RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction

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

Summary

The RPI Team developed a coarse-to-fine Valence-Arousal (VA) ensemble system for subtask 1 of SemEval-2026 Task 3 (DimABSA), focusing on aspect-level VA prediction. Their approach integrates a pair of LoRA-tuned Qwen 3 8B LLMs, which predict coarse VA bins between 1 and 8, generating ordinal VA guidance signals and distributional features. These LLM-derived features then inform the training of an instruction-style, multilingual E5 encoder model equipped with a multitask head, designed to produce continuous VA predictions. During inference, the trained LLMs generate guidance signals for the test set, which are subsequently fed into the trained encoder. This architecture utilizes the LLM as a high-level prior for initial estimation, while the encoder refines predictions for precise calibration across six languages and five domains. The system achieved an RMSEVA of 1.20 and demonstrated superior performance over separate valence and arousal models, particularly in arousal correlations.

Key takeaway

For Machine Learning Engineers developing advanced sentiment analysis systems, this ensemble approach offers a robust strategy. If your goal is precise aspect-level Valence-Arousal prediction across multiple languages and domains, consider integrating coarse LLM guidance with a calibrated encoder. This method, demonstrated by an RMSEVA of 1.20, suggests you can achieve superior performance, especially for arousal correlations, by utilizing LLMs for high-level priors and encoders for fine-tuned, continuous outputs.

Key insights

An LLM-encoder ensemble uses coarse LLM guidance to calibrate a multilingual encoder for precise continuous Valence-Arousal prediction.

Principles

• Combining LLMs and encoders improves VA prediction.
• Coarse LLM priors enhance fine-grained encoder calibration.
• Joint VA modeling outperforms separate valence/arousal.

Method

Train LoRA-tuned Qwen 3 8B LLMs for coarse VA bins (1-8). Use LLM outputs as guidance for training a multilingual E5 encoder with a multitask head for continuous VA prediction.

In practice

• Apply LLM-encoder ensembles for complex sentiment tasks.
• Use LLMs for initial coarse predictions.
• Integrate multitask heads for related continuous outputs.

Topics

• Valence-Arousal Prediction
• Sentiment Analysis
• LLM-Encoder Ensemble
• Qwen 3 8B
• E5 Encoder
• SemEval-2026 Task 3
• LoRA Tuning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• 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.
• What Makes Words Hard? Sakura at BEA 2026 Shared Task on Vocabulary Difficulty Prediction — Soft-target loss fine-tuning significantly improves LLM and MLM performance in continuous value prediction tasks.
• Multi-View Attention Multiple-Instance Learning Enhanced by LLM Reasoning for Cognitive Distortion Detection — Combining LLMs with MIL and ELB components improves cognitive distortion detection by enhancing interpretability and reasoning.
• Beyond Acoustic Emotion Recognition: Multimodal Pathos Analysis in Political Speech Using LLM-Based and Acoustic Emotion Models — LLM-based multimodal analysis captures political emotion better than acoustic models, despite acoustic utility for Arousal.
• Do We Still Need Humans in the Loop? Comparing Human and LLM Annotation in Active Learning for Hostility Detection — LLM labels can match human-level F1 scores for hostility detection but exhibit distinct error profiles.
• Deep Temporal Modeling and Ensemble Fusion for Multimodal Emotion Recognition from Physiological Signals — Combining deep temporal models with sensor and ensemble fusion significantly boosts physiological emotion recognition accuracy.

Open in AIssential →

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