The First Spoken Spanish Treebank in Universal Dependencies: Annotation, Distributional Analysis, and Parsing Evaluation

Resumen

This paper presents the first treebank for spoken Spanish within the Universal Dependencies (UD) framework, built from transcriptions of the Corpus Oral y Sonoro del Español Rural (COSER). The treebank comprises sentences from 17 Spanish regions, manually annotated using Arborator-Grew. We provide a distributional analysis of dependency relations revealing differences between spontaneous speech and written text, including higher frequencies of discourse adverbs, disfluencies, and coordination, alongside reduced nominal complexity. Evaluation using UDPipe 2 as a pre-trained baseline yields a UAS of 75.51% and LAS of 66.48% on the orthographic version of the treebank. Fine-tuning a spaCy transformer via 10-fold cross-validation achieves a UAS of 77.94% and LAS of 67.52%, surpassing the pre-trained spaCy transformer baseline by +8.81 pp in LAS -the dominant gain attributable to learning speech-exclusive dependency labels- and surpassing the stronger UDPipe 2 baseline by +1.04 pp in LAS. Error analysis reveals that word order variability (obj<->nsubj confusion) and spoken-specific labels (reparandum, discourse) constitute the primary challenges for written-trained parsers.