Building the PDT-Vallex valency lexicon

In our contribution, we relate the development of a richly annotated corpus and a computational valency lexicon. Our valency lexicon, called PDT-Vallex (Hajič et al., 2003) has been created as a “byproduct” of the annotation of the Prague Dependency Treebank (PDT) but it became an important resource for further linguistic research as well as for computational processing of the Czech language. We will present a description of the verbal part of this lexicon (more than 5300 verbs with 8200 valency frames) that has been built on the basis of the PDT corpus. Rigorous approach and the linking of each verb occurrence to the valency lexicon has made it possible to verify and refine the very notion of valency as introduced in the Functional Generative Description theory (Sgall et al., 1986; Panevová, 1974-5). Every occurrence of a verb in the corpus contains a reference to its valency frame (i.e., to an entry in the PDT-Vallex valency lexicon). The annotators insert the verbs (verb senses) found in the course of the annotation and their associated valency frames into the lexicon, adding an example (or more examples) of its usage (directly from the corpus). They also insert a note that refers to another verb that has one of its valency frames related to the current one (a synonym/antonym, an aspectual counterpart, etc.). A functor as well as its surface realization is recorded in every slot of each valency frame. The mapping between the valency frame and its surface realization is generally quite complex (Hajič and Urešová, 2003). The surface realizations through the morphemic case, preposition and a case, and subordinate sentence are the most common. The valency frame is fully formalized to allow for automatic computerized processing of the valency dictionary entries. Verb complementations are marked for obligatoriness, and their surface realization is attached. The realization of inner participants (arguments) is always given in full, since there is no “standard” or “default” realization; free modifications’ (adjuncts’) realization need not be specified. The PDT-Vallex is available as part of the PDT version 2 published by the Linguistic Data Consortium (http://www.ldc.upenn.edu, LDC2006T01).