An intelligent underwriting method, device and system

Abstract

The application discloses an intelligent underwriting method, device and system, and solves the technical problem that an existing computer system cannot perform semantic-level analysis and matching on heterogeneous data in insurance data. The method comprises the following steps: obtaining insurance data, wherein the insurance data comprises structured data and heterogeneous data; determining an underwriting knowledge point matched with the insurance data; analyzing the heterogeneous data and extracting risk description data; converting the risk description data into a first feature vector, converting a preset standard underwriting knowledge point into a second feature vector, calculating the similarity of the two vectors, and determining matching when the similarity exceeds a preset threshold; calling a risk assessment model according to the matched underwriting knowledge point to calculate a risk level and generate an underwriting conclusion. The application realizes automatic analysis of heterogeneous data through vectorized semantic matching, and improves the understanding ability of the computer system for non-standardized expressions and the automation level of underwriting processing.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent underwriting method, apparatus and system. Background Technology

[0002] With the rapid development of the insurance industry, underwriting, as one of the core links in insurance business, directly affects the operating costs and customer satisfaction of insurance companies. The traditional underwriting process relies on manual review. Underwriters manually consult underwriting manuals and product terms based on their personal professional knowledge and experience, and review each item of the paper or electronic documents submitted by the policyholder (such as application forms, medical reports, financial statements, etc.). This process is not only time-consuming and labor-intensive, but also prone to inconsistencies or errors due to human factors.

[0003] To improve efficiency, some insurance companies have introduced rule engines for automated underwriting. Underwriting rules (such as "BMI > 30, refer to human agent" or "reject coverage for history of hypertension") are predefined as computer-executable logical rules. The system automatically matches these rules to the structured data input by the customer and provides a conclusion. While this method improves the processing speed of some standardized processes, it suffers from high rule maintenance costs, cannot handle complex, unstructured data, and lacks flexibility.

[0004] In recent years, underwriting assistance solutions have emerged that utilize machine learning models (such as logistic regression and decision trees) to score customer risk. These models, trained on historical underwriting data, can uncover potential risks from more multi-dimensional structured data, providing underwriters with valuable insights. However, these models are typically "black boxes," lacking interpretability, and still struggle to effectively utilize unstructured text and image data such as medical examination reports and medical records.

[0005] The rise of Artificial Intelligence Generated Content (AIGC) technology has brought new opportunities to the insurance industry. AIGC technology, through deep learning, natural language processing, and other techniques, can automatically extract valuable information from massive amounts of data and generate high-quality content. Therefore, applying AIGC technology to the establishment of underwriting knowledge bases and the underwriting decision-making process has become an important research direction for improving underwriting efficiency and accuracy.

[0006] However, existing technical solutions still have the following drawbacks in practical applications:

[0007] (1) Inefficiency: Manual underwriting is time-consuming, which affects customer experience. Although the rule engine is fast, it still needs to be switched to manual underwriting in complex situations, which creates a new bottleneck.

[0008] (2) High subjectivity: Differences in underwriter experience may lead to inconsistent risk assessments. Different underwriters may have different judgments on the same risk, resulting in inconsistent underwriting conclusions, which may lead to customer complaints and operational risks.

[0009] (3) Insufficient data utilization: Traditional methods are difficult to fully explore customer historical data, behavioral data and external data, especially unstructured text and image data such as physical examination reports and medical records, resulting in one-sided risk assessment.

[0010] (4) Poor dynamic adaptability: Traditional rule engines struggle to adjust underwriting strategies in real time when faced with complex risk scenarios (such as health insurance). Existing rule engines are based on static decision trees, and rule updates rely on manual modification of code or configuration, making it impossible to adjust rule weights or structures in real time based on newly entered underwriting cases. Insurance products, medical knowledge, and underwriting rules are constantly being updated, and rule engines require manual updates, resulting in significant lag.

[0011] (5) Insufficient intelligence: In the existing technology, some insurance companies use rule engines or simple machine learning models for automated underwriting, but there are still problems such as insufficient intelligence and limited dynamic decision-making capabilities, which cannot achieve end-to-end automated processing and continuous optimization of human-machine collaboration.

[0012] Therefore, there is an urgent need for an AIGC-based intelligent underwriting method to improve underwriting efficiency, enhance risk identification capabilities, make full use of multi-source heterogeneous data, and achieve dynamic optimization and continuous evolution of underwriting strategies. Summary of the Invention

[0013] This application proposes an intelligent underwriting method, apparatus, and system to solve the problem that existing computer systems cannot perform semantic-level parsing and matching of heterogeneous data in insurance application data.

[0014] In a first aspect, embodiments of this application provide an intelligent underwriting method, comprising the following steps:

[0015] Obtain insurance application data; the insurance application data includes structured data and heterogeneous data;

[0016] Identify the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions, and medical assessments;

[0017] The heterogeneous data is parsed to extract risk description data; the risk description data is converted into a first feature vector; the preset standard underwriting knowledge points are converted into a second feature vector; the similarity between the first feature vector and the second feature vector is calculated; when the similarity exceeds a preset threshold, it is determined that the risk description data matches the standard underwriting knowledge points.

[0018] Based on the matched underwriting knowledge points, the risk assessment model is invoked to calculate the risk level, and an underwriting conclusion is generated based on the risk level.

