PDT-VALLEX: Creating a Large-coverage Valency Lexicon for Treebank Annotation

The valency theory as a part of the theory of Functional Generative Description ([16]) of language meaning has been around for some time ([14]). However, it is for the first time that a large-scale corpus (the Prague Dependency Treebank (PDT, [4]) has been fully annotated with valency information based on this theory, i.e., with fully referenced valency lexicon at each relevant verb, noun or adjective.1 The PDT is a long-term research project, whose main aim is a complex manual annotation of (roughly) a one-million-word part of the Czech National Corpus.2 It is being annotated on three layers. On the lowest, morphological layer the lexical entry (usually represented by a lemma) and values of morphological categories (person, number, tense, gender, voice, aspect, . . . ) are assigned to each word. At the analytical layer, a sentence is represented as a dependency tree. Nodes of the tree represent tokens (i.e. word forms and punctuation marks) as they are found in the original sentence. No node is added or deleted. Edges usually (where it makes sense) represent relation of formal dependency. In addition, an analytical function capturing the type of dependency relation between the child and its parent is added. The highest (or, “deepest” depending on the point of view) layer is the tectogrammatical layer ([7]). It captures the deep (underlying) structure of a sentence. Nodes represent only autosemantic words; synsemantic (i.e., auxiliary) words and punctuation marks are not represented by nodes, they may only affect values of attributes of the autosemantic words which they are attached to. At this layer, several attributes are assigned to each node, one of the most important ones