Sarcasm Detection Using Learning Techniques in a Social Engineering Context: A PRISMA-Based Systematic Review
Sarcasm—the deliberate expression of a meaning opposite to the literal content of an utterance—is pervasive in social-media communication and constitutes a major source of error for sentiment analysis, opinion mining and manipulation-detection systems deployed in social-engineering contexts. Because sarcasm inverts surface polarity, its automatic detection remains a hard, context-dependent natural-language-processing (NLP) problem. This paper presents a systematic review conducted in accordance with the PRISMA 2020 guidelines. Eight bibliographic databases were searched for the period 2010–2026; from 1,320 identified records, a four-phase identification–screening–eligibility–inclusion process yielded 40 primary studies for synthesis. Beyond the review protocol, the paper provides an extended theoretical foundation covering pragmatic theories of irony, text representation, classical machine-learning classifiers, recurrent and convolutional architectures, the attention and Transformer formalism, pre-trained language models, multimodal fusion and evaluation metrics. The synthesis is organised around four research questions addressing (RQ1) benchmark datasets and social-media text sources, (RQ2) NLP pre-processing techniques, (RQ3) machine-learning and deep-learning models, and (RQ4) transformer-based and context-aware models, with a dedicated comparison at each question and a per-study questionnaire-style extraction form. Results reveal a clear methodological progression from feature-engineered classifiers, through CNN/RNN and attention hybrids, to transformer and multimodal graph architectures, with contextual embeddings consistently improving performance. Persistent gaps include weak modelling of extra-utterance context, noisy distant-supervision labels, severe class imbalance in intended-sarcasm corpora, limited cross-domain and code-mixed generalisation, and poor interpretability. On this basis we outline a precision-aware, context-integrating sarcasm-prediction algorithm and an evaluation protocol for validation against the state of the art.