Knowledge graph-based insurance customer demand matching and service recommendation method and equipment

Abstract

The invention discloses an insurance customer demand matching and business recommendation method and device based on a knowledge graph, and the method comprises the steps: carrying out the targeted preprocessing of structured data and unstructured data in an insurance business scene, and extracting standardized entity, relation and attribute data; s2, converting the standardized data extracted in the step S1 into an RDF / RDFS format of a unified semantic framework to obtain a multi-source independent initial knowledge graph; through an entity alignment method combining semantic standardization and semantic matching, redundancy and noise of a multi-source initial knowledge graph are eliminated, and a global customer knowledge network is formed through fusion; based on the fused global customer knowledge network, customer portraits are perfected through rule reasoning, and potential association relationships among entities are identified; and realizing accurate matching and personalized recommendation of customer demands and insurance businesses through risk level evaluation, hidden demand mining and claim settlement risk check in combination with the improved customer portrait and the inference result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an insurance customer demand matching and business recommendation method and device based on a knowledge graph. BACKGROUND

[0002] The current insurance industry has sustained growth in premium income, and customer data presents a multi-source heterogeneous characteristic, covering various types such as policy data, claim records, medical examination reports, and medical record reports. However, the traditional data management mode has significant problems. Data is scattered in various systems, forming serious data islands, and there are breakpoints in cross-department information flow, leading to fragmented customer portraits and difficulty in accurately reflecting customer needs and risk conditions.

[0003] Current industry mainstream solutions mostly rely on traditional relational databases to store data or use simple data integration tools to process information, but these solutions have obvious defects. Traditional relational databases cannot achieve semantic-level fusion of multi-source data, data redundancy and noise are prominent, and it is difficult to identify potential relationships between entities; simple integration tools lack standardized entity alignment and knowledge reasoning capabilities, and cannot effectively mine customer implicit needs, risk assessment and business recommendation rely on experience, and the accuracy is insufficient.

[0004] The core difficulties that need to be solved at present are: how to achieve semantic unification and efficient fusion of multi-source heterogeneous data and eliminate data islands; how to accurately identify potential relationships between entities and mine customer implicit needs; how to achieve scientific risk level assessment and personalized business recommendation based on complete customer portraits, and promote the transformation of insurance business from "experience-driven" to "data-driven". SUMMARY

[0005] The present application provides an insurance customer demand matching and business recommendation method and device based on a knowledge graph to solve the above problems.

[0006] In one aspect, the present application provides an insurance customer demand matching and business recommendation method based on a knowledge graph, which comprises the following steps: Step S1: The structured data and unstructured data in the insurance business scenario are respectively pre-processed for targeted extraction of standardized entities, relationships and attribute data; Step S2: The standardized data extracted in step S1 is converted into the RDF / RDFS format of a unified semantic framework to obtain an initial knowledge graph of multiple independent sources; Step S3: Through an entity alignment method combining semantic standardization and semantic matching, the redundancy and noise of the multi-source initial knowledge graph are eliminated, and a global customer knowledge network is formed by fusion; Step S4: Based on the fused global customer knowledge network, the customer portrait is perfected through rule reasoning, and the potential correlation between entities is identified. Step S5: Combining the improved customer profile and reasoning results, through risk level assessment, implicit needs mining and claims risk verification, achieve accurate matching and personalized recommendations between customer needs and insurance business.

[0007] In one implementation of this application, in step S1, the structured data includes policy data and claims data. The policy data includes information related to the policy number, insurance type, and premium amount. The claims data includes information related to the claims number, claims amount, and claims reason. The preprocessing specifically involves using relevant functions of the pandas library to convert the structured data into a dictionary-formatted JSON file, and setting up an exception handling mechanism to repair and standardize missing fields and data with format errors.

[0008] In one implementation of this application, the unstructured data in step S1 includes policy terms, medical examination reports and medical records. Knowledge extraction of the unstructured data adopts the BERT model in the field of NLP. Through this model, entity recognition, relation extraction and attribute extraction are performed on the unstructured data, and finally the unstructured data is converted into a structured JSON file.

[0009] In one implementation of this application, in step S2, the format conversion adopts a technical solution combining the Jena API and the open-source framework json2rdf to convert the JSON format data corresponding to each data source into an initial knowledge graph in RDF / RDFS format.

