“AGI” Team at SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays

Harsh Rathwa

Abstract:

This paper describes our submission to
SemEval-2026 Task 2: Predicting Variation in
Emotional Valence and Arousal. We combine
RoBERTa-Large text encoding with a uni-
directional GRU for temporal modeling and
gated user embeddings for personalization. A
four-phase staged training curriculum employs
ordinal regression for absolute affect predic-
tion and a zero-inflated delta model for change
detection. Our approach achieves competitive
performance on Subtask 1 (longitudinal affect
assessment) with composite correlation r =
0.600 for valence and r = 0.452 for arousal.
However, we observe systematic degradation
in Subtask 2A (state change detection) with
negative correlations (r=−0.167 for valence,
r =−0.147 for arousal), revealing a fun-
damental trade-off between stability-oriented
representations and change sensitivity. We
provide detailed empirical analysis of these
failure modes, contributing insights into the
challenges of modeling emotional dynamics in
ecological data. Code and trained checkpoints
are publicly available.