Treebank

A treebank is a linguistics resource—a corpus of text that has been annotated with syntactic structure. In most treebanks, sentences are annotated with either constituency parse trees or dependency graphs, often accompanied by part-of-speech tags, named entities, morphological features, and sometimes semantic roles. Treebanks are used to study syntax and to train and evaluate parsing algorithms.

The best known example is the Penn Treebank (PTB), created at the University of Pennsylvania and the Linguistic Data Consortium in the 1990s. It annotated the Wall Street Journal portion of the WSJ corpus with constituency trees and became a standard benchmark for statistical parsers. Other well-known treebanks include the Chinese Treebank, the Prague Dependency Treebank, and the multilingual Universal Dependencies collection, which provides cross-linguistic dependency annotations.

Treebanks differ in their annotation schemes, domains, sizes, and licensing. PTB uses bracketed constituency representations; PDT uses rich dependency annotations with functional relations; UD applies a single universal dependency scheme across languages. Treebank projects typically publish annotation guidelines, inter-annotator agreement measures, and revision histories to ensure consistency and reproducibility.

Treebanks are created through manual annotation guided by formal guidelines, often aided by automatic pre-annotation followed by human correction. They are used to train and evaluate parsers, study syntactic phenomena, perform cross-linguistic comparison, and support downstream NLP tasks such as information extraction and machine translation. Availability ranges from open-access corpora to licensed datasets.