Method, device and equipment for generating dynamic insurance policy and storage medium

Abstract

The embodiment of the application discloses a dynamic insurance policy generation method, device, equipment and storage medium. It relates to the fields of medical treatment and financial insurance. The method comprises the following steps: obtaining open user information and environment information; extracting risk keywords from the user information and the environment information; mining potential correlation in the keywords to obtain a preliminary risk graph containing a risk transmission path, the weight of the edges of the graph changing with the user information and the environment information; mapping the feature vector corresponding to the graph from risk representation to insurance policy structure to obtain an insurance policy draft; and adjusting the insurance policy draft in response to a user modification instruction. In the application, potential dependent correlation in the user information and the environment information is mined, the risk graph obtained from the correlation is mapped from risk representation to insurance policy structure, and deep adaptation is achieved. In addition, real-time modification feedback signals are integrated to realize closed-loop optimization, adaptive driving evolution, and truly realize intelligent dynamic insurance policies that adapt to thousands of people and thousands of faces at different times.

Description

Technical Field

[0001] This application relates to the fields of medical and financial insurance technology, and in particular to methods, devices, equipment and storage media for generating dynamic insurance policies. Background Technology

[0002] In the fields of healthcare and financial insurance, traditional insurance product design has long followed the static, standardized logic of the industrial age. The product architecture is characterized by a one-time price, fixed coverage, and annual renewal. Its underlying assumptions are the homogeneity of risk and the predictability of user behavior. However, against the backdrop of the accelerating evolution of the digital economy and a more individualized society, this model is increasingly revealing structural mismatches. In reality, users are highly heterogeneous and dynamically evolving in terms of health status, career paths, family structures, consumption habits, geographical migration, and even psychological preferences. Furthermore, external factors such as climate change continuously disrupt the risk profile, making it difficult to accurately capture true risk exposure.

[0003] While there have been some attempts in recent years to introduce personalized mechanisms, such as health insurance or car insurance based on wearable devices, these explorations have mostly focused on single behavioral dimensions such as step count and frequency of sudden braking. Existing methods often treat risk factors as independent variables. The current mainstream still relies on manual underwriting or static scoring rule engines based on preset static thresholds, whose decision-making logic is rigid and response cycles are long. Existing insurance policies are "one-size-fits-all," unable to adapt to dynamically changing business scenarios in real time. In other words, existing technologies present a technical problem that prevents current insurance products from providing personalized solutions for users. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for generating dynamic insurance policies, which can solve the technical problem that existing insurance products cannot provide users with personalized adaptation solutions.

[0005] In a first aspect, embodiments of this application provide a method for generating a dynamic insurance policy, the method comprising:

[0006] In response to the policy generation command, the system obtains user information authorized by the user and environmental information from the environment side, wherein the user information includes at least one of user behavior logs, user physiological data, and user financial information.

[0007] Risk keywords are extracted from the user information and the environmental information based on a pre-trained language model;

[0008] Based on a preset rule template and a relational graph neural network, potential correlations in the keywords are mined to obtain a preliminary risk map containing risk transmission paths. The nodes in the risk map include risk nodes corresponding to the risk keywords, and the edges in the risk map are assigned a dynamic weight that changes with the user information and the environmental information.

[0009] The risk map is encoded into a feature vector, wherein the topological structure information of the risk map is preserved during encoding;

[0010] Based on a preset conditional generation model, the feature vector is mapped from risk representation to policy structure to obtain the structured policy draft corresponding to the feature vector.

[0011] In response to the user's instruction to modify the draft policy, the draft policy is adjusted and updated to obtain the target policy.

[0012] In some embodiments, adjusting and updating the draft policy in response to a user's modification instruction to obtain the target policy includes:

[0013] In response to the user's modification instruction on the policy draft, multiple first policy drafts with multiple types are initially generated, wherein the first policy drafts are structured and the types include conservative, balanced or aggressive.

[0014] Multiple draft first policies are used as the initial population of generation 0, where the population genes include coverage, pricing logic, service components, or terms and conditions.

[0015] Based on the probability of future policy renewals by users, underwriting profitability, user satisfaction, and potential compliance risks, an evaluation function is determined to assess the current draft policy.

[0016] The population genes are subjected to mutation operations, wherein the mutation operations include at least one of random insertion / deletion operations, noise application operations, or replacement operations;

[0017] The genes of the population are cross-linked to generate a second draft insurance policy;

[0018] The second policy draft is evaluated based on the evaluation function, and the second policy draft is modified and adjusted by mutation and crossover operations based on the evaluation results until the preset evolution cutoff condition is met, so as to obtain the final second policy draft.

[0019] The final draft of the second policy will be used as the target policy.

[0020] In some embodiments, the conditional generation model includes a conditional diffusion module, which corresponds to a temperature parameter. The temperature parameter is used to control the randomness of the model output by adjusting the sharpness of the probability distribution. The preset conditional generation model maps the feature vector from risk representation to policy structure to obtain a structured policy draft corresponding to the feature vector, including:

[0021] To obtain information about users' subjective preferences and objective economic capabilities;

[0022] Based on the subjective preference information and the objective economic capacity information, temperature parameters matching the user are generated.

[0023] The structured policy draft is obtained by mapping the feature vectors in the latent space based on the conditional diffusion module. During the mapping process, the output includes a probability distribution containing all possible options. The probability distribution is calibrated based on the temperature parameter. The smaller the temperature parameter, the sharper the probability distribution, and the less randomness the output of the conditional generation model has.

[0024] In some embodiments, the step of obtaining user information authorized by the user side and environmental information from the environment side in response to the policy generation instruction includes:

[0025] In response to the policy generation command, the user behavior log authorized by the user is obtained from the preset user interaction device. The user behavior log includes actively filled explicit information and / or implicit behavior information in the user interaction device. The explicit information includes at least one of smoking history, family hereditary diseases, and occupation. The implicit behavior information includes at least one of the user's time spent on the critical illness insurance page, the operation of adjusting the coverage amount slider, and the point of abandoning the insurance application.

[0026] The user physiological data authorized by the user is obtained from a preset user monitoring device, wherein the user monitoring device includes a wearable device, and the user physiological data includes at least one of resting heart rate, heart rate variability, sleep stage distribution, blood oxygen saturation, and activity intensity.

[0027] Obtain the user's financial information authorized by the user from a pre-set financial platform, wherein the user's financial information includes at least one of the following: income flow, debt status, credit score, and consumption records;

[0028] The environmental information is obtained from a pre-accessed third-party data platform and is authorized for public access. The environmental information includes at least one of the following: regional influenza data, weather forecasts, unemployment rate reports, housing disaster risk ratings, and interest rate adjustment information.

[0029] In some embodiments, after responding to the policy generation instruction and obtaining the user information authorized by the user side and the environment information from the environment side, the process includes:

[0030] The user information and the environmental information are detected.

