I want to use it to classify free text input as “Question”, “Statement”, or “Command”, and to help characters answer questions about the game universe. Do you guys think it would be acceptable to have things like adversarial or hostile characters? All of the actual responses that users would see are picked from a conversation tree, I am only using GPT-3 to understand player input (identifying the subject, verb and such of player input).
I had this same idea yesterday! I did some prompt tests, creating a NPC to answer questions about a magical sword in a cave. But it was a bit hard to make it ignore all other questions, so the players can still keep chatting about strange stuff which would off the game “bounds”.
Interesting idea, though, would love to explore it further.
One disadvantage, is that in terms of game design, it would be more boring for a player to have to type everything, instead of choosing from pre-selected dialog options. But might be worth exploring!
You could use a separate prompt to filter out of tangent user questions.
“Is the above message a question about X”
And then only pass the message on to your conversation engine if the initial classifier says yes.
Check out the “intent” prompts from my book. The TLDR is yeah, GPT-3 is really good at inferring intent from unstructured chat. Also, you’re thinking of this as too low level of an NLP task. Don’t focus on the grammar or anything, just ask GPT-3 what the player/user wants.