Solving sentiment uncertainty using newly proposed sentiment similarity measure for single-valued neutrosophic sets

The accurate detection of sentiment polarity in social media content is a persistent challenge, particularly in the absence of large labeled datasets and inherent linguistic ambiguity. Existing machine learning and fuzzy set-based approaches often struggle to capture the indeterminate and complex nature of real-world sentiment. In this work, we introduce a novel similarity measure for Single-Valued Neutrosophic Sets in a three-dimensional sentiment space, enabling a more expressive and robust representation of positive, negative, and indeterminate sentiment components Building upon this foundation, we propose SentiSim, an unsupervised and domain-neutral sentiment analysis framework that leverages the Grey Wolf Optimizer to identify optimal sentiment prototypes and enhance classification performance. SentiSim processes textual data using lexicon-based mapping to neutrosophic vectors, optimizes sentiment cluster centers via metaheuristic search, and classifies sentiment by maximizing similarity to these centers. Extensive experiments on benchmark datasets including SemEval 2017, Sentiment140, and Apple X demonstrate that SentiSim consistently outperforms state-of-the-art methods, with statistical analyses confirming the significance and robustness of its improvements. The proposed framework enhances the interpretability and reliability of sentiment analysis under uncertainty and indeterminacy.