MKJ at SemEval-2026 Task 9: A Comparative Study of Generalist, Specialist, and Ensemble Strategies for Multilingual Polarization

Abstract: We present a systematic study of multilingual polarization detection across 22 languages for SemEval-2026 Task 9 (Subtask 1), contrasting multilingual generalists with language-specific specialists and hybrid ensembles. While a standard generalist like XLM-RoBERTa suffices when its tokenizer aligns with the target text, it may struggle with distinct scripts (e.g., Khmer, Odia) where monolingual specialists yield significant gains. Rather than enforcing a single universal architecture, we adopt a language-adaptive framework that switches between multilingual generalists, language-specific specialists, and hybrid ensembles based on development performance. Additionally, cross-lingual augmentation via NLLB-200 yielded mixed results, often underperforming native architecture selection and degrading morphologically rich tracks. Our final system achieves an overall macro-averaged F1 score of 0.796 and an average accuracy of 0.826 across all 22 tracks. Code and final test predictions are publicly available at: https://github.com/Maziarkiani/SemEval2026-Task9-Subtask1-Polarization.

• The paper introduces a language-adaptive pipeline that selects optimal architectures (generalist, specialist, ensemble) based on development set performance.
• It achieves a macro-averaged F1 of 0.796 across 22 languages, with notable improvements in underrepresented and morphologically complex scripts.
• Results reveal that native model selection outperforms naive cross-lingual augmentation, emphasizing challenges in multilingual polarization detection.

Comparative Analysis of Generalist, Specialist, and Ensemble Strategies for Multilingual Polarization Detection

Problem Formulation and Motivation

The paper "MKJ at SemEval-2026 Task 9: A Comparative Study of Generalist, Specialist, and Ensemble Strategies for Multilingual Polarization" (2604.21370) presents a comprehensive approach to polarization detection across 22 typologically diverse languages within the POLAR benchmark, focusing on Subtask 1 of SemEval-2026 Task 9. Polarization detection, as distinguished from conventional sentiment and stance analysis, requires models to capture subtle, ideologically divisive rhetoric in multilingual and multi-script settings. The study directly addresses three fundamental questions: the adequacy of large-scale multilingual generalists (e.g., XLM-RoBERTa), the conditions under which monolingual specialists or alternative generalists outperform such baselines, and the efficacy of cross-lingual data augmentation in low-resource regimes.

Methodological Framework

The proposed methodology abandons a uniform architecture in favor of a language-adaptive pipeline, which empirically selects among generalist, specialist, and ensemble models for each language based on development set performance:

• Generalists: High-capacity multilingual transformers such as XLM-RoBERTa and mDeBERTa-v3 serve as strong baselines, particularly for languages with adequate subword coverage and typological proximity to high-resource scripts.
• Specialists: For languages where tokenization misalignment or script-specific features cause generalists to falter, monolingual models pretrained on native corpora (e.g., AraBERT for Arabic, Metythorn Khmer-XLMR for Khmer) are prioritized.
• Hybrid Ensembles: In cases of complementary error profiles, soft-voting and weighted ensembles (e.g., DeBERTa-v3-Large/BERTweet for English) are applied, leveraging domain adaptation and capacity to offset the limitations of individual systems.
• Architecture Selection Policy: Non-baseline architectures are adopted only if they achieve at least a 2-point Macro-F1 improvement or yield a better precision–recall trade-off on the development set.
• Threshold Calibration: For certain tracks, class imbalance and prediction bias are mitigated through tuned decision thresholds based on observed bias or over-prediction, with $\tau$ adjusted accordingly.

Moreover, the paper implements cross-lingual data augmentation, translating English training data into other languages using NLLB-200. However, results from this approach suggest instability and frequent performance drops in morphologically rich target languages, making native architecture selection preferable.

Experimental Results

The final system secures an overall macro-averaged F1 of 0.796 and an average accuracy of 0.826 across 22 tracks. Notably:

• In 18 out of 22 languages, specialized or alternative generalist architectures improved over the XLM-R baseline, often by substantial margins (e.g., Odia: +10.6%, Khmer: +8.1% Macro-F1).
• The model is within 4 points of the public leaderboard SOTA for 13 languages, maintaining robust consistency across both high-resource and challenging low-resource scripts.
• Weighted ensembles notably benefit tracks with strong domain adaptation needs or error diversification (e.g., Hindi, Persian, English).
• Contradictory finding: Naive cross-lingual translation augmentation generally reduced performance in highly inflected languages (e.g., Polish, Russian, Urdu), often underperforming compared to the best architecture chosen without augmentation.

Failure analysis reveals several consistent patterns:

• Recall Trap: In languages like Khmer or Urdu, the model collapses into majority-class prediction, failing to recover the neutral class.
• Specialist Trap: Overfitting to small validation splits leads specialists (e.g., GBERT for German) to outperform on dev but not generalize to test.
• Semantic Overfitting: Large ensembles on extremely small validation sets (e.g., Punjabi) are prone to learning dataset-specific artifacts, not general attributes.

Analysis of dev–test shift further highlights tracks where the pipeline exhibits robust generalization (e.g., Telugu, Burmese), high stability, and cases of performance regression due to overfitting or label bias.

Implications for Multilingual NLP

The findings provide rigorous empirical evidence that universal multilingual models, despite their scale, are insufficient for robust cross-lingual polarization detection. Tokenization inefficiencies and domain misalignment present critical bottlenecks, particularly for underrepresented scripts, morphologically complex, and resource-scarce languages. Monolingual specialists and high-capacity multilinguals with script coverage (e.g., mDeBERTa-v3) deliver clear gains when appropriately matched to the data typology and domain.

The demonstrated instability of naive forward translation for data augmentation—especially its negative effect on languages with complex morphology—has direct implications for cross-lingual transfer research. The results suggest that augmentation strategies for polarization detection must account for both target language morphology and the translation model's capacity for inflectional precision.

Practically, the approach demonstrates the superiority of language-adaptive architecture selection over monolithic scaling or undifferentiated augmentation. The ensemble strategies validate the utility of combining domain-adaptive specialists with robust generalists to counter individualized error modes and class imbalance.

Future Directions

The research identifies several avenues for further work:

• Systematic multi-seed and large-scale ablation studies to quantify validation set sensitivity.
• Fine-grained analysis of ensemble weight perturbations and the incorporation of advanced regularization, such as focal loss, to address class collapse in highly imbalanced tracks.
• Exploration of more sophisticated augmentation (e.g., back-translation, paraphrastic generation) conditional on robust inflectional mapping and domain fit.
• Extension of the tokenization fragmentation framework, potentially guiding automatic model selection for unseen or unseen-like languages.

In the context of AI developments, these findings underscore that scaling LLMs alone is unlikely to solve nuanced sociopolitical tasks at the multilingual frontier without explicit architectural, tokenization, or training policy adaptations. They also highlight a persistent need for resources and community efforts in low-resource and morphologically unique languages.

Conclusion

This work establishes that multilingual polarization detection, as posed by the POLAR benchmark in SemEval-2026, resists universal architectures. Performance is intricately determined by the interaction of tokenizer fit, domain adaptation, ensemble calibration, and sensitivity to resource sparsity. The failure of naive cross-lingual augmentation in several tracks demonstrates the non-triviality of data transfer in real-world NLP system design. The systematic language-adaptive approach presented here offers a robust, empirically validated blueprint for high-performance multilingual NLP systems under non-uniform conditions (2604.21370).