[0031] In response to the instruction to fill in the detected missing information, context-aware filling processing is performed on the time-series data in the user information and / or the environmental information based on a preset time-series prediction model.

[0032] For the discrete data in the user information, obtain the user profile data corresponding to the user, and infer and populate the data based on the user profile data and the profile dataset of the neighboring user groups.

[0033] In response to the abnormal handling instruction for the detected abnormal information, a secondary confirmation process for the abnormal information by the user is triggered, wherein, upon receiving the user's operation instruction for the abnormal information, the abnormal information is updated based on the operation instruction;

[0034] In response to the confirmation command that the detection is complete, a secondary derivation is performed based on the user information and environmental information after the detection confirmation to obtain user information and environmental information with derived characteristics.

[0035] In some embodiments, the step of mining potential correlations among the keywords based on a preset rule template and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths includes:

[0036] Based on a preset rule template, entities are extracted from the keywords, and entity relationships are established between the entities, wherein the entity relationships include at least one of causal relationships, triggering relationships, and collaborative relationships;

[0037] Based on the relational graph neural network, potential topological relationships of the entity relationships are mined and relation completion processing is performed to obtain a preliminary risk graph in which the keywords are interconnected. The nodes in the risk graph include risk entity nodes, attribute nodes and environmental event nodes. The edges in the risk graph represent the relationships between the nodes. The edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0038] In some embodiments, the step of mining potential correlations among the keywords based on a preset rule template and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths includes:

[0039] Based on the user information and the environmental information, the disease risk transmission path and the financial risk transmission path are determined.

[0040] Based on the aforementioned disease risk transmission path and the aforementioned financial risk transmission path, the correlation strength, time decay factor, and environmental impact factor are determined.

[0041] The dynamic weights are determined based on the correlation strength, the time decay factor, and the environmental impact factor.

[0042] Based on the dynamic weights, preset rule templates, and relational graph neural networks, potential relationships among the keywords are mined to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0043] Secondly, embodiments of this application also provide a dynamic policy generation apparatus, the apparatus comprising:

[0044] The acquisition module is used to acquire user information authorized by the user side and environmental information from the environment side in response to the generation instruction of generating the insurance policy. The user information includes at least one of user behavior logs, user physiological data and user financial information.

[0045] The keyword extraction module is used to extract risky keywords from the user information and the environmental information based on a pre-trained language model.

[0046] The graph construction module is used to mine the potential correlations in the keywords based on the preset rule template and the relation graph neural network to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include the risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0047] An encoding module is used to encode the risk map into a feature vector, wherein the topological structure information of the risk map is preserved during encoding;

[0048] The mapping module is used to map the feature vector from risk representation to policy structure based on a preset condition generation model, so as to obtain the structured policy draft corresponding to the feature vector.

[0049] The policy adjustment module is used to respond to the user's modification instruction on the policy draft and adjust and update the policy draft to obtain the target policy.

[0050] Thirdly, embodiments of this application also provide a dynamic policy generation device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0051] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, can implement the above-described method.

[0052] This application provides a method, apparatus, device, and storage medium for generating dynamic insurance policies. In this application, in response to a policy generation command, user information authorized by the user and environmental information are obtained. The user information includes at least one of user behavior logs, user physiological data, and user financial information. Risk keywords are extracted from the user information and environmental information based on a pre-trained language model. Potential relationships among the keywords are mined using a preset rule template and a relational graph neural network to obtain a preliminary risk graph containing risk transmission paths. Nodes in the risk graph include risk nodes corresponding to the risk keywords, and edges in the risk graph are assigned dynamic weights that change with the user information and environmental information. The risk graph is encoded into feature vectors, retaining its topological structure information during encoding. The feature vectors are mapped from risk representation to policy structure based on a preset conditional generation model to obtain a structured policy draft corresponding to the feature vectors. In response to a user's modification command for the policy draft, the policy draft is adjusted and updated to obtain the target policy. Unlike existing technologies that treat risk factors as independent variables, this application mines the potential dependencies and correlations between user information and environmental information. It then maps the risk map obtained from these correlations from risk representation to policy structure, achieving deep adaptation. Furthermore, it integrates real-time modification feedback signals to achieve closed-loop optimization, adaptively driving evolution. This truly realizes a personalized, real-time adapted intelligent dynamic policy. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating the method for generating dynamic insurance policies provided in this application embodiment;

[0055] Figure 2A schematic block diagram of a dynamic insurance policy generation device provided in an embodiment of this application;

[0056] Figure 3 This is a schematic diagram of the structure of a computer device according to one embodiment of this application;

[0057] Figure 4 This is another structural schematic diagram of a computer device in one embodiment of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0060] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0061] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0062] Traditional insurance product design has long followed the static, standardized logic of the industrial age. The product architecture is characterized by a one-time price, fixed coverage, and annual renewal. Its underlying assumptions are the homogeneity of risk and the predictability of user behavior. However, against the backdrop of the accelerating evolution of the digital economy and a more individualized society, this model is increasingly revealing structural mismatches. In reality, users are highly heterogeneous and dynamically evolving in terms of health status, career paths, family structures, consumption habits, geographical migration, and even psychological preferences. Moreover, external factors such as climate change continuously disrupt the risk profile, making it difficult to accurately capture true risk exposure.

[0063] While some attempts have been made in recent years to introduce personalized mechanisms, such as health insurance or car insurance based on wearable devices, these explorations have largely focused on single behavioral dimensions like step count and frequency of sudden braking. Existing methods often treat risk factors as independent variables, ignoring the nonlinear coupling and dynamic dependencies between them. For example, the progression of chronic diseases may amplify income risks from occupational changes, and frequent extreme weather events can simultaneously affect the payout ratios of property and health insurance. This fragmented perception means that personalization remains only at a superficial level, failing to achieve a deep reconstruction of risk profiles. Current mainstream approaches still rely on manual underwriting or static scoring rule engines based on preset static thresholds, whose decision-making logic is rigid and response cycles are long. Existing insurance policies are "one-size-fits-all," unable to adapt to dynamically changing business scenarios in real time. In other words, existing technologies present a technical problem that prevents current insurance products from providing personalized solutions for users.

[0064] This application integrates and models multi-source heterogeneous risk factors. Unlike existing technologies that treat risk elements as independent variables, this application adapts the risk profile at a deep level based on the nonlinear coupling and dynamic dependencies between them, achieving a deep reconstruction of the risk profile. Furthermore, it integrates real-time feedback signals to achieve closed-loop optimization. The insurance product uses an adaptive-driven evolutionary model as its design framework, enabling iterative processing. This transforms the policy from a static contract into a dynamic service, providing full-cycle risk management and truly realizing a new paradigm of intelligent insurance that is personalized and adaptable to every individual.

