A transformer that reads feelings, and tells you why.
"I do not feel happy at all."
→ joy, 100% sure. It had only learned to spot keywords.
"I do not feel happy at all."
→ sadness, and it knows the word not is what matters.
Real data
Google's GoEmotions: Reddit comments labelled by people, grouped into joy, sadness, anger, fear, surprise, disgust and neutral. Class weights stop the model from ignoring rare emotions like fear.
Two transformers
DistilBERT and the 6× smaller BERT-mini, fine-tuned on a GPU, then compressed to 8-bit with no loss in accuracy. Switch between them at the top right.
Why, not just what
Each word is removed in turn and the sentence is read again. Words that change the answer the most light up. Everything runs in your browser; your text never leaves your device.
Honest limits. Sarcasm still fools it ("Oh great, another Monday" reads as joy). It only reads English. And people often disagree about emotions too, even GoEmotions' own annotators.