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

TL;DR

This paper presents a multilingual sentiment regression system using fine-tuned XLM-RoBERTa, outperforming large language models in predicting continuous valence-arousal scores across multiple domains and languages.

Contribution

It introduces a fine-tuning approach with dual regression heads for multilingual DimABSA, demonstrating superior performance over LLM prompting methods.

Findings

Fine-tuning XLM-RoBERTa outperforms LLM prompting in DimABSA.

Separate models for each language-domain pair improve accuracy.

Merged training sets enhance final test predictions.

Abstract

Dimensional Aspect-Based Sentiment Analysis (DimABSA) extends traditional ABSA from categorical polarity labels to continuous valence-arousal (VA) regression. This paper describes a system developed for Track A, Subtask 1 (Dimensional Aspect Sentiment Regression), aiming to predict real-valued VA scores in the [1, 9] range for each given aspect in a text. A fine-tuning approach based on XLM-RoBERTa-base is adopted, with dual regression heads with sigmoid-scaled outputs for valence and arousal prediction. Separate models are trained for each language-domain pair (English and Chinese across restaurant, laptop, and finance domains), and training and development sets are merged for final test predictions. In development experiments, the fine-tuning approach is compared against several large language models under a few-shot prompting setting, demonstrating that task-specific fine-tuning outperforms these LLM-based methods across all evaluation datasets.