Highly imbalanced data set for multiclass classification, sklearn class_weight='balanced' resulting in model predicting always the same class

Question

and happy new year, I have a multiclass problem with a highly imbalanced data set (the goal is to predict which interjection could be used in the answer to a given utterance