Comparative Study on Sentiment Analysis Using Advance Deep Learning Models

Sentiment Analysis (SA) involves collecting opinions regarding a particular subject and classifying them based on emotional content; it refers to opinion mining, which examines product reviews and online feedback. Capturing relationships of long sequences, contextual meaning, and effectively achieving Cross-Lingual Sentiment Classification can be considered key challenges of the SA. This paper performs a comparative analysis of traditional sequence models, such as Gated Recurrent Unit (GRU) and Bidirectional GRU (Bi-GRU), and state-of-the-art transformer-based models, including Bidirectional Encoder Representations from Transformers (BERT) and XLM-RoBERTa (XLM-R). The traditional sequence models are enhanced with pre-trained Word2Vec embeddings. Firstly, all the models are trained and evaluated using the English IMDB dataset. Then, the XLM-R is assessed on the Spanish IMDB dataset without further fine-tuning. The results indicate that both BERT and XLM-R surpass GRU and Bi-GRU in sentiment classification accuracy. XLM-R obtained the best results, with an accuracy of 93.21%, slightly outperforming BERT, which achieved an accuracy of 93.11% in English. XLM-R was also utilized for zero-shot cross-lingual sentiment classification, achieving a sentiment accuracy of 93.12% in Spanish, close to its performance in English.