Upload labeled data. Train a model. See why it thinks what it thinks.
These are the labeled training comments most similar to what you typed, measured by cosine similarity in DistilBERT's embedding space. The model's prediction is shaped by patterns learned from data like this.
Each word was removed one at a time to measure its effect on the prediction. Positive = removing this word decreased confidence (it mattered). Negative = removing it increased confidence (it was working against the prediction).