Three-Way Decisions Based RNN Models for Sentiment Classification

Recurrent neural networks (RNN) has been widely used in sentiment classification. RNN can memorize the previous information and is applied to calculate the current output. For sentiment binary classification, RNN calculates the probabilities and then performs binary classification according to the probability values, and some emotions near the median are forcibly divided. But, it does not consider the existence of some samples that are not very clearly polarized in sentiment binary classification. Three-way decisions theory divides the dataset into three regions, positive region, negative region, boundary region. In the process of training classification, the probabilities of some samples belonging to different categories are very close, and three-way decisions can divide them into the boundary region by setting thresholds. Reasonable processing of the boundary region can get better results for binary classification by adjusting the probability of samples in the boundary region. Therefore, in this paper, we propose three-way decisions based RNN models for sentiment classification. Firstly, we use basic RNN models to classify the data. Secondly, we apply three-way decisions theory to set the thresholds, divide the boundary region based on probability. Finally, the probabilities of samples in the boundary region are adjusted and applied in the next round of training. Experiments on four real datasets show that our proposed models are better than corresponding basic RNN models in terms of classification accuracy.