UzDeriv: A morphologically segmented dataset of Uzbek derivational word forms with lexeme-level part-of-speech annotation
Annotatsiya
UzDeriv is a dataset of 7,819 derived Uzbek word forms segmented into theirconstituent morphemes and annotated for part of speech. Each record gives thewritten form, its segmentation into an optional prefix, a root and up to foursuffixes, the part of speech of the root, and the part of speech of the derivedform. Character-level IOB labels aligned to the written form are provided forthe 7,762 records whose morphemes concatenate to the surface string; theremaining 57 records, in which a sound alternation applies at a morphemeboundary, are listed separately together with the string the morphemes wouldotherwise produce. The records originate from the knowledge base of the Morphoanalyzer system, arule-based morphological analyser for Uzbek completed in 2021 and deployed atthree library information centres. The archive also contains a lexeme inventoryof 4,858 entries keyed on the pair (root, part of speech), which resolves 4,733root strings and yields a list of 122 roots corresponding to more than onelexeme; an affix inventory of 217 entries recording frequency, slotdistribution and category transition; training, development and test portionsof 6,257, 783 and 779 records that are disjoint by root; and baseline scoresfor four segmentation systems and three classifiers. Documentation covers the schema, the annotation guidelines, a datasheet in theformat of Gebru et al., and a quality report. The scripts that regenerate everyderived file from the source database export are included, together with anaudit log of the 22 corrections applied. Not included: the source root lexicon of 76,978 strings, and an aggregate ofits subject-field codes. That lexicon carries abbreviations and residual glossfragments consistent with digitisation from a published explanatory dictionary,and its provenance has not been established.