Aspect-level sentiment analysis in social media using a hybrid deep transfer learning approach

Abstract

In recent years, researchers have become interested in aspect-level sentiment analysis. In the traditional sentiment analysis of documents or sentences, a label was assigned to the entire sentence or document. Whereas a sentence or document can have aspects with different sentiments. Although deep learning models have succeeded in aspect-level sentiment analysis, these models require rich labeled datasets in different domains to extract text features and sentiment analysis. This paper uses deep transfer learning for sentiment analysis of aspect-level sentiment analysis (AHDT) of social network data. The backbone of the AHDT model is a version of RoBERTa’s pre-trained deep neural network specially trained to work on social data. The features extracted from the pre-trained RoBERTa network for sentiment analysis are injected into the Bi-GRU deep neural network and then the attention layer. BI-GRU can process sequences from both sides (left to right and vice versa) and extract hidden relationships. In addition, the attention layer allows the model to pay attention to the more influential aspects of the text and provide a better interpretation. Also, this article uses the Class imbalance method to balance for training the model with almost the same polarities. The test results of the AHDT model on four SemEval datasets for the aspect-sentiment analysis task show that the model has improved the F1-score value in Resturan2014, 2015, and 2016 datasets by 0.63, 27.01, and 15.93, respectively. Also, this model has increased the accuracy value in Resturan2015 and 2016 datasets to 9.21 and 0.54, respectively. In addition, the results of experimental tests in all datasets show that the obtained values of accuracy and F1-score are close to each other, which indicates the stability of the AHDT model.