[0065] Figure 1 This is a flowchart illustrating the method for generating a dynamic insurance policy provided in this application embodiment. The method for generating a dynamic insurance policy is applied to a dynamic insurance policy generation device. Figure 1 As shown, the method includes the following steps S110-S160:

[0066] S110. In response to the policy generation command, obtain user information authorized by the user side and environmental information from the environment side, wherein the user information includes at least one of user behavior logs, user physiological data and user financial information.

[0067] The dynamic policy generation device of this application connects to multiple data sources. Upon receiving a policy generation instruction, it acquires multi-source heterogeneous data based on the connected data sources, achieving unified access to this data. The collected information includes not only user information from the user side but also environmental information from the external environment side. It should be noted that access permissions must be requested before any data collection. Only after obtaining data access authorization can the acquisition of user information and environmental information be implemented.

[0068] This involves monitoring users using pre-set user monitoring devices and monitoring the environment using pre-set environmental monitoring devices. User monitoring items can include subjective behavioral logs or objective physiological data. Environmental monitoring items can be of many types, without specific limitations, including but not limited to macro-level objective variable data provided by government open platforms or commercial data service providers.

[0069] In step S110, in response to the instruction to generate a policy, the steps of obtaining user information authorized by the user side and environment information from the environment side include steps S1101-S1104:

[0070] S1101. In response to the policy generation command, obtain the user behavior log authorized by the user from the preset user interaction device. The user behavior log includes actively filled explicit information and / or implicit behavior information in the user interaction device. The explicit information includes at least one of smoking history, family hereditary diseases, and occupation. The implicit behavior information includes at least one of the user's dwell time on the critical illness insurance page, the operation of adjusting the coverage amount slider, and the point of abandoning the insurance application.

[0071] As an example, the user interaction device can be an app or mini-program on the user's terminal, and prior authorization has been obtained for the user interaction device. The authorization grants access and copying permissions to the device generating the dynamic policy in this application. User behavior logs can be logs of user interactions within the app or mini-program.

[0072] As an example, user behavior logs can record explicit information that users actively submit: such as a user's smoking history, family history of hereditary diseases, and occupational risk level.

[0073] As an example, user behavior logs can also contain implicit behavioral information, such as the duration of time spent on critical illness insurance pages, multiple adjustments to the coverage amount slider, and instances of abandoning the application process.

[0074] The explicit information and implicit behavioral information submitted by users can be used to infer their risk awareness level and the strength of their willingness to take precautions.

[0075] S1102. Obtain the user physiological data authorized by the user side from a preset user monitoring device, wherein the user monitoring device includes a wearable device, and the user physiological data includes at least one of resting heart rate, heart rate variability, sleep stage distribution, blood oxygen saturation, and activity intensity.

[0076] As an example, user monitoring devices can be wearable devices worn by users, such as smartwatches and fitness trackers.

[0077] For example, the wristband records the user's physiological indicators at all times, including but not limited to resting heart rate, heart rate variability, sleep stage distribution, blood oxygen saturation, and activity intensity. These physiological indicators are synchronized in real time via Bluetooth, Wi-Fi, or the manufacturer's open API.

[0078] It also supports automatic identification of device types and standardization of data formats to ensure cross-brand compatibility.

[0079] S1103. Obtain the user's financial information authorized by the user from a preset financial platform, wherein the user's financial information includes at least one of income flow, debt status, credit score and consumption records;

[0080] As an example, this involves acquiring user-authorized financial information from pre-set financial platforms such as banks, credit reporting agencies, or payment platforms. With explicit user authorization, and through compliant data cooperation mechanisms, it obtains information on the stability of income streams; debt structure based on credit card and loan balances; and high-frequency consumption categories such as credit score trends, medical expenses, and travel expenses.

[0081] S1104. Obtain the authorized and open environmental information from a pre-accessed third-party data platform, wherein the environmental information includes at least one of the following: regional influenza data, weather forecast, unemployment rate periodic report, housing disaster risk rating, and interest rate adjustment information.

[0082] In this embodiment, environmental information serves to calibrate the relative position of individual risk within a group and geographical context.

[0083] Obtain authorized and open environmental information from pre-connected third-party data platforms such as government open platforms or commercial data service providers.

[0084] As an example, information such as regional influenza indices obtained from the Centers for Disease Control and Prevention, extreme weather warnings obtained from the Meteorological Bureau, monthly industry unemployment rate reports obtained from the Ministry of Human Resources and Social Security, housing disaster risk ratings obtained from the Ministry of Housing and Urban-Rural Development, and interest rate adjustment information are obtained from the Ministry of Housing and Urban-Rural Development.

[0085] All the data obtained is first preprocessed.

[0086] S110, in response to the policy generation command, after obtaining the user information authorized by the user side and the environment information from the environment side, includes steps A1-A5:

[0087] A1. Detect the user information and the environmental information;

[0088] The user information and the environmental information are detected. If any missing or abnormal data is detected, preprocessing is performed.

[0089] A2. In response to the instruction to fill in the detected missing information, context-aware filling processing is performed on the time-series data in the user information and / or the environmental information based on a preset time-series prediction model.

[0090] When missing data is detected, it needs to be filled, and a fill instruction is generated to fill in the missing information.

[0091] If there are time-series data gaps in user information and / or the environmental information, such as daily step count data in user information, then in response to the instruction to fill in the detected missing information, context-aware filling is performed on the daily step count in user information using a Long Short-Term Memory (LSTM) time series prediction model.

[0092] A3. For the discrete data in the user information, obtain the user profile data corresponding to the user, and infer and fill the data based on the user profile data and the profile dataset of the neighboring user group.

[0093] As an example, for discrete information such as education level in user information, the similarity of user profiles is combined with clustering or graph embedding to perform collaborative reasoning from neighboring user groups to fill in the gaps, avoiding the bias introduced by simple means.

[0094] A4. In response to the abnormal handling instruction for the detected abnormal information, a secondary confirmation process for the abnormal information by the user is triggered, wherein, upon receiving the user's operation instruction for the abnormal information, the abnormal information is updated based on the operation instruction;

[0095] When anomalies are detected, anomaly handling is required. The system deploys a lightweight online anomaly detection model to mark and backtest data points that are abruptly changed but unreliable. If the authenticity cannot be confirmed, the data is downgraded or a user secondary confirmation mechanism is triggered.

[0096] A5. In response to the confirmation command for the completion of the detection, perform secondary derivation based on the user information and environmental information after the detection confirmation to obtain user information and environmental information with derived characteristics.

