Semi-supervised learning with generative model for sentiment classification of stock messages

Classification of investors' sentiments in stock message boards has attracted a great deal of attention. Since the messages are usually short, we propose a semi-supervised learning method to make full use of the features in both train and test messages. The generative emotion model takes message, emotion and words into consideration simultaneously. Based on the facts that words are of different ability in discriminating sentiments, they are categorized into three classes in the model with different emotion strength. Training the generative model can transform the messages into emotion vectors which finally feeds to a sentiment classifier. The experiment results show that the proposed model and learning method are efficient for modeling sentiment in short text, and by properly selecting the amount of train data and the percent of test samples, we can achieve higher classification accuracy than traditional ones. The results indicate that the generative model is effective for short message sentiment classification, and provides a significant approach for the implementation of semi-supervised learning which is a typical expert and intelligent information processing method. (C) 2020 Elsevier Ltd. All rights reserved.