Evaluating Embeddings for One-Shot Classification of Doctor-AI Consultations