DifStoryGen: Diffusion-Based Storytelling Algorithm with Distributed Attention

Storytelling involves creating a narrative by sequentially connecting documents, where the chain of documents follows the evolution of events and entities of a specific topic or subject. Early storytelling algorithms consecutively connected pairs of documents with overlapping contents to form a story chain. With the advent of contextual embeddings and distributed attention, the formation of story chains can be much more guided by the flow of concepts to replace mere overlaps of words to generate meaningful stories. This paper introduces a novel storytelling algorithm, DifStoryGen (Diffusion-based Story Generator), that uses a diffusion model and a distributed attention to construct a story from a document collection, such that the story (1) carries a theme from start to end, (2) captures the evolution of an event and associated entities, (3) allows connecting documents without word overlaps but having contextual similarity, and (4) generates hypothetical intermediate documents when required to bridge two documents are temporally distant. Our experimental results show that DifStoryGen outperforms other storytelling methods, producing significantly more meaningful, contextually rich, and coherent narratives.