PDTB

The Penn Discourse Treebank (PDTB) is a linguistically annotated corpus designed to capture discourse relations in English text. Developed by researchers at the University of Pennsylvania as part of the Penn Discourse Treebank project, the resource provides annotations of discourse connectives and their semantic relations across multiple genres, including newswire, fiction, and blogs.

The annotation focuses on discourse relations signaled by explicit connectives (such as because, however) as well as implicit relations where no overt connective is present. For each relation, PDTB records the connective (when present), the two discourse units it links (Arg1 and Arg2), and a sense label from a hierarchical taxonomy.

The top-level senses in the PDTB taxonomy are Expansion, Contingency, Temporal, and Comparison, with finer-grained subtypes under each category. The taxonomy is designed to cover cross-sentence and cross-textual relations, noting that Arg1 can be unrelated to the connective's phrase order. The PDTB also marks the presence or absence of an explicit connective and includes annotation of argument spans.

PDTB serves as a foundational resource for automatic discourse parsing and related natural language processing tasks. It provides training data and evaluation benchmarks used to develop systems for identifying discourse relations, segmenting discourse, and inferring coherence in text.

Release history includes major versions such as PDTB 2.0 and PDTB 3.0, with expanded coverage and revised guidelines. The corpus and accompanying guidelines are publicly available for research, along with documentation on annotation procedures and evaluation metrics.