ALPS-Lab at SemEval-2026 Task 3: A Multilingual Generative LLM Approach for Dimensional Aspect Sentiment Analysis

AbstractWe propose a SFT approach for the DimABSA shared task, which predicts aspect-level sentiment intensities using large language models. The approach uses Gemma-3 27B with QLoRA for efficient fine-tuning on multilingual datasets. Merging data across languages improves performance, especially in low-resource domains. Post-processing removes duplicate outputs for accurate evaluation.