptive, requiring a delicate balance between surprise and logic.
Conclusion
CNGCI demonstrates that framing narrative conflict through defeasible logic allows AI to move beyond simple coherence toward genuine storytelling. By treating obstacles as causal weakeners, the framework provides a structured way to generate drama. Future work will focus on scaling these techniques to larger LLMs and exploring more nuanced forms of conflict, such as internal character dilemmas.
Summary updated