This lesson introduces a compact dialog act classifier for short task-oriented turns. Each synthetic example contains a small speaker-tagged dialog snippet and a label for the next-turn act such as greeting, request, or confirmation.
- synthetic dialog-turn generation with diverse intent-like act labels
- pooled token embedding encoder with a multiclass classification head
- cross-entropy training with top-1 accuracy for dialog act prediction
python -m tracks.nlp.lesson_32_toy_dialog_act_prediction.train --device cpu --epochs 1