Modern language models are remarkably fluent, yet they often struggle with the "connective tissue" of a story. While they can predict the next likely word, they frequently fail to understand the underlying logic that makes a sequence of events a narrative. Research across several decades suggests that the key to bridging this gap is event causality: the web of cause-and-effect relationships that explain why characters act and how their actions lead to specific consequences.
The Linear Sequence vs. The Causal Web
Most standard language models treat a story as a linear sequence of tokens. However, as Yidan Sun and Boyang Li argue in Event Causality Is Key to Computational Story Understanding, human-like comprehension requires viewing a story as an interconnected causal graph. In this view, an event isn't just something that happens after another; it is something that is necessitated by what came before.
Without this causal framework, models suffer from "topic drift" or "hallucination" in open-ended generation. Piji Li and colleagues note in Event Transition Planning for Open-ended Text Generation that models often lack a long-term plan because they do not explicitly model how one action logically leads to the next. This results in stories that are grammatically correct but narratively hollow.
The "Why" Gap in Machine Understanding
The difficulty of narrative reasoning is most visible in Question Answering (QA) benchmarks. While models are excellent at identifying "who" did "what," they struggle with "why" and "how."
- Fact Retrieval vs. Inference: In the FairytaleQA dataset, Dakuo Wang and Ying Xu found that while models can retrieve explicitly stated feelings, they fail at explaining causal relationships.
- Situational Logic: Makarand Tapaswi and Antonio Torralba demonstrated in MovieQA that answering abstract questions about character motivations requires aggregating information into meaningful "episodes" rather than just matching isolated frames or words.
- Semantic Relations: The ESTER dataset, introduced by Rujun Han and Qiang Ning, specifically tests for causal, conditional, and counterfactual relations, revealing that even state-of-the-art models lack the "narrative intelligence" to grasp how discrete events form a cohesive whole.
Building the Causal Skeleton
To move beyond surface-level patterns, researchers are developing methods to infuse models with explicit causal knowledge.
One approach is event-centric pretraining. Instead of predicting the next word, models like EventBERT, developed by Yucheng Zhou, are trained to understand "event correlation reasoning." By treating multi-word spans as single logical units (predicates and arguments), these models learn the typical flow of situational transitions.
Another approach involves backward reasoning. In Automated Story Generation as Question-Answering, Mark Riedl and Jonathan C. Balloch introduced EDGAR, a system that generates stories by working backward from a target ending. By asking "why" an event happened and generating a preceding event as an answer, the system ensures that every part of the story serves as a causal justification for the conclusion.
Why It Works: Scripts and Commonsense
The intuition behind these systems dates back to early AI research like Richard E. Cullingford’s SAM (Script Applier Mechanism) in Controlling Inference in Story Understanding. SAM used "Scripts"—pre-defined knowledge structures about common scenarios—to fill in missing details.
Modern versions of this idea, such as KnowSemLM by Haoruo Peng, integrate external knowledge graphs of cause-and-effect event pairs. This allows models to recognize "narrative common sense"—for example, knowing that "dropping a glass" is a contingent event for "glass breaking." By grounding language models in these logical skeletons, researchers are moving toward AI that doesn't just mimic the sound of a story, but understands the reason for it.