Tagsets

A tagset is a collection of labels used to annotate linguistic data in corpora and natural language processing. Each tag represents a category such as a part of speech, a morphological feature, a named entity class, or a semantic role, and a token may receive one or more tags depending on the scheme. Tagsets enable consistent annotation, data retrieval, and linguistic analysis across corpora and NLP systems.

Tagsets vary in granularity and language focus. Part-of-speech tagsets range from coarse to fine-grained classifications. Morphological tagsets encode features such as tense, number, gender, case, and aspect. Named-entity tagsets distinguish types like person, organization, and location. Some tagsets are language-specific, while others aim for cross-linguistic compatibility. An example of the latter is the Universal Dependencies POS tagset, which provides a common set of tags and morphological features usable across languages.

Notable examples include the Penn Treebank II POS tagset, widely used for English corpora, and the Brown corpus tagset. Universal Dependencies offers language-agnostic morpho-syntactic features. For named entities, tagsets used in CoNLL and related shared tasks often define categories like PER, LOC, and ORG. Tagging schemes such as BIO or IOB are commonly employed to mark multi-token entities within a sequence.

Design considerations for tagsets include choosing the desired level of granularity, providing precise and publicly available definitions, documenting guidelines, and versioning. Interoperability is enhanced by adopting standard tagsets or mappings between tagsets. Tagsets play a central role in corpus annotation, evaluation of automatic taggers, and cross-language comparative linguistics.