lamanhnguyen at SemEval-2026 Task 2: Uncovering Lexical Bias and Momentum Lag in Longitudinal Emotion Prediction using Multi-task DeBERTa

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

Summary

Lam Anh Nguyen's system for SemEval-2026 Task 2 addressed the prediction of emotional Valence and Arousal variation. The approach involved fine-tuning a weighted ensemble of DeBERTa-v3-base models. This system achieved the second-highest Valence composite correlation and ranked 5th in the overall V&A average for Subtask 1. Beyond performance, the work provides an empirical analysis of the model's behavior in longitudinal tasks, revealing significant inverse correlations. It quantifies the "Venting Effect," which describes the model's systematic over-indexing on negative lexical cues even when self-reported relief is present. The analysis also explored the structural trade-off between Mean Absolute Error and Pearson correlation resulting from smoothing techniques.

Key takeaway

For NLP Engineers developing longitudinal emotion prediction systems, you should critically evaluate model biases like the "Venting Effect," where negative lexical cues might be over-indexed despite actual emotional shifts. Your model's performance on Valence and Arousal in time-series data may show inverse correlations, necessitating careful analysis beyond standard metrics. Consider how smoothing techniques impact the Mean Absolute Error and Pearson correlation trade-off in your evaluation.

Key insights

Longitudinal emotion prediction models can exhibit lexical bias and inverse correlations, like the "Venting Effect."

Principles

• Emotion prediction models may over-index on negative lexical cues.
• Longitudinal emotion prediction can show significant inverse correlations.
• Smoothing techniques impact MAE and Pearson correlation trade-offs.

Method

Fine-tuning a weighted ensemble of DeBERTa-v3-base models for emotional Valence and Arousal prediction.

Topics

• SemEval-2026 Task 2
• Emotion Prediction
• DeBERTa-v3-base
• Lexical Bias
• Venting Effect
• Longitudinal Analysis

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning 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.
• The Autocorrelation Blind Spot: Why 42% of Turn-Level Findings in LLM Conversation Analysis May Be Spurious — Uncorrected autocorrelation in LLM conversation analysis leads to significantly inflated statistical significance.
• Do Emotions Influence Moral Judgment in Large Language Models? — LLMs' moral judgments are systematically biased by emotional context, unlike human judgments.
• Frozen Multimodal Embeddings for Personality and Cognitive Ability Assessment in Asynchronous Video Interviews — Frozen multimodal embeddings effectively assess personality and cognitive ability from asynchronous video interviews, outperforming baselines.
• Assessing Y-Axis Influence: Bias in Multimodal Language Models on Chart-to-Table Translation — Y-axis biases in chart datasets significantly impact Multimodal Language Model performance in chart-to-table translation.
• Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation — D³-Net combines SDR-based training with LTMLE targeting to robustly estimate longitudinal treatment effects, mitigating error propagation.

Open in AIssential →

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