[0019] In one embodiment, constructing a knowledge base to determine underwriting knowledge points specifically includes the following steps:

[0020] Obtain underwriting knowledge documents;

[0021] The underwriting knowledge document is parsed and segmented to generate underwriting knowledge points;

[0022] The underwriting knowledge points are vectorized and stored in the knowledge base.

[0023] In one embodiment, the step further includes:

[0024] In response to the fact that the similarity between the risk description data and all preset standard underwriting knowledge points is lower than a preset threshold, supplementary inquiry data is generated and sent.

[0025] In one embodiment, the step further includes:

[0026] In response to receiving a modification instruction for the underwriting conclusion, a feedback signal is generated;

[0027] The feedback signal is input into the reinforcement learning model to optimize the parameters of the risk assessment model.

[0028] In one embodiment, parsing the heterogeneous data includes the following steps:

[0029] When the heterogeneous data contains semi-structured data, extract the tags and corresponding data content from the semi-structured data;

[0030] When the heterogeneous data contains unstructured data, text recognition is performed on the unstructured data to extract the text content;

[0031] The extracted data content, the text content, and the structured data are combined to form the risk description data.

[0032] In one embodiment, the step further includes:

[0033] Obtain confirmation or modification instructions for underwriting conclusions from historical underwriting records;

[0034] Calculate the adjustment amount of the current preset threshold based on the confirmation instruction or modification instruction;

[0035] The preset threshold is updated according to the adjustment amount.

[0036] Secondly, embodiments of this application also provide an intelligent underwriting device for implementing the intelligent underwriting method described in any embodiment of the first aspect, comprising: an acquisition module for acquiring insurance application data; the insurance application data includes structured data and heterogeneous data; a determination module for determining underwriting knowledge points matching the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms, and medical evaluations; and further for parsing the heterogeneous data to extract risk description data; converting the risk description data into a first feature vector; converting preset standard underwriting knowledge points into a second feature vector; calculating the similarity between the first feature vector and the second feature vector; and determining that the risk description data matches the standard underwriting knowledge points when the similarity exceeds a preset threshold; and a generation module for calculating a risk level by calling a risk assessment model based on the matched underwriting knowledge points, and generating an underwriting conclusion based on the risk level.

[0037] Thirdly, embodiments of this application also provide an intelligent underwriting system, comprising: an underwriting server, deployed with the intelligent underwriting device as described in the second aspect embodiment, or configured to execute the intelligent underwriting method as described in any one of the first aspect embodiments; a data acquisition server, connected to the underwriting server, for providing insurance application data to the underwriting server; the insurance application data includes structured data and heterogeneous data; a knowledge base server, connected to the underwriting server, for providing underwriting knowledge points to the underwriting server; the underwriting knowledge points include underwriting rules, insurance terms, and medical evaluations; and a management terminal, connected to the underwriting server, for receiving underwriting conclusions generated by the underwriting server and returning confirmation or modification instructions for the underwriting conclusions to the underwriting server.

[0038] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any one of the embodiments of the first aspect.

[0039] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the method as described in any one of the embodiments of the first aspect.

[0040] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0041] This application's embodiments achieve semantic-level parsing and matching of heterogeneous data in insurance application data by converting risk description data into a first feature vector and pre-set standard underwriting knowledge points into a second feature vector, and then calculating the similarity between the two for matching. This enables the computer system to understand the correspondence between non-standardized expressions and standard underwriting knowledge points, solving the technical problem that traditional computer systems cannot perform semantic understanding of unstructured data. By generating and sending supplementary query data in response to the similarity between risk description data and all standard underwriting knowledge points being lower than a preset threshold, the computer system can automatically identify information missing scenarios and generate targeted follow-up questions, improving the automation processing capability of the computer system without manual intervention. By generating feedback signals in response to receiving modification instructions for underwriting conclusions, and inputting the feedback signals into a reinforcement learning model to optimize the parameters of the risk assessment model, the computer system can continuously learn from manual corrections, achieving dynamic optimization of model parameters and self-evolution of system performance. By classifying heterogeneous data, extracting tags and content from semi-structured data, and performing text recognition on unstructured data, the extracted results are merged with structured data to form risk description data. This achieves unified processing and integrated utilization of multi-source heterogeneous data, improving the computer system's ability to process complex insurance data. By calculating and updating preset thresholds based on confirmation or modification instructions in historical underwriting records, the computer system can adaptively adjust matching thresholds, improving its adaptability and accuracy in different business scenarios. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a flowchart of an intelligent underwriting method according to an embodiment of this application;

[0044] Figure 2 This is a flowchart of an intelligent underwriting method including a supplementary inquiry step, according to an embodiment of this application.

[0045] Figure 3 This is a flowchart of an intelligent underwriting method including reinforcement learning optimization steps, according to an embodiment of this application.

[0046] Figure 4 This is a flowchart of an intelligent underwriting method including a threshold dynamic adjustment step, according to an embodiment of this application.

[0047] Figure 5 This is a structural diagram of an intelligent underwriting device according to an embodiment of this application;

[0048] Figure 6 This is a structural diagram of an intelligent underwriting system according to an embodiment of this application. Detailed Implementation

[0049] 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.

[0050] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0051] Figure 1 This is a flowchart of an intelligent underwriting method according to an embodiment of this application.

[0052] In a first aspect, embodiments of this application provide an intelligent underwriting method, including the following steps: 110~140.

[0053] Step 110: Obtain insurance data; the insurance data includes structured data and heterogeneous data.

