SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)

Liang-Chih Yu; Jonas Becker; Shamsuddeen Hassan Muhammad; Idris Abdulmumin; Lung-Hao Lee; Ying-Lung Lin; Jin Wang; Jan Philip Wahle

cs.CLApr 8, 2026

SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)

Authors: Liang-Chih Yu, Jonas Becker, Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Lung-Hao Lee, Ying-Lung Lin, Jin Wang, Jan Philip Wahle, +9 more

Organizations: Yuan Ze University · University of Göttingen · Imperial College London · University of Pretoria · National Yang Ming Chiao Tung University · Central Police University · Yunnan University · Lomonosov Moscow State University · Skoltech · AIRI · Maseno University · UCLA · National Research Council Canada

Abstract

We present the SemEval-2026 shared task on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which improves traditional ABSA by modeling sentiment along valence-arousal (VA) dimensions rather than using categorical polarity labels. To extend ABSA beyond consumer reviews to public-issue discourse (e.g., political, energy, and climate issues), we introduce an additional task, Dimensional Stance Analysis (DimStance), which treats stance targets as aspects and reformulates stance detection as regression in the VA space. The task consists of two tracks: Track A (DimABSA) and Track B (DimStance). Track A includes three subtasks: (1) dimensional aspect sentiment regression, (2) dimensional aspect sentiment triplet extraction, and (3) dimensional aspect sentiment quadruplet extraction, while Track B includes only the regression subtask for stance targets. We also introduce a continuous F1 (cF1) metric to jointly evaluate structured extraction and VA regression. The task attracted more than 400 participants, resulting in 112 final submissions and 42 system description papers. We report baseline results, discuss top-performing systems, and analyze key design choices to provide insights into dimensional sentiment analysis at the aspect and stance-target levels. All resources are available on our GitHub repository.

Figures & tables

Figure 2: Countries of official affiliations of participants. Larger dots indicate more participants. A total of 24 countries are represented.

Dataset	Team	Score	Dataset	Team	Score	Dataset	Team	Score	Dataset	Team	Score
Subtask 1
eng-rest	LogSigma	1.1035	eng-lap	LogSigma	1.2408	jpn-hot	TeleAI	0.5561	jpn-fin	TeleAI	0.6581
	Bert Kittens	1.1812		TeleAI	1.2425		PALI	0.6237		PALI	0.7532
	Baseline(KimiK2)	2.1461		Baseline(KimiK2)	2.1893		Baseline(KimiK2)	1.7553		Baseline(KimiK2)	1.6396
	Baseline(Qwen3-14B)	2.6427		Baseline(Qwen3-14B)	2.8089		Baseline(Qwen3-14B)	2.2906		Baseline(Qwen3-14B)	1.8964
rus-rest	PAI	1.2190	tat-rest	PAI	1.5294	ukr-rest	PAI	1.1888	zho-fin	HUS@NLP-VNU	0.4841

Table 2: Track A (DimABSA) results across all subtasks. Subtask 1 is evaluated using RMSE, while Subtasks 2 and 3 are evaluated using cF1. The top two teams and the official baseline systems are reported for each dataset.

Dataset	Team	Score	Dataset	Team	Score
Subtask 1
eng-env	LogSigma	1.4734	deu-pol	LogSigma	1.3417
	hllwan	1.5122		NTNU-SMIL	1.3467
	Baseline (Mistral-3-14B)	1.6431		Baseline (Mistral-3-14B)	1.5914
	Baseline (mBERT)	2.6985		Baseline (mBERT)	2.3294
zho-env	YangS_team	0.5468	pcm-pol	CYUT	1.1024

Table 3: Track B (DimStance) results. Evaluation is based on RMSE. The top two teams and the official baseline systems are reported for each dataset.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Entity Labels

LAPTOP, DISPLAY, KEYBOARD, MOUSE, MOTHERBOARD, CPU, FANS_COOLING, PORTS, MEMORY, POWER_SUPPLY, OPTICAL_DRIVES, BATTERY, GRAPHICS, HARD_DISK, MULTIMEDIA_DEVICES, HARDWARE, SOFTWARE, OS, WARRANTY, SHIPPING, SUPPORT, COMPANY

Attribute Labels

GENERAL, PRICE, QUALITY, DESIGN_FEATURES, OPERATION_PERFORMANCE, USABILITY, PORTABILITY, CONNECTIVITY, MISCELLANEOUS

Appendix

Table 5

Entity Labels

RESTAURANT, FOOD, DRINKS, AMBIENCE, SERVICE, LOCATION

Attribute Labels

GENERAL, PRICES, QUALITY, STYLE_OPTIONS, MISCELLANEOUS

Appendix

Table 6

Entity Labels

HOTEL, ROOMS, FACILITIES, ROOM_AMENITIES, SERVICE, LOCATION, FOOD_DRINKS

Attribute Labels

GENERAL, PRICE, COMFORT, CLEANLINESS, QUALITY, DESIGN_FEATURES, STYLE_OPTIONS, MISCELLANEOUS

Appendix

Table 7

Task	Input	Output	Prediction Type	Metric
DimASR	text + aspects	V#A	Regression	RMSE
DimASR	The food was excellent	8.00#8.12	Regression	RMSE
DimASTE	text	(A, O, V#A)	Extraction	cF1
DimASTE	Service at the bar was a little slow	Service, a little slow, 4.10#4.30)	Regression	cF1
DimASQP	text	(A, C, O, V#A)	Extraction	cF1
	Their sodas are usually expired and flat	(sodas, DRINKS#QUALITY, usually expired, 1.90#7.20)	Classification

