Font Size: a A A

Research On Sentiment Analysis Method Of Evaluation Objects For Case-related Public Opinio

Posted on:2023-05-19Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y XiangFull Text:PDF
GTID:1527306797978869Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
On social media platforms,netizens’ discussions gather and spread around different opinion targets of emergency cases in a short period of time.This will cause public opinion involved cases,affect the fair handling of cases and cause social problems.It is of great significance to mine opinion targets and analyze sentiment polarities of the opinion targets in online comments.However,the opinion target-based sentiment analysis of the comments involved in law cases faces with difficult problems such as fast data updating and difficult annotation,large noise,divergence of target terms,lacking of explicit expression of opinion targets,and implied sentiment expression,and so on.This thesis focuses on the opinion target-based sentiment analysis for online public opinion involved in cases.By effectively using various types of relationships within the public opinion data,some semi-supervised learning models are constructed to solve a series of problems.The main contributions and innovations of this thesis are summarized as follows:(1)An adversarial dual decision-based model for case related opinion sentences recognition is proposed.The rapid identification of the opinion sentences related to cases from the online comments is the basis for judging the quality of the online case related data,and also the basis of the subsequent sentiment analysis.Opinion representations based on deep neural networks require sufficient annotation data that are lacked in a new cases.Under this conditions,the identification of opinion sentences in a new case can be completed through feature migration with the annotation data of existing cases.Therefore,this thesis proposes a case-related opinion sentence identification method based on adversarial dual decision,which makes full use of the high correlation between the case background description text and opinion sentences.The model uses two generators to encode two different case-related characteristics,and adopts two classifiers to make decisions based on the two characteristics.Through the adversarial decision process of generator and classifier,the sentence category feature alignment of the existing case and the new case is completed,so as to realize the sentence identification of the new case.The experimental results show that the proposed method can obtain better recognition performance than other domain adaptation methods and pre-training methods,in the case of no annotation data or certain annotation data.(2)A variational dual topic representation model for opinion targets mining is proposed.It can directly show the different aspects of public opinion to mine the explicit opinion target terms from comments,and class these terms and comments into the corresponding categories.Current deep topic representation based methods for opinion targets mining have the deficiencies that they do not effectively use prior knowledge,the subject representation is fixed,and the final opinion target classification relies on artificial inference.To this end,this thesis proposes a method that uses the relationship between topic representation,comment representation and word representation.By encoding and reconstructing comments twice using the variational encoders,we obtain and combine two different topic representations to reconstruct comment.Meanwhile,we use a small number of label samples to guidance the network for fine-tuning,finally to realize the mining of target terms and automatic classification of target categories.Compared with multiple unsupervised and weak supervised methods,the proposed method can find more diverse and interpretable target terms,and have better performance in target classification.(3)An event graph neural network for opinion target classification is proposed.In addition to the explicit target terms,there are also implicit opinion targets in comments,and it is useful to class and summarize those opinion targets.The semi-supervised classification methods based on graph convolution network provide ideas for opinion targets classification,but they are insufficient to effectively learn implied category features.Therefore,this thesis effectively excavates and uses correlation between texts,including the case keyword correlation,reply association,similarity association,to construct a complete event graph.The comment representation are learned through graph convolution,and which help to complete the classification.Experiments show that the proposed method achieves good results on very few annotated training data,and outperforms a variety of the state of the art methods.(4)An target-category sentiment classification method based on tensor graph convolutional networks is proposed.The sentiment classification of the opinion targets can show the public attitude to different aspects in the case.The task is faced with the difficulties such as insufficient representation of the target categories and implied sentiment expression.For this problem,we propose a hybrid node graph tensor learning model,using different relationships between comments under sentiment words and target categories constraints.We expand target categories representation with corresponding terms,and treat comments and their target categories as a hybrid nodes.We construct graph tensors using the semantic similarity and sentiment correlation between hybrid nodes.The sentiment characteristics of hybrid nodes are obtained through the learning mode of graph tensors to complete the classification.Experiments show that the proposed model can significantly improve the performance of sentiment classification compared with other baseline models.and can still guarantee relatively accurate classification results under the condition of very few annotated data.
Keywords/Search Tags:Public opinion involved in cases, Opinion targets, Sentiment analysis, Social media, Semi-supervision learning
PDF Full Text Request
Related items