Prague Dependency Treebank - Consolidated 2.0: Enriching a Complex Annotation Scheme

The Prague Dependency Treebank - Consolidated 2.0 (PDT-C 2.0) represents the culmination of a 30-year project to create a unified, high-quality linguistic resource for the Czech language. By consolidating previous versions into a single, coherent package, the authors provide a massive, manually annotated dataset that supports research ranging from basic morphology to complex discourse analysis. The resource is designed to be genre-diversified, covering written, translated, spoken, and user-generated texts, making it a versatile tool for both linguistic theory and the development of modern NLP systems.

A defining feature of PDT-C 2.0 is its hierarchical, interlinked annotation scheme. This architecture allows researchers to navigate between different levels of linguistic representation:

• Morphological Layer: Assigns lemmas and detailed morphological tags to every token.
• Surface Syntax Layer: Captures dependency structures, identifying syntactic functions like subjects and objects.
• Deep Syntax and Semantics Layer: Represents the underlying meaning of sentences, including valency, coreference, and discourse relations.

By explicitly linking these layers, the treebank enables researchers to trace how specific surface-level tokens contribute to deep semantic structures, providing a robust framework for studying language at multiple levels of abstraction simultaneously.

PDT-C 2.0 is distinguished by its reliance on manual annotation across all layers, ensuring high consistency and accuracy. The corpus includes specialized annotations such as speech reconstruction for spontaneous spoken data, which cleanses transcripts to meet written-text standards. Furthermore, the inclusion of discourse relations and coreference across all datasets provides a rich foundation for studying inter-sentential phenomena. The corpus is fully compatible with external lexical resources, such as the MorfFlex morphological dictionary and the PDT-Vallex valency lexicon, which further support the consistency of the annotations.