From left: Po-Chun Chu, Ding-Xiang Lin and Associate Professor Lung-Hao Lee of NYCU’s Institute of Intelligent Systems.

A Global Challenge Spanning 22 Languages

SemEval, the International Workshop on Semantic Evaluation, is one of the natural language processing field’s leading international evaluation forums. Task 9 this year focused on detecting multilingual, multicultural and multievent online polarization.

The challenge required AI systems to determine whether online content was polarized, identify the type of polarization involved and recognize how that polarization was expressed. The evaluation covered 22 languages, including Chinese, English, German, Spanish, Arabic, and Burmese, and drew participants from universities and research institutions worldwide.

Some of the languages included in the competition are considered low-resource languages because they have relatively small speaker populations or limited amounts of digital data available for training AI systems. “The challenge is not only the limited training data,” Lee said. “Political, cultural and social contexts also vary considerably from one country and event to another, making it much more difficult for AI to interpret the content accurately.”

Rather than optimizing its system for a single language or one specific subtask, the NYCU team set out to develop a broadly applicable method that could deliver reliable performance across languages and assignments.

 

Three AI Models Working as One

During development, the researchers compared 10 recently released open-weight language models. Based on their experiments, they selected Google’s Gemma 3, Mistral Small 3.2 from the French AI company Mistral AI and Microsoft’s Phi-4. The team then combined the three models using a technique known as model stacking. The approach integrates predictions from multiple models, much like bringing together specialists with different strengths to solve the same problem.

By allowing the models to complement one another, the system achieved greater consistency across languages and different types of polarization analysis. It delivered particularly strong results in Chinese, Spanish and German, helping NYCU secure second place overall.

The system recorded 38 top-three finishes across the competition’s language-specific evaluations — the highest number achieved by any participating team. It placed first in 15 evaluations, second in 12 and third in 10.

From Competition Results to Real-world Applications

Lee said online polarization can leave users with one-sided information, deepen divisions between social groups and shrink the space available for reasoned public debate. “Online polarization has become an important new area of international natural language processing research,” he said. “The technology could eventually support cross-language analysis of online discourse, social risk monitoring and research into public issues.”

Such systems could help researchers and public institutions better understand how online divisions emerge across different cultures, languages and major events — and provide a clearer view of the forces shaping public debate in an increasingly connected world.

A presentation slide lists NYCU-NLP as the competition’s second-best-performing system, behind the University of Tokyo’s Tsuruoka Laboratory.