CuriosAI at SemEval-2026 Task 2: Predicting Emotion using RoBERTa-large model

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

Summary

CuriosAI presents a novel method for predicting continuous emotion dimensions, Valence and Arousal, from textual data. The approach integrates affective intermediate training with multi-task learning, utilizing a RoBERTa-large model as its backbone. This two-phase training process first involves pre-training on external emotion datasets, followed by fine-tuning with task-specific data. The model incorporates independent regression heads for each subtask. Experimental results from SemEval-2026 Task 2 Subtask 1 show Pearson correlation coefficients of 0.68 for Valence and 0.45 for Arousal. This method demonstrates stable performance, notably in its ability to capture subtle inter-user differences in emotional expression.

Key takeaway

For NLP engineers developing emotion detection systems, this research suggests a robust two-phase training strategy. Implementing affective intermediate pre-training before multi-task learning with a RoBERTa-large backbone can significantly improve Valence and Arousal prediction accuracy. Consider adopting independent regression heads for each emotion dimension to better capture nuanced inter-user differences in expression.

Key insights

A two-phase training method using RoBERTa-large improves continuous emotion prediction from text.

Principles

• Combine pre-training with multi-task learning.
• Use independent heads for distinct subtasks.
• Affective intermediate training enhances emotion models.

Method

The approach involves intermediate pre-training on external emotion datasets, followed by multi-task learning on task-specific data, using RoBERTa-large with independent regression heads.

In practice

• Apply two-phase training for emotion tasks.
• Consider RoBERTa-large for text emotion.
• Design separate regression heads.

Topics

• Emotion Prediction
• RoBERTa-large
• Multi-task Learning
• Affective Computing
• Valence-Arousal Model
• Natural Language Processing

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• 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.
• Exploration of symbolic methods for emotion detection for Portuguese — Symbolic emotion detection in Portuguese shows stable relative performance despite domain-specific absolute variations.
• 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.
• Two-Stage Multimodal Framework for Emotion Mimicry Intensity Prediction — Staged multimodal fusion with independent encoder pre-training improves robustness for continuous emotion intensity prediction.
• A Chinese Education Broadcast Emotion Corpus — A new Chinese education broadcast emotion corpus aids emotion recognition research.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.

Open in AIssential →

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