Public Opinion Evolution Analysis With Integration Of Topic And Sentiment

With the rapid development of Internet technology,online media has become an important platform for users to release and obtain news information and it also has became a distribution center for netizens to spread and discuss news events.Under this background,research enthusiasm on public opinion evolution is increasing.As two import complementary parts,topic extraction can show the evolution process of related topics and sentiment analysis can exhibit the polarity and intensity of user's sentiment.However,there are two shortcomings in related researches: firstly,traditional topic models simply use keywords to represent topics,thus degrading the semantic integrity of topics;secondly,the deficiency of sentence level pretraining task caused poor performance of language model in logical perception and semantic expression.To solve these problems,this thesis proposes a topic extractive model based on text summarization,which can effectively improve the semantic integrity of topic representation.Then,a transfer learning based sentimental polarity model is proposed to enhance the logical perception and semantic expression ability of language model so as to improve the classification results.Furthermore,a system with integration of topic and sentiment has been designed and implemented to show the public opinion evolution process in a more accurate and multidimensional way.Research details mainly includes the following three aspects:(1)Aiming at the problem that traditional topic models cannot provide clear and complete topic representation,this thesis proposes a topic representation model based on extractive summarization.This model selects key sentences from news,which can effectively avoid the shortcomings of traditional methods.Moreover,this thesis improves scoring mechanism of extractive summarization model based on Orthogonal Decomposition policy to ensure important sentences more likely to be selected.Experiments on CNN/DM and NLPCC datasets show the proposed model can extract better summarization than baseline.(2)Focusing on the problem that language models lack sentence level pre-training tasks,which may lead to low quality of sentence features,this thesis proposes a transfer learning based sentimental polarity analysis model.The proposed model trains word vectors and encoder parameters on text ranking task,and then transfers parameters to sentimental polarity analysis model.Through text ranking,the model achieves higher logical perception and semantic expression ability,which lays a good foundation for sentimental polarity analysis.Comparative experiments on public datasets show transfer learning can effectively increase the performance of sentiment analysis model.(3)This thesis designs and implements a prototype system of public opinion evolution analysis with integration of topic and sentiment.Adequate design and verification are carried out firstly based on proposed models.Then,by taking public opinion data of coronavirus pneumonia as an example,the effect of the system has been demonstrated.