[0054] The insurance data refers to the various information submitted by the policyholder when applying for insurance.

[0055] The structured data refers to data with a fixed format and predefined patterns, such as information like region, age, occupation, and income in a form filled out by the policyholder. This data is usually stored in a relational database as fields.

[0056] Heterogeneous data refers to data with inconsistent formats that are difficult to describe using a fixed pattern, including semi-structured data and unstructured data.

[0057] Semi-structured data refers to data with certain labels but no strict pattern, such as medical records stored in XML or JSON format, which contains labels and corresponding data content.

[0058] Unstructured data refers to data without a predefined format, such as medical report text, physical examination report PDF files, and medical image images. This heterogeneous data comes from diverse sources and has varying formats, making it difficult for traditional computer systems to directly parse and process it.

[0059] The insurance data may also include external data such as credit data, historical insurance records, and claims records, which are obtained from third-party systems through interfaces.

[0060] In practice, insurance application data can be obtained through various channels. For example, when an applicant fills out an online application form, the system directly collects structured fields such as age and occupation. Simultaneously, PDF medical examination reports or medical image files uploaded by the applicant are received via an interface and stored as unstructured data. In scenarios involving hospital system integration, electronic medical records in XML format may be received and processed as semi-structured data.

[0061] Step 120: Determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions and medical assessments.

[0062] The underwriting knowledge points refer to the professional knowledge units used for underwriting decisions.

[0063] The underwriting rules refer to the operational guidelines in underwriting practice, such as the various risk handling methods stipulated in underwriting and reinsurance practices.

[0064] The underwriting rules specifically include risk classification standards and handling rules for various diseases, occupations, and insured amounts in underwriting and reinsurance practices.

[0065] The insurance terms refer to the specific stipulations in the insurance product contract.

[0066] For example, the insurance terms include specific provisions such as insurance liability, exclusions, waiting period, and deductible.

[0067] The medical evaluation points refer to the key points of underwriting assessment for various diseases in the medical field, such as the underwriting principles for various diseases in the disease evaluation database.

[0068] For example, the medical assessment specifically includes medical evaluation data such as disease name, classification criteria, risk factors, probability of complications, and impact on life expectancy.

[0069] Identifying underwriting knowledge points that match the insurance application data refers to finding relevant underwriting rules, insurance terms, and medical assessments from a knowledge base based on key information in the application data.

[0070] For example, when the insurance application data includes disease-related information, the system retrieves the corresponding medical assessments, relevant underwriting rules, and the terms and conditions regarding the disease from the knowledge base. This information will serve as the basis for subsequent risk assessments.

[0071] In one embodiment, constructing a knowledge base for determining the underwriting knowledge points in step 120 specifically includes the following steps:

[0072] Step 100-1: Obtain underwriting knowledge documents.

[0073] The underwriting knowledge documents refer to original documents containing professional knowledge related to underwriting, including underwriting and reinsurance practices, insurance terms, medical examination exemption tables, medical reviews, etc. These documents can be automatically uploaded via an interface.

[0074] Step 100-2: Parse and slice the underwriting knowledge document to generate underwriting knowledge points.

[0075] The parsing refers to converting underwriting knowledge documents into a processable data format.

[0076] The slicing process refers to dividing the parsed content into independent underwriting knowledge points, with each knowledge point corresponding to a complete underwriting rule, an insurance clause, or a medical assessment.

[0077] The slicing process divides the data according to semantic boundaries. For example, the complete chapter on "hypertension" in the underwriting manual is divided into a knowledge point to ensure that each knowledge point contains complete contextual information.

[0078] Step 100-3: Vectorize and store the underwriting knowledge points in the knowledge base.

[0079] Vectorization refers to converting underwriting knowledge points into vector form for storage, so as to facilitate subsequent semantic retrieval and matching.

[0080] The vectorization method maps underwriting knowledge points in text form to a high-dimensional vector space through an embedding model, so that texts with similar semantics are close in distance in the vector space.

[0081] In one embodiment, the knowledge base stores a list of knowledge documents such as underwriting and reinsurance practices, insurance terms, medical examination exemption tables, and medical reviews. It supports dynamic management and automatic slicing of the knowledge base, continuously updates the latest medical and disease review points, and provides a knowledge base for automatically generating underwriting suggestions and conclusions in combination with product requirements.

[0082] In one embodiment, the knowledge base includes a rigid rule base and a flexible rule base. The underwriting knowledge points in the rigid rule base maintain fixed parameters during model optimization and are unaffected by feedback signal updates; the underwriting knowledge points in the flexible rule base allow for parameter optimization based on feedback signals.

[0083] In one embodiment, the continuous updating of the knowledge base includes the following steps:

[0084] Step 100-4: Regularly obtain newly added underwriting knowledge documents.

[0085] Step 100-5: Parse and slice the newly added underwriting knowledge file to generate new underwriting knowledge points.

[0086] Step 100-6: Vectorize the new underwriting knowledge points and store them in the knowledge base to achieve dynamic updates of the knowledge base.

[0087] In one embodiment, the triggering conditions for knowledge base updates include: immediate parsing triggered by new file uploads, timed incremental updates, and manual updates; after the updated knowledge points are reviewed and confirmed, an effective time is set, and the new version is implemented according to the effective time.

