That would be a good way to implement connotation in an AI. I don't think we are that accurate. To retrieve sufficient Bayesian evidence for general reasoning regarding B on being reminded of A, supposing you already had P(A), you'd have to retrieve any 2 of P(A|B), P(B|A), and P(B), right? But most current models assume concept activation retrieves one number per related concept (activation level), not two.