Aaron at SemEval-2026 Task 9: Multilingual Polarization Detection using Transformer-Based Models with Class Weighting and Threshold Tuning

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

Summary

Aaron Anampiu's submission to SemEval-2026 Task 9 presents a method for detecting multilingual, multicultural, and multievent online polarization across English and Swahili. The approach addresses three subtasks: binary polarization detection, polarization type classification, and manifestation identification. It employs transformer-based models, specifically RoBERTa-base for English and AfroXLMR-base for Swahili. To counter severe label imbalance, the method integrates class-weighted loss functions and utilizes per-label threshold tuning to optimize multi-label classification. On the test set, the system achieved F1 macro scores of 0.7901 for English and 0.7910 for Swahili in Subtask 1, 0.4615 (English) and 0.4808 (Swahili) in Subtask 2, and 0.4791 (English) and 0.5830 (Swahili) in Subtask 3. These results demonstrate competitive performance, though error analysis highlighted difficulties in detecting dehumanization and lack of empathy.

Key takeaway

For NLP Engineers developing multilingual content moderation systems, consider integrating class-weighted loss functions and per-label threshold tuning. This approach, demonstrated with RoBERTa-base and AfroXLMR-base, significantly improves performance in detecting online polarization across languages like English and Swahili, especially with imbalanced datasets. Be aware that current models still struggle with nuanced aspects such as dehumanization and empathy detection, requiring further refinement.

Key insights

Transformer models with class weighting and threshold tuning effectively detect multilingual online polarization despite label imbalance.

Principles

• Address label imbalance with class-weighted loss.
• Optimize multi-label classification via per-label threshold tuning.
• Transformer models adapt across languages (English, Swahili).

Method

The method involves fine-tuning transformer models (RoBERTa-base, AfroXLMR-base) with class-weighted loss functions and applying per-label threshold tuning for multi-label classification in polarization detection.

In practice

• Apply RoBERTa-base for English text analysis.
• Use AfroXLMR-base for Swahili language tasks.
• Implement class weighting for imbalanced datasets.

Topics

• Multilingual Polarization Detection
• Transformer Models
• Class Weighting
• Threshold Tuning
• RoBERTa-base
• AfroXLMR-base

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.
• From TF-IDF to Transformers: A Comparative and Ensemble Approach to Sentiment Classification — A soft voting ensemble of diverse machine learning and transformer models significantly enhances sentiment classification accuracy, particularly for nuanced text.
• An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect — A hybrid framework combining transformer embeddings with classical classifiers enables effective rumour detection in low-resource Algerian dialect.
• 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.
• Indigenous Letters to Brazil: Multi-Label Classification — Multi-label classification of indigenous letters presents challenges due to corpus imbalance and semantic overlap.
• Model-Agnostic Meta Learning for Class Imbalance Adaptation — HAMR dynamically addresses class imbalance and data difficulty in NLP via adaptive weighting and neighborhood-aware resampling.

Open in AIssential →

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