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

1. Problem Definition

Traditional Aspect-Based Sentiment Analysis (ABSA) and Stance Detection typically rely on categorical polarity labels (e.g., Positive, Negative, Neutral). The authors argue that this coarse-grained approach fails to capture the nuance of human emotion, which is better represented in psychology and affective science as continuous dimensions.

The paper addresses two main challenges:

1. Dimensional Sentiment Representation: Moving from categorical labels to a Valence-Arousal (VA) space.
   • Valence: Ranges from negative (1) to positive (9).
   • Arousal: Ranges from low/sluggish (1) to high/excited (9).
2. Domain Expansion: Extending ABSA beyond consumer reviews (e.g., restaurants, laptops) to public-issue discourse (e.g., politics, climate change, energy), where stance targets function as aspects.

2. Methodology & Task Organization

The task is organized into two tracks, each with specific subtasks and evaluation metrics.

Track A: Dimensional Aspect-Based Sentiment Analysis (DimABSA)

This track focuses on extracting sentiment elements and predicting their VA scores. It includes three subtasks:

1. DimASR (Dimensional Aspect Sentiment Regression): Given a sentence and specific aspect terms, predict the VA scores for each aspect.
2. DimASTE (Dimensional Aspect Sentiment Triplet Extraction): Jointly extract Aspect terms (A), Opinion terms (O), and predict their VA scores.
3. DimASQP (Dimensional Aspect Sentiment Quadruplet Prediction): Extend DimASTE by also predicting the Aspect Category (C) (e.g., FOOD#QUALITY).

Track B: Dimensional Stance Analysis (DimStance)

This track treats stance targets (e.g., a political figure or policy) as aspects.

• Subtask: Given a text and a target, predict the VA scores representing the speaker's stance toward that target. This reformulates stance detection from a categorical classification problem (Favor/Against/Neutral) to a regression problem in the VA space.

Datasets

The organizers constructed multilingual, multi-domain datasets:

• Track A: 6 languages (English, Japanese, Russian, Tatar, Ukrainian, Chinese) across 4 domains (Restaurant, Laptop, Hotel, Finance). Total: ~77k aspect instances.
• Track B: 5 languages (English, German, Chinese, Nigerian Pidgin, Swahili) across 2 domains (Environmental Protection, Politics). Total: ~11.7k stance targets.
• Annotation: Annotated by native speakers with VA scores (1–9). Agreement was measured using RMSE for VA and F1 for tuple extraction.

Evaluation Metrics

• RMSE (Root Mean Square Error): Used for regression subtasks (DimASR, DimStance) to measure the Euclidean distance between predicted and gold VA scores.
• cF1 (Continuous F1): A novel metric introduced for extraction subtasks (DimASTE, DimASQP).
  • Standard F1 only counts a tuple as a True Positive if categorical elements match exactly.
  • cF1 extends this by incorporating the VA prediction error. A predicted tuple is a "continuous True Positive" weighted by the normalized Euclidean distance of the VA scores. This allows for partial credit based on the accuracy of the sentiment intensity prediction.

3. Key Contributions

1. Paradigm Shift: The first major shared task to integrate Valence-Arousal dimensions into standard ABSA and Stance Detection frameworks, moving away from discrete polarity.
2. New Metric (cF1): Introduced a hybrid evaluation metric that jointly assesses structured extraction (categorical) and sentiment regression (continuous), addressing the limitation of standard F1 in dimensional tasks.
3. Multilingual & Cross-Domain Resources: Released high-quality, annotated datasets covering low-resource languages (Tatar, Swahili, Nigerian Pidgin) and diverse domains (finance, politics, travel), bridging the gap between consumer reviews and public discourse.
4. Baseline Systems: Provided strong baselines using Large Language Models (LLMs) like Kimi K2, Qwen3, and Mistral-3, alongside traditional Transformers (mBERT), establishing a benchmark for future research.

4. Results & Analysis

The task attracted 400+ participants resulting in 112 final submissions from 42 teams.

• Performance Trends:

  • High-Resource vs. Low-Resource: Systems achieved lower RMSE (better performance) on high-resource languages like Chinese and Japanese. Performance degraded significantly on low-resource languages (e.g., Tatar, Swahili), highlighting the data scarcity challenge.
  • Task Difficulty: Subtask 3 (Quadruplet) was harder than Subtask 2 (Triplet) due to the added complexity of aspect category classification, particularly in domains with long-tailed category distributions (e.g., laptops, hotels).
  • Stance Detection: Track B showed similar trends, with the best results on Chinese and the worst on Swahili.

• Top-Performing Systems:

  • Team PAI: Used distributional adaptation (Sinkhorn algorithm) to align predicted VA scores with training distributions while preserving inter-dimensional correlations.
  • Team TeleAI: Employed a single multilingual/multi-domain model (Qwen2.5-7B) with robust training techniques (Smooth L1 loss, R-Drop, adversarial training) and post-hoc linear calibration.
  • Team Takoyaki: Leveraged Retrieval-Augmented Generation (RAG) with BM25 to retrieve similar training examples for in-context learning with Gemini 3.0 Pro, followed by ensemble strategies.
  • Team LogSigma: Treated VA prediction as two separate regression tasks and used uncertainty-weighted multitask learning (learning log-variance parameters) to balance valence and arousal losses dynamically.

• Common Technical Strategies:

  • Architecture: 91% of systems used Transformer-based architectures or LLMs.
  • Training: Fine-tuning (Full or Parameter-Efficient/QLoRA) was dominant. Techniques like adversarial training, ensemble methods (multi-seed, cross-validation), and instruction prompting with few-shot examples were widely adopted.

5. Significance

• Nuanced Sentiment Understanding: By adopting the VA framework, the task enables applications requiring fine-grained emotional analysis, such as mental health monitoring, nuanced political analysis, and detailed customer feedback interpretation.
• Bridging Discourse Types: The inclusion of Track B (DimStance) successfully unifies the methodologies of ABSA and Stance Detection, demonstrating that stance targets can be effectively analyzed using aspect-based extraction techniques.
• Low-Resource Challenges: The results highlight the critical need for better multilingual models and data augmentation strategies for low-resource languages in dimensional sentiment tasks.
• Evaluation Innovation: The introduction of cF1 sets a new standard for evaluating hybrid tasks where both categorical structure and continuous values must be predicted accurately.

In conclusion, SemEval-2026 Task 3 successfully established a new benchmark for Dimensional ABSA and Stance Analysis, providing the community with the necessary data, metrics, and baselines to advance the field from coarse polarity classification to fine-grained, continuous emotional modeling.