Using multi-level models is necessary for relevant knowledge extraction during the analysis of large volumes of natural language texts. Such tasks are relevant for solving various problems in the field of analysis and generation of textual information. Such modelling of natural language texts requires big amount of texts with multi-layered annotations holding lexical, syntactical, semantic and narrative information. Annotated text collections are called text corpora; text corpora with syntactical annotations are called treebanks because syntactical information is stored in a tree form. The latest development in that field is semantic treebanks, that combine syntactical trees of sentences with the formal representation of their meaning in logical form. Many semantic treebanks, with shallow and deep semantic information, were developed in last years. Several approaches for manual and automatic building of semantic treebanks were developed. There are few generally accepted standards in that fast developing area, so different semantic banks vary significantly in the type of information they contain, especially on lexical level. In this paper, authors review semantic treebanks that could be used for text modeling, the characteristics the different semantic banks and the sort of data they contain on the semantic, narrative, syntactical and lexical levels, as well as the size and composition of relevant corpora and major tools for working with data in those banks. The approaches reviewed can be used for wide range of decision-making tasks in the field of analysis and generation of textual information, e.g. for information resources annotating and retrieval in the task of ontology-based collaborative development of domain information space for learning and scientific research [4], for generating and rewriting texts of different types (fiction, marketing, scientific etc.) and for different auditory [5] and many others.