Research Of Representation And Classification Of Multimodal Social Event Based On Max-margin Topic Model

Social events refer to specific behaviors that happen at specific time and place.With the rapid development of Internet,events in real life are rapidly spreading and developing on the Internet.There are more and more social media websites in the network,so that users can easily share views and disseminate information on these websites.Most data uploaded by users are related to specific event topics.Manually analyzing is not only time-consuming and laborious,but also easy to be influenced by artificial factors.Therefore,the automatic method of analyzing social events is of great significance for understanding the development and evolution of events.However,with the development of mobile Internet,all kinds of sensors are widely applied.The data on the Internet show multi-modal characteristics,making event analysis more challenging.The main work of this thesis focuses on the representation of multi-modal social events and the research on multi-modal social events.(1)Event representation is a basic work in the field of social event analysis.It is also the cornerstone of other social event analysis tasks,such as event detection,event tracking,event evolution analysis,etc.In the new era,multimedia data has multimodal characteristics,so multi-modal representation of social events has become an important research field.Based on the characteristics of multimedia data,a multi-modal event representation method is designed,and the textual and image are taken as an example to setup the experiment.In this thesis,two typical classification algorithms are used to verify the effectiveness of event representation method.The experimental results show that the multimodal event representation method based on the topic model can improve the performance of event classification and can mine multimodal topics of events.(2)Event classification is a hot work in the field of social event analysis.Although the algorithm of event classification based on single text modalities has achieved great success,it failed to combine the multi-modal information of network media.This thesis studies how to use the multi-modal information and supervised information of events to improve the performance of event classification.In order to achieve this goal,this thesis presents a multi-modal max-margin supervised topic model.On the one hand,this thesis exploits the topic model to mine the relationship between text and image at the topic level;on the other hand,this thesis uses the max-margin classifier to develop the label information of social events.Finally,the multimodal information and supervised information are integrated into a topic model.Qualitative and quantitative experimental results show that the Multi-modal Max-margin Supervised Topic Model for Social Event Analysis method can significantly improve the performance of event classification.