Dependency Parser Of Social Insurance Policy And House Fund

With the improvement of the insurance system,the number of policy and regulation texts in the four insurances and one gold field is increasing every year.These policies and regulations are not only huge in number,but also complex in language logic.For ordinary people who are not legal professionals,they want to find And it is very difficult to understand the texts of policies and regulations that you need.Therefore,a search engine for the four insurances and one gold field is urgently needed to help retrieve the required information.To retrieve the corresponding answers in the knowledge base,to complete these operations,you need the support of the knowledge graph,and if you want the computer to truly understand a sentence,and then complete the task of constructing the knowledge graph,you need to perform syntactic analysis and syntactic analysis on the sentence There are many methods,such as phrase structure syntax analysis and dependency syntax analysis.Compared with phrase structure syntax analysis,the dependency syntax analysis studied in this article has the advantages of easy understanding,convenient part-of-speech tagging,and concise and clear form.This paper suggests a modified dependency parsing analysis method on the basis of pointer-net neural network model,which a dependency syntax tree is constructed for sentences in the four social insurance and one housing fund policies and regulations.In the traditional method of dependency parsing analysis,it only focuses on the head-word in the analysis stack,which is used as the basis for decision making.For this purpose,this paper puts forward to use TreeLstm to encode the dependency subtree formed at each time point and input it into the training model,substituting it for the head-word in the original method as one of the features for the next decision judgment.Eventually,the purpose of enhancing the accuracy of the dependent parsing analysis is achieved.The results obtained from the dependency parsing analysis are available to us for pruning operations,so as to help realize the upper-level applications,such as term extraction of policy and regulation texts related to the four social insurance and one housing fund field.The research in this paper is mainly carried out from the following four aspects:(1)A method for dependency parsing analysis by means of a pointer-net neural network model is proposed.Unlike the traditional transfer-based dependency analysis method,this method involves no decision on transfer actions in decision making,rather it considers the similarity between the dependency parsing analysis problem and the convex hull problem solved by pointer-net,where the outputs are both heavily dependent on the inputs,and directly selects the appropriate dependency words from the input words by using the pointer-net network and scoring function.(2)Extraction of features from the dependency subtrees generated during the dependency analysis process is performed using TreeLstm.In the traditional dependency analysis method,it only takes into account the head-word feature at the root node position for the dependency subtree generated in the dependency analysis process,and overlooks the overall structural features of the whole dependency subtree.In this case,the structure of TreeLstm well fits the structure of the dependency subtree with a certain improvement in the richness of feature extraction.(3)In response to the technical terms in the texts of policies and regulations concerning the four social insurance and one housing fund fields,the word-string vector coding is adopted in the pre-training stage.Words tagged as out-of-set words are encoded by each word,the obtained word vectors are arithmetically averaged,and the resulting vector is taken as the word vector of that out-of-set word.Instead of using a uniform out-of-set word encoding,this approach allows the numerous out-of-set words in the corpus of four social insurance and one housing fund policies and regulations to be encoded separately.It facilitates the extraction of semantic features in the out-of-set words.(4)Conducting experimental validation of the above proposed method,the effectiveness and superiority of the integrated pointer-net and TreeLstm dependency parsing analysis approach proposed in this paper are demonstrated against the background of the knowledge graph application in the field of four social insurance and one housing fund.