[0088] Step 130: Analyze the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; when the similarity exceeds a preset threshold, determine that the risk description data matches the standard underwriting knowledge points.

[0089] The risk description data refers to information related to underwriting risks extracted from heterogeneous data, such as the description of medical history in medical report texts.

[0090] The risk description data specifically includes disease name, diagnosis time, test result values, medication use, surgical records, etc.

[0091] The first feature vector refers to the numerical vector representation obtained by transforming the risk description data through a vectorization model.

[0092] The standard underwriting knowledge points refer to the underwriting rules, insurance terms, and medical evaluations pre-set in the knowledge base.

[0093] The second feature vector refers to the numerical vector representation obtained by converting standard underwriting knowledge points through the same vectorization model.

[0094] The similarity refers to the degree of closeness between the first feature vector and the second feature vector in the vector space.

[0095] The preset threshold refers to a pre-set similarity threshold.

[0096] In practice, the heterogeneous data is first parsed, and NLP technology is used to extract key information from medical report texts, such as medical history, as risk description data. This risk description data is then input into a vectorized model, outputting the corresponding first feature vector. Simultaneously, pre-set standard underwriting knowledge points from the knowledge base are input into the same vectorized model, outputting the corresponding second feature vector.

[0097] Calculate the similarity between the first feature vector and the second feature vector. When the similarity exceeds a preset threshold, it is determined that the risk description data matches the standard underwriting knowledge points.

[0098] For example, once the disease description in the medical report text is successfully matched with the disease rating in the knowledge base, underwriting processing principles can be generated in conjunction with the product terms.

[0099] In one embodiment, after a match is determined, a matching result is generated that includes the customer's medical history and underwriting principles. The customer's medical history refers to a structured description of information such as disease history, medication history, and surgical history extracted from risk description data. The underwriting principles are determined based on the matched medical assessments and underwriting rules.

[0100] In one embodiment, step 130 involves parsing the heterogeneous data, including the following steps:

[0101] Step 130-1: When the heterogeneous data contains semi-structured data, extract the tags and corresponding data content from the semi-structured data.

[0102] The semi-structured data refers to data that has a certain structure but no strict pattern. Tags and corresponding data content are extracted and used as components of the risk description data.

[0103] In one embodiment, before parsing the heterogeneous data, sensitive personal information is identified, including name, ID number, mobile phone number, and detailed address; the identified sensitive personal information is de-identified to generate de-identified risk description data; only the de-identified data can enter the subsequent vectorized matching and model training stages.

[0104] Step 130-2: When the heterogeneous data contains unstructured data, perform text recognition on the unstructured data to extract the text content.

[0105] Unstructured data refers to data without a fixed format, such as medical report text, physical examination report PDF, or images. The process involves image recognition to convert PDF or image files into text and extracting the text content.

[0106] Step 130-3: Merge the extracted data content, the text content, and the structured data to form the risk description data.

[0107] The data content extracted from semi-structured data, the text content identified from unstructured data, and information such as region, age, and occupation from structured data are combined to form complete risk description data, which is then used for subsequent vectorized matching processing.

[0108] Step 140: Based on the matched underwriting knowledge points, call the risk assessment model to calculate the risk level, and generate an underwriting conclusion based on the risk level.

[0109] The risk assessment model refers to the model used to calculate the risk level.

[0110] The underwriting conclusion refers to the underwriting result determined based on the risk level.

[0111] Based on the matched underwriting knowledge points and risk description data, the risk assessment model is invoked to calculate the risk level. An underwriting conclusion is generated based on the risk level, such as standard coverage, coverage with increased premium, or exclusions. When the risk is controllable, underwriting is automatically approved; when uncertainty exists, manual review or adjustments to premiums and exclusions are triggered.

[0112] In one embodiment, the underwriting conclusion generated in step 140 includes standard coverage, coverage with additional premium, exclusions, deferred coverage, or rejection. When the risk is controllable, the system automatically approves the underwriting; when there is uncertainty, the system triggers manual review, or automatically adjusts the premium or exclusions based on the matched underwriting knowledge points before generating an underwriting conclusion.

[0113] In one embodiment, the customer submits an insurance application and uploads a medical examination report, which is in PDF or image format. The system converts the medical examination report into text using image recognition and transmits the text data to the underwriting intelligent agent interface. The AIGC engine parses the report text, combines key points of disease review with the terms of the insurance product to be purchased, and outputs an underwriting conclusion. The underwriting conclusion includes standard coverage, coverage with additional premium, or rejection.

[0114] In one embodiment, an underwriting report is generated based on the matched underwriting knowledge points and risk levels. The underwriting report includes an underwriting conclusion and a corresponding decision-making basis text. The decision-making basis text includes matching medical reviews, cited insurance clauses, and risk point identification criteria.

[0115] Furthermore, when generating the underwriting report, decision-making basis data is recorded simultaneously. This decision-making basis data includes: matching standard underwriting knowledge point identifiers, similarity scores, corresponding insurance clause numbers, original medical review texts, and risk description data sources. The decision-making basis data is then associated with the underwriting conclusion and stored in the underwriting record database.

[0116] Furthermore, the underwriting report employs a combination of a fixed framework template and dynamic content generation. The fixed framework template includes modules for policyholder information, risk overview, matching knowledge points, underwriting conclusion, and supporting evidence. The dynamic content is generated by the AIGC engine using natural language to describe the risk, provide medical explanations, and interpret the terms of the policy based on the characteristics of the case. This report can be reviewed by underwriters or directly output as the final underwriting result.

