UMUSP at SemEval-2026 Task 9: Mitigating Cross-Lingual Interference via Selective Multilingual and Multitask Specialization

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

UMUSP's selective multilingual and multitask fine-tuning strategy addresses online polarization detection, improving cross-lingual stability compared to fully joint training. This approach, presented at SemEval-2026 Task 9, covers three subtasks: polarization detection (POLARDETECT), polarization type classification (POLARTYPE), and rhetorical manifestation identification (POLARMANIFEST) across all 22 languages of the shared task. The method involves empirically grouping languages and subtasks, then fine-tuning separate specialist models for each subset. Restricting parameter sharing significantly enhances performance, even without ensemble averaging, a contrast to jointly trained models where ensembling failed to mitigate instability. The final specialist ensemble boosted Task 3 macro-F1 from 0.3330 to 0.4920 and reduced cross-lingual dispersion from a CV of 0.613 to 0.321. The system secured 7th place among 16 submissions with full multilingual and multitask coverage, staying within 5% of the top system in 37.70% of evaluation conditions.

Key takeaway

For NLP Engineers building multilingual models for online polarization detection, consider adopting a selective fine-tuning strategy. Your systems can achieve greater cross-lingual stability and improved performance by empirically grouping languages and subtasks, then training specialized models for each subset. Avoid relying solely on ensembling fully jointly trained models, as this approach proved less effective in mitigating instability. Implement controlled specialization to enhance macro-F1 scores and reduce cross-lingual dispersion in your applications.

Key insights

Selective multilingual and multitask fine-tuning improves cross-lingual stability for polarization detection.

Principles

• Restricting parameter sharing enhances multilingual model performance.
• Joint training ensembles may not mitigate cross-lingual instability.
• Empirical grouping of languages and tasks aids specialization.

Method

The method involves empirically grouping languages and subtasks, then fine-tuning separate specialist models for each subset to achieve controlled specialization for online polarization detection.

In practice

• Group languages and tasks for specialized models.
• Prioritize restricted parameter sharing over joint ensembling.
• Apply to online content analysis for polarization.

Topics

• Multilingual Models
• Cross-Lingual Interference
• Polarization Detection
• Multitask Learning
• Fine-tuning Strategies
• SemEval-2026 Task 9

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.
• KIT's Submission to Cross-Lingual Voice Cloning in IWSLT 2026 — Explicit language tags and lexical matching significantly improve cross-lingual voice cloning performance and reduce accent leakage.
• Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning — Multilingual code-switching instruction tuning improves LLM performance beyond bilingual settings.
• When English Isn't the Best Teacher: Source Language Effects in Cross-Lingual In-Context Learning — ICL cross-lingual transfer differs from fine-tuning, requiring new source language selection heuristics.
• Cross-Lingual Empirical Evaluation of Large Language Models for Arabic Medical Tasks — English-centric LLMs exhibit significant performance gaps and reliability issues in Arabic medical tasks.
• Multilingual Fine-Tuning via Localized Gradient Conflict Resolution — Bucket-Level MOO resolves multilingual LLM fine-tuning interference via localized gradient-based multi-objective optimization on parameter buckets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.