Research On Sentiment Classification Of Network Comments Based On Deep Neural Network

With the development and popularization of the Internet,people like to share their opinions and attitudes on social networking platforms,and are also used to sharing consumer experiences and shopping experiences on online shopping platforms.The user reviews of travel websites reflect the tourists' real experience and feelings about the scenic spots.It is a very meaningful subject to conduct sentiment classification research on these review data.In the study of text classification,traditional machine learning methods have some shortcomings in artificial feature selection and model construction,while deep neural networks have good performance in mining potential information in data.This article studies two issues,one is the optimization of the deep neural network model under unbalanced data,another is the construction of a fusion model based on CNN and RNN.The raw data has a category imbalance problem,in order to improve the classification effect of imbalanced data sets,this paper reshapes the loss function of the neural network and generalizes the Focal loss to the case of multiple classifications to solve the problem that the loss of hard samples is easily overwhelmed by easy samples.Experiments prove that the model based on the new loss function performs better on imbalanced data sets.In addition,compared with the models of traditional machine learning algorithms,such as Naive Bayes and Random Forest,the performance of deep neural networks in overall Accuracy,Recall and F1 score is also more prominent.Subsequently,this paper designs a hybrid model called CNN-GRU,which combines the characteristics of CNN that are good at capturing text segment information and GRU's ability to process sequence information.CNN-GRU overcomes the shortcomings of processing all inputs at once in feedforward network and ignoring the order of words in text.It is experimentally verified that the model can greatly shorten the training time compared with the GRU model.Moreover,this model has a greater improvement in Accuracy and overall F1-score than the simple RNN model.Anyway,it can provide a basis for managers to analyze tourists' satisfaction with scenic spots,and has certain positive significance for text sentiment classification research.