Research On Ranking Method Of Expert List Based On Deep Learning

Whether in research institutions or large companies needs to find experts who have a wealth of expertise and skills and represent the leading edge of development in the field, to lead the team to work and to improve work efficiency. So how to effectively find experts in the relevant fields is crucial. The traditional methods usually use search engine to retrieve experts. There are low accuracy and incomplete information and so on. So retrieval experts for this specific population is very necessary. Experts sorting is that identifies the degree of the correlation among experts given query topic with all kinds of documents and resources which can characterize specialist expertise, then sorts by correlation and displaies the list of experts result. Experts retrieval is the most effective way to get expert information, while experts ranking is the core of experts retrieval. The quality of experts ranking determines the precision of experts retrieval, therefore, experts ranking has important research value. This paper focuses on the issue of experts ranking, and has done in-depth studies and discussion in the following several aspects, and good results were obtained:(1) We propose an experts ranking method based on the list with feature hierarchy type information.The method thoroughly analyses the characteristic of experts ranking, then defines four types of features. They are the characteristics of the correlation between query and document, the page content characteristics, language model characteristics and experts related characteristics. The different types of features has different contribution degree for expert ranking. Therefore according to the size of the contribution degree defines the value of characteristics hierarchy type. Finally, the characteristics hierarchy type information combines with experts list to sort experts.(2) An experts ranking method based on the deep learning is proposed.The method for neural network due to the random initialization which easies to fall into local minimum, and training time is too long, and can not be a better approximation of the sort function,propose to build the model of deep belief nets to sort experts by restricted boltzman machine obtains the optimal parameters by unsupervised Self-training to assign to the weight one by one. (3) We propose an experts list ranking method based on the deep learningTraditional restricted boltzmann machine makes every document as a training instance when solve parameter by Self-training. It doesn’t take into account the correlation between the expert list, so restricted Boltzmann machine is improved.All expert documents corresponding to the query is formed expert training examples as the input of the restricted boltzmann machine. RBM update parameters when the training is end for all expert documents. And uses the cosine replace matrix simple subtraction to calculate the updated variableΔW(4) Design and implement the prototype system of the experts list ranking method based on the deep learning with the above research results.