NASIMLab at SemEval-2026 Task 9: A Comparative Analysis of Fine-Tuned Small Language Models vs. Generative Large Language Models for Multilingual Polarization Type Detection

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

Summary

NASIMLab's system for SemEval-2026 Task 9 addresses multilingual polarization type detection across 22 languages using the POLAR dataset, which comprises social media texts. The task involves identifying political, racial/ethnic, religious, gender/sexual, or other polarization types. Their approach fine-tunes language-specific small encoder-only models and compares them against large language models. The fine-tuned small models consistently outperform generative large language models, particularly in low-resource languages. This system achieved the top position on the leaderboard for Burmese (mya), ranked within the top 10 for 12 languages, and secured a spot within the top 20 for all other languages, demonstrating strong performance in this challenging task.

Key takeaway

For Machine Learning Engineers developing multilingual NLP systems, if you are weighing model choices for polarization detection or similar classification tasks, you should prioritize fine-tuning language-specific small encoder-only models. This approach consistently outperforms generative large language models, particularly for low-resource languages, offering a more efficient and accurate solution. You can achieve top-tier performance without the computational demands of larger models.

Key insights

Fine-tuned small encoder-only models consistently outperform generative large language models for multilingual polarization detection, especially in low-resource languages.

Principles

• Language-specific fine-tuning boosts performance.
• Small encoder-only models can surpass LLMs.
• Low-resource languages benefit most from SLMs.

Method

The system fine-tunes language-specific small encoder-only models on the POLAR dataset for polarization type detection across 22 languages.

In practice

• Apply fine-tuned SLMs for social media analysis.
• Prioritize SLMs for low-resource language tasks.
• Use encoder-only architectures for classification.

Topics

• Multilingual NLP
• Polarization Detection
• Small Language Models
• Large Language Models
• Fine-tuning
• Low-Resource Languages

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Lispector: Fine-tuning of Language Models for Grammar and Spelling Correction in Brazilian Portuguese — Supervised fine-tuning of smaller models outperforms few-shot prompting for grammar correction in Brazilian Portuguese.
• Large Language Models Make Sense When You See Why Previous AI Couldn’t Handle Language — Transformers' attention mechanism enabled LLMs to overcome sequential processing limitations, allowing for scalable language understanding and generation.
• Evaluating Monolingual and Multilingual Large Language Models for Greek Question Answering: The DemosQA Benchmark — Monolingual LLMs for under-resourced languages require specific evaluation against multilingual counterparts.
• Few Shades of Supervision for Discourse Segmentation — Classical fine-tuning of LLMs is most effective for EDU segmentation in spontaneous speech.
• Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats — Lightweight LLMs offer competitive biomedical NER performance, with specific output formats proving more effective than multi-format tuning.

Open in AIssential →

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