System and method for question-based content answering

This invention describes a computer system that answers questions by finding relevant parts of media content. It works by first training an AI model to understand questions and content, then indexing a large collection of media. When someone asks a question, the system searches for keywords in the indexed content to get an initial set of possible answers. It then uses the trained AI model to rank these possibilities and presents the best ones. The claims narrow the AI model to a specific type called a 'bidirectional encoder representations from transformers' (BERT) language model with at least 50 million parameters.