Good luck with Turkish. A very large lookup table may work for %90 cases but it will still fail. A basic `true` lemmatizer requires either a complex graph with rules, or an FST generated from it. input is searched through the graph and morphological disambiguator must be applied to result to pick the correct lemma.