Constraint-Aware Counterfactual Editing Targets the Aspect-Sentiment Shortcut Problem

arXiv·low signal

S M Rafiuddin, Vamsi Krishna Pavuluri, and Atriya Sen (arXiv 2607.13977, cs.CL) generate constraint-aware counterfactual edits for Aspect-Based Sentiment Analysis, attacking the failure mode where models fall back on global sentence polarity instead of reasoning about the specific aspect. The counterfactual-augmentation approach is a general recipe for breaking shortcut learning in any fine-grained classification task. Practical relevance is narrow — ABSA — but the shortcut-breaking pattern applies broadly to eval-set construction.