CITD@UIT at SemEval-2026 Task 2: Temporal Mixture-of-Experts for Longitudinal Valence and Arousal Prediction from Ecological Essays

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

Summary

CITD@UIT participated in SemEval-2026 Task 2, which addresses the longitudinal assessment and forecasting of emotional states from textual data. The challenge involved two main objectives: Subtask 1, focused on estimating continuous Valence and Arousal (V&A) scores across text sequences, and Subtask 2, dedicated to forecasting future emotional shifts, specifically State Change (2A) and Dispositional Change (2B). Their proposed unified framework utilizes cardiffnlp/twitter-roberta-base-sentiment-latest, a transformer architecture pretrained on 124 million tweets. For all subtasks, data was chronologically sorted by userid, employing a sliding window to capture longitudinal context. The system combined the pretrained RoBERTa model with Multilayer Perceptron (MLP) and Mixture-of-Experts (MoE) architectures, further enhancing feature extraction through both attention and mean pooling. This approach achieved competitive results, ranking 9th in Subtask 1 and 5th in Subtask 2A among participating teams.

Key takeaway

For NLP Engineers developing systems for longitudinal emotion prediction, consider integrating a temporal Mixture-of-Experts approach. Your models should process text chronologically with sliding windows to capture evolving emotional states. Experiment with pretrained RoBERTa models combined with MLP and MoE architectures, as this strategy demonstrated competitive performance in SemEval-2026 Task 2 for Valence and Arousal prediction.

Key insights

A temporal Mixture-of-Experts RoBERTa framework effectively predicts longitudinal emotional states from text.

Principles

• Longitudinal context improves emotion prediction.
• Combining RoBERTa with MoE enhances performance.
• Chronological sorting and sliding windows are key.

Method

A unified framework based on cardiffnlp/twitter-roberta-base-sentiment-latest combines RoBERTa with MLP and MoE. It sorts data chronologically by userid, uses a sliding window, and applies attention/mean pooling for feature extraction.

In practice

• Pretrain RoBERTa on large social media datasets.
• Implement sliding windows for temporal context.
• Experiment with MoE and MLP for fine-tuning.

Topics

• Emotion Prediction
• Valence and Arousal
• Longitudinal Analysis
• Mixture-of-Experts
• RoBERTa
• SemEval-2026

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• 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.
• Token-Level Pun Location Using Multi-Layer BERT with Mixture of Experts — A two-stage BERT-MoE model effectively locates puns in Portuguese by specializing expert components.
• 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.
• 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.
• A Mixture-of-Experts Model for Multimodal Emotion Recognition in Conversations — MiSTER-E uses a Mixture-of-Experts framework to fuse speech and text for robust emotion recognition in conversations.
• 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.

Open in AIssential →

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