AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis

Stavros Gazetas; Giorgos Filandrianos; Maria Lymperaiou; Paraskevi Tzouveli; Athanasios Voulodimos; Giorgos Stamou

cs.CLMar 5, 2026

AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis

Authors: Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou, Paraskevi Tzouveli, Athanasios Voulodimos, Giorgos Stamou

Organizations: School of Electrical and Computer Engineering, AILS Laboratory National Technical University of Athens

Abstract

In this paper, we present AILS-NTUA system for Track-A of SemEval-2026 Task 3 on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which encompasses three complementary problems: Dimensional Aspect Sentiment Regression (DimASR), Dimensional Aspect Sentiment Triplet Extraction (DimASTE), and Dimensional Aspect Sentiment Quadruplet Prediction (DimASQP) within a multilingual and multi-domain framework. Our methodology combines fine-tuning of language-appropriate encoder backbones for continuous aspect-level sentiment prediction with language-specific instruction tuning of large language models using LoRA for structured triplet and quadruplet extraction. This unified yet task-adaptive design emphasizes parameter-efficient specialization across languages and domains, enabling reduced training and inference requirements while maintaining strong effectiveness. Empirical results demonstrate that the proposed models achieve competitive performance and consistently surpass the provided baselines across most evaluation settings.

Figures & tables

Lang.	Domain	Dev			Test
		RMSE VA $\downarrow$	PCC V $\uparrow$	PCC A $\uparrow$	RMSE VA $\downarrow$	PCC V $\uparrow$	PCC A $\uparrow$
ENG	Laptop	$1.0166$	$0.9174$	$0.5768$	$1.4401$	$0.8520$	$0.5073$
ENG	Restaurant	$0.9208$	$0.9295$	$0.6882$	$1.3933$	$0.8757$	$0.6105$
ZHO	Finance	$0.5438$	$0.8241$	$0.5981$	$0.5425$	$0.8327$	$0.6436$
	Laptop	$0.7443$	$0.8863$	$0.7042$	$0.7457$	$0.8738$	$0.7133$
	Restaurant	$0.8007$	$0.8419$	$0.6748$	$1.0023$	$0.8258$	$0.5852$

Table 1: DimASR results per language and domain.

Lang.	Model	Domain	cF1 $\uparrow$
			Dev	Test
ENG	Llama 3.1 8B	Laptop	$0.5962$	$0.5311$
ENG	Llama 3.1 8B	Restaurant	$0.7668$	$0.6518$
ZHO	Qwen 2.5 7B	Laptop	$0.4193$	$0.4646$
ZHO	Qwen 2.5 7B	Restaurant	$0.5916$	$0.5042$
JPN	Qwen 2.5 14B	Hotel	$0.4879$	$0.5021$

Table 2: DimASTE results per language and domain.

Lang.	Model	Domain	cF1 $\uparrow$
			Dev	Test
ENG	Llama 3.1 8B	Laptop	$0.3556$	$0.2694$
ENG	Llama 3.1 8B	Restaurant	$0.7560$	$0.5988$
ZHO	Qwen 2.5 7B	Laptop	$0.3662$	$0.3703$
ZHO	Qwen 2.5 7B	Restaurant	$0.5665$	$0.4544$
JPN	Qwen 2.5 14B	Hotel	$0.4218$	$0.3747$

Table 3: DimASQP results per language and domain.

Lang.	Translate+Train	ENG-Model
ZHO	$0.36$	$0.26$
RUS	$0.31$	$0.31$
UKR	$0.24$	$0.25$
TAT	$0.20$	$0.22$

Table 4: Restaurant DimASQP (cF1), Dev set, Llama 3.1 8B: translation training vs. English-trained model.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Subtask	Example (JSON)
DimASR	{"ID": "rest26_aspect_va_dev_1", "Text": "Great diner food and breakfast is served all day", "Aspect_VA": [{"Aspect": "diner food", "VA": "7.25#6.75"}, {"Aspect": "breakfast", "VA": "7.25#6.75"}]}
DimASTE	{"ID": "rest26_aste_dev_2", "Text": "Customer service was fantastic and food was awesome", "Triplet": [{"Aspect": "Customer service", "Opinion": "fantastic", "VA": "7.33#7.33"}, {"Aspect": "food", "Opinion": "awesome", "VA": "7.67#7.67"}]}
DimASQP	{"ID": "rest26_asqp_dev_1", "Text": "Food and coffee are great", "Quadruplet": [{"Aspect": "Food", "Category": "FOOD#QUALITY", "Opinion": "great", "VA": "7.67#7.83"}, {"Aspect": "coffee", "Category": "DRINKS#QUALITY", "Opinion": "great", "VA": "7.67#8.00"}]}

Appendix

Table 5: Example JSON format for each Subtask.

