Integration Of Multiple Features Of Expert List Sorting Of Learning

Traditional search engines retrieve related pages with query through keywords combinations, but must go through manually selecting the information related to the query topic. Expert retrieval is a hot area of the current vertical information retrieval research, which is more accurate way of information retrieval for the expert characteristic, and it can provide various forms of queries related to the topic, and then return directly the expert list or the home page most related to the query topic, that is currently the most effective means of expert information acquisition. Expert ranking model is the core of expert retrieval, and the quality of expert ranking determines the performance of expert retrieval system directly. Therefore, constructing efficient expert ranking model has become critical. This thesis discussed some methods of expert ranking, dedicated to how to integrate expert evidence documents, expert relationships and expert metadata of these features information to build expert ranking model based on listwise approach, thus improve the expert ranking results. Specifically we made a deep study in the following aspects and had gained some achievements:(1) We analyze the factors that influence the expert ranking, and put forward three types of features for expert ranking. For the task of expert ranking, discuss the correlation between query and expert pages as well as evidence documents, analysis the impact of expert evidence documents, expert networks, expert metadata and other factors on expert ranking, extract similarity features, BM25score, expert page content features and expert relationship features. Experimental results show that integrating these features to ranking model can improve the ranking performance effectively.(2) We propose an expert ranking model based on the listwise approach with multiple features. Firstly, through the analysis of expert pages, we select similarity features between query and candidate expert pages, expert page content features and relationship features between expert pages; secondly, we integrate these features into ListNet ranking model, learn parameters through gradient descent, and then construct expert ranking model based on the listwise approach with multiple features; finally, expert ranking contrast experiment will be performed using the trained model, experimental results show that the proposed method has good effect, and the value of NDCG@1increase14.2%compared with the pairwise method with expert ranking, so the method based on the list with multiple features can improve the effect of expert ranking effectively.(3) We put forward an expert ranking listwise method based on relationship features. First of all, we build a model based on the expert evidence documents, a model based on the expert networks, a model based on the expert metadata, and then we proposed the expert-ListNet algorithm through the obtained three model based on relationship features. Finally, we get the expert ranking model based on relationship features by training. Experiments prove the validity and superiority of the proposed method.(4) Using the above research results, we designed and implemented the expert ranking prototype system based on listwise approach with multiple features.