Learning To Rank For QA Problem

With the development of Internet, more and more people like to search knowledge and resole doubts for it. So, a large number of Community Question Answering sites have emerged. If the answers posted to Community Question Answering sites can be shown to people as a ranking list, users will have a good experience when using. In recent years, supervised learning approaches for ranking is one of the hottest research topics in information retrieval. Supervised learning approaches can learn ranking models based on labeled data and applies the models to new unlabeled data for prediction. Ranking SVM is one of the implements of this approach. Ranking SVM formulizes the problem of ranking instances as that of binary classification on instance pairs and performs the classification using Support Vector Machines (SVM).In the QA ranking applications, we can find that the answers at the top of ranking are more important than the answers in the middle or the bottom because users may only read the top ranking. So, in ranking SVM, the answer pair which contains the top ranking answer should pay more attention. However, conventional Ranking SVM pay no attentions of this, and it think all answer pairs are equal. In this paper, we consider cost-sensitive into Ranking SVM, and propose cost-sensitive Ranking SVM and Position-dependent cost-sensitive Ranking SVM. In the new learning to rank approaches the different answer pairs are different treated. Experiment results show that the two new approaches are both more appropriate to QA ranking problem than Ranking SVM.