Source-linked AI summary
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, Terry Ruas, Natalia Loukachevitch, Alexander Panchenko, Ilseyar Alimova, Lilian Wanzare, Nelson Odhiambo, Bela Gipp, Kai-Wei Chang, Saif M. Mohammad
TL;DR
The paper addresses the limitation of categorical sentiment and stance labels by introducing dimensional valence–arousal analysis at aspect and stance-target levels. It defines the DimABSA and DimStance tracks with regression and structured extraction tasks, and reports broad participation alongside ongoing challenges, particularly for low-resource languages. The paper also identifies cross-language comparability and dataset valence imbalance as scope considerations.
Problem
Existing ABSA and stance approaches primarily use categorical labels, while the paper seeks finer-grained sentiment representation across aspect and public-issue discourse.
Method
The shared task introduces DimABSA and DimStance, combining valence–arousal regression with aspect, opinion, category, and stance-target extraction across multilingual, multidomain datasets.
Results
The task attracted over 400 participants, yielding 112 final submissions from 44 teams and 42 system description papers, while evaluation showed persistent challenges for low-resource languages.
Takeaways & Limitations
The task establishes a shared evaluation framework for dimensional sentiment analysis at aspect and stance-target levels and highlights challenges and opportunities for future systems.
Takeaways & Limitations
Valence and arousal interpretations can vary across cultures, so results should be interpreted as comparisons across language-community-domain settings.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper extends ABSA from categorical polarity to continuous valence–arousal representations and introduces a related dimensional stance task. The shared task evaluates structured extraction and regression across multilingual, multidomain settings.
- ABSA traditionally extracts aspect, category, opinion, and polarity elements, but commonly represents sentiment with coarse categorical labels.
- DimABSA integrates continuous valence–arousal scores with traditional ABSA elements in multilingual, multidomain datasets.
- DimStance treats stance targets as aspects and predicts their valence–arousal scores, extending analysis to political, energy, and climate discourse.
- The task comprises Track A with three DimABSA subtasks and Track B with only dimensional aspect sentiment regression for stance targets.
- Over 400 participants produced 112 final submissions from 44 teams and 42 system description papers.
- Evaluation results indicate that dimensional sentiment analysis remains challenging, especially for low-resource languages.
2 Related Work
Prior work developed categorical ABSA and stance resources, alongside dimensional sentiment resources at lexical, phrasal, and sentential levels. These strands motivate a dimensional formulation at aspect and stance-target granularity.
- Categorical ABSA benchmarks expanded from English restaurant and laptop reviews to multilingual datasets and richer triplet and quadruplet annotations.
- Categorical stance research broadened language coverage, dataset scale, and domain specificity beyond early English Twitter benchmarks.
- Dimensional sentiment resources assign affective scores at lexical, phrasal, and sentential granularities.
3 Task Description
Track A adapts ABSA to valence–arousal prediction by combining aspect, category, and opinion extraction with continuous scores. Track B applies the regression formulation to stance targets treated as aspects.
- Track A: Track A represents sentiment using aspect terms, aspect categories, opinion terms, and valence–arousal scores on 1–9 scales.
- Track A: The task defines three subtasks that adapt traditional ABSA formulations to dimensional sentiment analysis.
- Track A: DimASR predicts valence–arousal scores for given aspects, generalizing aspect sentiment classification to regression.
- Track A: DimASTE extracts (A, O, VA) triplets, while DimASQP extracts (A, C, O, VA) quadruplets with aspect categories.
- Track B: DimStance treats stance targets as aspects and predicts valence–arousal scores for each target in an utterance or post.
4 Datasets
The task provides multilingual, multidomain datasets for aspect sentiment and stance analysis, with annotations tailored to each subtask. Annotation combines structured tuple extraction, adjudication or voting, and averaged valence–arousal ratings.
- Dataset coverage: Track A covers six languages and four domains, providing 76,958 aspect instances across 42,590 sentences.
