Sentiment analysis has emerged as a prominent topic of research within the domain of natural language processing. The advancement of sentiment analysis techniques, particularly those based on dictionaries, has facilitated deeper insights into the sentiments expressed within textual data. Sentiment analysis uses a set of computational operations to identify the sentiment expressed in segment of words. By employing dictionary-based sentiment analysis techniques, researchers can automatically ascertain the polarity (positive, negative, or neutral) of textual content. This study presents a novel strategy for Semantic Analysis of Sentiment in social networks, which combines Lexicon-based methodology with semi-Supervised learning in order to improve sentiment analysis performance (SASLS). SASLS is to detect the polarity of words from twitter using personalized feature selection-based clustering and dictionary-based techniques. Our strategy can well deal with two common challenges in this problem, including the omnipresence of domain-specific vocabulary and the lack of labeled data in different domains. The proposed strategy has been evaluated on several datasets with different scales. Numerical findings show that SASLS significantly outperforms traditional supervised, unsupervised, semi-supervised, and deep learning approaches. Specifically, SASLS provides more than 2.5% more optimal Macro-F1 compared to the best existing state-of-the-art method. These results show that SASLS has good potential for semantic analysis of sentiments in social networks. © 2025 The Author(s)