Research And Application Of Scholar Profiling Based On Neural Network

The development of science and technology and many academic activities give birth to large demand for academic services.As one of the most important basic applications,scholar profiling aims to establish a multi-dimensional label model for scholars to support upper level services such as expert recommendation.Scholars' basic information extraction and scholars' interests mining are key tasks of open scholars profiling problem.To this,most of the solutions are based on the traditional method,which has the problems of high cost and limited performance.With the development of deep learning technology,it shows the advantages over traditional methods in many tasks.This thesis applies the deep learning-based methods to above key problems to optimize its performance on related tasks,reduce the dependence on Feature Engineering and the cost of constructing scholar profiling.The main contents are as follows:(1)For the task of scholars' basic information extraction,this thesis studies the existing strategies and designs a Bi-LSTM + CRF sequence tagging model based on deep learning.In view of the shortcomings of traditional methods in the representation of word vectors,a scholar information extraction model based on BERT is proposed to improve the overall performance.The advantages of the proposed method compared with the benchmark method are illustrated by experiments.(2)For the task of scholars' research interests mining,this thesis adopts a contentbased strategy to discover the scholars' interest labels from the publications,then transform the original task into multi label text classification task.For the dependencies between labels,this thesis designs a multi label classification model based on Seq2 Seq framework.Aiming at the problems of ignoring the semantic information of labels and the performance bottleneck of long tail labels.A label embedding based on label semantics is proposed to optimize the performance of original classifier.Finally,experiments show the advantages of the proposed method and the performance improvement for the downstream tasks.(3)A prototype system of scholar profiling based on deep learning methods is designed and implemented,and the proposed methods are applied to the real world data.Finally,this thesis demonstrates the effect of system,and main functional tests are carried out.