[0117] Figure 2This is a flowchart of an intelligent underwriting method including a supplementary inquiry step, according to an embodiment of this application, including steps 210-250.

[0118] Step 210: Obtain insurance data; the insurance data includes structured data and heterogeneous data.

[0119] This step and Figure 1 Step 110 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 110, which will not be repeated here.

[0120] Step 220: Determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions and medical assessments.

[0121] This step and Figure 1 Step 120 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 120, which will not be repeated here.

[0122] Step 230: Analyze the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; when the similarity exceeds a preset threshold, determine that the risk description data matches the standard underwriting knowledge points.

[0123] This step and Figure 1 Step 130 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 130, which will not be repeated here.

[0124] Step 240: Based on the matched underwriting knowledge points, call the risk assessment model to calculate the risk level, and generate an underwriting conclusion based on the risk level.

[0125] This step and Figure 1 Step 140 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 140, which will not be repeated here.

[0126] In one embodiment, the step further includes:

[0127] Step 250: In response to the fact that the similarity between the risk description data and all preset standard underwriting knowledge points is lower than a preset threshold, supplementary inquiry data is generated and sent.

[0128] The supplementary inquiry data refers to the supplementary questions sent to the policyholder in cases where information is missing or cannot be matched.

[0129] During the underwriting process, there may be situations where the risk description data cannot be matched with any of the standard underwriting knowledge points in the knowledge base.

[0130] For example, a medical report might mention a rare disease or a special examination, but the knowledge base doesn't yet include underwriting assessments for that disease or examination; or the risk description data might be too vague to determine which specific underwriting knowledge point it refers to. In such cases, the system calculates the similarity between the risk description data and all standard underwriting knowledge points. If all similarities are below a preset threshold, it indicates that the current information is insufficient to make an underwriting decision.

[0131] At this point, the system automatically generates supplementary inquiry data. This supplementary inquiry data consists of targeted questions designed to obtain more information from the policyholder.

[0132] For example, when a medical record mentions "surgical history" but does not specify the name and time of the surgery, the generated supplementary inquiry data can be an in-depth inquiry into the surgical history.

[0133] When a certain indicator in a physical examination report is abnormal but the description is vague, the generated supplementary inquiry data can be "Please provide a detailed examination report for this indicator".

[0134] After generating supplementary inquiry data, the system sends it to the policyholder, waiting for the policyholder to provide additional information. After the policyholder provides the supplementary information, the system re-executes steps 210 to 240 until a complete underwriting conclusion is obtained.

[0135] In one embodiment, when the similarity between the risk description data and all preset standard underwriting knowledge points is lower than a preset threshold, the AIGC engine is invoked to generate a temporary underwriting suggestion based on general medical knowledge and similar disease assessments, and the temporary underwriting suggestion is marked as pending manual confirmation; in response to the manual review and confirmation instruction, the risk description data and the confirmed underwriting conclusion are stored as new underwriting knowledge points in the knowledge base.

[0136] Figure 3 This is a flowchart of an intelligent underwriting method including reinforcement learning optimization steps, according to an embodiment of this application, including steps 310-360.

[0137] Step 310: Obtain insurance data; the insurance data includes structured data and heterogeneous data.

[0138] This step and Figure 1 Step 110 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 110, which will not be repeated here.

[0139] Step 320: Determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions and medical assessments.

[0140] This step and Figure 1Step 120 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 120, which will not be repeated here.

[0141] Step 330: Analyze the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; when the similarity exceeds a preset threshold, determine that the risk description data matches the standard underwriting knowledge points.

[0142] This step and Figure 1 Step 130 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 130, which will not be repeated here.

[0143] Step 340: Based on the matched underwriting knowledge points, call the risk assessment model to calculate the risk level, and generate an underwriting conclusion based on the risk level.

[0144] This step and Figure 1 Step 140 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 140, which will not be repeated here.

[0145] In one embodiment, the step further includes:

[0146] Step 350: In response to receiving a modification instruction for the underwriting conclusion, generate a feedback signal.

[0147] The modification instruction refers to the operation instruction given by the underwriter to correct the underwriting conclusion generated by the system.

[0148] The feedback signal refers to the data signal generated according to the modification instruction and used to train and optimize the model.

[0149] During the underwriting process, the underwriting conclusion generated by the system may require manual review. If the underwriter believes the conclusion is inaccurate when reviewing it, they can issue a modification instruction to correct the system conclusion to the correct underwriting conclusion.

[0150] For example, the system generates a conclusion of "insurance with increased premium" based on risk description data, but the underwriter, after review, believes it should be changed to "exclusion of liability," and issues a modification instruction. Upon receiving this modification instruction, the system converts it into a feedback signal that can be used for model training. This feedback signal includes the original input data, the system's erroneous conclusion, and the correct conclusion after manual correction.

[0151] Step 360: Input the feedback signal into the reinforcement learning model to optimize the parameters of the risk assessment model.

[0152] The reinforcement learning model refers to a machine learning model that optimizes decision-making strategies using reward signals. The optimization refers to adjusting the parameters of the risk assessment model so that its subsequent output more closely approximates the judgment of a human expert.

