SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)
Abstract Overview
This paper presents SemEval-2026 Task 3, a shared task on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which replaces categorical polarity labels in traditional ABSA with continuous valence–arousal (VA) scores. The task introduces two tracks: Track A (DimABSA) covering aspect-level regression, triplet extraction, and quadruplet extraction subtasks, and Track B (DimStance) treating stance targets as aspects and reformulating stance detection as VA regression. The benchmark spans multiple languages (six for Track A, five for Track B) and multiple domains. The paper details dataset construction, annotation procedures, evaluation metrics including a novel continuous F1 (cF1) metric, baselines, and an analysis of the 112 final submissions from 42 teams across both tracks.
Novelty
The key contribution is the reformulation of ABSA and stance detection in a continuous valence–arousal space rather than with discrete sentiment or stance labels. The work introduces DimStance as a new task formulation that treats stance targets as aspects with VA regression, and proposes the continuous F1 (cF1) metric to jointly evaluate structured extraction and continuous VA prediction.
Results
The shared task attracted over 400 participants, yielding 112 final submissions and 42 system description papers across two tracks. Systems achieved lower RMSE on Chinese and Japanese datasets, while the highest errors were observed on low-resource languages such as Tatar (Track A) and Swahili (Track B). Quadruplet extraction (DimASQP) proved harder than triplet extraction (DimASTE), with notable performance drops in domains with larger and more imbalanced aspect category sets such as laptop and hotel.
Key Points
- The task defines two tracks: DimABSA for aspect-level dimensional sentiment (with regression, triplet extraction, and quadruplet extraction subtasks) and DimStance for stance-target VA regression, supported by multilingual, multi-domain datasets.
- A new continuous F1 (cF1) metric is introduced to jointly evaluate structured extraction accuracy and VA regression quality by incorporating normalized VA distance into the F1 formulation.
- Top-performing systems rely on fine-tuned pretrained transformers and LLMs with techniques such as distribution calibration, adversarial training, retrieval-based in-context learning, and ensembling, while low-resource languages (Tatar, Swahili) remain the most challenging settings.