Sentiment Analysis Based on Hybrid Neural Network Techniques Using Binary Coordinate Ascent Algorithm

Sentiment analysis is a technique for determining whether data is positive, negative, or neutral using Natural Language Processing (NLP). The particular challenge in classifying huge amounts of data is that it takes a long time and requires the employment of specialist human resources. Various deep learning techniques have been employed by different researchers to train and classify different datasets with varying outcomes. However, the results are not satisfactory. To address this challenge, this paper proposes a novel Sentiment Analysis approach based on Hybrid Neural Network Techniques. The preprocessing step is first applied to the Amazon Fine Food Reviews dataset in our architecture, which includes a number of data cleaning and text normalization techniques. The word embedding technique is then used to capture the semantics of the input by clustering semantically related inputs in the embedding space on the cleaned dataset. Finally, generated features were classified using three different deep learning techniques, including Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), and Hybrid CNN-RNN models, in two different ways, with each technique as follows: classification on the original feature set and classification on the reduced feature set based on Binary Coordinate Ascent (BCA) and Optimal Coordinate Ascent (OCA). The experimental results show that a hybrid CNN-RNN with the BCA and OCA algorithms outperforms state-of-the-art methods with 97.91% accuracy.