NarraDetect: An annotated dataset for the task of narrative detection

AbstractNarrative detection is an important task across diverse research domains where storytelling serves as a key mechanism for explaining human beliefs and behavior. However, the task faces three significant challenges: (1) inter-narrative heterogeneity, or the variation in narrative communication across social contexts; (2) intra-narrative heterogeneity, or the dynamic variation of narrative features within a single text over time; and (3) the lack of theoretical consensus regarding the concept of narrative. This paper introduces the NarraDetect dataset, a comprehensive resource comprising over 13,000 passages from 18 distinct narrative and non-narrative genres. Through a manually annotated subset of ~400 passages, we also introduce a novel theoretical framework for annotating for a scalar concept of “narrativity.” Our findings indicate that while supervised models outperform large language models (LLMs) on this dataset, LLMs exhibit stronger generalization and alignment with the scalar concept of narrativity.