A Thorough Investigation of Conventional and Advanced Feature Extraction Methods in Natural Language Processing for Sentiment Analysis Employing SVM Kernels

The development of numerous feature extraction techniques, ranging from conventional to advanced, targeted at improving the accuracy of sentiment categorization, has been sparked by the rapid advancement of Natural Language Processing (NLP). In this study, we conduct a thorough investigation into the field of feature extraction with the aim of elucidating the subtleties of their performance in the context of sentiment analysis. To accomplish this, we carefully conduct a series of experiments using three conventional and three advanced feature extraction approaches, leveraging the reliable Benchmark IMDB dataset. Our investigation includes more advanced strategies like Word2Vec CBOW, Word2Vec Skip-gram, and Doc2Vec in addition to more conventional ones like Count Vectorization, Bigram Representation, and TF-IDF. The SVM classifier is used in conjunction with these feature extraction techniques, which make use of four different kernels—Linear, RBF, Poly, and Sigmoid. Our main goal is to assess how well different methods work using accuracy as the main parameter. In light of various kernel settings, this enables us to evaluate the efficiency of every feature extraction method. Notably, our research demonstrates that TF-IDF is a very adaptable performer, constantly outperforming all four kernels. Additionally, among the advanced methods, the Word2Vec Skip-gram stands out as a top performer.