Evaluation of Polarity Features for Sentiment Analysis on Product Reviews

Abstract

Customers express their opinions about products online, which influence potential buyers. This feedback is valuable for manufacturers to enhance their products. Sentiment analysis, which categorizes sentiments as positive, negative, or neutral, is crucial but challenging. Despite recent advancements, several research gaps persist. Firstly, prior studies have explored polarity features independently or in partial combinations, lacking comprehensive evaluation across annotated datasets. Secondly, most methods classify sentiments as merely positive or negative, overlooking nuances like intensity levels (e.g., strong positive, weak negative). Lastly, existing approaches often employ diverse classifiers on disparate datasets, lacking standardized comparison. To address these gaps, this research examines adjective, adverb, and verb polarity features both independently and in various combinations (Adjective-Adverb, Adjective-Verb, Adverb-Verb, and Adjective-Adverb-Verb). The findings demonstrate that adjectives can accurately classify sentiments into seven intensity levels. Notably, the Naïve Bayes classifier achieves high precision (0.984) when utilizing adjectives alone and (0.981) when combined with adverbs, outperforming other classifiers across six evaluated combinations.