The main advantage, to me, is that the input of Yacc is a (mostly) self-contained, (mostly) declarative representation of the grammar to be parsed. This format is portable between projects.
You can, for example, take Postgres’s SQL syntax parsing “engine” out of Postgres and use it in your own project, because it’s not actually an “engine”, but rather just a Yacc grammar.
How often are yacc/lex grammars actually portable? Truly portable too — usable in different threading contexts, for example.
You say they are declarative. My experience is admittedly very limited, but I would call them anything but declarative. There are declarative elements, but the actual payload of your code is very, very often imperative. (A lot of this is shaped by the API, imo).
I compare this to parser combinators (like parsec) and PEGs (like luapeg and rust-peg). I find these other options are more naturally side-effect free, both internally and externally.
You can, for example, take Postgres’s SQL syntax parsing “engine” out of Postgres and use it in your own project, because it’s not actually an “engine”, but rather just a Yacc grammar.