I am not yet 100% certain so I requested full-text access to read the paper, but I can say that a naive approach to string substitution will certainly be less performant than this, and offer less flexibility in replacement complexity without significant effort.

The reason being that in a naive approach, a vocabulary of size M and a document of token size N is an O(m*n) operation in the best case - while this is claimed to be an operation of O(2n).

Came here to ask the same question, I'm not clear on what functionality this offers beyond a simple k-v substitution.