La construcción del Corpus Oral y Sonoro del Español Rural-Anotado y Parseado (COSER-AP): avances en el etiquetado de partes del discurso

Abstract

This article presents the advances made for the construction of the ‘Annotated and Parsed Audible Corpus of Spoken Rural Spanish’. The methodology for building a treebank to evaluate the accuracy of state-of-the-art part of speech taggers: spaCy, Stanza and UDPipe is presented. It is shown that, when oral data is tagged the accuracy is 0.93; none of the regional varieties presents a significant difference in accuracy over the others. Regarding the grammatical categories, interjections, proper nouns, adjectives, and auxiliaries have the lowest F value. Finally, some examples of the polyfunctionality of some grammatical categories and the fuzzy boundaries between them are discussed, like the passive participle, and the ambiguity between adverbs and subordinate conjunctions that might affect accuracy.