Abstract
Text classification has evolved from machine learning to deep learning and pre-trained language models, but the relative merits of these paradigms remain dependent on task settings and deployment constraints. This study conducts a comparative review of 15 representative studies published between 2023 and 2025, examining machine learning, deep learning, and pre-trained language model approaches for text classification from the perspectives of accuracy, efficiency and interpretability. Based on a systematic literature screening process and multi-dimensional analysis, the review finds that no single paradigm is universally optimal across all scenarios. Machine learning methods remain competitive in small-scale, resource-constrained, and highly interpretable tasks. Deep learning models offer stronger semantic representation and are more suitable for medium-scale or noisy text classification problems. Pre-trained language models generally achieve the best performance in complex semantic and domain-specific tasks, but they also involve higher computational costs and weaker interpretability. The findings suggest that model selection for text classification should move beyond accuracy-centered evaluation and instead adopt a task-oriented balance among performance, efficiency, transparency, and practical deployability. Future research should focus on lightweight, interpretable, robust, and domain-adaptive models to support trustworthy text classification applications.
Recommended Citation
Meng, Yu; Cui, Xiling; Liu, Libo; Yuen, Man Ching; Yu, Xinyi; and Song, Yida, "Balancing Accuracy, Efficiency and Interpretability: A Comparative Review of Text Classification Paradigms" (2026). ICEB 2026 Proceedings (Beijing, China). 17.
https://aisel.aisnet.org/iceb2026/17