Trace back relevant words from claim predictions' feature importances

[jcezarms]

By identifying a prediction's best ranked EMB_# feature importances, it should be possible to trace back and get the most important words within a claim, from the model's perspective.
Say, if feature EMB_3 is the best ranked in a claim's robustness prediction, this would yield the 3rd word within the claim as the most relevant.

• Extract the 10 most relevant words from embedding importances, for each claim
• Store extracted words within the claims (as a list) and the discussions they belong to (as a Counter)
• Make a wordcloud for each discussion's extracted words
• Determine if the final result adds positively in interpretability, be it at claim or discussion level.