Research On Social Relationship Analysis Via Textual Information

With the development of the Internet, more and more information is presented to us nowadays. The origin of social networks is the online dating. People use the social network service to build and maintain the social relationships, develop new relationships. They participate in social activities for the purpose of feeling like a part of the community.The simplest social relationship among people is friendship. In addition to friendship, people can be categorized into different groups according to their interests or attributes. For example, people can be divided into the reading or dancing groups based on the interests. Moreover, people can also be split into the critics or audiences based on the specialism. Hence, detecting which groups people belong to is one of the research aspects in this study. On the other hand, personal information extraction is very useful for us to identify the social relationships. Hence, another research aspect of this study is to extract the representative personal information. The description of the aspects of this study is as follows:1. We extract two kinds of useful and representative personal information from personal profiles, of which one of them is the skill information. However, as the skill information is too simple to represent a person, we thus extract the summary from the profile to get information which is even more representative. We propose a joint factor graph model to extract both the skill and summary with both personal and skill connections. To be more specific, we use attribute function to represent different kinds of textual information, and use the factor function to connect people with personal and skill connections. The belief propagation algorithm is used to learn and predict the model.2. After we extract the personal information, we predict friendship among people with various kinds of textual information. Since friendship and interests are highly correlated, i.e. closely-connected friends tend to have similar interest, we explore various kinds of sentimental information to connect people with similar interest. Besides, researchers in sociolinguistics have found that social contexts influence language at different levels, namely structural, lexical, and syntactic level. Hence, personal written linguistic style can also be utilized for friendship prediction. Finally, factor graph model is used to incorporate the information above for predicting friendship on social networks.3. We divide users' group analysis into two sub-tasks. For extracting the interest-based groups, we propose a novel latent factor graph model to incorporate both explicit and implicit information. Since the focus on different kinds of people is different, we propose a matrix factorization based framework to predict the specialism between people with both textual and social information in order to separate critics from audience.