Performance Comparison of Deep Learning Text Embeddings in Sentiment Analysis Tasks with Online Consumer Reviews

In order to investigate the effect of various natural language processing models on different data processing, this paper adopted the consumer reviews of two well-known Internet retailing websites: Yelp and Zappos, and used four text embedding methods: word2vec, Glove, BERT, and GPT-2 and two text classification methods: SVM and Neural Network (NN) for text classification, in order to compare the performance of the combinations of these text mining techniques. The result shows that BERT is the best-performing text embedding method overall in both datasets when used with both SVM and NN. It is also found that NN is better than SVM for overall text classification. As an exploratory experiment, we aim to provide a three-dimensional comparison to find the most suitable algorithm for consumer review data, and the implication is that BERT and NN can achieve satisfactory results in most of the scenarios.