This framework adapts Invariant Risk Minimization (IRM) to natural language tasks by defining multiple training environments from available data and penalizing features whose predictive power varies across environments. It integrates with standard pretrained Transformer encoders and demonstrates improved robustness on out-of-domain NLI and sentiment benchmarks.
This page was last edited on 2024-04-16.
This page was last edited on 2024-04-16.