NLP-FSDM at SemEval-2026 Task 2: Temporal Smoothing and CCC-MAE Optimization for Balanced Longitudinal Affect Assessment

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

The NLP-FSDM system, developed for SemEval-2026 Task 2, Subtask 1, addresses longitudinal affect assessment by predicting Valence and Arousal (V & A) scores from sequences of ecological essays and feeling words. This system utilizes ModernBERT-large as its text encoder within a joint regression framework. Optimization is achieved through a Concordance Correlation Coefficient (CCC) loss, augmented by a lightly weighted Mean Absolute Error (MAE) term. To mitigate variance from fine-tuning large transformers on smaller user-specific datasets, a three-seed ensemble is employed. Additionally, a lightweight post-inference temporal smoothing mechanism is applied per user to enhance within-user consistency. The system achieved an rcomposite score of 0.546 for Valence and 0.453 for Arousal, demonstrating stable cross-dimensional performance without explicit sequential dependency modeling.

Key takeaway

For NLP Engineers developing longitudinal affect assessment systems, consider the NLP-FSDM approach to enhance prediction stability. You should integrate ModernBERT-large with a CCC-MAE joint regression loss and employ a three-seed ensemble to reduce fine-tuning variance. Additionally, applying a lightweight, per-user temporal smoothing mechanism post-inference can significantly improve within-user consistency for your time-series predictions, even without explicit sequential modeling.

Key insights

The NLP-FSDM system combines ModernBERT-large with CCC-MAE optimization, ensembling, and temporal smoothing for robust longitudinal affect assessment.

Principles

• CCC loss combined with MAE optimizes joint regression.
• Ensembling reduces variance in fine-tuning large transformers.
• Post-inference temporal smoothing improves within-user consistency.

Method

The NLP-FSDM method involves encoding text with ModernBERT-large, optimizing a joint regression with CCC-MAE loss, using a three-seed ensemble, and applying per-user temporal smoothing post-inference.

In practice

• Apply CCC-MAE for joint regression tasks.
• Use multi-seed ensembling to stabilize transformer fine-tuning.
• Implement per-user temporal smoothing for time-series predictions.

Topics

• Longitudinal Affect Assessment
• SemEval-2026 Task 2
• Valence Arousal Prediction
• ModernBERT-large
• Temporal Smoothing

Best for: AI Scientist, NLP Engineer, Research Scientist

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.
• Use of Machine Learning techniques and Large Language Models for automatic evaluation of Celpe-Bras exam texts — BERT adaptations achieved the best automatic evaluation for Celpe-Bras texts, but at a higher computational cost.
• Team MKC at CLPsych 2026: Capturing and Characterizing Mental Health Changes through Social Media Timeline Dynamics — An LLM-based pipeline analyzes social media timelines for early mental health change detection and continuous monitoring, integrating post and user-level assessments.
• How to Fine-Tune an SLM for Emotion Recognition — Fine-tuning SLMs for fine-grained emotion recognition on imbalanced datasets requires a multi-technique approach to data balancing and loss weighting.
• 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.
• Learning Robust Pair Confidence for Multimodal Emotion-Cause Pair Extraction — Robustly shaping pair confidence, by ensuring discriminative and stable predictions, is a highly effective training strategy for multimodal emotion-cause pair extraction.

Open in AIssential →

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