PhucNguyen@DravidianLangTech 2026: Political Multiclass Sentiment Analysis with XLM-RoBERTa and Low-Rank Adaptation

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Data Science & Analytics · Depth: Advanced, medium

Summary

A system developed by Dinh Khac Phuc Nguyen and Thìn Đặng Văn for the DravidianLangTech@ACL 2026 Political Multiclass Sentiment Analysis shared task achieved a macro-average F1-score of 0.3763, securing Rank 2 on the leaderboard. This approach categorizes code-mixed Tamil-English tweets into seven sentiment classes, addressing challenges like informal jargon, severe class imbalance, and distribution shifts. The system integrates XLM-RoBERTa with Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. To counter majority-class dominance, it employs a combination of random oversampling and automated hyperparameter optimization. Additionally, targeted preprocessing, including emoji demojization and noise removal, was crucial for preserving nuanced symbolic cues in the text.

Key takeaway

For NLP Engineers developing sentiment analysis models for challenging code-mixed languages like Tamil-English, you should consider integrating XLM-RoBERTa with Low-Rank Adaptation (LoRA). This approach, combined with random oversampling and automated hyperparameter optimization, effectively mitigates class imbalance and distribution shifts. Ensure your preprocessing includes emoji demojization and noise removal to preserve critical symbolic cues, potentially improving your model's macro-average F1-score and competitive ranking.

Key insights

Combining XLM-RoBERTa with LoRA, oversampling, and specific preprocessing effectively tackles political sentiment analysis in challenging code-mixed languages.

Principles

• Class imbalance requires macro-level balancing techniques.
• Parameter-Efficient Fine-Tuning (PEFT) is effective for complex tasks.
• Targeted preprocessing enhances nuanced symbolic cue preservation.

Method

The system fine-tunes XLM-RoBERTa using LoRA, applies random oversampling with automated hyperparameter optimization for class balance, and preprocesses text via emoji demojization and noise removal.

In practice

• Implement LoRA with XLM-RoBERTa for code-mixed sentiment.
• Use random oversampling to address severe class imbalance.
• Apply emoji demojization for nuanced symbolic cue preservation.

Topics

• Political Sentiment Analysis
• XLM-RoBERTa
• Low-Rank Adaptation
• Code-mixed Languages
• Class Imbalance
• Parameter-Efficient Fine-Tuning

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Can Machines Tell Whether Gamers Love or Hate a Game? — Fine-tuning Transformer models like RoBERTa significantly improves sentiment analysis accuracy on imbalanced, complex human language datasets.
• Quantization-Robust LLM Unlearning via Low-Rank Adaptation — Low-rank adaptation (LoRA) enables effective LLM unlearning that survives aggressive post-training quantization.
• A Lightweight LLM Framework for Disaster Humanitarian Information Classification — Parameter-efficient fine-tuning significantly outperforms prompting and RAG for disaster tweet classification, enabling efficient, accurate crisis intelligence.
• Low Kruskal-Rank Adaptation — LoKRA improves PEFT by using Kruskal rank to mitigate redundancy in LoRA update matrices, enhancing LLM adaptation.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific LoRA fine-tuning on Qwen2.5-14B with CCPoetry-49K significantly enhances classical Chinese poetry understanding and translation.
• Addressing Data Scarcity in Bangla Fake News Detection: An LLM-Based Dataset Augmentation Approach — LLM-based data augmentation significantly improves fake news detection in low-resource languages by generating diverse synthetic data.

Open in AIssential →

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