Narrative-UFET improves entity typing with controlled story generation

Ultra-fine entity typing (UFET) assigns highly specific categories to entity mentions (e.g., "doctor" vs. "surgeon"), but existing models falter on rare, long-tail types because they rely on sentence-level context. Disambiguating evidence often spans multiple sentences, yet all current UFET benchmarks are sentence-only. To test this hypothesis, the authors created Narrative-UFET, a controlled extension that pairs each entity mention with an automatically generated, coherent narrative. They designed two variants: Maintain (type stays constant) and Change (type shifts across the narrative).

Experimental results show that narrative context yields consistent improvements on long-tail types over sentence-level baselines, with the Change variant providing the strongest signal. Surprisingly, synthetic narratives outperformed naturally occurring contexts, suggesting that controlled discourse construction can surface implicit signals. Despite these gains, substantial room for improvement remains, pointing to open directions in discourse modeling and narrative generation.