[0153] The feedback signal generated in step 350 is input into the reinforcement learning model as training data. Based on the correction information in the feedback signal, the reinforcement learning model calculates the difference between the current risk assessment model output and the manually corrected conclusion, generating reward or penalty signals to adjust the parameters of the risk assessment model accordingly. Through multiple iterative optimizations, the parameters of the risk assessment model are continuously updated, its judgment accuracy gradually improves, and the underwriting conclusions output by the system become increasingly closer to the professional judgment of underwriters.

[0154] In one embodiment, when multiple modification instructions are received for the same underwriting conclusion and the content of the instructions is inconsistent, the credit weight of the underwriter corresponding to each modification instruction is obtained; the modification instructions are weighted according to the credit weight, and the opinion with the highest weighted score is used as the standard feedback signal; the credit weight is dynamically calculated based on the underwriter's job level, historical underwriting accuracy rate and professional qualifications.

[0155] In one embodiment, the parameters of the risk assessment model are updated periodically; the feedback signal corresponding to a single modification instruction is immediately added to the incremental parameter table for rapid response to similar cases; batch training is performed weekly, and the incremental parameters are merged into the main model.

[0156] In one embodiment, the step further includes:

[0157] Step 370: Optimize the prompt word strategy of the AIGC engine based on the feedback signal.

[0158] The prompt word strategy refers to the prompt word template used when calling the AIGC engine.

[0159] The feedback signal is input into the reinforcement learning model to optimize the prompt word strategy and adjust the relevant content in the prompt word template, so that the AIGC engine can more accurately understand the risk description data in subsequent processing.

[0160] In one embodiment, the optimization of the prompt word strategy is adjusted based on the type of comprehension bias identified in the feedback signal; the generated optimization suggestions are then manually confirmed to take effect.

[0161] Figure 4 This is a flowchart of an intelligent underwriting method including a threshold dynamic adjustment step, according to an embodiment of this application, including steps 410-470.

[0162] Step 410: Obtain insurance data; the insurance data includes structured data and heterogeneous data.

[0163] This step and Figure 1 Step 110 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 110, which will not be repeated here.

[0164] Step 420: Determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions and medical assessments.

[0165] This step and Figure 1 Step 120 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 120, which will not be repeated here.

[0166] Step 430: Analyze the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; when the similarity exceeds a preset threshold, determine that the risk description data matches the standard underwriting knowledge points.

[0167] This step and Figure 1 Step 130 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 130, which will not be repeated here.

[0168] Step 440: Based on the matched underwriting knowledge points, call the risk assessment model to calculate the risk level, and generate an underwriting conclusion based on the risk level.

[0169] This step and Figure 1 Step 140 in the illustrated embodiment is the same, and the specific implementation method is described in the description of step 140, which will not be repeated here.

[0170] In one embodiment, the step further includes:

[0171] Step 450: Obtain confirmation or modification instructions for the underwriting conclusion from historical underwriting records.

[0172] The historical underwriting records refer to all underwriting case data that the system has processed in the past.

[0173] The confirmation instruction refers to the instruction given by the underwriter to acknowledge the underwriting conclusion generated by the system.

[0174] The modification instruction refers to the instruction given by the underwriter to correct the underwriting conclusion generated by the system.

[0175] During system operation, each underwriting process generates an underwriting record, which includes the underwriter's feedback on the system's conclusions.

[0176] This step retrieves confirmation and modification instructions from these historical records as the basis for subsequent threshold adjustments. For example, the system can periodically retrieve confirmation and modification instructions issued by underwriters in all underwriting cases within the past month.

[0177] Step 460: Calculate the adjustment amount of the current preset threshold according to the confirmation instruction or modification instruction.

[0178] The preset threshold refers to the similarity threshold used in step 430 to determine whether the risk description data matches the standard underwriting knowledge points.

[0179] The adjustment amount refers to the value by which the threshold needs to be increased or decreased.

[0180] Based on confirmation and modification instructions in historical underwriting records, the system analyzes the rationality of the current threshold. If cases within a certain similarity range are frequently modified by underwriters, it indicates that the threshold setting for that range may need adjustment. The system calculates the threshold adjustment amount based on the distribution of modification instructions.

[0181] In one embodiment, the similarity is divided into continuous intervals, and the proportion of the core preservation conclusions that are manually modified is counted in each interval; the interval with the highest modification ratio is determined, and the current similarity threshold is unidirectionally fine-tuned in the direction of reducing the modification ratio of that interval by using a preset fixed step size.

[0182] Step 470: Update the preset threshold according to the adjustment amount.

[0183] The adjustment amount calculated in step 460 is applied to the current preset threshold to complete the threshold update.

[0184] The updated threshold will be used for similarity determination in step 430 of the subsequent underwriting process.

[0185] Figure 5 This is a structural diagram of an intelligent underwriting device according to an embodiment of this application.

[0186] Secondly, embodiments of this application also provide an intelligent underwriting device for implementing the intelligent underwriting method described in any embodiment of the first aspect, comprising:

[0187] The acquisition module 501 is used to acquire insurance data; the insurance data includes structured data and heterogeneous data.

[0188] The determination module 502 is used to determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and medical evaluations; it is also used to parse the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; when the similarity exceeds a preset threshold, it is determined that the risk description data matches the standard underwriting knowledge points.

[0189] The generation module 503 is used to call the risk assessment model to calculate the risk level based on the matched underwriting knowledge points, and generate an underwriting conclusion based on the risk level.

