DUTH at SemEval-2026 Task 3: Multilingual Transformer Models for Dimensional Stance Prediction Across Tracks

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

Summary

The DUTH system was developed for SemEval-2026 Task 3, specifically the Dimensional Aspect-Based Sentiment Regression (DimASR) subtask. DimASR involves predicting continuous Valence and Arousal (VA) scores for aspect terms in opinionated text and stance targets in public-issue discourse. DUTH employs a multilingual Transformer encoder, fine-tuned end-to-end, to jointly encode the input text and its corresponding aspect or stance target. A regression head then performs VA prediction. Evaluated on official multilingual and multidomain datasets, DUTH achieved competitive performance. It demonstrated improvements over the strongest official baseline in Track A and the mBERT baseline in Track B, consistently yielding stronger predictions for Valence compared to Arousal.

Key takeaway

For NLP Engineers developing multilingual sentiment or stance prediction systems, DUTH's approach demonstrates that fine-tuned multilingual Transformer encoders can effectively predict continuous Valence and Arousal scores. You should consider implementing joint encoding strategies for text and target to improve contextual understanding. Prioritize refining Valence prediction, as it consistently yields stronger results than Arousal across diverse datasets. This method offers competitive performance over established baselines.

Key insights

The DUTH system uses multilingual Transformers to jointly predict continuous Valence and Arousal scores for aspect-based sentiment and stance.

Principles

• Multilingual Transformers enhance cross-lingual sentiment.
• Joint text and target encoding improves VA prediction.
• Valence scores are more consistently predicted than Arousal.

Method

Fine-tune a multilingual Transformer encoder end-to-end to jointly encode input text and its aspect/stance target, then use a regression head for Valence and Arousal prediction.

In practice

• Apply multilingual Transformers for DimASR tasks.
• Design models for joint text-target encoding.
• Expect higher accuracy for Valence over Arousal.

Topics

• Multilingual Transformers
• Dimensional Sentiment Analysis
• Stance Prediction
• SemEval-2026
• Valence Arousal Regression
• Aspect-Based Sentiment

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• From Affect Prediction to Affect Forecasting: Evidence for Distinct Information Sources in Longitudinal Text — Current affect prediction relies on text semantics, while future affective change forecasting benefits from numeric trajectory dynamics.
• Automatic Evaluation of Enem Essays: A Comparative Analysis between Feature Engineering and Transformers — Regression formulations are generally more suitable for ordinal essay scoring than multiclass classification.
• Turkish Sentiment Analysis (Keras LSTM + BERT) — A hybrid NLP pipeline combines custom LSTMs and pre-trained BERT for robust Turkish sentiment analysis with real-time data acquisition.
• Exploration of symbolic methods for emotion detection for Portuguese — Symbolic emotion detection in Portuguese shows stable relative performance despite domain-specific absolute variations.
• 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.

Open in AIssential →

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