Sentiment forecasting based on past textual content and deep learning architectures

Abstract

In social networks, detecting opinions or sentiments by giving textual content has become a classical task. Usually, the concerned methods take as input a set of words in order to return the corresponding sentiment polarity, which can be called as a posteriori prediction. In this paper, we propose an a priori prediction model. Its objective is to predict the sentiment polarity of a user about a well-defined topic based only on textual history. To this aim, we suggest a deep learning-based approach which essentially depends on two deep learning architectures namely convolutional neural network (CNN) and long short-term memory (LSTM). To test our system, we have purposely created a collection from SemEval-2017 data. The results reveal that our deep learning method outperforms our previous approach which depends on sentiment intensity features with traditional machine learning algorithms. In addition, it outperforms the approach based on word embedding features with also traditional algorithms. On the other hand, the performance of our model is relatively close to that of ideal approaches (a posteriori prediction) that deal directly with texts carrying the sentiment.