Where is the Intelligence in a extractive Question-Answering model?

I’m trying to get the correct response from this context:

question = "What is the name of John's sister?"
context = ["the father of John is Michael", "Michael has a daughter called Maria"]

But “deepset/roberta-base-squad2” model doesn’t found any response. It just found response if I specifie that John’s sister is Maria. Where is the inteligence in this kind of AI models? I mean, I know that the can predict where start the answer and where finish, but there isn’t any way just to really understand the context?

Thanks to all!