The tokenizer nicely handles quoted column names, independent of dialect, e.g., [column name], "column name", and `column name` are all correctly recognized and the unquoted identifier column name is stored in the value slot of the corresponding Word. But then the parser's as_ident() restores the quotes by calling to_string() instead of just returning the value of self.value. I can't speak to all 7 lines that invoke this method but this seems the wrong implementation at least when parsing column names as well as parsing identifiers. If there are no objections, I'm gonna submit a pull request to fix this.