[0190] Furthermore, the acquisition module 501 includes a first acquisition unit for acquiring insurance data; the insurance data includes structured data and heterogeneous data.

[0191] The determining module 502 includes a first determining unit, used to determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and medical evaluations.

[0192] The determining module 502 further includes a second determining unit, used to parse the heterogeneous data and extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; calculate the similarity between the first feature vector and the second feature vector; and determine that the risk description data matches the standard underwriting knowledge points when the similarity exceeds a preset threshold.

[0193] The generation module 503 includes a first generation unit, which is used to call a risk assessment model to calculate the risk level based on the matched underwriting knowledge points, and generate an underwriting conclusion based on the risk level.

[0194] In one embodiment, the determining module 502 further includes a third determining unit for constructing a knowledge base for determining underwriting knowledge points, specifically including: acquiring an underwriting knowledge file; parsing and slicing the underwriting knowledge file to generate underwriting knowledge points; and vectorizing and storing the underwriting knowledge points in the knowledge base.

[0195] This embodiment is used to implement Figure 1 The function of building a knowledge base in steps 100-1 to 100-3.

[0196] In one embodiment, the generation module 503 further includes a second generation unit, used to generate and send supplementary inquiry data in response to the risk description data having a similarity to all preset standard underwriting knowledge points that is lower than a preset threshold.

[0197] This embodiment is used to implement Figure 2 The function of generating supplementary query data in step 250.

[0198] In one embodiment, the generation module 503 further includes a third generation unit, configured to generate a feedback signal in response to receiving a modification instruction for the underwriting conclusion; and input the feedback signal into a reinforcement learning model to optimize the parameters of the risk assessment model.

[0199] This embodiment is used to implement Figure 3 The function of optimizing model parameters in steps 350-360.

[0200] In one embodiment, the determining module 502 further includes a fourth determining unit, used to parse the heterogeneous data, specifically including: when the heterogeneous data contains semi-structured data, extracting tags and corresponding data content from the semi-structured data; when the heterogeneous data contains unstructured data, performing text recognition on the unstructured data to extract text content; and merging the extracted data content, the text content, and the structured data as the risk description data.

[0201] This embodiment is used to implement Figure 1 The function of parsing heterogeneous data in steps 130-1 to 130-3.

[0202] In one embodiment, the generation module 503 further includes a fourth generation unit, used to obtain confirmation instructions or modification instructions for underwriting conclusions in historical underwriting records; calculate the adjustment amount of the current preset threshold according to the confirmation instructions or modification instructions; and update the preset threshold according to the adjustment amount.

[0203] This embodiment is used to implement Figure 4 The threshold dynamic adjustment function in steps 450-470.

[0204] Figure 6 This is a structural diagram of an intelligent underwriting system according to an embodiment of this application.

[0205] Thirdly, embodiments of this application also provide an intelligent underwriting system, comprising:

[0206] Server-side 600, which specifically includes:

[0207] Underwriting server 601 is deployed with the intelligent underwriting device as described in the second aspect embodiment, or configured to execute the intelligent underwriting method as described in any embodiment of the first aspect, for implementing the following: Figures 1-4 The method steps are described below.

[0208] In one embodiment, the underwriting server 601 includes an intelligent decision-making unit that integrates a data acquisition module, an AIGC intelligent analysis engine, a risk assessment model, and a decision optimization module. These modules work together to complete the entire process from obtaining insurance data to generating underwriting conclusions.

[0209] The data acquisition module is used to integrate multi-source data, including customer information, medical records, credit data, and image data, to achieve... Figure 1 The function of obtaining insurance data in step 110.

[0210] The AIGC intelligent analysis engine is used to automatically parse customer data using natural language processing and generative AI technologies to achieve... Figure 1 The function of parsing and semantic matching heterogeneous data in step 130.

[0211] The risk assessment model is used to calculate customer risk levels in real time based on deep learning, and is used to achieve... Figure 1 The function of calculating the risk level in step 140.

[0212] The decision optimization module is used to combine reinforcement learning to optimize the underwriting strategy and dynamically adjust the underwriting conditions, in order to achieve... Figure 3 The function of optimizing model parameters in steps 350-360 and Figure 4 The function of dynamically adjusting the threshold in steps 450-470.

[0213] A knowledge base database 602, connected to the underwriting server 601, is used to store underwriting knowledge points; these knowledge points include underwriting rules, insurance terms, and medical evaluations. This database is used to implement... Figure 1 The data source function for determining underwriting knowledge points in step 120, and Figure 1 The knowledge base construction and storage functions in steps 100-1 to 100-3.

[0214] The insurance application terminal 700 is communicatively connected to the server side 600 and is used to submit insurance application data; the insurance application data includes structured data and heterogeneous data. The insurance application terminal is used to implement... Figure 1 The function of obtaining insurance data in step 110.

[0215] The management terminal 800, communicatively connected to the server 600, is used to receive the underwriting conclusion generated by the underwriting server 601 and return a confirmation instruction or modification instruction for the underwriting conclusion to the underwriting server 601. The management terminal is used to implement... Figure 3 The function of receiving modification instructions in step 350, and Figure 4 The function of obtaining confirmation or modification instructions in step 450.

[0216] In one embodiment, the server side 600 further includes an interface server 603 connected to the underwriting server 601, used to provide a data interaction interface to the outside world and receive query requests and operation instructions from the management terminal. This server is used to implement... Figure 2 The function of sending supplementary query data in step 250, and Figure 1 The function of obtaining underwriting knowledge documents through the interface in step 100-1.

