Narrative Context Improves Ultra-Fine Entity Typing

Ultra-fine entity typing (UFET) aims to assign highly specific types to entity mentions within text. However, current methods often struggle with less common, "long-tail" entity types, primarily because they rely on context limited to a single sentence. Disambiguating these entities often requires information spread across multiple sentences. Researchers have developed Narrative-UFET, a novel approach that extends UFET by pairing entity mentions with automatically generated, coherent short narratives. This allows for controlled experimentation to isolate the impact of broader discourse properties. Two variants were tested: one where the entity's type remains constant throughout the narrative, and another where it shifts. The results demonstrate that providing narrative context consistently improves the accuracy of typing long-tail entities compared to sentence-level baselines. The variant where the entity type shifts provided an even stronger signal. A comparison with naturally occurring contexts revealed that these synthetic narratives yielded greater gains, highlighting the potential of controlled discourse construction to surface implicit signals for UFET.