(!) -- 8th suggestion, after three apparently identical "upside-down 'i'" (including an "upside-down capital 'I' with a dot underneath")
(@) -- 1st suggestion
(#) -- not suggested. Top suggestion is "capital 'H' with stroke"
($) -- not suggested, although the 14th is an indistinguishable glyph "Canadian syllabics carrier sh" 0x165a, a phoenetic symbol for representing a Canadian aboriginal language.
(%) -- 1st suggestion
(^) -- 10th suggestion, to be fair this one is impossible
(&) -- not suggested (fried shrimp lol)
(*) -- 3rd suggestion
(?) -- 1st suggestion
(∫) -- 1st suggestion
(∂) -- 4th suggestion (1st one is same thing in boldface)
----------
This is suggests a really easy way to greatly improve the results: weight them by a prior probability (i.e., frequency of occurrence in a letter count). The OCR itself seems pretty good. Common math symbols are more likely than silly shrimps. Glyphs from common languages are more probable than esoteric ones, and real languages more than constructed ones. Plain glyphs are more common than variants (bold/italic) -- and these should be grouped together anyway. fried shrimp horseshoe.