Weibo Comments Sentiment Analysis Based On Deep Learning

With the rapid development of mobile Internet,the enthusiasm of Internet users participating in the discussion of social hot spots is constantly rising.Weibo has become an important platform for Internet users to express their views and feelings,On the other hand,the social networks based on Weibo largely reflect the social behavior and emotional tendencies of people.How to quickly dig out the hidden emotional information in Weibo and provide effective supplementary information for government and enterprise decision-making is becoming the intensive research in the field of Natural Language Processing.The traditional sentiment analysis method takes much more time to extract features from the corpus,and it often needs to be combined with grammatical rules in order to achieve better results.However,in the era of big data,it is difficult to extract features manually from the massive data character.This paper proposes a method to extract emotional information by using the word vectors and deep learning model.In this method,the unsupervised Word2vec and Glove models are used to train the data into word vectors and the word vectors will replace artificial features that extracted from data.This method can save the manpower and the deep learning model can be used to learn the emotion information of the word vectors automatically.Finally,through the contrast experiments,we can know that the deep learning model can achieve better results in sentence level sentiment analysis task.In this paper,we use Word2vec and the Glove to train Weibo comment data into two kinds of word vectors and put the word vectors into the shallow learning models,SVM,Logistic Regression,Naive Bayesian,and deep learning models,LSTM,CNN and LSMT+CNN respectively,and then the shallow learning models and deep learning model can learn the emotional information hidden in the word vectors.Finally,the results of classification can be obtained,according to the experimental results,we can get the accuracy,recall and other performance evaluation value.Among them,the highest accuracy rate of shallow learning model is close to 78.1%,and the highest accuracy rate of deep learning model is close to 84.5%.By comparing the experimental results we found that,compared with the shallow learning models,the LSTM can store long distance information and CNN can extract features from different dimensions,these functions can better dig out theemotional information hidden in the word vectors.While the shallow learning models lose semantic information between words when it is mining the emotional information hidden in the word vectors,which is a core reason for the performance degradation of shallow learning models.Compared with the Word2vec,the Glove can use global statistical information to store more emotional information into the word vectors.However,the Word2vec can only use local information,therefore,the effect of the Glove is better than the Word2vec.