[0097] In this embodiment, although abnormal and missing data have already been processed, to further mine the information hidden in the data, preprocessing can be performed before constructing the data map, which is more conducive to the accuracy of the map construction. Specifically, in addition to the original observations, derived features with business interpretability are constructed. Secondary derivation is performed based on the user information and environmental information after detection and confirmation. For example, the frequency of nighttime hypoxemia events in the past 30 days, the quarterly growth rate of medical expenditure, and the correlation coefficient between occupational changes and income fluctuations are used to obtain information with derived features.

[0098] S120. Extract risk keywords from the user information and the environment information based on a pre-trained language model;

[0099] First, the graph construction process uses pre-trained language models finely tuned for the medical and / or financial fields to extract risk keywords from text and aligns them to a standard ontology system by combining named entity links.

[0100] Natural language processing (NLP) techniques are introduced, and based on a pre-trained language model, symptom keywords such as "recently experiencing chest tightness" are extracted from free text input by users, and ICD encoding mapping is performed to extract risk keywords. This keyword extraction step can improve the usability of unstructured data.

[0101] S130. Based on a preset rule template and a relational graph neural network, the potential relationships in the keywords are mined to obtain a preliminary risk map containing risk transmission paths. The nodes in the risk map include the risk nodes corresponding to the risk keywords, and the edges in the risk map are assigned a dynamic weight that changes with the user information and the environmental information.

[0102] Subsequently, a combination of rule templates and relational graph neural networks is used for relation extraction, ensuring the accuracy of high-confidence causal chains while uncovering potential implicit associations. By mining the potential relationships within keywords, a preliminary risk graph containing risk transmission paths is obtained. This preliminary risk graph includes multiple nodes and edges. These nodes include risk nodes corresponding to risk keywords. The graph is ultimately persistently stored in a graph database, supporting millisecond-level subgraph queries and complex path reasoning.

[0103] As an example, multi-source heterogeneous user information data, including structured behavioral logs, user physiological data, financial records, and unstructured medical examination reports, as well as environmental information such as news and public opinion and policy announcements, are integrated into a risk knowledge graph that is interpretable, inferable, and dynamically evolving.

[0104] Crucially, each edge of the initial risk graph is assigned a dynamic weight that changes over time, adapting to changes in user and environmental information. Based on these dynamic edge weights, the constructed risk graph becomes more realistic, no longer static, but resilient to external and internal user-related shocks. This gives the graph real-time responsiveness, allowing it to adapt to dynamically changing business scenarios. It also provides a personalized adaptation solution for the next step of generating insurance policies.

[0105] The following details the dynamic weights assigned to each edge of the risk graph.

[0106] S130 is a step that mines potential correlations in the keywords based on a preset rule template and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths, including steps B1-B4:

[0107] B1. Based on the user information and the environmental information, determine the disease risk transmission path and the financial risk transmission path;

[0108] In this application, the dynamic weights corresponding to the edges of the risk graph serve one purpose: to improve the graph's real-time responsiveness. Therefore, the dynamic weights can be designed as a function. The functional relationship can be designed according to the policy's focus. For example, if the policy focuses on the user's health, a disease risk transmission path related to the user's health can be constructed; if the policy focuses on the user's finances, a financial risk transmission path related to the user's finances can be constructed.

[0109] Taking user A as an example, when the system detects that user A has been working overtime frequently recently, has elevated triglycerides in a physical examination, and is in an industry experiencing a wave of layoffs, it will automatically construct a dynamic disease risk transmission chain: "Increased work pressure → leading to → decreased sleep quality → exacerbation → risk of metabolic syndrome → increase → probability of serious illness".

[0110] Simultaneously, a financial risk transmission path is generated: "Industry layoff wave → Trigger → Decreased income stability → Cause → Premium payment ability risk → Threat → Insurance continuity".

[0111] B2. Based on the disease risk transmission path and the financial risk transmission path, determine the correlation strength, time decay factor and environmental impact factor.

[0112] Based on the disease risk transmission path and the aforementioned financial risk transmission path, the correlation strength, time decay factor, and environmental impact factor corresponding to each transmission path are determined.

[0113] B3. Determine the dynamic weight based on the correlation strength, the time decay factor, and the environmental impact factor;

[0114] As an example, the dynamic weighting (relevance strength, time decay factor, environmental impact factor) function comprehensively considers the static strength of medical and actuarial evidence, the timeliness of behavior, and the macro-environmental impact of recent events. Among them, the environmental impact factor characterizes the amplification effect of public health emergencies on specific risk transmission paths.

[0115] B4. Based on the dynamic weights, preset rule templates, and relational graph neural networks, the potential relationships among the keywords are mined to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include the risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0116] The algorithm mines potential relationships among the keywords. These relationships are revealed not only by the graph's topology but also by the dynamic weights of its edges. The graph is then persistently stored in a graph database, supporting millisecond-level subgraph queries and complex path reasoning.

[0117] S130, the step of mining potential correlations among the keywords based on a preset rule template and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths, includes steps S1301-S1302:

[0118] S1301. Extract entities from the keywords based on a preset rule template and establish entity relationships between the entities, wherein the entity relationships include at least one of causal relationships, triggering relationships, and collaborative relationships;

[0119] The preset rule templates are a series of patterns summarized from the experience or existing knowledge of domain experts, used to initially extract entities and relationships from text such as keywords. Entity relationships must include at least one of the following: causal relationship, triggering relationship, and collaborative relationship.

[0120] For example, the keyword "XX vulnerability leads to data leakage" can be extracted into two entities, "XX vulnerability" and "data leakage," using the rule template "A leads to B," and a "leads to" relationship can be established. Rule templates typically include regular expressions, keyword lists, dependency syntax patterns, etc., enabling rapid and accurate identification of known types of associations. This rule-based approach is highly interpretable, accurate, and can leverage prior domain knowledge.

[0121] S1302. Based on the relational graph neural network, the potential topological relationships of the entity relationships are mined, and relation completion processing is performed to obtain a preliminary risk graph in which the keywords are interconnected. The nodes in the risk graph include risk entity nodes, attribute nodes, and environmental event nodes. The edges in the risk graph represent the association relationships between the nodes. The edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0122] Graph neural networks are deep learning models specifically designed for processing graph-structured data. They are capable of learning higher-order and implicit relationships between entity nodes. By aggregating neighbor information and updating node representations through message passing mechanisms, they capture latent topological features in the graph.

[0123] Based on semantic ontology, this graph defines three types of core nodes: the first type is risk entity nodes, such as "hypertension risk", "asset liquidity risk", and "occupational interruption risk"; the second type is attribute nodes, such as "age > 50", "BMI = 28", and "industry = construction"; and the third type is environmental event nodes, such as "interest rate increase" and "regional influenza outbreak".

