Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU Models

Keyphrases

• Fewer Iterations 33%
• Generalization Accuracy 33%
• In-distribution 33%
• Language Comprehension 33%
• Learning Behavior 100%
• Long-tailed Distribution 100%
• Mitigation Framework 33%
• Out-of-distribution Data 33%
• Overconfident Prediction 33%
• Shortcut Learning 100%
• Training Model 33%
• Training Samples 33%
• Training Set 33%

Computer Science

• Data Distribution 33%
• Experimental Result 16%
• Language Understanding 16%
• Natural-Language Understanding 100%
• Training Model 16%
• Training Sample 16%

Engineering

• Experimental Result 100%
• Language Understanding 100%