This repository contains the full experimental pipeline, models, and analysis for SemEval 2026 Task 9 – Subtask 1: Multilingual Polarization Detection, a binary classification task across 22 languages.
Our study evaluates traditional multilingual transformer baselines, generative encoder–decoder models, parameter-efficient LoRA variants, and both zero-shot and few-shot LLM prompting.
Authors: Kamal Poshala, Kushi Reddy Kankar, and Rohan Mukka.
Polarization involves ideological hostility, adversarial rhetoric, and divisive language. The task requires building models that generalize across heterogeneous linguistic groups, including:
- High-resource languages
- Medium-resource languages
- Low-resource languages
- Noisy / morphologically complex languages
Our project analyzes these challenges and proposes a unified multilingual modeling pipeline.
Our study evaluates four major modeling families:
- Encoder-only model
- WordPiece tokenizer
- Performance: Strong on high-resource languages, weaker on noisy text.
- Reformulates classification as text generation
- Performance: Better on structured languages; unstable on noisy or morphologically complex languages.
- Variants: Multilingual LoRA, Per-language LoRA, MixLoRA (expert routing)
- Performance: Most efficient & best overall generalization. Top performer across resource tiers.
- Performance: Few-shot produces the highest accuracy for English. Useful for languages with strong LLM pretraining signals.
Overall performance across languages (Macro F1):
| Difficulty Tier | mBERT | mT5 | Few-shot LLM | Multilingual LoRA |
|---|---|---|---|---|
| High-resource | 0.70–0.75 | 0.75–0.85 | 0.80 (EN) | 0.82–0.87 |
| Medium-resource | 0.60–0.70 | 0.65–0.78 | 0.70–0.76 | 0.78–0.84 |
| Low-resource | 0.55–0.65 | 0.55–0.70 | 0.60–0.72 | 0.74–0.82 |
| Noisy/Complex | 0.50–0.60 | 0.50–0.68 | 0.65–0.70 | 0.72–0.80 |
🏆 Best Overall Method: Multilingual LoRA
- Most stable
- Best cross-lingual transfer
- Strongest performance in LR/NC languages
- Unicode normalization (NFKC)
- URLs & mentions masked
- Emojis, punctuation, emphasis markers preserved
- mBERT: WordPiece tokenizer
- mT5: SentencePiece tokenizer
- mBERT: lr
2e-5–5e-5, batch 16–32, epochs 3–5 - mT5: lr
1e-4(Adafactor), batch 8–16 - LoRA: rank 8, α = 16, lr
3e-4
- Macro F1, due to class imbalance and multilingual fairness.
- LoRA adapters preserve multilingual knowledge while specializing efficiently.
- LLM few-shot prompting is strongest when English dominates pretraining.
- mT5 behaves well on structured languages but is unstable on noisy text.
- mBERT saturates early, unable to capture complex ideological cues.
Polarization detection intersects with sociopolitical sensitivity. Important risks include:
- Misclassification bias across cultures
- Possible misuse for censorship
- Dialect and orthographic bias
- Privacy concerns in social media data
If you use this work, please cite the project authors:
Kamal Poshala, Kushi Reddy Kankar, Rohan Mukka
SemEval 2026 – Multilingual Polarization Detection (Task 9)