Research On Text Sentiment Analysis Method Based On Deep Learning

With the development of the Internet,a large number of text data are rapidly generated.How to analyze the publishers sentiment and psychological state through these text data is an important task.The outbreak of the new coronavirus 2019 has caused a huge blow to production and life all over the country.Therefore,it has affected the hearts of all people and attracted wide attention from all parties.Analyzing and mastering the emotional text data under this topic can timely grasp the psychological status of the people and timely guide them to avoid potential social harm.It is of great significance to social management.This paper analyzes and summarizes the existing research on text sentiment analysis,and takes public opinion tendency during the epidemic as the research object,and uses deep learning method to identify the polarity of emotional texts during the epidemic.The main work of this paper is as follows:(1)In view of the lack of high-quality data sets of epidemic emotions,we first crawled the text data about the epidemic by using the crawler network,and marked and screened the data set,marked out three categories of emotions,and constructed 6000 self-built data sets with balanced distribution.Secondly,the public online data sets are also collected,and the data cleaning,new word recognition,word segmentation and stop words removal are performed on the two data sets.(2)Traditional word2vec can only extract lexical-level semantics from grammatical structures,ignoring the meaning of ’word’ as an expression.In view of this insufficiency,this article first carries on the weighted combination to ’word’ granularity word vector and ’word’granularity word vector.Secondly,word2vec is modeled based on grammatical relations,and words with opposite emotions may have similar word vectors.Therefore,the emotional feature vector of vocabulary is constructed by emotional dictionary.Finally,these different word vectors are combined to improve and generate an emotion-word representation method,which provides a basis for subsequent emotion classification tasks.(3)Aiming at the shortcomings that the traditional emotion dictionary method is not suitable for big data tasks and the text features extracted by separate convolutional neural network and cyclic neural network are not comprehensive,a two channel emotion classification model with self attention mechanism is studied in this paper.Firstly,the convolutional neural network is used as channel 1,which can extract the spatial local features of text sentences.Secondly,the bidirectional long-term and short-term memory network with self attention mechanism is used as channel 2 to extract the time series features of the text.Two channel network is used to extract text depth features.The experimental results show that the recognition rate of this method is greatly improved compared with the traditional modeling method.(4)Word vectors generated for word2vec cannot effectively distinguish emotional words.The method of adversarial training is used to add disturbance to word embedding matrix,so as to distinguish emotional word vector more efficiently.Then the bagging idea in ensemble learning is used to improve the model.Firstly,three different convolution kernel base learners are trained,and each base learner is trained by random data sets without putting back sampling.Then the three learners are combined with an improved weighted voting strategy,Obtain final classification results.The experimental results show that the ensemble adversarial training model can obtain better classification results than the single base classifier method.