Efficient Method Based on Combination of Deep Learning Models for Sentiment Analysis of Text

People's opinions about a specific concept are considered as one of the most important textual data that are available on the web. However, finding and monitoring web pages containing these comments and extracting valuable information from them is very difficult. In this regard, developing automatic sentiment analysis systems that can extract opinions and express their intellectual process has attracted considerable attention in recent years. Sentiment analysis is considered as one of the most active research areas in the field of natural language processing which tries to classify a piece of text containing opinions based on its polarity and determine whether an expressed opinion about a specific topic, event or product is positive or negative. Since about a decade ago, many studies have been carried out to investigate the effects of traditional classification models, such as Support Vector Machine (SVM), Naïve Bayes, Logistic Regression, etc. in the task of sentiment analysis. Although machine learning models have achieved great success in this filed, they are still confronted with some limitations, notably manual feature engineering requirements. In other words, the classification performance of machine learning models is highly dependent on the extracted features and they play an important role in obtaining higher classification accuracy. To deal with these problems, deep learning models have been extensively employed as an alternative to traditional machine learning models and have achieved impressive results. It is worth mentioning that despite the remarkable performance of these methods, they are still confronted with some limitations and they are on their first steps of progress. Therefore, the goal of this paper is to propose a combinational deep learning model that can overcome their problems as well as utilizing their benefits. In this regard, an efficient method based on combination of convolutional and recursive neural networks is proposed in this paper that employs a generalized recursive neural network, where an intermediate feature is obtained by combining children's nodes, as an alternative of pooling layer in attention-based convolutional neural network with the aim of capturing long term dependencies and decreasing the loss of local information. Based on empirical results, the proposed method with the accuracy of 53.92% and 92.89% respectively on SST1 and SST2 datasets not only outperforms other existing models but also can be trained much faster.