home tagset/conversion edit page issue tracker

Tagset conversion tables to universal tags and features

Disclaimer: The UD tags have substantive definitions and are not simply equivalence classes of the part-of-speech (POS) tags of language-particular treebanks. That is, the situation is different from the pure equivalence class approach used in the original 2011 Google Universal POS Tags work. Some tags can only be mapped correctly if we also know the lemma or the syntactic context. As a concrete example, we now distinguish adpositions (ADP) from subordinating conjunctions (SCONJ). This distinction is not available in the Penn Treebank English POS tag set (where both are IN), and for words that can be used in either role, the correct UD POS tag can only be recovered by looking at the syntactic context in which the word appears. Such information is not used in the mappings made available here. These conversion tables were generated automatically via Interset. It uses only tags (+ features) as input; therefore the result is only an approximation of correct UD tags. Thus both the tables and the applications of them to POS-tagged text will likely require manual postprocessing in order to provide accurate and complete UD information. Nevertheless, in practice, most tags in most languages can still be mapped correctly without such use of syntactic content of morphological analysis, and hence this automatic conversion can be quite useful.