Users’ Sentiment Mining And Spreading Among Chinese Microblogs In The Context Of Specific Events

Microblogs and other online social media play important roles in the spread of public opinion. Users can express their views and perceptions of an event with text, images, video and other information forms. The information about an event can be quickly transmitted by a user to another one among different social networks within social media. In the context of a specific event, group polarization phenomenon often appears in social media and thus to cause mass incidents in actual life. Emotions expressed by users can not only influence the propagation speed of the event information, but also can infect other people in social media. Negative emotions can trigger the negative behavior of users and prompt events toward the negative direction. Therefore, it is necessary to analyze the sentiment of users in a particular event situation, identify the types of emotions and the intensity of emotional polarity and explore the factors that influence user’s emotional expression and transmission.This paper focuses on sentiment mining and sentiment spreading within Chinese microblogs when users face specific events. Three major research problems are solved:sentiment features recognition, sentiment description and sentiment spreading by three steps:Firstly, Chinese sentiment thesauruses, including an emotional thesaurus and an evaluational thesaurus are constructed. Sentiment words in thesaurus are selected from three existing sentiment words lists and microblog text corpus about specific events. Then these extracted sentiment words are classified and their polarity intensities are labeled according to HowNet Chinese repository. The emotional thesaurus contains3773classified words in12categories and32secondary classes. The evaluational thesaurus contains12844classified words in8categories and100secondary classes.Secondly, sentiment visualization and sentiment features statistics are carried out for two specific events. The relationships between sentiment words are determined through their co-occurrence frequency in the same microblog text, and then are displayed in a graph through location algorithm. The frequency and polarity of sentiment words are also displayed in graph through the size and color of the font. It is found that the high frequency words at the centre of graph can reflect the dominant sentiment types of events, while words which are more close to the edge of the graph reflect the emotion of the general public. Classification statistics for sentiment words can disclose the strength of all types of emotions expressed by the users on specific events. Time series statistics of emoticons can identify some messages posted by water armies. After removing these messages, it is found that the change trend of negative intensity is similar to positive intensity.Thirdly, the dissemination network of specific event information is constructed. Key users, information propagation distance and cluster degree of the information dissemination network are analyzed using social network analysis method. Users’ sentiments are embedded in the information dissemination network so as to visualize sentiment and help to understand the distribution of users’ emotions. The relationships between users’emotions and their roles played in dissemination networks are analyzed. It is suggested that policy makers should pay more attention to users who express emotions such as "excited","abuse","agree" or "against", because users with such emotions are more likely to play a key role in information dissemination.The innovation of this study mainly includes three aspects:firstly, Chinese sentiment thesaurus are improved. There are a few existing studies related with Chinese sentiment lexicon building, but most of these sentiment lexicons can’t be used publicly and a few available sentiment lexicons divide sentiment words into only two categories:positive words and negative words. Because of the limitation of previous sentiment lexicons, current sentiment analysis mostly focus on polarity judgment of sentences or texts, but cannot analyze specific types of emotions and their intensities. This study divides the whole sentiment vocabulary into emotional classification vocabulary and evaluational classification vocabulary, not only can analyze the polarity of emotions, but also can analyze specific types of sentiment. Secondly, visualization technology is used to visualize sentiment words and relationships between different words. At present, tag clouds are commonly used in the visualization of themes or keywords of texts, but very rarely applied in sentiment visualization. This study not only carries on the sentiment words visualization, but also visualizes the polarity intensity of words through graph. In addition, this study also visualizes users’ sentiment in event information transmission network. It can provide more intuitive sentiment information for decision makers through a variety of visualization technologies and algorithms. Thirdly, the research on sentiment spreading in Chinese online social network is prompted. There are some related studies in English, but most of them focus on users’ emotional interaction in daily communication network. In China there are few studies about sentiment spreading in the context of specific events within social media. In order to explore related factors of the emotional communication, this study also analyzes the relationships between users’ emotions and their roles in event information transmission.The theoretical implication of this dissertation is that it constructs a classification system for sentiment words based on psychology research and the HowNet ontology. This system will be available for reference for further research.The methodological implication of this dissertation is that it provides methods for emotional knowledge representation and sentiment spreading analysis. This will help to improve the current sentiment analysis and sentiment spreading research. In practical implication of this study is that it can help relevant government departments to understand public sentiment spreading during events and provide targeted information to avoid the agglomeration and polarization of public feelings. This study also can help enterprises or individuals to understand public emotional responses to events in micoblogs and to develop targeted strategies based on the regular patterns of public sentiment spreading, and can help governments and enterprises to improve the quality of their services or products by evaluating public mood in microblogs.