Artificial intelligence system for automated generation of realistic question and answer pairs

Abstract

Generally described, one or more aspects of the present application correspond to machine learning techniques for generating realistic question-answer (QA) pairs for populating an initial community ask feature of electronic store item detail pages. The machine learning model can use a shared encoder to generate an embedding of a seed sentence from existing description of an item, and then pass that embedding to a question decoder to generate a question. The embedding of the seed sentence can be combined with a state representation of the question and provided to an answer decoder, which can generate an answer to the generated question. This can help overcome the cold start problem, where customers are less likely to ask questions about items that have no existing QA set. This can also help surface relevant information about items in a concise QA format that is easy for customers to find and read.