NCL-BU at SemEval-2026 Task 3: Fine-tuning XLM-RoBERTa for Multilingual Dimensional Sentiment Regression

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

Summary

NCL-BU's system for SemEval-2026 Task 3, Subtask 1, addresses Dimensional Aspect-Based Sentiment Analysis (DimABSA) by predicting continuous valence–arousal (VA) scores within the [1, 9] range for given text aspects. The approach involves fine-tuning XLM-RoBERTa-base, utilizing dual regression heads with sigmoid-scaled outputs for VA prediction. Separate models are trained for each language–domain pair, specifically English and Chinese across restaurant, laptop, and finance sectors. Development experiments demonstrated that this task-specific fine-tuning method consistently outperforms several large language models operating under a few-shot prompting setting across all evaluated datasets. This highlights the efficacy of tailored model adaptation for specific sentiment regression tasks.

Key takeaway

For NLP engineers developing multilingual sentiment analysis systems, this research suggests prioritizing task-specific fine-tuning over few-shot large language model approaches. Your efforts in adapting models like XLM-RoBERTa with dedicated regression heads for valence-arousal prediction will likely yield superior performance. Consider training distinct models for each language and domain to maximize accuracy in continuous sentiment regression tasks.

Key insights

Task-specific fine-tuning of XLM-RoBERTa-base significantly outperforms few-shot LLM prompting for multilingual dimensional sentiment regression.

Principles

• Fine-tuning excels over few-shot LLMs for specific tasks.
• Separate models improve performance across language-domains.

Method

Fine-tuning XLM-RoBERTa-base with dual sigmoid-scaled regression heads to predict valence and arousal scores. Separate models are trained for each language-domain pair.

In practice

• Apply XLM-RoBERTa for continuous VA sentiment prediction.
• Consider domain-specific models for multilingual tasks.

Topics

• Dimensional Sentiment Analysis
• XLM-RoBERTa
• Fine-tuning
• Multilingual NLP
• Valence-Arousal Regression
• SemEval-2026

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.
• Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning — Multilingual code-switching instruction tuning improves LLM performance beyond bilingual settings.
• Beyond Sentiment Classification: A Generative Framework for Emotion Intensity Evaluation in Text — Generative LLMs fine-tuned on continuous emotion intensity scores outperform classification and generalize to unseen affective dimensions.
• Building an LLM fine-tuning Dataset — Building multi-speaker Reddit datasets for LLM fine-tuning is complex but achievable, with efficient QLoRA methods requiring minimal samples.
• Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models — LLM-derived sentiment from policy minutes significantly enhances inflation forecasting accuracy when aligned with human assessments.
• A Scalable Tool for Measuring Manner and Result Verbs in Developmental Language Research — A RoBERTa classifier, trained on LLM-generated annotations, accurately distinguishes manner and result verbs at scale.

Open in AIssential →

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