Sentiment Analysis on Book Reviews Using Convolutional Neural Network (CNN) Long Short-Term Memory (LSTM) Hybrid

Abstract

Sentiment analysis is one of the most prominent methods on gathering and analyzing insightful textual data from various sources. The information produced from such a method can be imperative in understanding the general public's sentiment on a certain product or service. Over the years, countless sentiment analysis models have already been established using known algorithms such as Naive Bayes, Support Vector Machine, and many more. However, with the advent of novel technologies and neural networking, recent studies have employed Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) Recurrent Neural Network together to formulate more efficient and modernized models (Rehman, Malik, Raza, & Ali, 2019). As such, the study proposed a similar model to analyze the sentiments of book user reviews from GoodReads categorized according to three distinct genres – children's, young adults', and romance. The paper also aimed to determine the viability and effects of amalgamating features such as Word2Vec, POS, and SenticNet to the overall accuracy (Ayutthaya & Pasupa, 2018). Once the model was trained to the procured dataset, the results suggested that combining Word Embedding, POS, and SenticNet features drastically improves its performance in contrast to other tested variations. Amalgamating the three features to a CNN-LSTM hybrid model yielded an F1-score of 90%; whilst other variants with lacking features or a standalone CNN or LSTM model only resulted to F1-scores around 86% below. Graphing the performance of all the constructed models to an ROC curve also indicated the effectiveness of the proposed model – having an AUC value of 0.9588.