In this paper, we propose a robust statistical semantic tagging model trained on completely unannotated data. The approach relies mainly on prior domain knowledge to counterbalance the lack of semantically annotated treebank data. The proposed method encodes longer contextual information by grouping strongly related semantic concepts together into cohesive units. The method is based on hidden Markov model (HMM) and offers high ambiguity resolution power, outputs semantically rich information, and requires relatively low human effort. The approach yields high-performance models that are evaluated on two different corpora in two application domains in English and German.