The ccomp/xcomp Boundary in Turkic Universal Dependencies: Evidence from Parallel Annotation and Low-Resource Perspectives
Year: 2026
Type: conference-paper
Abstract: Reliable multilingual parsing rests on consistent syntactic annotation across treebanks, yet the Universal Dependencies (UD) annotation of complement clause relations in Turkic languages has developed without explicit cross-treebank coordination. The structural root of this gap lies in the character of Turkic complementation: instead of the lexical subordinators characteristic of European grammars, Turkic languages realize complement arguments through inflected verbal nouns carrying case and possessive morphology – a configuration that the UD guidelines handle poorly in pro-drop environments. This challenge is sharpest for under-resourced varieties such as Kyrgyz, whose two UD treebanks provide approximately 25,000 annotated tokens, roughly one-tenth of Turkish UD coverage. The present study conducts a cross-treebank data quality audit of ccomp, xcomp, and csubj relations across a parallel corpus of 148 sentences in four Turkic languages (Turkish, Kyrgyz, Uzbek, Azerbaijani) and five large monolingual treebanks. Inter-annotator agreement, computed on all 44 sentences containing complement clauses, yields Fleiss’ κ = 0.29 and Krippendorff’s α = 0.30, alongside an observed cross-treebank divergence rate of 40.9% (18 of 44 sentences); the numerical proximity of the two figures is coincidental, as κ is chance-corrected while the divergence rate is a raw proportion. Possessive agreement marking on the nominalized verb is identified as the most reliable Turkic-specific diagnostic: 40.6% of ccomp tokens in Turkish bear a possessive suffix against 5.7% of xcomp tokens. A five-rule decision tree achieves 65– 67% agreement with existing annotations. A structurally distinct Kyrgyz desiderative construction is described whose behavior tentatively indicates a typologically motivated basis for csubj annotation where other Turkic languages use xcomp. Four concrete recommendations are proposed to harmonize Turkic UD guidelines and strengthen the Kyrgyz NLP resource base.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224892626