[0010] In one implementation of this application, the semantic standardization in step S3 is based on the enterprise's historical business terminology database to construct a core entity and relation dictionary. The dictionary is periodically updated according to the terms of new insurance products to ensure the timeliness of the terms. The semantic matching adopts the ensemble model PRASEMap combined with mainstream entity alignment algorithms to select entities that meet the preset scoring threshold for fusion. At the same time, the uniqueness of customer information is determined by entity-attribute features. For example, the uniqueness of customer information with the same name is distinguished by the combination of the name and the last few digits of the identity identifier.

[0011] In one implementation of this application, the knowledge reasoning in step S4 adopts a rule-based VLog reasoning engine. The input encoding format of the reasoning engine is optimized to support Chinese input. The reasoning rules include customer insurance liability determination rules and customer risk level determination rules. The customer profile is improved by reasoning in conjunction with customer medical record information.

[0012] In one implementation of this application, the risk level assessment in step S5 is achieved by setting a risk weight table, which specifies the weights corresponding to diseases, ages, and types of insurance. Based on the customer's medical history, age, and insurance information, the customer's risk level is determined by combining the weight calculation results.

[0013] In one implementation of this application, the system includes: in step S5, the implicit demand mining uses Path Embedding technology to convert the customer-product path into a fixed-dimensional vector, and determines the high demand association between the customer and the insurance product by the cosine similarity between the vectors.

[0014] In one implementation of this application, a knowledge storage step is also included, in which the fused global customer knowledge network is stored in the form of triples in the Neo4j graph database using Cypher statements. The database supports the storage and concurrent query requirements of entities and relationships in the insurance business scenario.

[0015] This application also provides a knowledge graph-based insurance customer demand matching and business recommendation device, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the aforementioned knowledge graph-based insurance customer demand matching and business recommendation method.

[0016] The knowledge graph-based insurance customer demand matching and business recommendation method and equipment provided in this application have the following beneficial effects: 1. Achieve standardized processing and fusion of multi-source heterogeneous insurance data. Through targeted preprocessing to extract standardized data and entity alignment to eliminate redundant noise, construct a full-domain customer knowledge network, solve the problems of traditional data silos and cross-departmental information flow breakpoints, and provide complete data support for subsequent demand matching.

[0017] 2. By relying on rule-based reasoning to improve customer profiles and identify potential connections between entities, and by combining risk weight tables to accurately assess customer risk levels, we can break through the limitations of traditional experience-driven approaches, improve the scientific nature and accuracy of customer risk identification, and provide an effective basis for risk control in insurance business.

[0018] 3. By leveraging implicit demand mining technology to identify high demand correlations between customers and products, and combining this with claims risk verification, we can achieve precise matching and personalized recommendations between customer needs and insurance business, promote the transformation of business from "experience-driven" to "data-driven", and improve service accuracy. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1A flowchart of a knowledge graph-based insurance customer demand matching and business recommendation method provided for embodiments of this application; Figure 2 A flowchart for knowledge graph construction provided in this application embodiment; Figure 3 A schematic diagram of a knowledge graph-based insurance customer demand matching and business recommendation device provided for an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a method and device for matching insurance customer needs and recommending services based on knowledge graphs. The technical solutions proposed in this application will be described in detail below with reference to the accompanying drawings.

[0022] Figure 1 A flowchart illustrating a knowledge graph-based insurance customer needs matching and business recommendation method provided in this application embodiment. Figure 1 As shown, the method mainly includes the following steps: Step S1: Perform targeted preprocessing on structured and unstructured data in the insurance business scenario to extract standardized entity, relationship, and attribute data; Step S2: Convert the standardized data extracted in Step S1 into the RDF / RDFS format of the unified semantic framework to obtain a multi-source independent initial knowledge graph; Step S3: By combining semantic standardization and semantic matching to form an entity alignment method, the redundancy and noise of the multi-source initial knowledge graph are eliminated, and the graphs are integrated to form a full-domain customer knowledge network. Step S4: Based on the fused omni-channel customer knowledge network, refine the customer profile through rule-based reasoning and identify potential relationships between entities; Step S5: Combining the improved customer profile and reasoning results, through risk level assessment, implicit needs mining and claims risk verification, achieve accurate matching and personalized recommendations between customer needs and insurance business.

