2018/10/31 by Hannah Rashkin, Eric M. Smith, Rashkin, Hannah +5
Computer Science · #Topic Modeling #Speech and dialogue systems #AI in Service Interactions
paper · pdf · doi:10.48550/arxiv.1811.00207
One challenge for dialogue agents is recognizing feelings in the conversation\npartner and replying accordingly, a key communicative skill. While it is\nstraightforward for humans to recognize and acknowledge others' feelings in a\nconversation, this is a significant challenge for AI systems due to the paucity\nof suitable publicly-available datasets for training and evaluation. This work\nproposes a new benchmark for empathetic dialogue generation and\nEmpatheticDialogues, a novel dataset of 25k conversations grounded in emotional\nsituations. Our experiments indicate that dialogue models that use our dataset\nare perceived to be more empathetic by human evaluators, compared to models\nmerely trained on large-scale Internet conversation data. We also present\nempirical comparisons of dialogue model adaptations for empathetic responding,\nleveraging existing models or datasets without requiring lengthy re-training of\nthe full model.\n