It works by deducing a "reasonable" (based on some heuristics) representation of what's in the object code. The degree of resemblance between what it produces and what was originally there tends to depend heavily upon how much information is contained in binary it starts from. If you start with basically a "pure" binary, it's generally stuck with just making up "reasonable" names for the variables, such as using things like i, j and k for loop indexes, and longer names for most others.
On the other hand, a language that supports introspection needs to embed a great deal more information about variable names, types, etc., into the executable. In a case like this, decompiling can produce something much closer to the original, such as typically retaining the original names for functions, variables, etc. In such a case, the decompiler can often produce something quite similar to the original -- possibly losing little more than formatting and comments.