[0124] The edges in the risk graph represent the relationships between the nodes, which are connected by various semantic relationships. These relationships include: 1) causal relationships such as "smoking → increased → lung cancer risk"; 2) triggering relationships such as "increased unemployment risk → trigger → premium payment ability risk"; 3) synergistic relationships; or inhibitory relationships such as "regular exercise → inhibition → cardiovascular disease risk". These relationships characterize the complex mechanisms of risk transmission and interaction.

[0125] Based on the aforementioned relational graph neural network, it predicts other possible relationships based on existing entities and partial relationships, mines potential topological relationships, and performs relationship completion processing. For example, it infers "A leads to C" from the known "A leads to B" and "B leads to C".

[0126] As a supplement, relationships can also be categorized to determine whether a certain relationship exists between two entities and the type of relationship, such as "containment" or "dependency".

[0127] S140. Encode the risk map into a feature vector, wherein the topological structure information of the risk map is retained during encoding;

[0128] Using graph representation learning algorithms, a subgraph containing multi-dimensional risk nodes and their complex semantic relationships, including health, finance, career, and external events, is encoded into a fixed-dimensional dense vector. This vector not only preserves the topological structure information of the original graph but also integrates risk intensity, transmission paths, and dynamic weights.

[0129] S150. Based on a preset conditional generation model, the feature vector is mapped from risk representation to policy structure to obtain the structured policy draft corresponding to the feature vector.

[0130] Building upon this, a conditional generative model architecture is employed, such as one incorporating a conditional variational autoencoder or a diffusion-based generative model. Using the feature vectors corresponding to the risk graph encoding as conditional input, the model learns the mapping distribution from risk representation to policy structure. The output is not natural language text, but a highly structured policy draft object containing four core elements:

[0131] List of coverage responsibilities; clearly define the covered scenarios and amounts, such as "a one-time payment of 1 million yuan for specific critical illnesses" and "payment based on the level of accidental disability";

[0132] Deductible and reimbursement rules; setting reimbursement thresholds and percentages, such as "annual deductible of 800 yuan for outpatient services, with 85% reimbursement for amounts exceeding this threshold";

[0133] Dynamic premium range: Based on the user's risk level and behavioral incentive mechanism, a floating price is given, for example, "the basic annual fee is ¥950, which can be reduced to ¥780 if the step count is met for 3 consecutive months";

[0134] Value-added service tags: Embedded health management ecosystem resources, such as "including 5 video consultations with top-tier doctors per year" and "providing gene testing discount coupons".

[0135] In some embodiments, a temperature parameter is introduced to adjust the degree of randomness in the model output. Correspondingly, this adjusts the degree of randomness in the generated policies. The greater the randomness, the higher the innovation of the generated policies, and the more they can break free from the inherent limitations of policy generation. The greater the difference between the generated policies and the original policies. Conversely, by adjusting the temperature parameter, the randomness of the model output is directly reduced, resulting in lower randomness in the corresponding policies and closer resemblance to the original policies. The conditional generation model includes a conditional diffusion module, which corresponds to the temperature parameter. The temperature parameter is used to control the degree of randomness in the model output by adjusting the sharpness of the probability distribution. Step S150, which maps the feature vector from risk representation to policy structure based on the preset conditional generation model to obtain the structured policy draft corresponding to the feature vector, includes steps S1501-S1503:

[0136] S1501. Obtain information about the user's subjective preferences and objective economic capabilities;

[0137] Before adjusting the randomness of the model output based on temperature parameters, it is necessary to first determine whether the user prefers a conservative, aggressive, or standard mode. These modes represent the user's subjective preferences. This approach can cater to different users' risk preferences and affordability by acquiring their subjective preferences and objective economic capabilities, and then determining the temperature parameters based on these information.

[0138] Conditional generative models refer to all models that generate data given additional information. Their components include: Conditional Variational Autoencoders (CVAEs), Conditional Generative Adversarial Networks (CGANs), and Conditional Diffusion Models.

[0139] The input to the conditional generation model is not only random noise, but also concatenated or fused with conditional vectors such as labels, text, and images.

[0140] The conditional diffusion module is a specific component or design pattern in the diffusion model architecture. It is the part that injects conditional information into the denoising network by modifying the network structure, such as by adding cross-attention layers or adaptive normalization layers, based on the standard diffusion model.

[0141] In generative models, the temperature parameter is a hyperparameter that scales the probability distribution during the sampling phase. Its core function is to control the randomness and creativity of the model's output by adjusting the sharpness of the probability distribution.

[0142] S1502. Based on the subjective preference information and the objective economic ability information, generate temperature parameters that match the user;

[0143] Based on the subjective preference information and the objective economic capacity information, temperature parameters matching the user are generated. The temperature parameters range from 0.8 to 1.5.

[0144] There are three states for the temperature parameter: First, the low-temperature mode, also known as the conservative mode. This is when the temperature parameter is between 0.1 and 0.8. It sharpens the probability distribution, amplifying high-probability options and suppressing low-probability options. This means the model tends to select the most probable basic atomic units, resulting in very low randomness in the output. The generated text or images exhibit mechanical repetition patterns. This is suitable for tasks requiring factual accuracy and rigorous logic. Second, the high-temperature mode, also known as the aggressive mode. This is when the temperature parameter is between 1.0 and 1.5. It smooths the probability distribution, allowing even low-probability options to be selected. The model dares to try rare combinations, generating more creative results. This is suitable for tasks requiring creativity and divergent thinking. Third, the standard mode, with a temperature parameter of 1.0. This maintains the original probability distribution unchanged and is usually used as the default baseline.

[0145] The insurance style is intelligently adjusted according to the user's risk preferences, truly achieving personalized insurance services tailored to each individual.

[0146] S1503. The feature vector in the latent space is mapped based on the conditional diffusion module to obtain the structured policy draft. During the mapping process, a probability distribution containing all possible options is output. The probability distribution is calibrated based on the temperature parameter. The smaller the temperature parameter, the sharper the probability distribution, and the less randomness the output of the conditional generation model is.

[0147] As an example, low temperature parameter values ​​tend to generate "conservative" solutions with high certainty and low risk exposure, such as high deductibles, low premiums, and basic coverage; high temperature parameter values, on the other hand, encourage the model to explore its boundaries, producing "aggressive protection" solutions such as zero deductibles, high sum assured, and additional high-end services; while medium temperature parameters correspond to a balanced approach. All generated variants undergo legality filtering through a pre-set compliance rule engine, such as a regulatory clause library and an actuarial constraint validator, to ensure that each draft complies with industry self-regulatory guidelines.

[0148] S160. In response to the user's instruction to modify the policy draft, the policy draft is adjusted and updated to obtain the target policy.

[0149] After generating the draft policy, the system also supports a human-machine collaborative optimization loop. Users can rate or modify the initially generated multiple drafts. For example, if a user requests a policy with the option to "reduce premiums while retaining critical illness coverage," the system can fine-tune hidden variables and quickly generate a new version.

