Abstract

Today, there is an abundance of opinions regarding products, services, institutions, and public events on online platforms that cannot be analyzed by any team of readers. Sentiment analysis solves this problem by classifying a given piece of text as positive, negative, or neutral. Machine learning algorithms are used for creating such classification models. This paper is a theoretical literature review of machine learning applications to sentiment analysis of text. It describes the standard process of text preprocessing (tokenization, stop-word removal, stemming, lemmatization, Bag-of-Words and TF-IDF), and gives the mathematical foundation of four popular text classifiers: Naive Bayes, Logistic Regression, Support Vector Machine and Random Forest. The reviewed studies are contrasted according to the representation of text, the classifier used, and the limitations mentioned. The trend to use deep learning and pre-trained transformers is described. The evidence from social media, e-commerce, and customer support is summarized with its limitations. Four persistent challenges of sentiment analysis are discussed: dependency on context, sarcasm and figurative language, multilingual and code-mixed texts, and noisy unstructured text. No machine learning model was created, and no experiment was conducted; thus, no performance results are given. The review concludes that classical models still hold their ground as cost-effective baselines for large and well-formed corpora, that transformers deal better with context at a higher computational cost, and that there is no solution that works well in all four cases.

Keywords

How to Cite This Article

Rights