The annotation of the Prague Dependency Treebank is realized in two sub-collections which differ in the subtlety of annotation (the large collection and the model collection). In the present paper, we focus on deletions of complementations of verbs, postverbal nouns and adjectives, from the point of view of the annotators of the model collection. We inquire into the issues of deletions of participants of verbs with respect to the coreferential relations between the restored node and its antecedent, and we introduce the new type of deletion where the deleted node has no concrete antecedent (we call these restored nodes vague, unspecified anaphoric elements). We also specify the differences between the newly introduced lemma Unsp(ecified) and the lemma Gen(eral), used for deletions of General Participants. After analysing the process of nominalization, we extend the description of the particular types of deletions also to the postverbal nouns denoting action. 1. Three-layer system of tags The annotation of the Prague Dependency Treebank (PDT in the sequel) is basically conceived of in accordance with the theoretical assumptions of the Functional Generative Description (FGD in the sequel, see Sgall, Hajicova, and Panevova, 1986; Hajicova, 1993). The present paper deals with the current phase of annotation of PDT at the so-called tectogrammatical level (which captures underlying syntactic structures of sentences). This level of annotation, in contrast to the two preceding phases of annotation, namely the morphemic and the so-called analytical levels (for a description of the annotation scheme of PDT see e.g. Hajic et al., 2001, Hajicova et al., 2001), is realized in two steps (in two subcollections) which differ in the subtlety of annotation. The first, basic step of annotation is represented by the so-called basic or large collection (LC; today it contains about 25,000 sentences). The second step of annotation, the so-called model collection (MC), should provide full, detailed information about the underlying structure of the sentence. Due to the fact that this way of annotation is more detailed, therefore more elaborate, the model collection contains only about 400 sentences today. 2. General principles of annotation at the tectogrammatical level Tectogrammatical tree structures (TGTSs) are based on dependency syntax; they have the shape of a dependency tree with the verb as the root of the tree and its daughter nodes representing nodes depending on the governor (on each layer of the tree). The two dimensions of the tree represent the syntactic structure of the sentence (the vertical dimension) and the topic-focus articulation of the sentence, based on the underlying word order (the horizontal dimension). The tagging at the tectogrammatical level can be described by the following principles: (a) a single node of a TGTS may be a representation of more than one word; only autosemantic words have a node of their own, while the correlates of functional words (auxiliaries, prepositions etc.) are attached to the autosemantic words to which they belong (auxiliary verbs and subordinating conjunctions to the verbs, prepositions to nouns, etc.); (b) in the cases of deletion in the surface shape of the sentence, nodes are introduced into the tectogrammatical tree to 'recover' a deleted word; The Prague Bulletin of Mathematical Linguistics 78, 2002 38 (c) no non-projective structures are admitted at the tectogrammatical level (projectivity is a counterpart to continuity of constituents; non-projectivity appearing in the surface shape of some sentences is supposed to be handled by movement rules between the tectogrammatical tree and the morphemic string); (d) not only the direction of the dependence on the governing node (dependence to the left, dependence to the right) is taken into account, but also sister nodes are ordered (from left to right); their order reflects the scale of communicative dynamism which is relevant for the description of topic-focus articulation.