Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal

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

Summary

The Draken system, developed for SemEval-2026 Task 2, Subtask 1, offers a lightweight and computationally efficient method for predicting longitudinal emotional valence and arousal from ecological essays. This approach utilizes frozen 768-dimensional contextual embeddings from BERT-base-uncased, obtained via mean-pooling sentence representations without fine-tuning the transformer. These embeddings are then input into a multi-output Ridge regression model to jointly predict normalized valence and arousal scores. The system prioritizes simplicity and reproducibility, deliberately avoiding complex temporal architectures, external lexicons, or user metadata. It achieved a strong valence prediction (r = 0.594) and moderate arousal prediction (r = 0.296). Evaluations revealed consistently higher between-user correlations and that valence is substantially easier to predict than arousal, suggesting this combination provides a competitive and interpretable baseline.

Key takeaway

For NLP Engineers developing affect prediction systems, consider adopting a lightweight approach using frozen transformer embeddings. Your team can achieve competitive valence prediction (r = 0.594) and moderate arousal prediction (r = 0.296) with a simple multi-output Ridge regression model. This avoids complex architectures, offering high reproducibility and computational efficiency. It serves as an excellent baseline for longitudinal affect tasks, especially when resources are constrained.

Key insights

Frozen BERT embeddings combined with Ridge regression offer a simple, efficient, and competitive baseline for longitudinal affect prediction.

Principles

• Simplicity and reproducibility enhance model utility.
• Valence is easier to predict than arousal.
• Between-user affect correlations are higher.

Method

Obtain mean-pooled 768-dimensional BERT-base-uncased embeddings from essays. Feed these frozen embeddings into a multi-output Ridge regression model to jointly predict normalized emotional valence and arousal scores.

In practice

• Use frozen BERT embeddings for efficiency.
• Apply Ridge regression for joint prediction.
• Prioritize valence prediction in affect tasks.

Topics

• SemEval-2026 Task 2
• BERT Embeddings
• Ridge Regression
• Emotional Valence
• Arousal Prediction
• Affective Computing
• Longitudinal Prediction

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP 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.
• 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.
• Fine-Tuning BERT for Sentiment Analysis: A Beginner-Friendly Guide — Fine-tuning pre-trained transformer models like BERT enables high-accuracy NLP tasks with minimal training data.
• EEGDancer: Dynamic Emotion Latent Space Masked Modeling with Reinforcement Learning for EEG Continuous Emotion Prediction — Continuous EEG emotion prediction benefits from a unified framework combining discrete latent space learning, masked temporal modeling, and reinforcement learning.
• 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.
• Beyond Sentiment Classification: A Generative Framework for Emotion Intensity Evaluation in Text — Generative LLMs fine-tuned on continuous emotion intensity scores outperform classification and generalize to unseen affective dimensions.

Open in AIssential →

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