Subtask	Lang.	Domain	Zero-shot		One-shot		Supervised Fine-Tuning
			GPT-5 mini	Kimi K2 Thinking	GPT-5 mini	Kimi K2 Thinking	Qwen 3 14B	Mistral 3 14B	Llama 3.3 70B	GPT-OSS 120B
DimASR	ENG	Restaurant	$2.9490$	$2.3432$	$2.3926$	$2.1461$	$2.6427$	$2.6316$	$2.5244$	$1.4605$
	ENG	Laptop	$3.2115$	$2.6546$	$2.5637$	$2.1893$	$2.8089$	$2.6258$	$2.7354$	$1.5269$
	JPN	Hotel	$3.1406$	$2.3294$	$2.1607$	$1.7553$	$2.2906$	$2.2999$	$2.6255$	$0.7188$
	JPN	Finance	$2.6760$	$2.3379$	$1.9243$	$1.6396$	$1.8964$	$2.0700$	$2.4191$	$1.0188$
	RUS	Restaurant	$2.5447$	$2.0630$	$2.0390$	$1.7768$	$2.1528$	$2.3617$	$2.5089$	$1.4775$

Appendix

Table 6: Combined Benchmark paper’s results across subtasks. DimASR values are RMSE VA ; DimASTE and DimASQP values are cF1.

Language	Model family	Model ID
ENG	DeBERTa	yangheng/deberta-v3-base-absa-v1.1 8 8 8 yangheng/deberta-v3-base-absa-v1.1
Japanese	DeBERTa	ku-nlp/deberta-v3-base-japanese 9 9 9 ku-nlp/deberta-v3-base-japanese
Chinese	RoBERTa	hfl/chinese-roberta-wwm-ext 10 10 10 hfl/chinese-roberta-wwm-ext
Russian	BERT-style	DeepPavlov/rubert-base-cased 11 11 11 DeepPavlov/rubert-base-cased
Ukrainian	XLM-R	FacebookAI/xlm-roberta-base 12 12 12 FacebookAI/xlm-roberta-base
Tatar	XLM-R	FacebookAI/xlm-roberta-base 13 13 13 FacebookAI/xlm-roberta-base

Appendix

Table 7: Language-specific encoder backbones used for DimASR.

Hyperparameter	Value
Learning rate ( $\eta$ )	$2e-5$
Batch size	16
Dropout	0.3
Warmup ratio	0.1
Early stopping patience	5
Max sequence length	128

Appendix

Table 8: Training hyperparameters for the DimASR models.

Hyperparameter	Value
Train epochs ( $E$ )	1
Per-device batch size	2
Gradient accumulation steps	4
Effective batch (per device)	$2\times 4=8$
Learning rate ( $\eta$ )	$2e-4$
Weight decay	$1\times 10^{-4}$

Appendix

Table 9: LLM fine-tuning configuration (4-bit loading + LoRA adapters) for DimASTE & DimASQP.

Lang.	Domain	Dev cF1 $\uparrow$
		Zero-shot	Few-shot
ENG	Laptop	$0.48$	$0.59$
ENG	Restaurant	$0.53$	$0.59$
ZHO	Laptop	$0.23$	$0.27$
ZHO	Restaurant	$0.37$	$0.43$
JPN	Hotel	$0.17$	$0.19$

Appendix

Table 10: DimASTE Dev set results in zero-shot and few-shot prompting (cF1) using Qwen 2.5 14B Instruct.

Lang.	Domain	Dev cF1 $\uparrow$
		Zero-shot	Few-shot
ENG	Laptop	$0.22$	$0.32$
ENG	Restaurant	$0.42$	$0.51$
ZHO	Laptop	$0.14$	$0.18$
ZHO	Restaurant	$0.23$	$0.31$
JPN	Hotel	$0.06$	$0.11$

Appendix

Table 11: DimASQP Dev set results in zero-shot and few-shot prompting (cF1) using Qwen 2.5 14B Instruct.

