Research On Emotional Tendency Of Weibo Comments Based On Deep Learning

In the modern Internet era,Weibo has become the main platform for modern netizens to express their views,opinions and emotions,and has taken up a large proportion in social media.Most of the comments on Weibo show netizens' emotional state towards an event,phenomenon or product.What kind of algorithm and processing method can be used to analyze these Weibo comments information more quickly and accurately,to timely obtain the emotional tendency of Weibo comments on a certain topic or event,and to obtain the tendency of topic public opinion has also become a research hotspot in the field of Natural Language Processing(NLP).In general,traditional affective analysis methods often use statistical features of texts or sentiment dictionary based method to represent a piece of text sentence.This method has the disadvantages of not being able to obtain semantic information of sentences and high classification error rate.However,with the increasing amount of data on the Internet,it is increasingly difficult to extract text features by statistical methods.Meanwhile,sentiment dictionary based sentiment analysis method(only counting emotion words)is relatively simple and has low robustness.And the traditional statistical machine learning method using the statistical characteristics of the text training,can not learn the semantic information of the text well.In order to solve the above problems,an improved deep learning algorithm based on word embedding to analyze sentiment of Weibo comments is proposed in this thesis.The main innovations of this thesis are as follows:(1)The neural network language model is used for training on a larger scale of corpus data to learn the hidden semantic features of words in a more convenient and unsupervised way.The word embedding vector is used to replace the traditional method of extracting text features,and then the features of this thesis are used for deep learning model supervised training;(2)In order to capture more semantic features of text during the training of deep model and complete the task of text sentiment analysis more effectively,attention mechanism that can change the weight of sequence information is introduced into the deep model.RNN-att,RCNN-att model based on attention mechanism are proposed;(3)In order to capture the local features of each part of the statement and the position dependence information of the statement at the same time,a C-RNN model with parallel hierarchy is proposed.While extracting the local features by CNN module,remember the long term position dependency by RNN module at the same layer of the model.In order to verify the effectiveness of the proposed method,this thesis trains language model to get word vector features by Word2Vec,and uses the statistical model,Random Forest,GBDT,SVM,Gaussian Naive Bayes and deep model TextCNN,RNN,RCNN and C-RNN to analyze sentiment of Weibo comments based on word vector.In this paper,420k comments on Weibo and other supplementary Chinese data sets are used as the training corpus of the language model.35k comments data were selected from the training corpus set and labeled as data set,which were divided into training set and test set at a ratio of 9:1.Accuracy,recall rate and fl-score were used as evaluation method for performance evaluation A comparison experiment was conducted between multiple models on the above data set.According to the experimental results,in statistical models,GBDT using word vector with dimension of 300 obtained good performance with precision of 0.8417,recall of 0.8416,and F1-Score of 0.8416.In deep models,RCNN-att using 200 dimensions word vector got the best performance with precision of 0.9266,recall of 0.9266,and F1-Score of 0.9266.C-RNN-att won the best performance in the group with F1-Score of 0.9153 when word vector dimension is 50.The experimental results show that the neural network model language can better learning text of semantic information,and deep models based on word embedding text feature in sentiment analysis task are better than statistical models.Furthermore,deep models can get 2‰?2%improvement by introducing mechanism.The improved depth algorithm studied in this thesis can be trained in a larger data set,and can further improve the reasoning ability of the algorithm in a higher dimensional language model.