[0150] At the same time, historical data is used to feed back into the generation model, continuously optimizing the match between insurance policies and users' real needs through reinforcement learning.

[0151] In some embodiments, the engine can also be integrated with underwriting and claims systems to achieve end-to-end intelligence and truly build a next-generation policy center centered on user risk profiles.

[0152] S160, the step of adjusting and updating the policy draft in response to the user's modification instruction to obtain the target policy, includes steps S1601-S1607:

[0153] S1601. In response to the user's modification instruction on the policy draft, multiple first policy drafts with multiple types are initially generated, wherein the first policy drafts are structured and the types include conservative, balanced or aggressive.

[0154] In terms of the evolution mechanism, in response to the user's modification instruction on the policy draft, multiple first policy drafts with multiple types are initially generated. The multiple first policy drafts refer to structured policy variants, such as conservative, balanced, and aggressive types.

[0155] S1602. Multiple drafts of the first insurance policy are used as the initial population of generation 0, wherein the population genes include coverage, pricing logic, service components or terms and conditions.

[0156] Each insurance policy is considered an individual, its "genes" encoded by modules such as coverage, pricing logic, service components, and terms and conditions. Subsequently, variants of multiple policy styles are used as the initial population of generation 0. The system will then perform two core genetic operations: crossover and mutation.

[0157] S1603. Based on the probability of future policy renewal by users, underwriting profitability of users, user satisfaction, and potential compliance risks, determine the evaluation function used to evaluate the current draft policy.

[0158] This embodiment uses a multi-dimensional weighted evaluation system to determine the evaluation function F used to evaluate the current policy draft.

[0159] Among them, the probability of a user renewing their policy in the future measures the likelihood of a user renewing or increasing their policy in the next 6–12 months, which can be predicted through historical behavior models;

[0160] User underwriting profitability (where the lower the ratio of claims payouts to premium income, the higher the user underwriting profitability score);

[0161] User satisfaction (Net Promoter Score) represents word-of-mouth potential and can be estimated based on feedback from simulated questionnaire tests.

[0162] Potential compliance risks are scored in real time by the built-in rules engine.

[0163] Weighting coefficients.

[0164] S1602. Perform mutation operations on the genes of the population, wherein the mutation operations include at least one of random insertion / deletion operations, noise application operations, or replacement operations;

[0165] Mutation operations introduce local perturbations to explore the neighborhood solution space.

[0166] For example, randomly adding or removing coverage items, such as adding "mental health counseling" or removing "overseas travel accidents");

[0167] For example, ±10% Gaussian noise can be applied to the risk coefficient in the premium calculation function.

[0168] S1604. Perform cross-linking operation on the genes of the population to generate a second draft insurance policy;

[0169] The genetic code of the population is cross-linked, which promotes the recombination and fusion of high-quality modules. For example, the comprehensive critical illness protection structure of parent A and the dynamic tiered pricing strategy of parent B are combined at the module level to generate a new individual second policy draft that combines coverage depth and cost flexibility. All operations are carried out within the preset compliance boundaries to ensure that the offspring after mutation and cross-linking still meet regulatory requirements.

[0170] S1605. The second policy draft is evaluated based on the evaluation function, and the second policy draft is modified and adjusted by mutation and crossover operations based on the evaluation results until the preset evolution cutoff condition is met, so as to obtain the final second policy draft.

[0171] Preset evolutionary cutoff conditions prevent infinite loops. These conditions can be set to allow the fitness improvement of the best individual in three consecutive generations to be less than 1%, or to allow the maximum number of generations to be reached.

[0172] The final draft of the second policy includes several high-quality proposals, each with its own advantages across different objective dimensions.

[0173] S1606. The final draft of the second policy shall be used as the target policy.

[0174] The goal of feedback fusion is not only to efficiently deliver the generated policy solutions to users, but more importantly, to enable the system to be self-adaptive and self-evolving through continuous collection of multi-source feedback signals, dynamic perception of the external environment, and intelligent triggering of re-optimization mechanisms, thereby achieving long-term collaborative optimization of risk protection and user experience in real business scenarios.

[0175] This application provides a method, apparatus, device, and storage medium for generating dynamic insurance policies. In this application, in response to a policy generation command, user information authorized by the user and environmental information are obtained. The user information includes at least one of user behavior logs, user physiological data, and user financial information. Risk keywords are extracted from the user information and environmental information based on a pre-trained language model. Potential relationships among the keywords are mined using a preset rule template and a relational graph neural network to obtain a preliminary risk graph containing risk transmission paths. Nodes in the risk graph include risk nodes corresponding to the risk keywords, and edges in the risk graph are assigned dynamic weights that change with the user information and environmental information. The risk graph is encoded into feature vectors, retaining its topological structure information during encoding. The feature vectors are mapped from risk representation to policy structure based on a preset conditional generation model to obtain a structured policy draft corresponding to the feature vectors. In response to a user's modification command for the policy draft, the policy draft is adjusted and updated to obtain the target policy. Unlike existing technologies that treat risk factors as independent variables, this application mines the potential dependencies and correlations between user information and environmental information. It then maps the risk map obtained from these correlations from risk representation to policy structure, achieving deep adaptation. Furthermore, it integrates real-time modification feedback signals to achieve closed-loop optimization, adaptively driving evolution. This truly realizes a personalized, real-time adapted intelligent dynamic policy.

[0176] Figure 2 This is a schematic block diagram of a dynamic insurance policy generation device provided in an embodiment of this application. Figure 2 As shown, corresponding to the above-described method for generating dynamic insurance policies, this application also provides a dynamic insurance policy generation apparatus 600. This dynamic insurance policy generation apparatus 600 includes a module for performing the generation of the aforementioned dynamic insurance policy, and can be configured in a terminal such as a desktop computer, tablet computer, or laptop computer. For details, please refer to... Figure 2 The dynamic policy generation device 600 includes an acquisition module 601, a keyword extraction module 602, a graph construction module 603, an encoding module 604, a mapping module 605, and a policy adjustment module 606, wherein:

[0177] The acquisition module 601 is used to acquire user information authorized by the user side and environmental information from the environment side in response to the generation instruction of generating the insurance policy. The user information includes at least one of user behavior logs, user physiological data and user financial information.

[0178] The keyword extraction module 602 is used to extract risky keywords from the user information and the environment information based on a pre-trained language model;

[0179] The graph construction module 603 is used to mine the potential correlations in the keywords based on the preset rule template and the relation graph neural network to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include the risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0180] Encoding module 604 is used to encode the risk map into a feature vector, wherein the topological structure information of the risk map is preserved during encoding;

[0181] The mapping module 605 is used to map the feature vector from risk representation to policy structure based on a preset condition generation model to obtain the structured policy draft corresponding to the feature vector.

