Towards Generating Contextual and Empathetic Response for Covid-related Queries

Sowmya Rasipuram, Anutosh Maitra, Bishal Shaw, Sriparna Saha

Abstract: This work addresses the vital need of keeping people informed with relevant, correct and essential information during the pandemic. Advanced NLP and machine learning mechanisms have been leveraged to generate responses to user queries through contextual conversation. In order to help people be discerning about what information they receive, a conversational system is proposed that identifies the correct intent of the query and a reinforcement Learning based generation model is used to proceed with conversation. We propose an end-to-end real-time text generation model that can respond to users queries on covid19. We created a new dataset with 1200+ covid-related questions from various sources and pre-processed them for a brief and direct answer. The dataset has also been manually observed to identify depressed questions and the responses are converted to be more empathetic. The dataset has been used to fine-tune DailoGPT, a GPT2-based transformer model to generate the responses related to COVID. COVID-related queries are bucketed into 15 categories to identify the exact intent of people. Our model generated both contextual and empathetic responses and achieved a human evaluation score of 3.48 (on a scale of 5) in terms of contextual relevance and a score of 2.12 (on a scale of 3).