hllwan at SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis via LLM Feature Fusion and Test-Time Adaptation

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

Summary

The hllwan team developed a system for SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA), which predicts continuous Valence and Arousal (VA) scores across diverse languages and domains. This task presents significant challenges due to data scarcity and cross-domain distribution shifts. Their robust framework incorporates a translation-based data augmentation strategy with precise HTML-tag alignment to mitigate low-resource constraints. It also features an unsupervised opinion extraction module utilizing syntactic dependency parsing to capture sentiment-bearing words. A Tripartite Feature Fusion architecture integrates both encoder-only (DeBERTa-v3) and causal LLM (Qwen2.5) models for dynamic aggregation of global and localized aspect-opinion embeddings. Finally, an unsupervised Test-Time Adaptation (TTA) mechanism calibrates normalization layers on the fly. The system achieved competitive performance and offered insights into LLM limitations in cross-lingual sentiment transfer.

Key takeaway

For NLP Engineers developing cross-lingual sentiment analysis systems, consider integrating a multi-faceted approach to overcome data scarcity and domain shifts. You should explore translation-based data augmentation with HTML-tag alignment and combine diverse LLM architectures like DeBERTa-v3 and Qwen2.5 for feature fusion. Implementing unsupervised Test-Time Adaptation can dynamically calibrate your models, enhancing performance in challenging cross-domain scenarios. This strategy offers a robust path for predicting continuous Valence and Arousal scores.

Key insights

A robust DimABSA framework combines data augmentation, opinion extraction, LLM feature fusion, and test-time adaptation to handle cross-lingual, cross-domain sentiment prediction.

Principles

• Address data scarcity with translation-based augmentation.
• Fuse encoder-only and causal LLM features.
• Dynamically calibrate models via Test-Time Adaptation.

Method

The framework uses translation-based data augmentation, unsupervised opinion extraction via syntactic dependency parsing, a Tripartite Feature Fusion with DeBERTa-v3 and Qwen2.5, and unsupervised Test-Time Adaptation for normalization layer calibration.

In practice

• Apply HTML-tag aligned translation for low-resource data.
• Integrate DeBERTa-v3 and Qwen2.5 for feature fusion.
• Implement TTA for dynamic model calibration.

Topics

• Dimensional Sentiment Analysis
• LLM Feature Fusion
• Test-Time Adaptation
• Data Augmentation
• Cross-lingual NLP
• SemEval-2026 Task 3

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

Related on AIssential

• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• Label-Aware Pseudo-Training Sample Generation for Text Classification — A new data augmentation method uses LLMs to subtly alter sentences, improving NLP performance with limited data.
• Enhancing Brazilian Inflation Forecasts through Sentiment Analysis Using Large Language Models — LLM-derived sentiment from policy minutes significantly enhances inflation forecasting accuracy when aligned with human assessments.
• Emotion-Aware Image Generation from Korean Diary Text via LLM-based Prompt Translation and LoRA Fine-Tuning — A pipeline integrates LLM-based sentiment recognition with LoRA-tuned T2I models for emotion-aware image generation.
• A NATO-accredited organization published a report concerning language-based disparities in LLM performance (Published April 25, 2026) — LLM performance varies significantly by language, but adaptation strategies can bridge the gap.
• Datasets for Verb Alternations across Languages: BLM Templates and Data Augmentation Strategies — This work addresses the underexplored ability of large language models (LLMs) to capture cross-sentence paradigmatic patterns, specifically verb alternations.

Open in AIssential →

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