treebanks

Treebanks are linguistically annotated corpora in which sentences are paired with syntactic structure representations. Broadly, they come in two traditions: constituency treebanks, which mark hierarchical phrase structure (such as NP, VP, and S), and dependency treebanks, which encode head–dependent relations between words. Some projects provide both views for the same data. Treebanks are foundational resources in computational linguistics and natural language processing, enabling systematic study of syntax and training of parsing models.

The creation of a treebank typically involves manual annotation guided by formal schemes or guidelines. Annotators assign syntactic structures and often part-of-speech tags and morphological features; in some cases semantic roles or functional annotations are added. To ensure quality, multiple annotators may work on the same data, with adjudication and measurements of inter-annotator agreement. The resulting resource is then mapped to a standard format and made accessible for research and development.

Treebanks serve multiple purposes. They provide training data for syntactic parsers, serve as benchmarks for evaluating parsing accuracy, and support linguistic research on cross-linguistic syntax, syntactic theory, and language variation. They also underpin downstream NLP tasks such as information extraction and machine translation, where explicit syntactic information improves performance.

Notable examples include the Penn Treebank for English constituency syntax, the Chinese Treebank, the Prague Dependency Treebank for Czech, and the Tiger Treebank for German. The Universal Dependencies project offers a large, multilingual, cross-lynt representation of dependency trees, emphasizing cross-language comparability and broad accessibility.