Abstract
Text classification systems are increasingly deployed in settings where a confident but incorrect prediction carries a real cost. Aggregate accuracy on a held-out split says little about whether such a system can be trusted. This systematic review examines how reliability is pursued and evidenced across emotion classification, sentiment analysis, and AI-generated text detection. Following PRISMA, seven electronic databases (ScienceDirect, Scopus, Web of Science, ACM Digital Library, IEEE Xplore, Compendex, and PubMed) were searched together with a supplementary Google Scholar search. The search covered text-only journal articles and conference papers published from January 2016 through July 2026 that addressed a relevant NLP task and evaluated or improved at least one reliability dimension. Of 1,797 records identified, 1,294 were screened after deduplication, 464 were assessed in full text, and 51 studies met all criteria. Each study was characterized using a five-dimensional taxonomy: Task Category, Model Type, Reliability Approaches, Reliability Evaluation Evidence, and Reliability Validation. Task Category is mutually exclusive; the remaining four dimensions are non-mutually exclusive. Dataset language was recorded separately. Across the 51 included studies, AI-generated text detection accounts for 28 studies, sentiment analysis for 15, and emotion classification for 8. Encoder-Based Pre-trained Language Models (PLMs) are the most common Model Type (35 studies), followed by Decoder-Based LLMs (25) and Other Deep-Learning Architectures (18). Reliability Approaches are led by Uncertainty Estimation (15 studies), followed by Selective Prediction (9), Calibration (7), Distance and Representation Methods (7), Statistical and Likelihood Methods (5), and Consistency and Agreement Methods (3). Their distribution differs sharply by task category: Under the operational coding definition used in this systematic review, Uncertainty Estimation appears in 12 sentiment studies and 3 emotion studies, while no AI-generated-text study met the criteria for this category. All 5 Statistical and Likelihood Methods studies occur in AI-generated text detection, a concentration that partly reflects the category definition, since likelihood-based provenance statistics are themselves a detection technique. Reliability Evaluation Evidence is most often Error-Rate and Operating-Point Evidence (19 studies), followed by Calibration Evidence (14), Selective Prediction and Coverage Evidence (10), and Uncertainty Evidence (9). Reliability Validation is also uneven: 24 studies include Out-of-Distribution Validation, 21 Adversarial and Perturbation Validation, 20 Cross-Generator Generalization, 4 Cross-Lingual Validation, while 18 are limited to In-Distribution testing. The review 2 contributes a unified five-dimensional framework that makes the application, model context, reliability mechanism, supporting evidence, and validation conditions directly comparable across the three task categories and exposes where evidence remains narrower than the reliability claim it is used to support.