Emotional Guidance Of Microblog Text Based On BERT

With the rapid development of social networks and the popularity of mobile network devices,the number of Internet users has increased significantly,and social media has gradually enriched and diversified.People are used to expressing their ideas and opinions through interactive data such as text,image,and video on various social media.Among many social media platforms,Sina Weibo has a wide range of loyal users.By analyzing the massive short text data generated by Sina Weibo social network,this paper examines the user's emotions and emotions behind the text.It clarifies the local environment of social networks that breeds negative public opinion,which has positive significance for China's ideological and political,public emergency response,public emotion prediction,public opinion dissemination prediction,and public opinion disposal.This study divided the research topic into two parts:short text sentiment analysis,and text-based public opinion guidance scheme.First,this paper takes Sina Weibo's historical short text data as the research object,explores its potential user emotion,and uses a deep learning language model to achieve a short text emotion classification task.The traditional affective analysis mainly includes dictionary-based and simple deep learning-based methods.Some problems appear in traditional methods.For example,the emotion classification method based on the dictionary depends on the quality of the constructed dictionary.The calculation accuracy of a simple deep learning model can not meet the engineering requirements.The model itself has defects,which can not solve the related tasks of context information semantic understanding.In this paper,a deep learning language model is used to study emotion analysis tasks,and a kind of pre-training model based on BERT?Bidirectional Encoder Representation From Transformers?^[1]is proposed.First of all,in the process of text data preprocessing,only Chinese characters are filtered out,and the length is less than 140 characters.According to the short text of Chinese microblog,a model is designed to recognize its emotion accurately,and the results of the model operation are used as the basis of subsequent experiments.Secondly,this paper improves the pre-training and fine-tuning process of the model.We update the pre-training parameters by using the corresponding microblog data set.In the mask language model?MLM?training task,we changed the proportion of partial masking to adapt the model to the data characteristics of emotion classification.This work also changed some training parameters of the original model in the process of fine-tuning.As a multi-task pre-training model,the higher the data quality and the more considerable the amount of data in the corpus,the better the effect can be achieved.These improvements can improve the classification accuracy in the final stage of emotional classification.Secondly,in the aspect of text public opinion guidance,this paper proposes a new model to correct the text emotion.First,in order to improve the effectiveness of the experiment,a single word,and doubleword mask operation is carried out before the input of the word vector,which is used as the input of the noise reduction self encoder.Then in the encoder,the long short term memory?LSTM?^[2]is used to extract the features of the input word vector,and the self-attention is used to weight the features.The decoder uses a similar structure to the encoder to predict the mask part of the output.Secondly,an emotional constraint module is added by using the output of a decoder.Bi-LSTM and attention represent its structure.The module can weigh the output of the encoder emotionally.Finally,the two parts of the vector results are summed through the connection layer and Softmax^[3]classification and standardized to get the sentences of directed emotion.In this paper,two experiments are carried out on five datasets to verify the validity of the model.