AI4PC-Howard University at SemEval-2026 Task 2: Fine-Tuning DistilBERT, DeBERTa and ModernBERT for Valence–Arousal Prediction and Change Estimation
Year: 2026
Type: conference-paper
Abstract: We present a merged system description for two participating teams, AI4PC-Howard University and BisonAI4PC, on longitudinal valencearousal (VA) prediction in the SemEval-2026 Task 2 text corpus.Using only the official data, we enforce user-disjoint splits to prevent leakage and evaluate three settings: essay-level VA state estimation, short-horizon VA change forecasting, and long-horizon disposition change prediction.Our best systems use DistilBERT for essay-level regression, ModernBERT-based history modeling with a GRU and a blended previous-delta baseline for short-horizon change, and pooled De-BERTa history embeddings with a compact MLP for disposition change.Reporting the best-performing official submission for each subtask, we achieve r comp = 0.665/0.468(valence/arousal) for Subtask 1, r = 0.597/0.413for Subtask 2A, and r = 0.046/0.348for Subtask 2B.* Corresponding author.Our approach follows two principles: (1) strict user-disjoint evaluation with inference-time-safe feature construction, and (2) compact, transparent pipelines that are easy to reproduce.In this paper, we merge two participating teams (AI4PC-Howard University and BisonAI4PC), hereafter Team 1 and Team 2, and report both teams' official submissions and leaderboard results across Subtasks 1/2A/2B.Across the two official submissions, our best-performing systems use a DistilBERT regressor for essay-level VA, ModernBERT-based representations with a GRU and a blended previousdelta baseline for short-horizon change, and pooled DeBERTa history embeddings augmented with summary features for disposition change.Our contributions are as follows:• Leakage-controlled, user-disjoint evaluation for reproducible longitudinal VA modeling.• Lightweight systems for Subtasks 1/2A/2B with explicit, reproducible inference pipelines.• Official evaluation results from both teams, with analysis of how backbone choice and temporal aggregation relate to performance.
Authors: Araj Shah, Utsav Shah, Saurav Aryal
Institution: Howard University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166840843