Abstract
Universal Dependencies (UD) is widely used as a multilingual infrastructure for morphosyntactic annotation and evaluation, and its benchmark scores are typically reported at the level of whole treebank splits. For CMC-oriented questions, this is limiting because many treebanks contain heterogeneous material, including web, social, nonfiction, and academic content. Building on earlier work that released sentence-level genre annotations for UD as a reproducible, versioned sidecar layer, this paper presents a CMC-facing POS-tagging showcase in which social is used as a practical proxy for CMC- and social-media-adjacent material, and specifically trained UDPipe models are evaluated on full, social-only, and non-social test slices within the same underlying treebanks. The results do not support a simple "social is harder" conclusion but show that sentence-level genre labels expose evaluation contrasts that remain hidden under conventional whole-treebank scoring. The main contribution of the paper is therefore infrastructural and methodological: demonstrating that UD resources can be repurposed as reusable genre-sliced evaluation sets for CMC-oriented analysis.