IMPROVING PRODUCT REVIEW SENTIMENT ANALYSIS VIA WORD-ALIGNED ATTENTION POOLING

Abstract

The product review gives critical data for both businesses and consumers, offering insights needed before buying a service or product. However, the existing methods has drawback of there is not understanding semantic relationship among adjacent characters. To overcome these limitations in this research proposed a Multi-source Word-aligned Attention Average Pooling based Bidirectional Encoder Representation from Transformers (MWAAP-BERT) model for product review sentiment analysis. The dataset utilized for the research is Amazon product review dataset and the data is per-processed by several methods. The Term Frequency-Inverse Document Frequency (TF-IDF) and Skip N-gram methods are used for feature extraction and proposed MWAAP-BERT model is used for product review sentiment classification. Performance of proposed method is estimated with performance measure of accuracy, precision, recall and f1-score. Proposed technique attained high accuracy of 97.5 %, precision of 96.5%, recall of 96.5% and f1-score of 96.5% which is superior than other existing methods like Recurrent Neural Network (RNN), Taylor-Harris Hawks Optimization driven Long Short-term Memory (THHO-BiLSTM) and Convolutional Neural Network- Long Short-Term Memory (CNN-LSTM).