Technologies Research On Public Event And Public Opinion Analysis In Microblogging

Microblogging is bringing huge impact to communication and has become a valuable source of public opinions in recently years. It is a kind of social media with clear we-media features, in which information disseminates rapidly, and the public is widely involved. Public events which happened in real world and public opinions about them are gathered in microblogging, conflicting and influencing each other, which makes microblogging become a huge and complex opinion field. The spread of event information in microblogging gives rise to public opinions, and influences their evolution. The gathering of the latter, in turn, affects the development of events in the real world. Therefore, microblogging is an excellent platform for technologies researches on public events, public opinion, and their interaction patterns. It also makes the study has a strong practical significance and social value.In this thesis, messages in microblogging are divided into two categories according to the targets of their descriptions. One is the objective information, which describes the properties and developments of events. The other is subjective ideas, which describes interests, sentiments and opinions of the public. Furthermore, two special kinds of messages in microblogging, expandable posts and their reposts, are selected as samples corresponding to the two types of messages above. Base on this, some key technologies on public event and public opinion analysis, including new event detection, event evolution analysis, public concern evolution analysis, and the measurement of influence of the public concern on the event information evolution, are studied in this thesis.(1) In order to detect new events from short and noisy messages in microblogging, we proposed a novel method for new event detection based on swarm intelligence. According to microblogging users and their relationships, message stream is split into multiple destinations. Then we constructed multiple individual models to simulate user’s knowledge structures when judging new events. The final conclusion of new event detection is drawn from an election by integration of each vote. Since the stream decomposition process, noises from individual differences are decomposed with messages. It makes that inherently similar information is gathered into same model and the influence of the noises to the final conclusion is weaken. The experiment shows that the method can obtains higher precision than classical single-model method. Moreover, several key problems about the method are discussed. For the problem that how to select voters, we proposed that active users could be selected in general tasks, while the followers of specialists were the best choice for pre-defined target tasks. For the problem of voting rules, we proposed “minority-vote veto” strategy, considering both the novelty and tolerance. For the problem of high spatial and temporal complexity, we proposed a distributed solution to improve them.(2) Since the granularity of events is hard to define in event evolution analysis, we introduced Information Extraction(IE) techniques and the concept of atomic event, and proposed an evolution analysis approach based on atomic event mining. Our method firstly extracts atomic events from each document, and then identifies their co-reference among the whole corpus, finally measures their relationships and builds the evolution graph. By introducing atomic event, the granularity of events in evolution analysis is regularized into unified level. Since the method breaks the boundaries of documents which traditional researches abided, more related events can be extracted from the expandable posts in microblogging. Therefore, the vision of event analysis is expanded. Meanwhile, the view that “event is smaller than document” makes some new features available in event correlation, which cannot be obtained in traditional methods. It allows our method to get finer granularity and stronger interpretability in the event evolution structure. However, the disadvantage of our method is the finer granularity may extract too many events from corpus, which are bad for people’s understanding and reading. Therefore, two filtering policies are presented to refine the event evolution graph through evaluation of the event importance.(3) The previous works for public concern analysis in microblogging require prior knowledge and corresponding keywords in advance. In order to overcome the disadvantages, an unsupervised public concern analysis method considering both the expandable posts and their reposts is proposed. Based on a definition of public concern at content-level, our method firstly reconstructs topic space by expandable posts, and then locates their reposts in the space. Thus, the task of tracking public concerns is transformed into tracking the movement of the reposts in the topic space. Because our method does not require any event frameworks of public concern in advance, it can be applied to more general situations. What’s more, a tripartite graph is not an intuitive solution, although it has enough capability to express all detailed relationships between every two objects in the expandable posts, their reposts, and the public concerns. Therefore, we drew on the Sequence Diagram in Unified Modeling Language, and proposed Public Concern Evolution Graph. In the experiments on a real corpus collected from Weibo, we found out five types of public concerns through the timing analysis, which verified the effectiveness of the method. They are respectively "Warming Concern", "Sustained Concern", "Saturated Concern", "Unheeded Concern" and "Expected Concern".(4) Public concern can influence on the evolution of event information, but there are few existing researches focusing on this work. We referred to the influence as Public Opinion Work, and proposed an approximate measurement to calculate it. Based on a concept of public opinion field and analogy with physical field, such as electric field or magnetic field, the states of the public opinion space, the state transfer, and the influence on event information are quantified. Then, we took above factors together, pro-posed Event Information Evolution Graph(EIE Graph), which depicts the evolution of event information as transfer between two states of public opinion space, and describes the reason of the transfer, which indicates whether the evolution is due to the public opinion work. Based on a corpus collected from Weibo, the EIE Graph is used to explain intuitively either "public concern led the evolution of event information" or "event information led the evolution of public concern". Therefore, the effectiveness of the method is verified.In summary, this thesis presents technical solutions to several essential issues of event information and public concern analysis in microblogging. The experiments demonstrate that our methods can properly achieve their goals. It is significant to the theoretical research and practical applications on public events and public opinion analysis.