Emotion Mirror · BERT (CBS4702 Final Project)

bert-base-uncased fine-tuned on dair-ai/emotion (16,000 training texts) to classify short English texts into six emotions: sadness, joy, love, anger, fear, surprise. Code and full evaluation: https://github.com/Cecile1015/emotion-mirror-cbs4702 (notebooks/06_bert.ipynb).

Training

Full fine-tuning, [CLS] -> pooler -> dropout -> linear; class-weighted cross-entropy; AdamW, learning rate 5e-05 (selected from 2e-05, 3e-05, 5e-05 on validation macro-F1), 10% linear warm-up, batch size 32, 4 epochs; checkpoint selected by validation macro-F1 (epoch 3.0); seed 42.

Results (official test split, 2,000 texts)

Per-class F1: sadness 0.969, joy 0.944, love 0.848, anger 0.941, fear 0.893, surprise 0.783.

Probabilities are calibrated by temperature scaling, T = 1.320 (stored in training_info.json; apply softmax(logits / T)).

Limitations

The dataset consists of English Twitter messages built around templates such as "i feel ...". The model recognises the emotion expressed in a text, not the writer's actual emotional state, and it is a course project, not a clinical or psychological instrument.