Humans know that, how does transform know that? Based on training data?

Sort of. Part of the training for a model includes telling it which parts of a sentence are important... a human points and clicks.

This is extremely important to know. That the relationships between words in the sentence are actually trained by human evaluation.

They are not.

No, thats incorrect. The connections are automatically deduced from the training data (which is just vast amounts of raw text).