Research On Knowledge Graph Construction For Automatic Evaluation Of Chinese Patent Valu

With the development of science and technology,technological innovation has gradually become an important component of national and social progress.As a carrier of invention and creation,patents contain rich innovative technologies and scientific research achievements.Deep mining and analysis of the knowledge contained in patents and the information that affects patent value evaluation can provide a basis for rapid understanding of patents,judgment of patent value,and improvement of patent level.However,most of the existing methods for patent information mining and value evaluation are based on traditional statistical analysis methods,which require a lot of manpower and resources.Their subjectivity is strong and they lack the mining of deep information and internal relationships of patents,making it difficult to effectively,scientifically,and automatically evaluate patent value.Therefore,this article utilizes technologies such as deep learning to mine and analyze the basic attributes,technical dimensions,policy dimensions,and other knowledge of patents.The aim is to construct a knowledge graph for automatic evaluation of patent value,express knowledge and link related information of patents,in order to comprehensively represent the core information of patents and provide a knowledge evaluation basis for patent value evaluation.The main research work is as follows:In terms of extracting technical dimensions that affect patent value evaluation,a two-stage patent technology point generation method is proposed.Considering the data characteristics such as scattered patent technology information and rich professional expression,the idea of combining text extraction and text generation is adopted to generate patent technology points.Firstly,Roberta combined with Dilate Gated Convolutional Neural Network(DGCNN)performs context learning and feature extraction on the patented text,so as to extract the core technical sentence information of the patented text.Then input the extracted results into the NEZHA+UNILM generation model to generate technology points.In the model adjustment stage,the Copy mechanism and external knowledge are introduced to polish the generated text to obtain the final patent technology point.The experiment on constructing a dataset shows that the ROUGE-L score of the proposed method reaches 68.94%,which is superior to existing mainstream methods and can effectively solve the problems of insufficient logic and readability of generated text.In terms of extracting policy dimensions that affect patent value evaluation,a multi-strategy fusion extraction method for patent-related policy documents is proposed.The theme and keyword of the patent is extracted,and then it is put into the relevant policy document search to extract the policy documents.In terms of topic extraction,aiming at the problems of diverse topics and low rate of displayed topics,a topic model based on Bert combined with prompt learning and fusion of multiple features is proposed to transform patent topic mining into patent topic classification to accurately extract topics.In terms of keyword extraction,in view of the characteristics of low interpretability and long term length of patent texts,this paper proposes a patent keyword extraction method that combines Roberta with head-tail recognizer(HTR)and fuses subject information at the same time.Firstly,Roberta learns the contextual information,passes the embedding into the recognizer for decoding to obtain the head-tail sequence,and then combines them to extract patent keyword information.In terms of extracting relevant policy documents,a trigger mechanism is constructed to input the topic and keywords for patent-related policy retrieval into the trigger to obtain policy documents.After experimental verification,the methods proposed at different stages can accurately extract patent themes,keywords,and relevant policy document information compared to current mainstream methods.In terms of knowledge graph construction,the basic attribute information of patents is extracted by means of self-defined template rules and other methods for semi-structured information such as patent inventors and invention time.For unstructured information,deep learning methods are used to extract information such as patented technology points,topics,keywords,and relevant policy documents.For redundant information in patent knowledge,knowledge fusion methods such as Synonym,Levenshtein,and Jaccard are used to uniformly align it.Integrate and process the acquired information to construct and format triplet data.Finally,Neo4j was used to store triples and a front-end interface was constructed for graph visualization.In terms of the knowledge graph system platform,different functional modules are designed to achieve visual display of patent knowledge.In summary,on the basis of focusing on the technical and policy dimensions affecting patent value assessment,this paper through the construction of knowledge graph for automatic patent value assessment by means of technology point generation,theme extraction,keyword extraction,relevant policy extraction,basic attribute extraction and knowledge fusion,which can effectively organize and store the basic attributes of patents and information affecting patent value assessment and provide technical and data support for the automatic assessment of patent value.