Unstructured Knowledge Source for Future Conversational agents | Tony Zhao

Encoding knowledge is a central requirements for any dialog systems. Traditional task-oriented dialog systems usually only interface with structured database, which limits the scope of requests a dialog agent can answer and imposes challenges in learning. In this talk, we would like to describe a new solution, SocoDB, that is developed at Soco Inc. Concretely, an AI database for natural language is developed that is able to directly treat large text corpus as a virtual knowledge base, and enable downstream applications, e.g. a chatbot, to retrieve information from it using both natural and structured language. We will show novel methods that enable efficient inference and how SocoDB can be used to power task oriented systems in complex business environments with dynamic knowledge.