Aligning NLP Models with Target Population Perspectives using PAIR: Population-Aligned Instance Replication

Abstract

NLP models trained on crowd-sourced annotations often fail to represent the perspectives of specific target populations. We introduce PAIR (Population-Aligned Instance Replication), a method that reweights training instances to align model predictions with target population perspectives. PAIR adjusts for differences between annotator pools and target populations, enabling models to better reflect the views of underrepresented groups.

Publication
Proceedings of the 4th Workshop on Perspectivist Approaches to NLP