QA Problem Oriented Coast Sensitive Learning Algorithm For Ranking

With the fast development of Internet, the network information shows anexplosive growth trend, search engine technology helps people to search forinformation they want from the mass of information. However the current searchengine technology sometimes can t help people access to the information immediatelyand accurately as they want. Answering system is a new retrieval technology which isdeveloped on the basis of the traditional search engine technology.In recent years, a large number of QA(Question Answer) problems frequentlyappears on various of websites, such as the QA communities and forums. The bigchallenge of apply QA thread as knowledge is how to automatically sort the answersof the question according to their qualities. Because the answers (QA thread) qualityis varies, and almost all of the QA community and forums do no treatment to theanswers. Obviously, this would have some adverse effects to user s experience. Withthe consideration of cost, our study mainly research sequential learning algorithm.Our works are as follow:First, with the "Zhidao.Baidu.cn" as corpus and the answers to the questions asfeature vector, we get a sort model. Based on the existing feature selection algorithm,the improved sort learning-oriented Championship Sort method is proposed.Second, after analyzing some problems of the sort SVM in the application of QASort field, then applied the cost sensitive factor to the traditional sort of learningalgorithm-sort support vector machine, and the cost sensitive sort of learningalgorithm and location sensitive sort of learning algorithm is put forward. The costsensitive sort of learning algorithm suppose the cost of error on the top of thesequence is higher than the one in the middle of the sequence or the end, so costsensitive sort of learning algorithm is about an ordered pair of front and rear elementsclassified. Location sensitive sort of learning algorithm not only consider an orderedpair of front and rear elements classified, but also consider the location information.Then, study proved that the convexity of loss function of the sort of costsensitive learning algorithm and location sensitive learning algorithm, the gradientdescent method is proposed to optimize them. Finally, the algorithms above are applied to the extracted data from"Zhidao.Baidu.cn" to train a sort model, and then the experiment results are analyzedand discussed. The results show that the new proposed feature selection method andsequential learning algorithm performance higher than the original method.