Doubt regarding linking based system

Is there a work done on creating a linking based system, for example, linking words with each other, like,

'red'
^
|
'apple' -> 'tree'
|
V
'taste'

and instead of updating word embedding, our neural network task is to update these links, that is create new links, or break existing links, what would be the parameters in this case?

You might want to look into link prediction for knowledge bases.

The basic goal is that, given an existing knowledge base, what new facts can be predicted. Not sure if this matches your use case.