[0182] The policy adjustment module 606 is used to adjust and update the policy draft in response to the user's modification instruction to obtain the target policy.

[0183] In some embodiments, the policy adjustment module 606, in response to a user's modification instruction on the policy draft, adjusts and updates the policy draft to obtain the target policy, specifically for:

[0184] In response to the user's modification instruction on the policy draft, multiple first policy drafts with multiple types are initially generated, wherein the first policy drafts are structured and the types include conservative, balanced or aggressive.

[0185] Multiple draft first policies are used as the initial population of generation 0, where the population genes include coverage, pricing logic, service components, or terms and conditions.

[0186] Based on the probability of future policy renewals by users, underwriting profitability, user satisfaction, and potential compliance risks, an evaluation function is determined to assess the current draft policy.

[0187] The population genes are subjected to mutation operations, wherein the mutation operations include at least one of random insertion / deletion operations, noise application operations, or replacement operations;

[0188] The genes of the population are cross-linked to generate a second draft insurance policy;

[0189] The second policy draft is evaluated based on the evaluation function, and the second policy draft is modified and adjusted by mutation and crossover operations based on the evaluation results until the preset evolution cutoff condition is met, so as to obtain the final second policy draft.

[0190] The final draft of the second policy will be used as the target policy.

[0191] In some embodiments, the conditional generation model includes a conditional diffusion module, which corresponds to a temperature parameter. The temperature parameter is used to control the randomness of the model output by adjusting the sharpness of the probability distribution. The mapping module 605 maps the feature vector from risk representation to policy structure based on the preset conditional generation model to obtain a structured policy draft corresponding to the feature vector, specifically for:

[0192] To obtain information about users' subjective preferences and objective economic capabilities;

[0193] Based on the subjective preference information and the objective economic capacity information, temperature parameters matching the user are generated.

[0194] The structured policy draft is obtained by mapping the feature vectors in the latent space based on the conditional diffusion module. During the mapping process, the output includes a probability distribution containing all possible options. The probability distribution is calibrated based on the temperature parameter. The smaller the temperature parameter, the sharper the probability distribution, and the less randomness the output of the conditional generation model has.

[0195] In some embodiments, the acquisition module 601, when executing the generation instruction in response to generating the policy, acquires user information authorized by the user side and environmental information from the environment side, specifically for:

[0196] In response to the policy generation command, the user behavior log authorized by the user is obtained from the preset user interaction device. The user behavior log includes actively filled explicit information and / or implicit behavior information in the user interaction device. The explicit information includes at least one of smoking history, family hereditary diseases, and occupation. The implicit behavior information includes at least one of the user's time spent on the critical illness insurance page, the operation of adjusting the coverage amount slider, and the point of abandoning the insurance application.

[0197] The user physiological data authorized by the user is obtained from a preset user monitoring device, wherein the user monitoring device includes a wearable device, and the user physiological data includes at least one of resting heart rate, heart rate variability, sleep stage distribution, blood oxygen saturation, and activity intensity.

[0198] Obtain the user's financial information authorized by the user from a pre-set financial platform, wherein the user's financial information includes at least one of the following: income flow, debt status, credit score, and consumption records;

[0199] The environmental information is obtained from a pre-accessed third-party data platform and is authorized for public access. The environmental information includes at least one of the following: regional influenza data, weather forecasts, unemployment rate reports, housing disaster risk ratings, and interest rate adjustment information.

[0200] In some embodiments, the dynamic policy generation device 600 further includes a preprocessing module. After responding to the policy generation command and obtaining user information authorized by the user side and environmental information from the environment side, the preprocessing module is specifically used for:

[0201] The user information and the environmental information are detected.

[0202] In response to the instruction to fill in the detected missing information, context-aware filling processing is performed on the time-series data in the user information and / or the environmental information based on a preset time-series prediction model.

[0203] For the discrete data in the user information, obtain the user profile data corresponding to the user, and infer and populate the data based on the user profile data and the profile dataset of the neighboring user groups.

[0204] In response to the abnormal handling instruction for the detected abnormal information, a secondary confirmation process for the abnormal information by the user is triggered, wherein, upon receiving the user's operation instruction for the abnormal information, the abnormal information is updated based on the operation instruction;

[0205] In response to the confirmation command that the detection is complete, a secondary derivation is performed based on the user information and environmental information after the detection confirmation to obtain user information and environmental information with derived characteristics.

[0206] In some embodiments, the risk map construction module 603 performs a preliminary risk map containing risk transmission paths by mining potential relationships among the keywords based on a preset rule template and a relational graph neural network, specifically for:

[0207] Based on a preset rule template, entities are extracted from the keywords, and entity relationships are established between the entities, wherein the entity relationships include at least one of causal relationships, triggering relationships, and collaborative relationships;

[0208] Based on the relational graph neural network, potential topological relationships of the entity relationships are mined and relation completion processing is performed to obtain a preliminary risk graph in which the keywords are interconnected. The nodes in the risk graph include risk entity nodes, attribute nodes and environmental event nodes. The edges in the risk graph represent the relationships between the nodes. The edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0209] In some embodiments, the risk map construction module 603, after performing the mining of potential correlations among the keywords based on a preset rule template and a relational graph neural network, obtains a preliminary risk map containing risk transmission paths, specifically for:

[0210] Based on the user information and the environmental information, the disease risk transmission path and the financial risk transmission path are determined.

[0211] Based on the aforementioned disease risk transmission path and the aforementioned financial risk transmission path, the correlation strength, time decay factor, and environmental impact factor are determined.

[0212] The dynamic weights are determined based on the correlation strength, the time decay factor, and the environmental impact factor.

[0213] Based on the dynamic weights, preset rule templates, and relational graph neural networks, potential relationships among the keywords are mined to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

[0214] This application provides a method, apparatus, device, and storage medium for generating dynamic insurance policies. In this application, in response to a policy generation command, user information authorized by the user and environmental information are obtained. The user information includes at least one of user behavior logs, user physiological data, and user financial information. Risk keywords are extracted from the user information and environmental information based on a pre-trained language model. Potential relationships among the keywords are mined using a preset rule template and a relational graph neural network to obtain a preliminary risk graph containing risk transmission paths. Nodes in the risk graph include risk nodes corresponding to the risk keywords, and edges in the risk graph are assigned dynamic weights that change with the user information and environmental information. The risk graph is encoded into feature vectors, retaining its topological structure information during encoding. The feature vectors are mapped from risk representation to policy structure based on a preset conditional generation model to obtain a structured policy draft corresponding to the feature vectors. In response to a user's modification command for the policy draft, the policy draft is adjusted and updated to obtain the target policy. Unlike existing technologies that treat risk factors as independent variables, this application mines the potential dependencies and correlations between user information and environmental information. It then maps the risk map obtained from these correlations from risk representation to policy structure, achieving deep adaptation. Furthermore, it integrates real-time modification feedback signals to achieve closed-loop optimization, adaptively driving evolution. This truly realizes a personalized, real-time adapted intelligent dynamic policy.