Lang.	Model	Domain	Dev cF1 $\uparrow$
ENG	Llama 3.1 70B	Laptop	$0.3373$
ENG	Llama 3.1 70B	Restaurant	$0.7441$
ZHO	Qwen 2.5 32B	Laptop	$0.3119$
	Qwen 2.5 32B	Restaurant	$0.5220$
	Qwen 2.5 72B	Laptop	$0.3675$
	Qwen 2.5 72B	Restaurant	$0.5508$

Appendix

Table 12: DimASQP development set results (cF1) after lora-tuning larger LLMs.

Figure 1: DimASR dataset statistics across splits: number of reviews (left) and aspects per review (right).

Figure 2: DimASTE and DimASQP dataset statistics across splits: number of reviews (left) and quadruplets per review (right).

Figure 3: DimASTE and DimASQP NULL analysis on Train set: NULL rate (left) and composition of NULL cases (right).

Figure 4: DimASTE and DimASQP distribution shift measured by PSI between Train and Dev/Test across features (review length, quadruplets/review, and category).

Explore similar work

May 11, 2026

ICT-NLP at SemEval-2026 Task 3: Less Is More -- Multilingual Encoder with Joint Training and Adaptive Ensemble for Dimensional Aspect Sentiment Regression

This paper describes our system to SemEval-2026 Task 3 Track A Subtask 1 on Dimensional Aspect Sentiment Regression (DimASR). We propose a lightweight and resource-efficient system built entirely on multilingual pre-trained encoders, without relying on LLMs or external corpora. We adopt joint multilingual and multi-domain training to facilitate cross-lingual transfer and alleviate data sparsity, introduce a bounded regression transformation that improves training stability while constraining predictions within the valid range, and employ an adaptive ensemble strategy via subset search to reduce prediction variance. Experimental results demonstrate that our system achieves strong and consistent performance, ranking 1st on zho-res, 2nd on zho-lap, and 3rd on jpn-hot, with all remaining datasets placed within the top half of participating teams.

Liyuan Huang, Jiawei He, Wutao Shen +2

1State Key Laboratory of AI Safety · Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences

Aspect-Based Sentiment Analysis

Ask PDF

Apr 10, 2026

NCL-BU at SemEval-2026 Task 3: Fine-tuning XLM-RoBERTa for Multilingual Dimensional Sentiment Regression

Dimensional Aspect-Based Sentiment Analysis (DimABSA) extends traditional ABSA from categorical polarity labels to continuous valence-arousal (VA) regression. This paper describes a system developed for Track A, Subtask 1 (Dimensional Aspect Sentiment Regression), aiming to predict real-valued VA scores in the [1, 9] range for each given aspect in a text. A fine-tuning approach based on XLM-RoBERTa-base is adopted, with dual regression heads with sigmoid-scaled outputs for valence and arousal prediction. Separate models are trained for each language-domain pair (English and Chinese across restaurant, laptop, and finance domains), and training and development sets are merged for final test predictions. In development experiments, the fine-tuning approach is compared against several large language models under a few-shot prompting setting, demonstrating that task-specific fine-tuning outperforms these LLM-based methods across all evaluation datasets.

Tong Wu, Nicolay Rusnachenko, Huizhi Liang

Independent Researcher · Centre for Applied Creative Technologies (CFACT+), Bournemouth University, UK · School of Computing, Newcastle University, Newcastle upon Tyne, UK

Sentiment Analysis Aspect-Based Sentiment Analysis

Ask PDF

Mar 8, 2026

QuadAI at SemEval-2026 Task 3: Ensemble Learning of Hybrid RoBERTa and LLMs for Dimensional Aspect-Based Sentiment Analysis

We present our system for SemEval-2026 Task 3 on dimensional aspect-based sentiment regression. Our approach combines a hybrid RoBERTa encoder, which jointly predicts sentiment using regression and discretized classification heads, with large language models (LLMs) via prediction-level ensemble learning. The hybrid encoder improves prediction stability by combining continuous and discretized sentiment representations. We further explore in-context learning with LLMs and ridge-regression stacking to combine encoder and LLM predictions. Experimental results on the development set show that ensemble learning significantly improves performance over individual models, achieving substantial reductions in RMSE and improvements in correlation scores. Our findings demonstrate the complementary strengths of encoder-based and LLM-based approaches for dimensional sentiment analysis. Our development code and resources will be shared at https://github.com/aaronlifenghan/ABSentiment

A. J. W. de Vink, Filippos Karolos Ventirozos, Natalia Amat-Lefort +1

LIACS, Leiden University, NL · Manchester Metropolitan University, UK · Leiden University Medical Center, NL

Aspect-Based Sentiment Analysis Language Model Ensembles

Ask PDF