Takoyaki at SemEval-2026 Task 3: Ensembling LLM Predictions using Demonstration Retrieval for Dimensional Aspect-based Sentiment Analysis

AbstractThis paper describes our system for SemEval-2026 Task 3 (DimABSA). We participate in Subtask 2 (DimASTE), which requires extracting triplets of aspect term, opinion term, and valence-arousal scores from review sentences, and Subtask 3 (DimASQP), which additionally requires aspect category classification to form quadruplets. Our proposed system consists of a multi-step pipeline: (1) retrieval-based in-context learning using BM25 to select relevant demonstrations for LLM inference, (2) agreement-based ensemble combining LLM predictions from multiple retrieval variants, and, for a subset of datasets, (3) error-pattern correction refining uncertain predictions using correction rule sets based on training data. Retrieval-based ICL and the agreement-based ensemble show consistent improvements across languages and domains. Error-pattern correction yields further improvement for the Japanese dataset. To further investigate output quality beyond automated evaluation metrics, we conducted human evaluation. The results suggest that LLM-based labeling achieves higher agreement with gold labels than human annotators, and additionally indicate a discrepancy between automated scores and practical output quality.