Controlled Evaluation of Document-Level Inter-Mention Links for Named-Entity Type Classification in Ukrainian Texts

Document-level links between repeated mentions can provide contextual evidence unavailable to a local entity-type classifier, a potentially important capability for Ukrainian because of its limited annotated resources and rich inflectional morphology. However, their specific contribution is difficult to isolate because graph-based systems often simultaneously alter the classifier architecture. We evaluate named-entity type classification with known mention boundaries on NER-UK 2.0 under controlled conditions, using frozen XLM-RoBERTa-base representations, 10 optimization seeds, and 270 graph-head training runs. The evaluation design combines a no-message baseline with matched head parameterization, degree-preserving topology shuffling, simple averaging, oracle diagnostics, and controlled context-weakening interventions; no novel graph architecture is claimed. On intervention-free data, the exact-repeat graph changed macro-F1 by +0.0041 (95% CI [−0.0042,0.0127]); thus, an overall benefit was not established, although the interval did not rule out a practically meaningful effect. A Ukrainian token-level lemmatization rule increased graph coverage from 34.41% to 42.07% but did not improve performance over exact repetition (Δ=−0.0033, 95% CI [−0.0135,0.0055]). When only the target mention’s context was artificially weakened, with no explicit degradation cue, graph refinement increased target accuracy by 0.0784 (95% CI [0.0544,0.1021]). The advantage reversed when the entire repeat component was weakened, while simple averaging performed at least as well as the learned graph head. Overall, the results support graph-based refinement as a conditional context-recovery mechanism rather than a universal replacement for local classification. The evidence remains controlled but not independently confirmatory and does not establish gains for end-to-end NER or robustness to naturally degraded context.