Visualize: Figure1.
Description: "Example rationales drawn from various datasets. Underlined tokens are rationales provided by humans. Human annotators can fail to produce faithful rationales (row 1 and 2), and fidelity metrics themselves can be misleading (row 3 and 4)."