Storyline Trees: Hierarchical Representations for Long-Form Narratives

The article introduces "storyline trees," a hierarchical representation designed to enhance long-context models' understanding of long-form narratives by organizing them from global themes to specific events. The method involves segmenting chapters into scenes and inferring storyline trees through a combination of top-down and bottom-up approaches, demonstrating improved performance in question answering tasks on three benchmarks compared to existing methods. This advancement is significant for AI practitioners as it facilitates adaptive retrieval, allowing models to navigate complex narrative structures more effectively, thus improving the efficiency of information retrieval in long-form content.