Reframing sentiment analysis in the LLM Era: a comprehensive review of methods, evaluation, and sustainability

Abstract

Sentiment analysis has evolved rapidly from lexicon-based pipelines to transformer and large language model centric systems, expanding its role across domains such as finance, healthcare, crisis monitoring, education, and public opinion analysis while exposing persistent challenges in reproducibility, calibration, fairness, operational reliability, and ecological sustainability. This survey provides a structured synthesis of an analytical and contextual reference corpus of 165 studies on sentiment analysis in the transformer and LLM era, selected under strengthened inclusion/exclusion criteria to support task, method, theory, language, domain, and evaluation coverage. This synthesis is complemented by a descriptive mapping of survey-oriented literature and a conservative reporting-prevalence audit of the 89 empirical technical papers. The PRISMA 2020 statement is cited separately as a methodological reporting reference. We make five contributions. First, we organize SA research across document-level polarity, aspect-based sentiment analysis, emotion recognition, stance, sarcasm, and conversational settings, including monomodal, multimodal, and multilingual scenarios. Second, we summarize an evaluation bundle that extends beyond headline accuracy or macro-F1 to include calibration, subgroup-wise fairness, operational metrics, and ecological indicators, while explicitly treating cross-study metric values as non-harmonized reference points rather than pooled estimates. Third, we review reproducibility and drift risks in LLM-based SA and identify reporting items such as model and version identifiers, prompt templates, inference dates, decoding settings, and operational context. Fourth, we discuss sustainability-aware reporting through latency, token cost, hardware/runtime information, energy or carbon estimates, and an illustrative sentiment-per-watt reporting ratio. Finally, we organize observed frontier methods and open challenges into a structured research agenda connecting knowledge-enriched modeling, socio-cultural robustness, continual learning, privacy-preserving training, and documented evaluation practice. Across domains and methodologies, the evidence suggests that future progress in sentiment analysis depends not only on model scale but also on documented integration of knowledge, calibration, fairness, and lifecycle-aware evaluation. The review therefore frames SA as a methodological area in which task performance, reproducibility, accountability, and sustainability need to be evaluated jointly.