Various methods like Linear Discriminant Analysis, k-Nearest Neighbors, Support Vector Machines or Decision Trees are used to successfully solve many classification problems. However, there is no single classifier that works the best in all classification problems. In pervasive adaptive systems the classification of human cognitive, emotional and physical states may be somewhat specific and should be closer to the way humans actually recognize these states. As an attempt in this direction we propose a probabilistic semantic classifier, which is based on discretization, structure identification and semantic optimization. Furthermore, this classifier supports three types of learning characteristic for humans: by repetition, by generalization and by specialization.