As I put into a sub-comment, I'm not sure it is always true that micro-F1 = accuracy. That statement assumes that every instance/object must be assigned to some class. In information extraction (e.g. NER), not all words need to be assigned any class. So if you fail to extract a "person" name, it would be a false negative for the person class but it would not also be a false positive for any other class.