We use large unlabeled data (n-gram model, density estimates of features) to improve quality of surface coherence evaluation in Czech texts. Along with new features across layers of language description, the additions significantly improve the performance of the system.