It's used to allow writing a "resolver" and an "interpreter" that both know how to handle every type of node in the AST. It's almost a functional approach, and I actually think it might increase boilerplate because Java is more designed around the object-oriented approach of adding `resolve` and `interpret` methods directly to each node type.
The two approaches result in very different code organisation, though - maybe the visitor approach was easier to explain in prose.