Research On Emotional Classification Based On Short Text(Sentence Level)

Text sentiment classification,also known as opinion mining or sentiment orientation analysis.Since the beginning of the 21 st century,emotional classification has developed into a research hot spots in the field of natural language processing.In this field,according to the different granularity of processing text,it is roughly divided into three levels: chapter level,sentence level,aspect level.This paper focuses on short texts,namely the study and analysis of sentiment classification at the sentence level.In today's society,with the continuous development of Internet technology,various e-commerce platforms and social networks have also developed rapidly.More and more people like to express their opinions on the Internet,or target a social hot spots or online shopping products.Usage experience,etc.Among them,Weibo has quickly become the birthplace and concentration of domestic network public opinion because of its simplicity,convenience,and high real-time information sharing.Internet public opinion is closely related to the public's life,and it also affects the stability and development of society.By analyzing and researching the network public opinion,it is possible to prevent major events from happening and help the government make decisions accurately and quickly.Therefore,this paper conducts an emotional orientation analysis on Weibo to better monitor network public opinion and provide support for government decision-making.In this paper,a microblog data set with a length of no more than 140 bytes is selected to improve the traditional small batch gradient descent algorithm.A training batch cycle change strategy based on hot restart and cosine annealing is proposed.The method is not monotonous or Randomly change the batch_size,but make the batch_size cycle between reasonable boundary values.Training using loop batch_size instead of fixed values can be used to accelerate model convergence,improve model accuracy,and improve model diversity.Then based on the Sigmoid function and the ReLU function,a new activation function SReLU is proposed to reduce the offset value and gradient disappearanceproblem.Finally,based on TextCNN,the short text of microblog is studied by SGDR algorithm and SReLU function,and compared with various neural network models and activation functions,and finally more accurate experimental results are obtained.First,this article uses Google's word2 vec to build and extract the word vector model used in this article.First of all,this paper uses the basic neural network model to set the convolution core channel number,convolution kernel width,learning rate and other text convolutional neural network parameters.The best model for this dataset is obtained.Based on this model,the proposed training batch cycle change strategy SGDR based on hot restart and cosine annealing is validated,and then the validity of the activation function proposed in this paper is verified by comparing other activation functions.Finally,combined with the method and activation function proposed in this paper,the experiment is carried out,and compared with other neural network models,the macro precision rate is91.66%,the macro check rate is 96.24,the accuracy rate is 97.41%,and the macro F1 value is 93.21 % of the experimental results.