[0023] In this embodiment, by converting traditional data storage into graph database storage and combining it with physical examination and medical record reports, the limitations of traditional data management are effectively overcome, enabling semantic-level fusion of multi-source data and promoting the transformation of insurance business from "experience-driven" to "data-driven." The entire graph database is constructed using a bottom-up approach, with the process mainly divided into four aspects: knowledge extraction, knowledge fusion, knowledge processing, and knowledge storage.

[0024] I. Knowledge Extraction.

[0025] The purpose of knowledge extraction is to identify, extract, and transform heterogeneous data scattered across various systems to form standardized entity, relation, and attribute data, laying the foundation for subsequent integration and processing.

[0026] 1. Data preprocessing.

[0027] (1) Structured data: This includes policy data (policy number, insurance type, premium amount, etc.) and claims data (claim number, amount, reason, etc.). This type of data has a clear structure. In this project, the json_dump function in the pandas library is directly used to convert it into a dictionary-formatted JSON file with the format {key1:value1,key2:value2,...,key1,value2,value3,value4,value5,value6,value7,value8,value8,value9 ... n :value n For example, `{name:Customer A,age:52,gender:Male,bd:{Policy Number 1,Policy Number 2]}`, for subsequent processing. An exception handling mechanism is also set up to repair missing fields (such as policy numbers) using pre- and post-term interpolation, and to validate and standardize erroneously formatted data (such as non-numerical premium amounts) using regular expressions, ensuring data integrity ≥98%.

[0028] (2) Unstructured data: including textual description documents such as insurance policy terms, physical examination reports and medical records. For such documents, this project uses the BERT model in NLP to perform entity recognition, relation extraction and attribute extraction. The training dataset consists of more than 500,000 medical records from the insurance industry and more than 200,000 policy terms. After multiple comparative experiments, the final model parameters were set as follows: learning rate 5e-5, iteration rounds 30, batch size 32, entity recognition accuracy ≥92.5%, relation extraction F1 value ≥89.6%; and finally converted into structured JSON files.

[0029] 2. Format conversion.

[0030] This project stores data extracted from each data source as a separate dataset. After collecting all datasets, the project uses the Jena API and the open-source framework json2rdf to convert the independent datasets in JSON format into an initial knowledge graph in RDF / RDFS format. For example: <rdf:description rdf:about="”http: / / localhost / 客户A”"> <ns1:age rdf:resource="”http: / / localhost / 52” / "> <ns1:gender rdf:resource="”http: / / localhost / 男” / "> < / ns1:gender> < / ns1:age> < / rdf:description> II. Knowledge Integration.

[0031] Each independent dataset extracts knowledge to construct an independent knowledge graph. These independent knowledge graphs need to be merged into a unified knowledge graph. However, the original data is often full of redundancy and noise, so knowledge fusion technology is needed to merge knowledge graphs from multiple different sources. Knowledge fusion integrates multi-source data through a unified semantic framework, eliminating data silos, building a comprehensive customer knowledge network, and solving the problem of information flow breakpoints across departments. The most critical step in knowledge fusion is entity alignment. This project uses a combination of methods based on standardized dictionaries and semantic matching for entity alignment.

[0032] 1. A method based on standardized dictionaries Different data sources may represent the same entity and attribute in different ways, requiring semantic standardization and the establishment of cross-data source semantic mapping rules. This project first constructs a total of 2,352 core entities (such as "customer" and "critical illness insurance") and 5,368 relationships (such as "insurance application" and "claims") using relevant regulations and the company's 10-year business terminology database. It iterates quarterly based on new insurance terms and regulatory policies to ensure the timeliness of the terminology. If an entity in the initial knowledge graph matches a standardized entity in the dictionary, it is aligned to the dictionary entity and retained in the unified new knowledge graph. For example, "user" and "policyholder" are standardized to "customer".

[0033] 2. Semantic matching-based methods.

[0034] (1) Model selection: The standardized dictionary can only solve a small part of the instances. Due to the different information storage methods and usage, not all entities in the initial data can be mapped to entities in the dictionary, especially customer medical record information extracted from unstructured data. Therefore, this project will align the remaining entities that do not match the standardized dictionary using a semantic matching method and select the relatively advanced ensemble model PRASEMap.

