Text Sentiment Analysis Based On Multiple LSTM Structures

With the rapid development of the internet,more and more users are posting their comments and remarks on the internet,which express their emotional tendencies.Businesses must fully understand users'comments and remarks on their goods and services to improve products quality and marketing strategy.However,there are so many comments that traditional ways based on artificial questionnaire survey are unable to match these needs.Hence,it has been an active research subject to automatically extract users ' emotional tendencies from texts.In this paper,we establish two text sentiment analysis models to classify Chinese text sentiment based on Long Short-Term Memory(LSTM)structure and one of its variants:Gated Recurrent Unit(GRU)structure,respectively.The traditional Recurrent Neural Network(RNN)can theoretically process sequential data of any length.However,in practice,as the interval between the relevant information and the current prediction position increases,RNN will lose the ability of learning long distance information,due to the gradient disappearance and gradient explosion problems,which is overcome by the LSTM structure.We also propose a pseudo gradient descent method for model hyper-parameters adjustment.Common approaches for hyper-parameters adjustment include grid search and random search.However,the grid search is highly inefficient and the random search introduces great uncertainty.Numerical experiments show that our pseudo gradient descent method can achieve high accuracy in shorter time.