In-Context Learning (ICL) has emerged as a powerful method for Large Language Models (LLMs) to perform tasks by utilizing prompts without additional training. Few-shot learning, a subset of ICL, enhances model performance by providing a small number of task-specific examples. However, the effectiveness of few-shot learning depends heavily on the selection of examples. Therefore, it is crucial to choose representative few-shot instances. In this study, we propose a novel few-shot selection method that leverages similar tasks instead of directly using the target task data. The target task in this paper is aspect term extraction on sentiment analysis. We introduce two selection strategies as similar tasks: one based on named entity recognition (NER) utilizing Predictive Entropy (PE) and another based on positive-negative classification using Information Gain (IG). In the experiment, the IG-based selection significantly outperformed baseline approaches: random selection and confidence-based selection. Our findings highlight the importance of choosing few-shot instances with small uncertainty. This suggests that informative and clearly structured instances contribute to better performance in LLM-based few-shot learning. The proposed method provides an effective way to optimize few-shot selection without relying on target task data.