Explaining Language Models' Predictions with High-Impact Concepts