NYCU-NLP at SemEval-2026 Task 9: Stacking Small Language Models for Multilingual, Multicultural and Multievent Polarization Detection

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

Summary

The NYCU-NLP system participated in SemEval-2026 Task 9, focusing on online polarization analysis. Their approach utilized instruction-tuned small language models (SLMs) including Phi-4 (14B), Mistral-small-3.2 (24B), and Gemma-3 (27B). These models were combined using a stacking ensemble strategy, enhanced by task-specific prompting, to leverage their complementary strengths. The system was evaluated across 22 languages and three subtasks, achieving macro-averaged F1 scores of 0.8071 for Polarization Detection (Subtask 1), 0.6108 for Polarization Type Classification (Subtask 2), and 0.5111 for Polarization Manifestation Identification (Subtask 3). This method secured top ranks, including first in 15, second in 12, and third in 10 of 62 language-specific leaderboards, demonstrating its effectiveness in multilingual contexts.

Key takeaway

For NLP engineers developing multilingual content moderation or social media analysis tools, consider implementing stacking ensembles of instruction-tuned small language models. This approach, demonstrated by NYCU-NLP's success in SemEval-2026 Task 9, offers robust performance across diverse languages and polarization subtasks. You can achieve competitive results by combining models like Phi-4, Mistral-small-3.2, and Gemma-3 with tailored prompting strategies.

Key insights

Stacking instruction-tuned small language models effectively detects multilingual online polarization.

Principles

• Stacking ensembles leverage complementary capacities of diverse SLMs.
• Task-specific prompting enhances SLM performance in specialized tasks.

Method

Combine instruction-tuned SLMs (e.g., Phi-4, Mistral-small-3.2, Gemma-3) with task-specific prompting, then apply a stacking ensemble for improved multilingual polarization detection.

In practice

• Utilize Phi-4, Mistral-small-3.2, or Gemma-3 for polarization tasks.
• Apply stacking ensembles for robust multilingual NLP.

Topics

• Small Language Models
• SLM Stacking
• Multilingual NLP
• Polarization Detection
• SemEval-2026
• Ensemble Methods

Best for: AI Engineer, 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.
• Beyond Cascaded Pipelines: Building a Native Spoken Language Model Prototype — Native Spoken Language Models (SLMs) overcome cascaded pipeline limitations by directly processing audio, preserving nuance, reducing errors, and lowering latency.
• Understanding LLM Behavior in Multi-Target Cross-Lingual Summarization — LLMs perform cross-lingual summarization and translation jointly in later layers, not sequentially.
• TeamHerald@CHIPSAL 2026: Hate Speech Detection and Sentiment Analysis of Nepali Memes using Transformer-based Architectures and Ensemble Learning — Ensemble strategies for meme analysis perform differently for binary vs. multi-class tasks, requiring task-specific aggregation.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Can Small GenAI Language Models Rival Large Language Models in Understanding Application Behavior? — Small GenAI models can rival larger LLMs in malware detection, offering efficiency for resource-constrained environments.

Open in AIssential →

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