<p>Sentiment analysis is a fundamental task in natural language processing with broad applications. While pre-trained language models have achieved remarkable success in sentiment analysis, these approaches typically require large training data. In sentiment analysis, labeled data may be scarce as the problem expands to new areas, such as novel modes of expression on social media, emerging topics, and reviews on new products. Moreover, given the dynamic nature of real-world applications, continuously updating the data with new information is impractical. This highlights the need for efficient models that perform well with limited labeled data. In this paper, we propose Sentiment Lexicon Integrated Meta-training (SLIM), a novel framework designed to optimize pre-trained language models for sentiment analysis in data-scarce settings. SLIM integrates the knowledge of a sentiment lexicon into self-supervised meta-training tasks. Specifically, we propose three methods, SLIM_boolean, SLIM_mixed, and SLIM_polar, to generate sentiment-aware meta-training tasks by exploiting the semantic relationship between words. SLIM_boolean distinguishes sentiment-bearing expressions from neutral ones, SLIM_mixed captures variations among sentiment expressions, and SLIM_polar explicitly contrasts opposing sentiment polarities. We then meta-train several SLIM models using these task generation methods along with two model initialization settings. Experimental results on 11 benchmark datasets show that SLIM significantly improves performance in few-shot settings, demonstrating the capability of SLIM models to adapt effectively with minimal training examples. When combined with pre-existing sentiment knowledge, our model outperforms recent meta-learning models and sentiment knowledge-enhanced pre-trained language models. Furthermore, we conduct comprehensive ablation studies to examine the impact of specific design choices within the SLIM framework.</p>