[0035] (2) Algorithm selection: PRASEMap integrates the current mainstream entity alignment algorithms: ItransE, BootEA, NAEA, and TransEdge. The evaluation metrics adopted in this project follow the existing entity alignment algorithms. Hits@k(1-10) and the average inverse rank are used to conduct experiments on the four entity alignment algorithms. The experimental results are as follows:

[0036] Finally, PRASEMap+TransEdge was selected for alignment, and entities with a score of 0.9 or higher were selected for alignment. The aligned knowledge graph was used as the dataset for this project. The dataset size is as follows: 120,000 entities (customers, products, diseases) and 650,000 relationships (insurance, diagnosis, relatives).

[0037] In the entity set, in order to better distinguish customer information with the same name and comply with regulatory requirements, this project uses entity-attribute features to determine uniqueness, such as name + last four digits of ID number: Zhang San 3114.

[0038] III. Knowledge Processing.

[0039] Knowledge processing is based on the integrated knowledge graph. Through knowledge reasoning and business applications, the constructed knowledge graph is transformed into actionable business data.

[0040] 1. Knowledge Reasoning. Knowledge graph reasoning can deduce potential relationships between entity pairs, inferring implicit knowledge from existing knowledge to correct errors and improve the knowledge graph. By combining customer medical record information, knowledge reasoning can build a more complete customer profile. This project uses the rule-based reasoning engine VLog as the reasoning support tool and optimizes the input encoding format to support Chinese input. Example of reasoning rules: Enjoy liability (? customer, ? liability): - Insured (? customer, ? product), including liability (? product, ? liability).

[0041] Medium- to high-risk (client): - Medical history (client, hypertension), medical history (client, diabetes), insurance (client, critical illness insurance).

[0042] 2. Demand Matching. By combining the inference results with customer profiles, differentiated services can be generated, enabling business recommendations that better match customer needs.

[0043] (1) Risk level assessment: Set up a risk weight table and clarify the weight of disease, age and insurance type (e.g., hypertension 0.4, diabetes 0.3, age ≥60 years 0.2, total weight ≥0.7 → medium to high risk, 0.5-0.7 → medium risk). If it is known from the medical record data that customer A has a history of "hypertension" and "diabetes" and has purchased "critical illness insurance", then it can be inferred that "customer A's health risk level is medium to high".

[0044] (2) Implicit Relationship Mining: The customer-product path is converted into a 128-dimensional vector using PathEmbedding technology. Cosine similarity ≥ 0.8 is used to determine high demand association. For example: a. It is inferred that customer A and customer B are relatives or colleagues. If customer B has purchased critical illness insurance and received a claim, then customer A will be much more likely to accept critical illness insurance than other ordinary customers; b. If customers who purchase product C have basically also purchased product D, then it can be calculated from the similarity path that customers who have purchased product C have a greater demand for product D.

[0045] (3) Secondary reasoning for claims: If the same agent has purchased the same critical illness insurance for ≥5 people within 15 days and the accident occurred in the same prefecture-level city, the probability of triggering fraud risk is ≥90%, and the claim will be automatically pushed to manual verification.

[0046] IV. Knowledge Storage.

[0047] The knowledge graph constructed in this project is represented in the form of triples. Choosing an appropriate storage method is crucial for the later use of the knowledge graph. Relational databases are the most traditional type of database, and triples are usually stored in a two-dimensional table with n rows and 3 columns, which is also the simplest and most direct method. However, when querying information about complex diseases, it is necessary to combine multiple tables to search for information, resulting in high query costs. Neo4j is an open-source NoSQL graph database that stores triples on a graph network instead of tables, and has complete transaction features and high performance. This project chooses Neo4j graph database as the storage method for the knowledge graph and writes triples into the Neo4j graph database using Cypher statements. The current graph contains over 1 million entities and over 5 million relations, supporting an average of 100,000 concurrent queries per day.

[0048] Through the above process, the knowledge graph constructed in this application embodiment realizes the transformation of multi-source heterogeneous customer data "from scattered to integrated, from fragmented to complete, and from data to value," providing data-driven capabilities for sustained premium income growth, efficient data interoperability for information sharing and cross-departmental flow, and comprehensive support for refined customer management and risk control. However, the knowledge graph constructed in this project is a static knowledge graph and does not achieve dynamic updates, so there is a time lag in information updates. In the future, it can be connected to various systems to update data in real time, and combined with customer behavior information such as the number of times the same information is viewed and rich knowledge content from third-party data such as credit reports, to construct a more complete industry knowledge graph.

