Sentiment Analysis: Theoretical Foundations and Conceptual Overview 13 9

Abstract

Sentiment analysis or opinion mining is a computational and interpretive method of detecting, isolating, and quantifying emotional expressions, subjective views, and expression of affection in text or other multimodal data. This conceptual paper gives a theoretical analysis of sentiment analysis, its interdisciplinary origins in the fields of linguistics, psychology, computer science, and social sciences, without referring to particular empirical applications. The discussion describes three basic levels of sentiment analysis: document-level, sentence-level and aspect-based methods. Sentiment is considered at the document level as a holistic property and sentiment-level analysis at the sentence level in the analysis of mixed polarity in a single text; and sentiment analysis at the aspect level, is focused on opinion detail at the attribute / entity level. Sentiment is theorized in the dimensions of polarity (positive, negative, and neutral), intensity, and types of emotion (e.g. happiness, anger, sadness) based on psychological models, including Valence-Arousal-Dominance theory and emotion appraisal theories.
This paper describes the parallels between lexicon-based models, based on a set of predefined sentiment dictionaries and valence scores, and machine learning models, such as supervised, unsupervised, and deep learning models, which use contextual embeddings and transformer models. Some of the difficulties encountered comprise, sarcasm detecting, dealing with negation, domain dependency, cross cultural differences in expression of emotion, and the subjectivity nature of the sentiment tags. They point out ethical implications, including the amplification of bias in training data and the privacy aspects of a large-scale opinion mining as the problems of serious theoretical concern. Conclusively, this theoretical inquiry places the sentiment analysis as a mediating point between the emotional experience of humans and computational modelling, and it also highlights the importance of the rule-based linguistic knowledge combined with data-driven methods. This paper establishes the basis of building more context-sensitive, interpretable, and theoretically rigorous sentiment analysis paradigm in future studies by demystifying important theoretical constructs and methodological trade-offs