I think you might be underestimating the amount of intellectual input
involved in creating a translation, but I take your point. The intellectual
responsibility should be shared between the author and the translator.
I feel that the problem is the rather rigid approach created by the FRBR
model. In reality you have record created for one instance, along comes
another instance which needs to be related in some way. All that's needed is
to express that relationship in a way which the end-user will find useful
when reading the record. In this case the end-user needs to see that "b" is
a translation of "a" in a particular language. Whether a and b are works,
expressions, instances, is irrelevant.
In the world of linked data the relationships between bibliographic entities
which are in some way derived one from the other should be as flexible as
they need to be to clearly inform the end-user what that relationship is.