Neural Network Models Incorporating Sentiment Information For Short Text Sentiment Classification

With the rapid development of the Internet,the era of information explosion has arrived.Humans are surrounded by various kinds of information and everyone has become accustomed to this.Humans begin to find effective ways to use this information.For example,in the online shopping,in addition to the introduction of the product,we also refer to the buyers'evaluation;when going to the cinema to watch a movie,we will make a choice according to the film's review.The great value hidden behind of the information is the use of sentiment expressed by the text.Thus,text sentiment analysis which is a task of natural language processing has been proposed.The purpose of text sentiment analysis is to automatically mine the emotions expressed by the text.There are many methods,from the classification based on rules and lexicon to traditional machine learning such as Bayesian classification,decision trees,support vector machines(SVM),and now deep learning,such as convolutional neural network(CNN),recurrent neural network(RNN),long short-term memory(LSTM),recursive autoencoders,etc.Various network models of deep learning have got certain achievements in the sentiment analysis task.Compared with traditional machine learning,the accuracy rate is improved and the generalization ability is enhanced.However,there are still many shortcomings.As a result,it is far from satisfying people.This paper deal with the neural network models incorporating sentiment information applied to short text sentiment classification.The specific works are as follows:Sentiment knowledge such as sentiment lexicon,negation words,intensity words and conjunction words all have prior information,which is an important part of sentiment analysis in the early stage of natural language processing.However,in deep learning,sentiment knowledge is rarely seen,which wastes the prior information.Therefore,this paper attempts a way to combine neural networks with the sentiment information to improve the accuracy of sentiment classification.Recursive autoencoders and tree-structured long short-term memory(Tree-LSTM)incorporate syntactic information,which greatly improves the accuracy of the model.However,these models require expensive phrase-level annotations,which limits their application.This paper attempts to combine tree-structured long short-term memory(Tree-LSTM)with sentiment information,so that the combination of sentiment information and syntactic information could replace the phrase-level annotation and improve the accuracy of the tree-structured model on the sentence-level annotated datasets.The neural network models incorporating sentiment information will be verified on two English public datasets and compared with various neural networks to prove the advantage of the models.Although deep learning has developed rapidly,the application about special populations is still insufficient.Therefore,the models proposed in this paper will be applied on the messages of prisoners.The messages contain the potential emotions of the prisoners.The sentiment analysis of the messages helps to improve the prison guards' understanding of the prisoners,thus enabling the prison guards to help the prisoners more effectively.