Research On Domain-oriented Indicator Mining Method And System Implementation

With the development of innovation drive,science and technology have received unprecedented attention all over the world.In recent years,with the increasing number of patent applications,a large number of unstructured patent text have emerged on the Internet platform,which contains many key indicators of scientific and technological development.How to effectively mine indicators from patent text data and provide reference for patent comparative evaluation and enterprise technology evaluation has become the focus of this paper.Therefore,the paper proposes the research and system implementation of domain-oriented indicator mining method,and the specific measures are as follows:By investigating the research status of patent text data and text indicator word mining at home and abroad,and combining with the attribute feature of other scholars’ research on patent text content mining and text indictor word mining,this paper summarizes the feature selection forms of domain-oriented patent text indicator word mining,using the common features of text mining,the position information of patent text indicator words and the characteristics of surrounding modifiers,Seven Attribute Features such as N-gram,part of speech,verb+noun,verb+adjective+noun,verb+adjective+adverb+noun,verb+adjective+adverb+noun,verb+adjective+quantifie r+noun are constructed as the research features of patent text indicator word mining,and feature engineering is processed by combining existing dependency syntactic parsing and word2 vector technology.The selection direction of method is mined for domain indicators.In the model of indicator mining method based on BP neural network,the nonlinear mapping ability and self-adaptive ability are used to predict indicators.The number of neurons in the hidden layer are calculated the range of according to the empirical formula,and find the number of neurons when the MSE is minimum and adjust different learning rates to optimize the model.In the research of indicator mining method based on Bayesian neural networks,the good robustness that network weight is randomly distributed and not easy to fall into over-fitting are used to predict indicator words,the number of network layers is determined according to experience,the number of hidden layer neurons is obtained by formula and the optimization model of hidden layer neurons is found when MSE is minimum.Then,the experimental results of the two models are compared and analyzed,proving that the research method proposed is effective.The research method proposed is applied to practice,and a domain-oriented indicator mining system is built by using java language in the paper,which provides auxiliary basis for patent researchers to evaluate patent and enterprise technology.