I think understanding the conversation without seeing what the speakers do will be very limiting. The concepts will probably need to be built from the ground https://en.wikipedia.org/wiki/Lincos_language unless we have a large amount of data on which we can train GPT