"Sentiment Analysis: Four-Class Classification Models (Positive, Negative, Neutral, and Sarcastic) and Artificial Intelligence Challenges"

Abstract

 Sentiment Analysis (SA) is a fundamental application of Natural Language Processing (NLP), historically relying on a three-way classification (positive, negative, neutral) to evaluate public opinion trends. However, sarcasm poses a critical challenge to these systems due to "polarity inversion," leading to misleading classifications that rely on the superficial lexical tone of the text while ignoring the true intent. This paper presents a systematic literature review of recent studies to explore methodologies for overcoming this challenge, tracing the methodological evolution from traditional mechanisms to Transformer-based architectures and Multi-Task Learning (MTL) strategies. The findings demonstrate the superiority of models like RoBERTa in English (achieving an F1-score of up to 99%), while Arabic models like AraBERT recorded an accuracy approaching 84%. Furthermore, MTL frameworks proved highly effective in enhancing sarcasm detection performance through auxiliary tasks. Concurrently, the paper highlights the research gap in low-resource languages, particularly Arabic, with its morphological complexity, dialectal diversity, and non-standard writing phenomena (Franco-Arabic). The paper concludes with actionable recommendations centered around the urgent need to build rich, manually annotated Arabic datasets and employ advanced contextual models to achieve a deeper understanding of sarcastic texts.