UAlberta at SemEval-2026 Task 2: Temporal Fusion Models for Predicting Affect Over Time

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

Summary

UAlberta's systems for SemEval 2026 Task 2 focused on predicting variation in emotional valence and arousal from ecological essays. Their approach combined language model embeddings with a Recurrent Neural Network (RNN) for predicting affect in single instances and forecasting dispositional change. For predicting state changes between timesteps, the team employed time-series forecasting techniques. These integrated methods led to significant results, with their systems ranking first globally for forecasting dispositional change and third for forecasting state change over time. The code for their successful systems has been made publicly available, offering a resource for further research and development in temporal affect prediction.

Key takeaway

For NLP engineers developing systems to predict emotional affect over time, you should consider integrating temporal fusion models. UAlberta's approach, combining language model embeddings, RNNs, and time-series forecasting, demonstrated top performance in SemEval 2026 Task 2 for both dispositional and state change prediction. This suggests a robust architecture for complex temporal sentiment analysis tasks, offering a strong baseline for your own model development.

Key insights

Temporal fusion models combining LMs, RNNs, and time-series forecasting excel at predicting emotional affect over time.

Principles

• Combine diverse models for complex temporal prediction.
• Time-series forecasting is effective for state changes.

Method

Utilizes language model embeddings and RNNs for single-instance affect and dispositional change, complemented by time-series forecasting for state changes between timesteps.

In practice

• Apply temporal fusion models for time-series emotional analysis.
• Integrate RNNs with time-series methods for affect forecasting.

Topics

• SemEval 2026 Task 2
• Temporal Fusion Models
• Affect Prediction
• Recurrent Neural Networks
• Time-Series Forecasting
• Language Models

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

Related on AIssential

• Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction — Multi-dimensional LLM sentiment, beyond polarity, improves WTI crude oil futures return prediction.
• Choosing ER Time Series Models (part 2): How to Fairly Compare ARIMA and XGBoost? — Fair comparison of time series models requires careful feature engineering and consistent evaluation metrics.
• Time-TK: A Multi-Offset Temporal Interaction Framework Combining Transformer and Kolmogorov-Arnold Networks for Time Series Forecasting — Multi-offset temporal embedding combined with KANs and Transformers significantly improves time series forecasting accuracy.
• From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting — Importance-aware compression and PRM-guided retrieval improve news-enhanced time series forecasting by managing context and selection.
• 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.
• MP3: Multi-Period Pattern Pre-training forSpatio-Temporal Forecasting — MP3 resolves "temporal mirage" in spatio-temporal forecasting by learning multi-period patterns from long time series.

Open in AIssential →

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