💬 NLP

YEZE at SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization via Heterogeneous Ensembling

The paper presents YEZE, a system for SemEval-2026 Task 9 that detects online polarization across 22 languages by employing a heterogeneous ensemble of XLM-RoBERTa-large and mDeBERTa-v3-base models, where independent task modeling combined with class weighting proved most effective for handling severe label imbalance.

Original authors: Fengze Guo (University of Tübingen), Yue Chang (University of Tübingen)

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fengze Guo (University of Tübingen), Yue Chang (University of Tübingen)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet as a massive, global town square where people from 22 different countries are shouting, whispering, and arguing all at once. Sometimes, these arguments get heated and turn into "polarization"—where people stop listening to each other and only listen to their own side.

The paper you shared describes a team of researchers (YEZE) who built a special set of "digital ears" to listen to this town square and figure out three things:

Is this argument heated? (Binary detection)
Who or what are they fighting about? (Target classification: politics, race, religion, etc.)
How are they fighting? (Manifestation: Are they using insults, stereotypes, or dehumanizing language?)

Here is how they did it, explained simply:

1. The Problem: A Noisy, Unbalanced Crowd

The researchers faced two big hurdles:

The Language Barrier: They had to understand 22 different languages, from English and Spanish to Amharic and Burmese. It's like trying to understand a conversation where everyone is speaking a different dialect.
The "Rare Event" Problem: In the data, most posts were calm. Only a few were actually polarized. It's like trying to find a single red marble in a bucket full of white marbles. If you just guess "white" every time, you'd be right most of the time, but you'd miss the red ones. This is called "label imbalance."

2. The Solution: A "Two-Person Detective Team"

Instead of building one giant brain to do everything, the team built a heterogeneous ensemble. Think of this as hiring two different types of detectives to solve the same case:

Detective A (XLM-RoBERTa): A veteran who has read almost everything in the world. They are great at understanding the general vibe and context of a sentence in any language.
Detective B (mDeBERTa): A specialist who pays very close attention to the specific arrangement of words and how they relate to each other.

How they worked together:
They didn't just let them vote; they gave them a weighted score. If Detective A is 70% sure and Detective B is 30% sure, the final answer leans heavily on Detective A. This combination made the system more robust, like having a safety net.

3. The Strategy: Treating Each Clue Separately

The team realized that trying to solve all three mysteries (Is it polarized? Who is it about? How is it said?) at the same time was like trying to juggle while riding a unicycle. The clues for "who" and "how" were so rare that they got lost in the noise of the "is it polarized" clue.

So, they decided to split the job:

They trained separate models for each specific question.
For the rare clues (like specific types of hate speech), they used a special trick called Weighted Cross-Entropy. Imagine a teacher grading a test. If a student gets a rare, hard question right, the teacher gives them extra points. If they get a common question right, they get normal points. This forced the AI to pay extra attention to the rare, important cases it usually ignored.

4. What They Found

Separation works: Trying to do everything at once (Multi-Task Learning) actually made the system worse because the rare clues confused the common ones. Keeping the tasks separate was the winning move.
The "Translation" Trap: They tried translating posts into other languages to create more practice data (Data Augmentation), but it didn't help much. It's like translating a joke; the words might be right, but the cultural humor and nuance get lost, making the AI confused.
The Results: Their system was very good at spotting the "red marbles" (polarized content) across all 22 languages. In many languages, they finished in the top 10 of all the teams competing. However, they still struggled a bit with the most specific, rare types of insults (the "manifestations"), showing that even the best AI finds it hard to spot the subtlest forms of hate when there are very few examples to learn from.

In a Nutshell

The paper is about building a smart, multi-lingual security system for social media. Instead of using one giant, confused brain, they used a team of two specialized AI models working together. They taught these models to care more about the rare, dangerous posts than the boring ones. While they didn't solve every problem (especially the very rare types of hate speech), they proved that splitting the tasks and combining different types of AI brains is the best way to handle a messy, multilingual world.

