Abstract
With the advent of natural language enabled applications, there has been a growing appetite for conversational question answering systems. This demand is being largely satisfied with the help of such powerful language models as Open AI’s GPT models, Google’s BERT, and BigScience’s BLOOM. However, the astounding amount of training data and computing resources required to create such models is a huge challenge. Furthermore, for such systems, catering to multiple application domains typically requires the acquisition of even more training data. We discuss an alternative approach to the problem of open-domain conversational question answering by utilizing knowledge graphs to capture relevant information from a body of text in any domain. We achieve this by allowing the relations of the knowledge graphs to be drawn directly from the body of text being processed, rather than from a fixed ontology. By connecting this process with SPARQL queries generated from natural language questions, we demonstrate the foundations of an open-domain question answering system that requires no training and can switch domains flexibly and seamlessly.