| Community Question Answering (CQA) services, such as Yahoo! Answers and Baidu Zhidao, provide a platform for a great number of users to ask and answer for their own needs. In recent years, the efficiency of CQA services for question solving and knowledge learning, however, is challenged by a sharp increase of questions raised in the communities. To facilitate answerers access to proper questions and help askers get information more efficiently, in this thesis we propose a computational framework for question processing in CQA services.;The framework consists of three components: popularity analysis and prediction, routing, and structuralization. The first component analyzes the factors affecting question popularity, and observes that the interaction of users and topics leads to the difference of question popularity. Based on the findings, we propose a mutual reinforcement-based label propagation algorithm to predict question popularity using features of question texts and asker profiles. Empirical results demonstrate that our algorithm is more effective in distinguishing high-popularity questions from low-popularity ones than other state-of-the-art baselines.;The second component aims to route new questions to potential answerers in CQA services. The proposed question routing (QR) framework considers both answerer expertise and answerer availability. To estimate answerer expertise, we propose three models. The first one is derived from the query likelihood language model, and the latter two models utilize the answer quality to refine the first model. To estimate answerer availability, we employ an autoregressive model. Experimental results demonstrate that leveraging answer quality can greatly improve the performance of QR. In addition, utilizing similar answerers' answer quality on similar questions provides more accurate expertise estimation and thus gives better QR performance. Moreover, answerer availability estimation further boosts the performance of QR.;Expertise estimation plays a key role in QR. However, current approaches employ full profiles to estimate all answerers' expertise, which is ineffective and time-consuming. To address this problem, we construct category-answerer indexes for filtering irrelevant answerers and develop category-sensitive language models for estimating answerer expertise. Experimental results show that: first, category-answerer indexes produce a much shorter list of relevant answerers to be routed, with computational costs substantially reduced; second, category-sensitive language models obtain more accurate expertise estimation relative to state-of-the-art baselines.;In the third component, we propose a novel hierarchical entitybased approach to structuralize questions in CQA services. Traditional list-based organization of questions is not effective for content browsing and knowledge learning due to large volume of documents. To address this problem, we utilize a large-scale entity repository, and construct a three-step framework to structuralize questions in "cluster entity trees (CETs)". Experimental results show the effectiveness of the framework in constructing CET. We further evaluate the performance of CET on knowledge organization from both user and system aspects. From a user aspect, our user study demonstrates that, with CET-based organization, users perform significantly better in knowledge learning than using list-based approach. From a system aspect, CET substantially boosts the performance on question search through re-ranking.;In summary, this thesis contributes both a conceptual framework and an empirical foundation to question processing in CQA services. |