Technical Summary: YEZE at SemEval-2026 Task 9

Problem Definition
This paper presents the "YEZE" system developed for SemEval-2026 Task 9, which focuses on detecting multilingual, multicultural, and multievent online polarization. The task involves analyzing social media posts across 22 languages (including Amharic, Arabic, Chinese, English, Hindi, and others) to identify polarized content. The challenge is structured into three subtasks:

Binary Detection: Determining whether a post is polarized (Subtask 1).
Target Classification: Identifying the specific targets of polarization (e.g., political, racial/ethnic, religious) via multi-label classification (Subtask 2).
Manifestation Identification: Identifying the specific linguistic manifestations of polarization (e.g., stereotyping, vilification, dehumanization) via multi-label classification (Subtask 3).

The system faces significant challenges, including severe label imbalance, cross-lingual variation in label distributions (prior shift), and extreme sparsity in fine-grained labels for Subtasks 2 and 3.

Methodology
The authors propose a system based on heterogeneous ensembling of multilingual pretrained models, treating each subtask as an independent problem rather than a joint multi-task learning (MTL) objective.

Model Architecture: The system utilizes two complementary multilingual encoders:
- XLM-RoBERTa-large: Selected for its broad cross-lingual coverage and stable transfer performance.
- mDeBERTa-v3-base: Chosen for its use of disentangled attention to encode content and relative position separately.
Training Strategy:
- Independent Modeling: Unlike traditional MTL approaches that share an encoder across tasks, the authors train separate models for each subtask. Experiments indicated that shared MTL training introduced negative transfer, where sparse fine-grained labels competed with the dominant binary objective, destabilizing fine-grained predictions.
- Imbalance-Aware Optimization: To address severe class imbalance, particularly in multi-label subtasks, the system employs Binary Relevance (BR), decomposing multi-label problems into independent binary tasks. It utilizes Weighted Binary Cross-Entropy (WBCE) loss, where class weights are inversely proportional to label frequency. This approach proved more effective than Focal Loss for low-resource settings.
- Data Augmentation: The team investigated translation-based data augmentation (using machine translation to generate synthetic samples for languages like Hausa) but found it did not yield significant performance gains, likely due to pragmatic shifts and the obscuring of region-specific idioms.
Ensembling: The final submission combines the predictions of XLM-R and mDeBERTa using weighted probability averaging. The ensemble weight ( $\alpha$ ) was optimized on the development set to maximize Macro-F1, resulting in a weight of 0.7 for XLM-R and 0.3 for mDeBERTa. A global threshold of $\tau = 0.5$ was applied for all predictions.

Key Results
In the official evaluation, the system achieved competitive results, ranking in the Top 10 for the majority of languages across all subtasks:

Subtask 1: Top 10 in 11/22 languages.
Subtask 2: Top 10 in 16/22 languages (including Top 5 in Amharic, Urdu, Odia, and Polish).
Subtask 3: Top 10 in 17/18 languages (including Top 5 in Amharic, Arabic, English, Khmer, Spanish, and Urdu).

Ablation studies revealed that:

WBCE significantly outperformed standard BCE and Focal Loss in low-resource scenarios (e.g., a +0.212 Macro-F1 increase for Telugu).
Independent modeling was superior to MTL, as shared representations often led to negative transfer for sparse labels.
Translation augmentation failed to improve performance, suggesting that synthetic data may not capture the nuanced socio-political context required for polarization detection.

Significance and Claims
The paper claims that the primary bottlenecks in multilingual polarization detection are not merely cross-lingual variation, but specifically fine-grained label sparsity, calibration instability, and cross-task inconsistency.

The authors argue that their findings demonstrate that independent per-subtask modeling combined with imbalance-aware optimization is a more robust practical alternative to shared multi-task learning, particularly when dealing with severe label sparsity. They highlight that while high-coverage labels (e.g., "Political" targets, "Vilification") are learnable, sparse manifestation labels remain fragile and prone to miscalibration (high recall but low precision).

The work underscores the difficulty of fine-grained polarization analysis in low-resource and imbalanced settings, suggesting that future improvements may require lightweight hierarchical calibration or culturally grounded data synthesis rather than simply scaling model size or relying on synthetic translation data.

1. The Problem: A Noisy, Unbalanced Crowd

2. The Solution: A "Two-Person Detective Team"

3. The Strategy: Treating Each Clue Separately

4. What They Found

In a Nutshell

More like this