Abstract. Political communication has migrated to social platforms, where opinions are expressed through short, code-switched, and rhetorically dense messages. Conventional sentiment analysis systems, tuned for product reviews, transfer poorly to this domain because political polarity is entangled with stance, irony, and discourse structure. This article surveys the principal families of sentiment analysis methods—lexicon-based, classical machine learning, deep neural, and transformer-based approaches—and introduces PoliSent, a hybrid discourse-aware architecture for political sentiment analysis on the Internet. PoliSent unifies three complementary signals: multilingual contextual embeddings produced by an XLM-RoBERTa encoder, a domain-specific political lexicon (PoliLex) with valence shifters and negation handling, and a graph-attention module operating over Universal Dependencies parses. A learned gating mechanism fuses these signals, and a multi-task objective jointly predicts sentiment, stance, and rhetorical tactic. We evaluate the framework on PoliWeb-CA, a newly annotated corpus of 18,500 Kazakh- and Russian-language political messages collected from microblogs, messaging channels, and news comment threads. PoliSent attains a macro-F1 of 0.904 and a Matthews correlation coefficient of 0.857, surpassing a fine-tuned KazRoBERTa baseline by 4.6 macro-F1 points and outperforming all lexicon, classical, and deep baselines by a wide margin (p < 0.01, paired bootstrap). Ablation studies confirm that each architectural component contributes measurably, with the dependency-graph attention and lexicon augmentation jointly responsible for a 5.8-point improvement over the contextual backbone. The results demonstrate that explicit linguistic structure remains valuable even in the era of large pretrained encoders, particularly for low-resource and morphologically rich political discourse.

Keywords: sentiment analysis, political discourse, opinion mining, natural language processing, transformer models, graph attention networks, stance detection, low-resource languages, Kazakh; multilingual NLP.