This concerns:

• Interpretation - trying to get a better understanding of triggers for specific (text, prediction) combinations
• Visualization - visualizing the attributions of words / tokens
• Automatically detecting possible problems with the model / data / etc
• Finding an effective / efficient "cure"