Towards Empathetic Open-domain Conversation Models: a New Benchmark and Dataset

One challenge for dialogue agents is recognizing feelings in the conversation and replying accordingly, a key communicative skill. While it is for humans to recognize and acknowledge others' feelings in a, this is a significant challenge for AI systems due to the paucity suitable publicly-available datasets for training and evaluation. This work a new benchmark for empathetic dialogue generation and, a novel dataset of 25k conversations grounded in emotional. Our experiments indicate that dialogue models that use our dataset perceived to be more empathetic by human evaluators, compared to models trained on large-scale Internet conversation data. We also present comparisons of dialogue model adaptations for empathetic responding, existing models or datasets without requiring lengthy re-training of full model.