Aatman at SemEval-2026 Task 9: Transfer Learning for Multilingual Polarization Detection

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

Summary

Aatman's system for SemEval-2026 Task 9, POLAR, addresses multilingual polarization detection as a binary classification problem across 22 diverse languages. The team explored three approaches: supervised fine-tuning of multilingual encoder-only transformer models like mDeBERTa, zero- and few-shot classification using large language models (LLMs), and transfer learning from related harmful language tasks such as hate speech and toxicity. While mDeBERTa demonstrated the strongest baseline performance among supervised models, and prompt-based LLM methods showed limited effectiveness, particularly in zero-shot settings, the most successful results came from transfer learning. This method, involving initial fine-tuning on related task datasets before adaptation to polarization, achieved a Macro-F1 score of 0.81. These findings highlight the continued effectiveness of supervised multilingual encoders and the substantial performance gains from integrating related harmful language tasks for nuanced polarization detection.

Key takeaway

For NLP Engineers developing multilingual polarization detection systems, prioritize transfer learning from related harmful language tasks. Your models, especially supervised multilingual encoders like mDeBERTa, can achieve a Macro-F1 score of 0.81 by first fine-tuning on datasets for hate speech or toxicity. Avoid relying on zero-shot prompt-based methods with open-weight LLMs, as they demonstrated limited effectiveness in this context. Focus on utilizing existing annotated data from similar tasks to substantially improve performance on nuanced polarization expressions.

Key insights

Transfer learning from related harmful language tasks significantly boosts multilingual polarization detection performance.

Principles

• Supervised multilingual encoders remain effective.
• Related harmful language tasks improve detection.
• LLM prompt-based methods have limited zero-shot utility.

Method

Fine-tune models on related harmful language datasets, then adapt to the specific polarization detection task.

In practice

• Use mDeBERTa for strong multilingual baselines.
• Incorporate hate speech data for transfer learning.
• Prioritize supervised fine-tuning over zero-shot LLMs.

Topics

• Multilingual Polarization Detection
• Transfer Learning
• Harmful Language Detection
• Transformer Models
• Large Language Models
• SemEval

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, 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.
• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.
• Multilingual Detection of Alzheimer's Disease from Speech: A Cross-Linguistic Transfer Learning Approach — Cross-language transfer learning effectively detects Alzheimer's Disease from speech across multiple languages, achieving strong generalization and rapid inference.
• Comparison of Modern Multilingual Text Embedding Techniques for Hate Speech Detection Task — Supervised models with multilingual embeddings and gradient boosted trees excel at hate speech detection.
• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.

Open in AIssential →

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