APL expressions, just like mathematical formula, don't just float alone in a void in practice. They exist within a context, some problem domain, surrounding code and/or equations, etc. Just how the equation "a^2 + b^2 = c^2" can encode equivalence relations between different structures, depending on context, APL expressions take on interesting, semantically useful meanings when applied to a particular problem domains.
The reason I call this a strong point of APL is that, in "normal" languages we tend to re-encode that extra domain-specific information all over the place at the micro level—within function names, type signatures, and more pertinently in the abstractions we make. APL enables (but doesn't demand) us to let that information manifest directly at the macro level, without erasing the unused potential semantics of the raw algebraic expression.
This is something that practicing mathematicians get intuitively, I believe. Take category theory, for instance, where commutative diagrams in their most abstract form really are Abstract Nonsense™, but their power comes from instantiating them into specific categories, without having to erase the information about the fact that the same diagram also applies within other categories. That kind of thing really tickles the natural abilities of human pattern matching in a productive way.