- Dataset coverage: Track B covers five languages and two public-issue domains, containing 11,746 stance targets across 7,365 texts.
- Dataset statistics: Table 1 reports dataset sources, subtask type, and train, development, and test sizes as text/instance counts.
- Annotation design: Datasets use (A, VA) pairs for regression-only settings and full (A, C, O, VA) quadruplets where all subtasks are supported.
- Annotation process: Track A tuple extraction uses two annotators with third-party adjudication, whereas Track B stance targets are validated by five annotators through majority voting.
- Annotation quality: Final valence–arousal scores are averaged across five annotators, while agreement is assessed with tuple-level F1 and valence/arousal RMSE.
5 Evaluation
The evaluation uses RMSE for VA regression and a continuous F1 metric for jointly evaluating categorical extraction and VA prediction. Baselines, infrastructure, and two-phase submission procedures support both tracks.
- Metrics: DimASR is evaluated with RMSE over prediction error in the valence-arousal space.The metric uses predicted and gold valence and arousal values across instances.
- Metrics: DimASTE and DimASQP use cF1 because standard F1 cannot jointly assess categorical prediction and VA regression.The metric extends categorical true positives with VA prediction error.
- Metrics: Continuous true positives require exact categorical matches and are discounted by normalized Euclidean distance in VA space.The distance is normalized using the maximum possible distance on the [1, 9] scale, ensuring values in [0, 1].
- Baselines: Two baseline systems are provided for each track, using combinations of Kimi K2 Thinking, Qwen3-14B, mBERT, and Mistral-3-14B.Qwen3-14B and Mistral-3-14B use QLoRA, while mBERT is fully fine-tuned.
- Task organization: The task used Codabench, pilot data, a GitHub starter kit, and support activities for participants.Support included beginner resources, a Q&A session, and a writing tutorial.
- Task organization: The competition had development and evaluation phases, with up to 999 development submissions and four evaluation runs per participant.The final official ranking used each participant’s last evaluation submission.
6 Participating Systems and Results
The shared task drew broad participation across tracks and subtasks, while results reveal language- and task-dependent difficulty. Top systems combine calibration, multilingual training, retrieval, adaptive objectives, ensembling, or geometry-informed signals.
- Participation: Over 300 participants joined Track A and over 100 joined Track B, with English receiving the most submissions while every language received at least 20 per track.Development and evaluation phases together recorded 2664 and 177 Track A submissions, and 357 and 67 Track B submissions.
- Participation: 112 submissions from 42 unique teams were reported, including 39 teams in Track A and 13 teams in Track B.Ten teams participated in both tracks.
- Track A: Track A attracted 36 teams for DimASR, 22 for DimASTE, and 20 for DimASQP.Table 2 reports the top two systems and official baselines for each dataset across all three subtasks.
- Track A: DimASR systems achieve lower RMSE on Chinese and Japanese data and the highest RMSE on low-resource Tatar data.DimASTE likewise has its highest cF1 on English datasets and lowest cF1 on Tatar; DimASQP is harder because it adds domain-dependent category classification.
- Top systems: Top systems use distributional calibration, multilingual robust training, retrieval-based prompting, and agreement-based correction strategies across Track A datasets.Reported systems include PAI, TeleAI, and Takoyaki, with each achieving first place on specified dataset-subtask combinations.
- Track B: Track B systems achieve the lowest RMSE on Chinese stance data and the highest RMSE on low-resource Swahili data.Track B contains 13 participating teams, with results and baselines reported in Table 3.
- Top systems: Track B leaders balance valence and arousal objectives with learned log-variance parameters, aspect-aware dual heads, ensembling, or geometry-derived auxiliary signals.These approaches are associated with the reported first-place systems LogSigma, YangS_team, and CYUT.
7 Analysis and Discussion
Participating systems predominantly use pretrained transformers and LLMs, often enhanced with ensembling, specialized training losses, auxiliary signals, adapters, and prompting strategies.
