Research On Some Key Theories And Applications Of Network Public Opinion Analysis

With the rapid development of Internet in China,more and more social media such as forums,micro-blog and official accounts provide a platform for netizens to express their opinions.More and more netizens use these social platforms to express their views on various social phenomena and social things.In this era of unprecedented rapid information dissemination,the demand for online public opinion expression is increasingly diversified.If network public opinion is not properly guided,negative public opinion may pose a serious threat to social public security.Network public opinion reflects the social situation and public opinion in essence.One of the important links to strengthen social governance is to strengthen the monitoring of network public opinion.This thesis focuses on some key technologies of network public opinion analysis.The main research contents are as follows:(1)Aiming at the problem of named entity recognition of low resource languages with insufficient annotation data,a knowledge distillation learning model based on knowledge alignment and association mapping in word embedding space is proposed.The teacher model of source language training with rich annotation data is used to guide the named entity recognition of low resource languages;(2)Aiming at the problem of insufficient training data in NLP task,a method of generating and enhancing data set by using natural language interpretation is proposed,and an emotion analysis model for learning external knowledge from natural language interpretation is implemented;(3)Aiming at the problem of rumor detection in social media networks,a rumor classification model using two-way GCN heterogeneous networks and integrating microblog semantic information is proposed,and a method of early rumor detection is proposed;(4)Using the key technology of network public opinion analysis,a multilingual public opinion monitoring system is designed and implemented.The specific research is as follows:(1)To better solve the problem of named entity recognition in low-resource languages with poorly annotated data or without annotated data,the cross-language named entity recognition task must effectively utilize the knowledge learned from the source language with rich annotation data.This thesis proposes a teacher-student learning model My NER via knowledge distillation to solve this challenge.Specifically,firstly,a NER teacher model?_src is trained on the source language dataset,then word-by-word translation technology is used to convert the source language into the embedding space of the target language to generate a fake dataset of the target language,and a teacher model?_trans is trained.Using knowledge distillation,using teacher model?_src and teacher model?_transto train student model together,the soft label of prediction is obtained.Once again,the unlabeled target language dataset is converted into the source language fake dataset by using word-by-word translation technology,and the predicted hard label is obtained.The student model needed for supervision training combined with soft labels and hard labels.On the NER benchmark data set published by Co NLL-2002 and Co NLL-2003 and the Uyghur dataset constructed by our laboratory,English is used as the source language,and the other three low-resource languages are used as the target language.The My NER model is used for training and prediction.Experiments show that the My NER model performs better than the baseline method in three low-resource target languages at the same time.Finally,the application details of the model are demonstrated in a real scene.(2)In order to solve the problem of lack of tagging data and data imbalance in multi language environment,this thesis proposes an affective analysis model SANLE,which can learn knowledge from natural language interpretation.According to the pre-provided natural language explanation examples,the SANLE framework can generate labels with natural language explanations for each unlabeled sample to construct a large number of weakly supervised data sets and train sentiment classifiers.The model consists of three components:semantic parser,filter bank,and label aggregator.The semantic parser converts the pre-provided natural language explanation into a plurality of programmed label functions,On specific unlabeled samples,the filter bank removes redundant,erroneous,and duplicate label functions,the label aggregator combines potential conflicting and overlapping labels into one label for each example,and then uses the obtained label examples to train a discriminant model based on attention mechanism and Bi LSTM.On two customer review datasets of Semeval2014-Task4,natural language explanation is used as an external knowledge to train the sentiment analysis model jointly.Experiments show that the SANLE model has better performance than the baseline method.Finally,the application details of the model are demonstrated in a real scene.(3)To solve the problem that current rumor detection methods only rely on searching clues from user generated content,user account information or widely spread structure,but ignore the effective combination of rumor spread structure graph and text semantics,this thesis proposes a rumor detection model KRumor.It combines the semantic information of microblog text with the heterogeneous graph representation of two-way propagation,and uses the tag function to generate tags for the data from the effective experience of artificial rumor detection,so as to jointly train the rumor classifier.Specifically,this thesis uses the attention mechanism to learn the semantic representation of microblog text and introduces the bidirectional GCN in the direction of propagation and diffusion to capture the global and local relationship representation among all source microblogs,forwards,and users.Then text semantics and propagation heterogeneous graphs are combined effectively to train the rumor detection classifier.The model is used for training and prediction on Sina Weibo,Twitter15,and Twitter16public rumor detection datasets.Experiments show that the performance of the KRumor model is superior to the baseline approach.Finally,the application details of the model are demonstrated in a real scene.(4)Using the model proposed in this thesis,combined with the key technology of network public opinion analysis,a multilingual public opinion analysis system for Xinjiang region is implemented.In system architecture,the system is divided into data acquisition layer,data mining layer and information service layer;In terms of system function,the system is divided into key station monitoring,key person monitoring and hot topic monitoring.The system has a wide application prospect in Xinjiang.The government regulatory departments can understand the social situation and public opinion and strengthen the efficiency of social governance by using the system;Enterprises and institutions can use the system to understand the public opinion on the unit image or brand,and effectively enhance the corporate image.