Research On Patent Design Goal Classification Method Based On Word Sequence Kernel Function

Innovation is the key to maintain the stable development of modern society.When the existing technology is difficult to improve,you need to change the way of thinking.Design-ers are limited by thinking inertia and knowledge.The rich innovation information and kno-wledge in patent provide a convenient source of innovative thinking for people.It is difficu-lt to directly extract design goals,problem that want to be solved,the implementation of in-novative information and other information,even if the patent has a relatively fixed format.How to get the patent information from the patent has become the hots pot research of mod-ern society.This paper analyzes the structure and format of the Chinese patent text.By using lexic-al analysis and syntactic analysis,we can get the mark words of design goals sentences and their sentence features.Through the mark words and sentence features,we can get the design goals of candidate sentences from the patent text.The candidate sets are transformed into th-e samples by feature selection.The three classification algorithms,Naive Bayes,Logistic R-egression and SVM,are used to do the comparative experiments.The samples are classifie-d into patent design target sentences.Experiment results show that the accuracy of the SV-M algorithm is 2 percents higher than the other two algorithms,and it's overall accuracy is over 70%.In the study of the classification of design goals,the semantic support of the traditional word sequence kernel is insufficient.The support of semantic information is improved by Gaussian kernel function,and the new word sequence kernel are improved by linear weight-ing.The modified kernel function of word sequence and SVM is used to classify the design target.Comparing it with the traditional SVM based on Gaussian kernel function.First,the sample data are transform-ed into feature vectors by lexical and syntactic analysis.Secondl-y,the dependency relations in the syntactic analysis are added to the feature vectors as the semantic features.Finally,the improved kernel and Gaussian kernel were used to predict.The experiment results show that the improved kernel method improves the accuracy by 4percentage points compared with the SVM method based on Gaussian kernel.