| In the recent years, since the development of internet technology, social networks emerged in large numbers. As one among the things, online Question Answering system is a website for users to questioning and answering, realizing information and knowledge sharing through its unique way. There are some bad phenomena in existing online Question Answering systems which have strong impact on user experience, such as questions questioned waiting for long time to be answered or even never be answered, the accuracy of answers not be guaranteed and imbalance of user’s questions and answers. The primary reason leading to those phenomena may be that users glance over all questions to finding those they are interested in and enabled to provide the answers. They will need to spend a lot of time to search and can’t share their knowledge with others who desire it timely and effectively. For this reason, the introduction of the expert finding in online Question Answering systems is necessary.Existing studies of expert finding can be divided into two aspects, one is based on probabilistic language model and topic model, another one is based on link analysis. At the same time, expert finding in online Question Answering systems is also focused by many researchers. This thesis put forward a new way combining topic and link analysis, finding expert through realizing topic location first and score propagation later. The object of our study is the focused online Question Answering system called Stack Overflow, which is about programming. The goal is to realize that when one question is entered, a series of experts ranked who are fit to answer the question be returned. Specific research contents of our study are as follows:1) Existing studies of expert finding in online Question Answering systems are summarized. By analyzing flaws in existing study we put forward new schema and ideas. At the same time, the process of data crawling from Stack Overflow is introduced and statistic and analysis on the database and some preprocessing work are completed.2) Topic model is built based on all questions in our database for topic decomposition and a method to measure the distance between questions and topics we obtained is put forward, realizing clustering for all training questions based on taking the shortest distance. Therewith,2000questions are tested for the effect of clustering, the result show that our topic model and method of clustering are good for topic decomposition and clustering.3) Score propagation model is built based on users’relation network, which is built on questioning and answering records. The scores are defined as Authority property score and Hub property score. For a unique topic, the model returns the rank of users in it. Therewith,2000questions are tested for ranking and the result show that our score propagation model is effective for ranking users with professional ability.4) The process realizing the goal that finding experts ranked for a question inputted is introduced. By combining semantic similarity between questions and users with semantic similarity among questions, latent links for the final work of expert finding are established. Then the model of score propagation was applied on the new users’ relation network. The result shows that that the mining of latent links improves the result for returning experts for the question inputted.5) The process of building the Graphical User Interface (called GUI) of expert finding is introduced. The GUI is built based on the module of Python called wxPython, which is a module includes various components. The GUI can visualize the expert finding work accomplished in our study. While a user key a question in the entry, the system will return a series experts ranked and displayed them on the interface for the user to reference and selection. |
