Accurate cross-narrative clause extraction

Improve cross-narrative clause extraction so that overlapping and conflicting clauses can be extracted accurately and at a satisfactory level from complete narratives.

Background

The paper introduces Overlap–Unique–Conflict (OUC) extraction, which requires a model to identify overlapping, conflicting, and unique clauses directly from two complete narratives. Experiments with 14 open-source LLMs show that unique information is substantially easier to extract than overlap and conflict relations.

Although task-specific supervision improves performance considerably, overlap and conflict extraction remain substantially below satisfactory levels. The authors therefore characterize accurate cross-narrative clause extraction as an unresolved challenge, motivating methods that can better discover and pair corresponding information across narratives.