Empathetic Dialogue Generation with Pre-trained RoBERTa-GPT2 and External Knowledge

AbstractOne challenge for dialogue agents is to recognize the feelings of conversation partner and respond accordingly. In this work, RoBERTa-GPT2 proposed empathetic generation, where pre-trained auto-encoding RoBERTa utilized as encoder auto-regressive GPT-2 decoder. With combination GPT-2, our model realizes a new state-of-the-art emotion accuracy. To enable ability model, we propose commonsense knowledge emotional concepts extractor, in which commonsensible context are extracted The experiment results demonstrate that generation benefits from both encoder-decoder architecture external knowledge.