PolAR Bears at SemEval-2026 Task 9: Parameter-Efficient Fine-Tuning and Cross-Lingual Augmentation for Multilingual Polarization Detection
Year: 2026
Type: conference-paper
Abstract: This paper describes our system for SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization.We focus on four low-resource Indian languages (Hindi, Bengali, Telugu, and Odia) across three subtasks: Polarization Detection, Type Classification, and Manifestation Identification.To address data scarcity, we employ cross-lingual data augmentation using IndicTrans2, expanding our dataset fourfold.Our unified architecture leverages Qwen3-4B-Instruct optimized via QLoRA, training a linear classification head on masked mean-pooled hidden states.Our system achieved an average Macro F1 of 0.813 across all languages in Subtask 1.For the multi-label frameworks of Subtasks 2 and 3, our results expose a significant pre-training bias within foundational LLMs; while Hindi maintained strong F1 scores of 0.7008 and 0.7248, performance dropped considerably for the other three languages, highlighting ongoing challenges in cross-lingual transfer for nuanced rhetorical techniques.
Authors: Vinay Ulli, Jyoti Kumari
Institution: Banaras Hindu University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166848998