AI story generation trails story understanding in its use of narrative theory

Large language models have become tools for both writing and analysing stories, and some researchers are trying to ground that work in the long scholarly tradition of narrative theory. A survey published in Artificial Intelligence Review by David Y. Liu, Aditya Joshi and Paul A. Dawson of UNSW Sydney maps this territory and reports an imbalance. Research that uses narrative theory to help machines understand stories is, in several respects, further along than research that uses it to help them generate stories.

The survey sits where two fields meet. Natural language processing (NLP) is the branch of computing concerned with getting machines to handle human language, and large language models (LLMs) are now its dominant tool. Narratology is the branch of literary and cultural study that analyses how narratives are built: who tells them, in what order events are revealed, and how time and perspective are handled. The authors examine how NLP research uses LLM methods to engage with concepts from narrative studies. They organise the work they review using distinctions that narratologists themselves have long relied on.

One such distinction is between fabula and discourse. The fabula is the bare sequence of events in chronological order, or what happens. Discourse is the telling: the arrangement, emphasis and point of view through which those events reach a reader. A thriller that opens with a body and then flashes back has the same fabula as a version told in order, but a different discourse. Narrative theory also describes levels beyond these two. According to the survey, generation research has so far done less than understanding research to address them.

That is one of several gaps the authors identify. Generation work also trails, they report, in how it applies theory, in its use of post-training methods, and in its exploration of non-fiction narratives. Post-training methods are techniques for adjusting a model after its initial training, such as fine-tuning it on targeted examples. The survey also finds that the narrative texts studied come from many sources beyond literature. And it finds that such research can do more than borrow from theory. Theoretical synthesis and validation are potential outcomes, meaning that experiments with LLMs might help combine, test or refine ideas about narrative rather than simply apply them.

On where the field should go, the authors argue against pursuing a single, generalised benchmark, or standardised test, for ‘narrative quality’. They favour four other efforts. The first is defining and improving theory-based metrics for individual narrative attributes. The second is continuing large-scale, theory-driven literary, social and cultural analysis. The third is generating narratives in situated contexts. The fourth is running experiments whose outputs can be used to validate or refine narrative theories.

The survey is an overview rather than an experiment. Its conclusions describe the shape of a research field, not the abilities of any particular model. The authors present it as a contextual foundation for more systematic and theoretically informed narrative research in NLP. Two questions remain open: whether theory-based measures of individual narrative qualities can be made to work, and whether machine-generated stories can sharpen theories about narrative. The authors pose both as directions for future work rather than settling them.