Research On Several Issues Of Text Sentiment Classification Based On Deep Learning

With the development and popularization of the Internet,more and more people begin to express various opinions on platforms such as microblog,forums,social network,television and shopping websites,etc.to share their moods and opinions.Meanwhile,these published contents may contain different emotion just like positive or negative,supporting or opposing.Analyzing the emotional tendencies of these text messages is of great significance to individuals,businesses,governments,etc.In addition,with the rise of artificial intelligence,machine learning and deep learning methods have received widespread attention.More and more researchers have proposed many practical and effective text sentiment classification models,and the deep learning models have gradually become an important method to solve the problem of sentiment classification because of its strong feature learning ability.Besides ignoring some emotional resources and useful features,existing deep learning models also ignore the way of network optimization.Therefore,there is still some methods to improve the performance of sentiment classification.Based on this,this paper mainly develops two aspects of work:(1)This paper proposes a feature fusion model based on the fully connection network to solve the problem that the sparsity of short comment texts and the insufficient utilization of features in deep learning network.Firstly,this paper extracts the syntactic features of the comment text based on the dictionary,mainly including the basic features and the polarity features of the emotional phrase,the basic features are the number of positive and negative emotional words,degree adverbs,negative words,turning words,etc.;Secondly,extracting the the lexical features of emotional words' weight based on the emotional dictionary and TF_IDF.At the same time,making corresponding rules according to the location of emotional words and expanding the weight of emotional words to get the final TF_IDF value,which is used as the lexical features of comment text;Then,extracting the word vector features of comment text by word2vec;Finally,normalize the above three categories of features and connect the three features through full concatenation so that they can be fully learned in the back propagation.(2)In order to solve the problems of slow convergence and weak robustness caused by randomly setting network weights and hidden unit thresholds in deep learning models,an improved quantum particle swarm optimization algorithm is proposed to optimize the BGRU.This method optimizes and improves the quantum particle swarm algorithm while using quantum particle swarm to optimize the network weights and hidden unit thresholds of the input layer of the network.First of all,this paper proposes an adaptive quantum particle swarm algorithm to solve the problem that the low efficiency of quantum particle swarm.Secondly,in order to improve the robustness of the model and the generalization ability of the network,this paper considers the input and output sensitivity information during the network training process.Finally,the experimental results on the sentiment classification of movie reviews verify the effectiveness of the proposed method.