Research On The Construction Of Patent Knowledge Graph Based On Natural Language Processing

In recent years,the United States launched a trade war against China and once again wielded the big stick of intellectual property protection to exert pressure on China.Intellectual property is the core competitiveness of the United States.In order to maintain its own technological and economic advantages,the United States will not hesitate to curb competitors' technological catch-ups.As a latecomer chaser,China has been targeted by the United States everywhere,especially patent as a representative of high-tech in the field of intellectual property rights.China's patent protection development time is short,and the situation is complex.Due to the unstructured patent text and the high professional level of patent field,and the huge amount of data of patent text,patent analysis work relies heavily on professionals,raising the threshold of patent analysis and protection.As a powerful analysis and retrieval tool in the big data era,knowledge graph has been widely used in various fields.Therefore,this paper attempts to introduce knowledge graph into the field of patent protection,construct knowledge graph based on patent text to assist analysis and reduce the threshold of patent analysis and protection.In terms of the construction of the patent knowledge graph,this paper studies the input method of the existing model.Based on the combination of the existing summation and splicing of the two most effective character&words,the Bert model is combined to combine the character&words improved,and got four model input methods combining character&words.On this basis,this paper designs a reasonable comparative experiment.According to the experimental results,the sequence of Bert operations has a great influence on the experimental results.Performing the Bert operation on the word vector at the same time will eliminate the feature differences between the word vectors,and has no obvious effect on the result.Performing the Bert operation separately improves the F1 value of the existing model by nearly 3%.This paper applies the improved input Bert-BiLSTM-CRF model and Bert-BiLSTM-Attention model to the named entity extraction and entity relationship extraction of the subsequent knowledge graph of this thesis.In terms of the construction of the patent knowledge graph retrieval system,this article collected and sorted out more than 4000 patent documents in the electrical field of Chinese patent websites,cleaned the data,sorted out and summarized the cleaning process,and realized the automatic cleaning of patent documents.Improved words combined with input Bert-BiLSTM-CRF model for entity and relationship extraction.On this basis,this paper designs and implements a patent knowledge graph retrieval system that includes data model management module,core function module,retrieval and visualization module.The data model management module realizes the management function of system data,and the core function module includes data cleaning,word segmentation,entity extraction and relation extraction.In terms of retrieval,based on the patent text search,it realizes the function of visual browsing and searching patent knowledge graphs,and achieves the use of patent knowledge graphs to help lower the threshold of patent protection.