Towards Generating Topic-Driven and Affective Responses to Assist Mental Wellness.

Conversational Agents have great potential to serve as a low-cost, effective tool to support mental well-being when equipped with affective and contextual dialogue. However, it is challenging to build effective chatbots that can handle user free-text responses. In this work, we present a Topic-driven and Affective Conversational Agent that aims to tackle the emotional and contextual relevance properties of a chatbot supporting mental well-being. We leverage the transfer learning based scheme using a large pre-trained language model with a multitask objective to train a conversational model. Additionally, in order to keep the dialogue contextual to the topics of mental well-being without retraining, we use topic-based classifier models to achieve controlled response generation. We evaluate this model on two metrics- emotional and contextual relevance with human annotation. To further validate this approach, we integrate this scheme to MoEL, an empathetic conversational agent, and show its improvement on the two relevant metrics. Our results show that the generated responses achieve significant emotional relevance and are contextually relevant to the conversation topic. Our relatively unexplored approach of using the topic classifier to control the topic in a conversation can aid chatbots on mental well-being and beyond.