This paper presents an approach to SemEval-2026 Task 3, which focuses on dimensional aspect-based sentiment analysis. The work investigates methods for predicting fine-grained, real-valued scores for sentiment valence and arousal rather than traditional categorical labels.
- The team participates in Subtask 1, predicting scores for given aspects, and Subtask 3, extracting full sets of sentiment details including aspects, categories, opinions, and scores.
- For the regression task, the approach uses a weighted ensemble of transformer-based encoder models.
- For Russian language inputs, the method enhances data by using a large language model to generate synthetic sentiment descriptions.
- For the extraction task, a decoder LLM is fine-tuned to perform structured prediction, identifying sentiment elements and estimating numerical scores simultaneously.