[0049] In this embodiment of the application, the knowledge graph construction process is as follows: Figure 2 As shown.

[0050] The above describes the insurance customer demand matching and business recommendation system based on knowledge graphs provided in this application. Based on the same inventive concept, this application also provides an insurance customer demand matching and business recommendation device based on knowledge graphs. Figure 3 A schematic diagram of a knowledge graph-based insurance customer demand matching and business recommendation device provided in this application embodiment is shown below. Figure 3 As shown, the device mainly includes: at least one processor 301; and a memory 302 communicatively connected to the at least one processor; wherein the memory 302 stores instructions that can be executed by the at least one processor 301, and the instructions are executed by the at least one processor 301 to enable the at least one processor 301 to complete the aforementioned knowledge graph-based insurance customer demand matching and business recommendation method.

[0051] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0052] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0053] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A knowledge graph-based method for matching insurance customer needs and recommending services, characterized in that: The method includes the following steps: Step S1: Perform targeted preprocessing on structured and unstructured data in the insurance business scenario to extract standardized entity, relationship, and attribute data; Step S2: Convert the standardized data extracted in Step S1 into the RDF / RDFS format of the unified semantic framework to obtain a multi-source independent initial knowledge graph; Step S3: By combining semantic standardization and semantic matching to form an entity alignment method, the redundancy and noise of the multi-source initial knowledge graph are eliminated, and the graphs are integrated to form a full-domain customer knowledge network. Step S4: Based on the fused omni-channel customer knowledge network, refine the customer profile through rule-based reasoning and identify potential relationships between entities; Step S5: Combining the improved customer profile and reasoning results, through risk level assessment, implicit needs mining and claims risk verification, achieve accurate matching and personalized recommendations between customer needs and insurance business.

2. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, In step S1, the structured data includes policy data and claims data. The policy data includes information such as policy number, insurance type, and premium amount. The claims data includes information such as claim number, claim amount, and claim reason. The preprocessing specifically involves using relevant functions of the pandas library to convert the structured data into a dictionary-formatted JSON file, and setting up an exception handling mechanism to repair and standardize missing fields and data with format errors.

3. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, The unstructured data mentioned in step S1 includes policy terms, medical examination reports and medical records. Knowledge extraction of unstructured data adopts the BERT model in the field of NLP. Through this model, entity recognition, relation extraction and attribute extraction are performed on the unstructured data, and finally the unstructured data is converted into a structured JSON file.

4. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, In step S2, the format conversion adopts a technical solution that combines JenaAPI with the open-source framework json2rdf to convert the JSON format data corresponding to each data source into initial knowledge graphs in RDF / RDFS format.

5. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, The semantic standardization in step S3 is based on the enterprise's historical business terminology database to build a core entity and relation dictionary. The dictionary is regularly updated according to the terms of new insurance products to ensure the timeliness of the terminology. The semantic matching adopts the integrated model PRASEMap combined with mainstream entity alignment algorithms to select entities that meet the preset scoring threshold for fusion. At the same time, the uniqueness of customer information is determined by entity-attribute features.

6. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, In step S4, rule-based reasoning uses a rule-based VLog reasoning engine. The input encoding format of the reasoning engine is optimized to support Chinese input. The reasoning rules include customer insurance liability determination rules and customer risk level determination rules. The customer profile is improved by reasoning in conjunction with customer medical record information.

7. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, The risk level assessment described in step S5 is achieved by setting a risk weight table. The risk weight table specifies the weights corresponding to diseases, ages, and types of insurance. Based on the customer's medical history, age, and insurance information, the customer's risk level is determined by combining the weight calculation results.

8. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, In step S5, the implicit demand mining uses Path Embedding technology to convert the customer-product path into a fixed-dimensional vector, and determines the high demand association between the customer and the insurance product by the cosine similarity between the vectors.

9. The knowledge graph-based insurance customer demand matching and business recommendation method according to claim 1, characterized in that, It also includes a knowledge storage step, in which the fused full-domain customer knowledge network is stored in the form of triples to the Neo4j graph database using Cypher statements. The database supports the storage and concurrent query requirements of entities and relationships in the insurance business scenario.

10. A knowledge graph-based insurance customer demand matching and business recommendation device, characterized in that, The device includes: At least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the knowledge graph-based insurance customer demand matching and business recommendation method as described in any one of claims 1-9.

Images