Abstract
Emotional language generation is one of the keys to human-like artificialintelligence. Humans use different type of emotions depending on the situationof the conversation. Emotions also play an important role in mediating theengagement level with conversational partners. However, current conversationalagents do not effectively account for emotional content in the languagegeneration process. To address this problem, we develop a language modelingapproach that generates affective content when the dialogue is situated in agiven context. We use the recently released Empathetic-Dialogues corpus tobuild our models. Through detailed experiments, we find that our approachoutperforms the state-of-the-art method on the perplexity metric by about 5points and achieves a higher BLEU metric score.