In this article we focus on valency, which belongs to the core phenomena being captured in the underlying level of the Prague Dependency Treebank (PDT). We present a summary of the basic principles of the applied theoretical framework including proposals for suitable refinement relevant to NLP. The current status of description of valency behavior of verbs, nouns and adjectives is outlined. We present two branches of manual creation of a valency lexicon: (i) the PDT-VALLEX created during the annotation of the PDT and used primarily to obtain consistent annotation, and (ii) the Complex Valency Lexicon VALLEX, where the whole verbal lexemes are processed and other syntactically relevant information is assigned to particular valency frames.