Browse Teach Curate

Debiasing Entity Shortcuts in Relation Extraction via Entity Replacement

Relation extraction models often exploit a pseudo-correlation between named entities and relation types, predicting the relation from the entities alone. This concept covers breaking that shortcut by replacing entities to build a debiased benchmark, and a combined data-augmentation plus model-level debiasing method that restores a model's ability to generalize from context.

Rosie Text

Rethinking Relation Extraction: Beyond Shortcuts to Generalization with a Debiased Benchmark Liang

Relation extraction models often exploit entity-mention shortcuts, predicting a relation from the named entities alone rather than the surrounding context, so they score well on standard benchmarks w…