Question Recommendation Mechanism In User-Interactive Question Answering Systems

User-Interactive Question Answering (UIQA) system is now a popular social network application in the age of Web2.0. It provides a platform for people in online communities to seek information and share knowledge through the way of questions and answers rather than lists of documents from search engine. However, in existing UIQA service, users need to passively wait several hours or even a few days for other users to access the portal website and provide answers. On the other hand, some answer repliers with lower expertise level provide lots of spam answers may just want to earn incentives.To overcome the problems of existing large amount of zero-replied questions and lower quality answers in UIQA systems, we propose a question recommendation mechanism which automatically recommending the unsolved questions to appropriate answer experts. We give a definition of question recommendation in UIQA systems and depict the question recommendation model. Two different kinds of question recommendation strategies are presented, which one is recommending questions to domain answer experts and the other one is recommending questions to specific answer experts.In the first strategy, we propose to find answer experts for each category in UIQA systems. Questions in the same category will be recommended to those domain answer experts. We firstly construct a user question-answer interaction graph with different kinds of semantic links,In this graph, each node represents a user, and the link between users represents question-answering relationship. Different sources of semantic information are extracted from user interaction behaviors and answer contents in question sessions. The extracted semantic information is utilized to model the weights of relation links in the graph. A link analysis method called propagation is then performed on the graph to generate the first semantic link analysis. The higher score a user obtains, the higher expertise level he/she is regarded. Meanwhile, another approach, semantic language model (SLM), is proposed by incorporating the extracted semantic information into the traditional language model. Experiments on a collected Yahoo! Answers data demonstrate that the proposed methods of semantic link analysis and SLM reach much higher performance than baseline methods. It also verifies the effectiveness of extracted semantic information in our approaches. In the second strategy, we propose to find a particular answer expert who can answer a specific question. Different from the first strategy, the second strategy aims to identify the exact answer experts for the question in a fine grained level. In this strategy, a topic-based user interests (TUI) model, which effectively and correctly describes the topic distribution of user interests, is introduced at first. A probability calculation of whether a user is answer expert to a specific question is then proposed based on the TUI model. The higher probability means the more expertise the user may have to answer the specific question. Experiments on a collected Yahoo! Answers data demonstrate the high efficiency of our proposed two question recommendation strategies. In addition, we discuss and compare the two different strategies according to the experimental results. From the results, we can conclude that the first strategy of question recommendation in UIQA systems gives a better performance than the second one. The reason for this phenomenon might be that the task of finding domain answer experts might cover most of professional users. While recommending question to specific answer experts might only identify a few proficient users. This is a significant guidance in question recommendation in UIQA systems.