Sentiment Classification of Opinions based on Multi-source Transfer Learning Using Structural Correspondence Learning

Sentiment classification of opinions is a field of Natural Language Processing which has been considered in recent years by researchers due to popularity of Internet stores and the possibility of expressing opinions about sold goods or services. To train classifier models, we need labeled datasets, but as there are not rich labeled samples and as labeling is a difficult and time-consuming process, we must employ labeled samples of other domains. In this article, a new method for binary classification of opinions is proposed based on multi-domain transfer learning. The proposed method tries to adapt different domains by using Structural Correspondence Learning; and based on repetitive procedure of the boosting algorithm, a weight is assigned to classified samples of different domains and the class of each opinion is specified by merging these classifiers. Weighting the dataset samples to boost the process of classification based on the Adaboost algorithm and combining it with the Structural Corresponding Learning is the most important innovation of the current research. The Amazon dataset of four different domains, each one containing 1000 positive and 1000 negative opinions is used for training the proposed model. Accuracy measures of %89.64, %93.97, %92.39 and %90.17 are obtained for Electronics, DVD, Books and Kitchen domains, respectively. It illustrates that the proposed method is very effective compared with the similar methods.