Tagset

A tagset is a defined collection of tags used to label items in a data annotation scheme. In linguistics and natural language processing, a tagset commonly labels words with grammatical categories, such as part of speech, and may also encode additional morphosyntactic features like tense, number, or case. In markup and data representation, the term refers to the set of tags or element names that a language, tool, or standard recognizes.

In NLP, tagsets are central to corpus annotation, tagging, parsing, and information extraction. They vary in granularity from coarse schemes that distinguish broad classes (nouns, verbs, adjectives) to fine-grained sets that differentiate subtypes (proper vs common noun, past vs present tense verb). Well-known examples include the Penn Treebank tagset, with about 45 POS tags, and the Universal Dependencies tagset, which is designed to be cross-linguistically applicable and compact. The choice of tagset affects tagging accuracy, interoperability, and the transfer of models across languages. Design goals typically balance coverage with consistency and ease of annotation.

Tagset design also involves practical considerations such as annotation guidelines, training costs, and inter-annotator agreement. Researchers often provide conversion mappings between tagsets to enable comparisons or cross-system applications. Beyond linguistic tagging, tagsets appear in markup languages, where a fixed set of element names defines a document’s structure, and in information extraction for labeling named entities, relations, or events.