The Technological Evolution of Sentiment Analysis—A ComparativeStudy from SVM to Large Language Models

Abstract

As the ecosystem encompassing social media and product reviews grows ever more intricate, emotional expression presents prominent traits including subtlety, sarcasm, fragmentation and multimodal fusion. Traditional machine learning models (e.g., SVM), which rely on manual feature engineering, encounter bottlenecks in recognition accuracy when dealing with irony, metaphor, and long-distance sentiment dependencies. Sentiment analysis of reviews is thus trapped in the dual predicament of "semantic noise" and "shallow understanding." This paper focuses on the advantages and accuracy verification of Large Language Models (LLMs) in tackling high-difficulty sentiment analysis of reviews. This study abandons the mere enumeration of single accuracy values and instead delves into the cognitive breakthroughs of LLMs across three key dimensions: context-aware ambiguous meaning resolution, which deeply interprets ironic and euphemistic connotations; fine-grained sentiment element extraction, accurately identifying the polarity of praise or criticism toward specific product attributes; and closed-loop verification through sentiment generation and explanation, providing traceable justifications via chain-of-thought mechanisms.