The paper proopses a new approach to constructing POS tagsets. It argues for disociating morphosyntactic information from purely syntactic or semantic information implicit in tagsets based on traditional parts of speech (cf. `pronoun', `numeral') and for formulating a clear and unambiguous set of morphosyntactic criteria for belonging to a given POS class. The paper includes a feasibility study, i.e., it presents a tagset for Polish, a morphologically-rich Slavic language, constructed according to this approach.