Appendix

Table 8

		VA error distance
Prediction/Gold	$TP_{cat}$ (A)	Raw (B)	Normalized $(C)=(B)/\sqrt{128}$	cTP (A)-(C)
P: (food, good, 8.00#8.00)	1	$\sqrt{2}$	$\frac{\sqrt{2}}{\sqrt{128}}=0.125$	0.875
G: (food, good, 7.00#7.00)
P: (soup, spicy, 7.50#7.50)	1	$\sqrt{32}$	$\dfrac{\sqrt{32}}{\sqrt{128}}=0.5$	0.5
G: (soup, spicy, 3.50#3.50)
P: (staff, friendly, 7.00#7.00)	0	-	-	0

Appendix

Table 9

Tracks	Team	Affiliation	Paper
A	AILS-NTUA	Artificial Intelligence and Learning Systems Laboratory, School of Electrical and Computer Engineering, National Technical University of Athens, Greece	Gazetas et al. (2026)
A	ALPS-Lab	Fujian University of Technology, China	Dai and Lin (2026)
A	Bert Kittens	Individual researcher	Sukhodolsky et al. (2026)
B	CLRG	Bangladesh University of Engineering and Technology, Bangladesh; BRAC University, Bangladesh	Iqbal et al. (2026)
B	CYUT	Chaoyang University of Technology, Taiwan	Wu et al. (2026a)
A, B	DUTH	Department of Electrical & Computer Engineering, Democritus University of Thrace, Greece	Arampatzis and Arampatzis (2026)

Appendix

Table 10

S/N	Team	eng-rest	eng-lap	jpn-hot	jpn-fin	rus-rest	tat-rest	ukr-rest	zho-rest	zho-lap	zho-fin
1	AILS-NTUA	1.3933	1.4401	0.7484	0.9635	1.7236	2.1144	1.6724	1.0023	0.7457	0.5425
2	Bert Kittens	1.1812	1.2769	0.7267	0.9675	1.5828	2.2118
3	DUTH		1.5924
4	HUS@NLP-VNU	1.2745	1.4109	0.6386	0.8296	1.3075	1.8220	1.3538	0.9595	0.6663	0.4841
5	Habib university	1.3049	1.3654	0.6680	0.8907	1.4344	1.6041	1.4661	0.9898	0.7311	0.5333
8	LogSigma	1.1035	1.2408

Appendix

Table 9: DimABSA results (Track A, Subtask 1).

S/N	Team	eng-rest	eng-lap	jpn-hot	rus-rest	tat-rest	ukr-rest	zho-rest	zho-lap
1	AILS-NTUA	0.6518	0.5311	0.5021	0.4988	0.3874	0.4725	0.5042	0.4646
2	ALPS-Lab	0.0000	0.0000	0.0000	0.5414	0.4798	0.5613	0.5247	0.4935
3	Bert Kittens	0.5628	0.4469	0.4202	0.3137	0.1692
5	HUS@NLP-VNU	0.6391	0.5304
6	Habib university	0.5202	0.4770	0.3311	0.5492	0.4839	0.5324	0.4622	0.4159
7	ICT-NLP	0.6174	0.5622	0.3152	0.4622	0.3088	0.4355	0.2756	0.3019

Appendix

Table 10: DimABSA results (Track A, Subtask 2).

S/N	Team	eng-rest	eng-lap	jpn-hot	rus-rest	tat-rest	ukr-rest	zho-rest	zho-lap
1	AILS-NTUA	0.5988	0.2694	0.3747	0.4369	0.3306	0.4154	0.4544	0.3703
2	ALPS-Lab	0.6202	0.3395	0.3617	0.5042	0.4404	0.5163	0.4853	0.3968
3	Bert Kittens	0.5162	0.2578	0.2845		0.1479
4	HUS@NLP-VNU	0.5871	0.2587
5	Habib university	0.0000	0.0000	0.1853	0.3029	0.2500	0.2938	0.4199	0.3139
7	NYCU Speech Lab							0.5521	0.4824

Appendix

Table 11: DimABSA results (Track A, Subtask 3).

S/N	Team	eng-env	deu-pol	zho-env	pcm-pol	swa-pol
1	CLRG	2.0654	2.0923	0.6170	1.9114	2.1320
2	CYUT	1.6331	1.4827	0.6771	1.1024	2.1042
3	DUTH	2.1964
4	HUS@NLP-VNU	1.6899	1.4108	0.5826	1.4269	1.8713
5	LogSigma	1.4734	1.3417	0.6460	1.1269	1.7959
6	NTNU-SMIL	1.5207	1.3467	0.5561	1.5674	1.9602

Appendix

Table 12: DimStance results (Track B).

Figure 3: Model architectures used by participants.

Figure 4: LLMs used by participants.

Figure 5: Training techniques used by participants.

Figure 6: Fine-tuning strategies used by participants.

Figure 7: Prompting strategies used by participants.

Figure 8: External data used by participants.

Explore similar work

Mar 5, 2026

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

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.

Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou +3

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

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