[0217] In one embodiment, the server side 600 further includes a file server 604 connected to the underwriting server 601, used to store heterogeneous data files such as medical examination reports (PDF) and medical image files uploaded by the policyholder. This server is used to implement... Figure 1 The function of acquiring unstructured data in step 110, and Figure 1 The data storage function in step 130-2 is performed on unstructured data before text recognition.

[0218] In one embodiment, the server-side 600 further includes an underwriting record database 605, connected to the underwriting server 601, for storing historical underwriting records, including application data, risk description data, matching underwriting knowledge points, risk levels, and underwriting conclusions. This database is used to implement... Figure 4 The function of obtaining historical underwriting records in step 450.

[0219] In one embodiment, the server-side 600 further includes a feedback record database 606, connected to the underwriting server 601, for storing underwriters' confirmation and modification instructions regarding underwriting conclusions, as well as generated feedback signals. This database is used to implement... Figure 3 The storage function of the feedback signal in step 350, and Figure 4 The function of obtaining confirmation instructions and modification instructions in step 450.

[0220] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] Therefore, this application also proposes a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the methods described in any embodiment of this application.

[0222] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0223] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0224] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0225] Furthermore, this application also proposes an electronic device (or computing device) including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any embodiment of this application.

[0226] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, a network interface, and memory. Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media. Computer-readable media includes both permanent and non-persistent, removable and non-removable media; information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of a program, or other data.

[0227] 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 process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0228] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0229] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0230] The above description is merely an embodiment of this application and is not intended to limit the scope of 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 principles of this application should be included within the scope of the claims of this application.

Claims

1. An intelligent underwriting method, characterized in that, Includes the following steps: Obtain insurance application data; the insurance application data includes structured data and heterogeneous data; Identify the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and conditions, and medical assessments; The heterogeneous data is parsed to extract risk description data; the risk description data is converted into a first feature vector; the preset standard underwriting knowledge points are converted into a second feature vector; and the similarity between the first feature vector and the second feature vector is calculated. When the similarity exceeds a preset threshold, it is determined that the risk description data matches the standard underwriting knowledge points; Based on the matched underwriting knowledge points, the risk assessment model is invoked to calculate the risk level, and an underwriting conclusion is generated based on the risk level.

2. The intelligent underwriting method according to claim 1, characterized in that, Building a knowledge base that defines underwriting knowledge points involves the following steps: Obtain underwriting knowledge documents; The underwriting knowledge document is parsed and segmented to generate underwriting knowledge points; The underwriting knowledge points are vectorized and stored in the knowledge base.

3. The intelligent underwriting method according to claim 1, characterized in that, It also includes the following steps: In response to the fact that the similarity between the risk description data and all preset standard underwriting knowledge points is lower than a preset threshold, supplementary inquiry data is generated and sent.

4. The intelligent underwriting method according to claim 1, characterized in that, It also includes the following steps: In response to receiving a modification instruction for the underwriting conclusion, a feedback signal is generated; The feedback signal is input into the reinforcement learning model to optimize the parameters of the risk assessment model.

5. The intelligent underwriting method according to claim 1, characterized in that, The process of parsing the heterogeneous data includes the following steps: When the heterogeneous data contains semi-structured data, extract the tags and corresponding data content from the semi-structured data; When the heterogeneous data contains unstructured data, text recognition is performed on the unstructured data to extract the text content; The extracted data content, the text content, and the structured data are combined to form the risk description data.

6. The intelligent underwriting method according to claim 1, characterized in that, It also includes the following steps: Obtain confirmation or modification instructions for underwriting conclusions from historical underwriting records; Calculate the adjustment amount of the current preset threshold based on the confirmation instruction or modification instruction; The preset threshold is updated according to the adjustment amount.

7. An intelligent underwriting device for implementing the intelligent underwriting method according to any one of claims 1-6, characterized in that, include: The acquisition module is used to acquire insurance data; The insurance data includes structured data and heterogeneous data; The determination module is used to determine the underwriting knowledge points that match the insurance application data; the underwriting knowledge points include underwriting rules, insurance terms and medical evaluations; it is also used to parse the heterogeneous data, extract risk description data; convert the risk description data into a first feature vector; convert the preset standard underwriting knowledge points into a second feature vector; and calculate the similarity between the first feature vector and the second feature vector. When the similarity exceeds a preset threshold, it is determined that the risk description data matches the standard underwriting knowledge points; The generation module is used to call the risk assessment model to calculate the risk level based on the matched underwriting knowledge points, and generate an underwriting conclusion based on the risk level.

8. An intelligent underwriting system, characterized in that, Include: An underwriting server is deployed with the intelligent underwriting device as described in claim 7, or is configured to execute the intelligent underwriting method as described in any one of claims 1-6; A data acquisition server, connected to the underwriting server, is used to provide insurance data to the underwriting server; The insurance data includes structured data and heterogeneous data; A knowledge base server, connected to the underwriting server, is used to provide underwriting knowledge points to the underwriting server; the underwriting knowledge points include underwriting rules, insurance terms, and medical evaluations; The management terminal is connected to the underwriting server and is used to receive the underwriting conclusion generated by the underwriting server and return a confirmation instruction or modification instruction for the underwriting conclusion to the underwriting server.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-6.

Images