Symptoms-Disease Detecting Conversation Agent using Knowledge Graphs

Conversational agents have become extraordinarily popular over the last few years, with accelerated adoption due to COVID-19. Even though a lot of work has been done to devise a real-time agent very few of them focus on dynamic responses. The challenges for automatic medical diagnosis not only include issues for topic transition coherency and question understanding but also issues regarding the context of medical knowledge and symptoms of disease relations. In this paper, we propose a conversational agent that not only generates answers to specific medical questions but also makes more natural and human-like conversations and can adapt to the context and evolve over time. We propose an End-to-End knowledge-routed Relational Dialogue System that would incorporate a rich medical knowledge graph into the topic transition in dialogue management, and make it accommodative with NLU (Natural Language Understanding) and NLG (Natural Language Generation). A knowledge-routed graph for topic decision-making is used, which helps to identify relationships between symptoms and symptom-disease pairs. However, there are constraints on the extent of questions that knowledge graphs can address independently. To overcome these, we have used a fine-tuned GPT-3 model. While knowledge graphs organize data as interconnected entities, GPT-3 generates human-like text using learned patterns from large datasets. This approach enhances responses to intricate queries.