This study comprehensively examines the interactions among predictive performance, explainability, and computational efficiency in sentiment analysis and empirically analyzes the trade-offs among these factors. While previous sentiment analysis studie...
This study comprehensively examines the interactions among predictive performance, explainability, and computational efficiency in sentiment analysis and empirically analyzes the trade-offs among these factors. While previous sentiment analysis studies have primarily focused on improving predictive accuracy, real-world applications additionally require explainability to interpret model decisions and computational efficiency under resource-constrained environments. Recent advances in large language models and foundation models have expanded the applicability of sentiment analysis; however, predictive performance and explanation consistency may vary depending on prompt design, model architecture, explanation methods, and lightweight strategies. Accordingly, this study integrates prompt design, model architecture, explanation methods, contrastive explanation approaches, and representation compression-based lightweight techniques into a unified analytical framework to systematically elucidate the relationships among performance, explainability, and efficiency in sentiment analysis models. To achieve this objective, sentiment analysis performance and explanation characteristics were comparatively analyzed across different prompt designs and model architectures. In addition, a contrastive explanation method, CCCE (Contrastive Contextual Contribution Explanation), was proposed to evaluate its effectiveness in improving explanation stability and discriminability. Representation compression-based lightweight techniques were also applied to analyze the relationship between efficiency and explainability. The results indicate that predictive performance and explanation consistency vary according to model types and prompt design strategies. Although lightweight approaches can improve computational efficiency while minimizing predictive performance degradation, explainability appears to be sensitive to subtle changes in internal representations. In contrast, CCCE provided relatively stable explanations even in sentiment classification environments characterized by high uncertainty. Overall, this study extends beyond conventional approaches that evaluate performance, explainability, and efficiency independently by providing an integrated analysis of the interactions and trade-offs among these factors in sentiment analysis. Furthermore, by empirically demonstrating the effects of prompt design, model architecture, explanation methods, and lightweight strategies on predictive performance and explanation reliability, this study offers practical implications for the design and selection of sentiment analysis systems that jointly consider predictive accuracy, explanation reliability, and computational efficiency in real-world applications.