KEEP CHATTING! An Attractive Dataset for Continuous Conversation Agents

Ongoing conversation is crucial for conversational agents to build long-term connections with users. However, users tend to quickly lose interest if the conversational agent's responses are not engaging enough. In this paper, we introduce a novel task aimed at increasing users' willingness to continue interacting with the agent. We create a dataset named CONTINUOUSCHAT by: (i) collecting and revising personas, then expanding them into detailed personas through experiences, daily life, future plans, or interesting stories; (ii) transforming detailed personas into dialogues infused with emotions and feelings; (iii) rewriting the dialogues in specific styles using few-shot prompts, conditioned on handwritten style-specific examples. We benchmark Large Language Models (LLMs) on the CONTINU-OUSCHAT dataset using both fine-tuning and in-context learning settings. Experiments with publicly available models show that while there is substantial room for improvement in generating style-specific dialogues, our CONTINU-OUSCHAT dataset is valuable for guiding conversational agents to produce more engaging dialogues and increase users' willingness to continue conversations. 1