Research And Realization Of Emotion Analysis Model In Social Network

With the popularization of information network applications,users participate in communication,publishing,and exchanging all kinds of information through social networks.The exchange of these massive information offers possibilities for sentiment analysis.The sentiment analysis in this thesis is limited to sentiment analysis or sentiment classification for social media data.Social network data is mainly stored in the form of text.Its text is short and refined,features are sparse,and semantics are complex.The general approach to such information is to analyze words in terms of dimensions,ignoring the relationship between words in texts.This article mainly proposes two sentiment analysis methods based on the Weibo and Wikipedia English and Chinese corpus:1)According to Weibo Chinese corpus,a weighted doc2vec model is proposed based on the weighted word2vec model,which can directly train sentence vectors and omit the steps of word vector transformation sentence vector.The main idea is to combine the TF-IDF algorithm with the pre-trained word vectors,and use the TF-IDF algorithm to weight the learned word vectors,and then train the sentence vectors as the input of the doc2vec model.In the experiment,compared with the other sentiment analysis,it can be seen that this method performs more experimentally.2)For Wikipedia English corpus,this thesis introduces the relational information model based on the doc2vec model,and proposes a Relationship Information Sentence Vector Model(RISV).In the training,the doc2vec model is used to train the sentence vector,while the Relationship Information Sentence Vector Model is used to train the word relational information,which is equivalent to adding the relationship supervising information for the doc2vec model training in combination.In addition,for the relational information*The model adds the weighting parameters to balance the effect of the two models on sentence vector learning.Finally,we verify the validity of the learned sentence vectors in the tasks of document classification and short text semantic similarity.The experimental results show that the RISV model has more advantages in these two tasks than the traditional method of sentiment analysis.Finally,this thesis proposes a pre-training scheme for two sentiment analysis methods.The use of pre-training brings regularization similar to the prior distribution to the training of sentence vectors.Using RCM model as a pre-training model also supplements knowledge of relation information to a certain extent.In training,pre-training model and sentence vector training model share word vectors and parameters.In the experiment,it can be found that pre-training can improve experimental performance to some extent.However,pre-training is not effective for any task,especially for some networks with low number of layers and few nodes.