- Model Architecture: Pretrained transformers and LLMs form the dominant architectural basis, while some teams combine models through heterogeneous or seed-based ensembles.The reported architectures include RoBERTa-family models, Qwen, and a syntax-aware GCN.
- Training Techniques: Most systems fine-tune pretrained models, using full or parameter-efficient adaptation alongside techniques such as Smooth L1 loss and log-variance weighting.Other approaches add auxiliary learning signals or adjust prediction distributions.
- Training Techniques: Per-language adapters can capture language-specific VA distributions while reducing the number of required models.Team PALI used this approach and ranked first on zho-lapST2 in Track A.
- Prompting Strategies: Instruction prompting with few-shot demonstrations is widespread, while retrieval-based demonstrations and self-consistency target prediction consistency and reliability.Self-consistency aggregates multiple stochastic predictions and retains tuples reaching consensus.
8 Conclusions
The shared task extends ABSA and stance detection with valence–arousal representations across regression and structured extraction settings, and evaluates systems on multilingual, multidomain datasets.
- Contributions: The task introduces DimABSA and DimStance tracks with subtasks spanning VA regression and hybrid structured extraction with regression.The framework covers dimensional aspect sentiment and stance-target analysis.
- Evaluation: The cF1 metric unifies categorical and continuous evaluation for structured extraction with VA prediction.It is designed to assess both extraction correctness and dimensional regression.
- Analysis: The paper reports multilingual and multidomain system results, analyzes top-performing systems, and summarizes key design choices.These analyses identify challenges and opportunities at aspect and stance-target levels.
Limitations
The datasets and comparisons are bounded by cross-cultural differences in VA interpretation and uneven valence distributions across languages.
- Cross-lingual comparability: Cross-cultural differences in valence and arousal interpretation can affect crosslingual comparability despite native-speaker annotation, shared scales, and common guidelines.Results should therefore be interpreted as comparisons across language-community-domain settings.
- Data distribution: Uneven concentrations of positive or negative valence in some datasets may bias training and inflate performance in majority regions of VA space.The paper encourages reweighting or stratified sampling when comparing models across languages.
Ethical Considerations
The paper emphasizes that inferred sentiment reflects perceived rather than necessarily true attitudes and warns that sentiment technologies can be misused, especially against vulnerable people.
- Interpretation: Human annotations capture perceived sentiment and attitudes, which may differ from speakers’ true attitudes.Perceived opinions may still correlate with actual opinions at an aggregate level.
- Interpretation: Individual inferred opinions should be used cautiously for decisions about people, particularly in high-stakes settings.The caution follows from person-to-person variation and the distinction between perceived and true attitudes.
- Misuse: ABSA and stance detection can be abused to identify preferences and manipulate behavior, with heightened concern for children and older adults.The paper expressly forbids commercial use of its data.
- Further considerations: The paper directs readers to dedicated discussions of ethical considerations in automatic sentiment and emotion detection.
B Overview of Subtasks with Examples
The task evaluates dimensional aspect sentiment through regression and structured extraction, using valence–arousal scores and the continuous F1 metric. Examples illustrate aspect-level outputs, sentiment expressions, categories, and VA predictions.
- Dimensional aspect sentiment triplet extraction: Triplet extraction identifies an aspect, opinion expression, and VA score, evaluated with cF1.The example extracts Service and “a little slow” with a 4.10#4.30 prediction.
- Dimensional aspect sentiment quadruplet extraction: Quadruplet extraction adds an aspect category to the aspect, opinion, and VA score structure.The example represents sodas with category DRINKS#QUALITY and the opinion “usually expired.”
- Dimensional aspect sentiment quadruplet extraction: A single sentence can yield multiple aspect-category-opinion records with distinct VA predictions, as shown for “usually expired” and “flat.”
- Evaluation: VA scores range from 1 to 9, and cRecall/cPrecision become standard recall and precision when prediction distance is zero.