LogSigma at SemEval-2026 Task 3: Uncertainty-Weighted Multitask Learning for Dimensional Aspect-Based Sentiment Analysis

A research team led by Baraa Hikal, Jonas Becker, and Bela Gipp has developed LogSigma, a novel AI system that won first place in SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA). Unlike traditional sentiment analysis that outputs simple positive/negative labels, DimABSA requires predicting continuous scores for Valence (pleasantness) and Arousal (intensity) on a 1-9 scale. The central innovation is the use of learned homoscedastic uncertainty, where the model automatically learns task-specific log-variance parameters during training. This allows it to dynamically balance the regression objectives for Valence and Arousal, which vary significantly in difficulty across different languages and domains.

LogSigma's architecture combined this uncertainty-weighting technique with language-specific encoders and multi-seed ensembling. The results were decisive: the system achieved 1st place rankings across five different datasets in the competition. The research revealed that the learned variance weights for balancing Valence and Arousal tasks varied dramatically by language, from 0.66x for German to 2.18x for English. This proves that the optimal weighting for multitask learning in sentiment analysis is inherently language-dependent and cannot be effectively determined through manual, one-size-fits-all approaches.

The success of LogSigma highlights a significant advancement in making AI models more adaptive and efficient for complex, real-world NLP tasks that involve multiple, correlated objectives. By teaching the model to assess and weight its own uncertainty, the researchers have created a more robust framework for multilingual sentiment analysis that can automatically adjust to the unique challenges presented by different languages.