Performance of classification models on original language (untranslated), and English-translated conversations for each topic of interest.

Blue and orange fill indicate traditional-ML, and transformer architecture language models, respectively. Bold text indicates an F1 score meeting our performance cutoff (≥0.7). Topics are ordered by best-performing model F1 performance score (descending) when trained with original language (untranslated) conversations as ultimately used to label the full corpus.