In recent years, the tasks of automatic sentiment detection in text have gained particular relevance due to the growing volume of user-generated content in social media, blogs, product and service reviews, political discussions, and other digital communication environments. The advent of deep neural networks has introduced new approaches to extracting complex linguistic patterns and contextual dependencies, significantly improving sentiment classification quality across a wide range of benchmark datasets. However, despite these advancements, several challenges remain in applying machine learning and neural networks to sentiment analysis. Firstly, the heterogeneity and contextual variability of sentiment-related vocabulary pose a serious obstacle. Secondly, sarcasm, irony, and implicitly expressed attitudes present a significant difficulty: standard machine learning methods often fail to effectively interpret such nuances, especially without incorporating pragmatic context. In addition, the quality of input data plays a crucial role in building reliable sentiment detection systems. Of particular note is the challenge of recognizing emotions that go beyond the simple positive/negative polarity spectrum. The objective of the article is to analyze existing machine learning methods for sentiment analysis in text in order to identify promising directions for further research. A range of existing text analysis methods applicable to sentiment detection has been reviewed. The current challenges have been identified. These are limited availability of domain-specific training data, class imbalance in sentiment categories, and restricted access to representative datasets. Directions for future research have also been outlined, including ambiguity in language, idiomatic expressions, and the expansion of detectable emotional categories. The results of the analysis can be applied in the design of systems focused on identifying authors' opinions regarding specific issues or objects, as well as emotional responses to particular events or entities. These findings may support the early stages of system design and technical requirements analysis in selecting the most appropriate methodological approach.
Presnov et al. (Mon,) studied this question.