PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation for Multilingual Polarization Detection

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

Summary

The PolAR Bears system, developed for SemEval-2026 Task 9, addresses Multilingual Polarization Detection across four low-resource Indian languages: Hindi, Bengali, Telugu, and Odia. It tackles three subtasks: Polarization Detection, Type Classification, and Manifestation Identification. To combat data scarcity, the system employs cross-lingual data augmentation using IndicTrans2, expanding the dataset fourfold. Its unified architecture leverages Qwen3-4B-Instruct, optimized via QLoRA, training a linear classification head on masked mean-pooled hidden states with only ~33M trainable parameters. The system achieved competitive results in Subtask 1, with an average Macro F1 of 0.813 across all languages, peaking at 0.8668 for Telugu. However, for the complex multi-label frameworks of Subtasks 2 and 3, results exposed significant pre-training bias within foundational LLMs, with Hindi maintaining strong F1 scores of 0.7008 and 0.7248, but performance dropping considerably for the other three languages.

Key takeaway

For NLP Engineers developing multilingual systems in low-resource Indian languages, you should prioritize cross-lingual data augmentation with tools like IndicTrans2 and parameter-efficient fine-tuning (e.g., QLoRA on Qwen3-4B-Instruct) to achieve competitive initial performance. However, be prepared for significant pre-training bias in foundational LLMs when tackling complex multi-label tasks, necessitating further domain-specific adaptation or bias mitigation strategies for nuanced rhetorical analysis.

Key insights

Cross-lingual augmentation and PEFT address data scarcity for multilingual polarization detection, but LLM pre-training bias affects nuanced cross-lingual transfer.

Principles

• Cross-lingual data augmentation expands datasets fourfold for low-resource languages.
• QLoRA optimizes Qwen3-4B-Instruct with ~33M trainable parameters for efficiency.
• Foundational LLMs exhibit pre-training bias in complex multi-label cross-lingual tasks.

Method

The system employs IndicTrans2 for cross-lingual data augmentation, expanding datasets fourfold. It uses Qwen3-4B-Instruct optimized via QLoRA, training a linear classification head on masked mean-pooled hidden states.

In practice

• Apply IndicTrans2 for low-resource language data augmentation.
• Utilize QLoRA for efficient fine-tuning of large LLMs like Qwen3-4B-Instruct.
• Evaluate LLM pre-training bias for complex cross-lingual rhetorical tasks.

Topics

• Multilingual NLP
• Polarization Detection
• Parameter-Efficient Fine-Tuning
• QLoRA
• Cross-Lingual Augmentation
• Low-Resource Languages

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.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.
• LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification — Iterative imbalance-aware fine-tuning of LLMs, combined with augmentation and post-processing, significantly improves rare-class NLP performance.
• Low-resource Language Discrimination Towards Chinese Dialects with Transfer learning and Data Augmentation — Transfer learning and data augmentation effectively overcome low-resource challenges in Chinese dialect discrimination.
• Datasets for Verb Alternations across Languages: BLM Templates and Data Augmentation Strategies — This work addresses the underexplored ability of large language models (LLMs) to capture cross-sentence paradigmatic patterns, specifically verb alternations.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.

Open in AIssential →

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