[0215] Specific limitations regarding the dynamic policy generation device can be found in the above description of the dynamic policy generation method, and will not be repeated here. Each module in the aforementioned dynamic policy generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0216] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for generating dynamic insurance policies based on artificial intelligence.

[0217] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an artificial intelligence-based dynamic policy generation method.

[0218] In one embodiment, a computer device is provided, 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 perform the following steps:

[0219] In response to the policy generation command, the system obtains user information authorized by the user and environmental information from the environment side, wherein the user information includes at least one of user behavior logs, user physiological data, and user financial information.

[0220] Risk keywords are extracted from the user information and the environmental information based on a pre-trained language model;

[0221] Based on a preset rule template and a relational graph neural network, potential correlations in the keywords are mined to obtain a preliminary risk map containing risk transmission paths. The nodes in the risk map include risk nodes corresponding to the risk keywords, and the edges in the risk map are assigned a dynamic weight that changes with the user information and the environmental information.

[0222] The risk map is encoded into a feature vector, wherein the topological structure information of the risk map is preserved during encoding;

[0223] Based on a preset conditional generation model, the feature vector is mapped from risk representation to policy structure to obtain the structured policy draft corresponding to the feature vector.

[0224] In response to the user's instruction to modify the draft policy, the draft policy is adjusted and updated to obtain the target policy.

[0225] Therefore, this application also provides a storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the following steps:

[0226] In response to the policy generation command, the system obtains user information authorized by the user and environmental information from the environment side, wherein the user information includes at least one of user behavior logs, user physiological data, and user financial information.

[0227] Risk keywords are extracted from the user information and the environmental information based on a pre-trained language model;

[0228] Based on a preset rule template and a relational graph neural network, potential correlations in the keywords are mined to obtain a preliminary risk map containing risk transmission paths. The nodes in the risk map include risk nodes corresponding to the risk keywords, and the edges in the risk map are assigned a dynamic weight that changes with the user information and the environmental information.

[0229] The risk map is encoded into a feature vector, wherein the topological structure information of the risk map is preserved during encoding;

[0230] Based on a preset conditional generation model, the feature vector is mapped from risk representation to policy structure to obtain the structured policy draft corresponding to the feature vector.

[0231] In response to the user's instruction to modify the draft policy, the draft policy is adjusted and updated to obtain the target policy.

[0232] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0233] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other storage media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0234] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0235] It should be noted that any software tools or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0236] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for generating dynamic insurance policies, characterized in that, The method includes: In response to the policy generation command, the system obtains user information authorized by the user and environmental information from the environment side, wherein the user information includes at least one of user behavior logs, user physiological data, and user financial information. Risk keywords are extracted from the user information and the environmental information based on a pre-trained language model; Based on a preset rule template and a relational graph neural network, potential correlations in the keywords are mined to obtain a preliminary risk map containing risk transmission paths. The nodes in the risk map include risk nodes corresponding to the risk keywords, and the edges in the risk map are assigned a dynamic weight that changes with the user information and the environmental information. The risk map is encoded into a feature vector, wherein the topological structure information of the risk map is preserved during encoding; Based on a preset conditional generation model, the feature vector is mapped from risk representation to policy structure to obtain the structured policy draft corresponding to the feature vector. In response to the user's instruction to modify the draft policy, the draft policy is adjusted and updated to obtain the target policy.

2. The method according to claim 1, characterized in that, The step of responding to a user's modification instruction on the draft policy and adjusting and updating the draft policy to obtain the target policy includes: In response to the user's modification instruction on the policy draft, multiple first policy drafts with multiple types are initially generated, wherein the first policy drafts are structured and the types include conservative, balanced or aggressive. Multiple draft first policies are used as the initial population of generation 0, where the population genes include coverage, pricing logic, service components, or terms and conditions. Based on the probability of future policy renewals by users, underwriting profitability, user satisfaction, and potential compliance risks, an evaluation function is determined to assess the current draft policy. The population genes are subjected to mutation operations, wherein the mutation operations include at least one of random insertion / deletion operations, noise application operations, or replacement operations; The genes of the population are cross-linked to generate a second draft insurance policy; The second policy draft is evaluated based on the evaluation function, and the second policy draft is modified and adjusted by mutation and crossover operations based on the evaluation results until the preset evolution cutoff condition is met, so as to obtain the final second policy draft. The final draft of the second policy will be used as the target policy.

3. The method according to claim 1, characterized in that, The conditional generation model includes a conditional diffusion module, which corresponds to a temperature parameter. This temperature parameter is used to control the randomness of the model output by adjusting the sharpness of the probability distribution. The model maps the feature vector from risk representation to policy structure based on the preset conditional generation model, obtaining a structured policy draft corresponding to the feature vector, including: To obtain information about users' subjective preferences and objective economic capabilities; Based on the subjective preference information and the objective economic capacity information, temperature parameters matching the user are generated. The structured policy draft is obtained by mapping the feature vectors in the latent space based on the conditional diffusion module. During the mapping process, the output includes a probability distribution containing all possible options. The probability distribution is calibrated based on the temperature parameter. The smaller the temperature parameter, the sharper the probability distribution, and the less randomness the output of the conditional generation model has.

4. The method according to claim 1, characterized in that, The process of responding to the policy generation command involves obtaining user information authorized by the user side and environment information from the environment side, including: In response to the policy generation command, the user behavior log authorized by the user is obtained from the preset user interaction device. The user behavior log includes actively filled explicit information and / or implicit behavior information in the user interaction device. The explicit information includes at least one of smoking history, family hereditary diseases, and occupation. The implicit behavior information includes at least one of the user's time spent on the critical illness insurance page, the operation of adjusting the coverage amount slider, and the point of abandoning the insurance application. The user physiological data authorized by the user is obtained from a preset user monitoring device, wherein the user monitoring device includes a wearable device, and the user physiological data includes at least one of resting heart rate, heart rate variability, sleep stage distribution, blood oxygen saturation, and activity intensity. Obtain the user's financial information authorized by the user from a pre-set financial platform, wherein the user's financial information includes at least one of the following: income flow, debt status, credit score, and consumption records; The environmental information is obtained from a pre-accessed third-party data platform and is authorized for public access. The environmental information includes at least one of the following: regional influenza data, weather forecasts, unemployment rate reports, housing disaster risk ratings, and interest rate adjustment information.

5. The method according to claim 1, characterized in that, After responding to the policy generation command and obtaining the user information authorized by the user side and the environment information from the environment side, the process includes: The user information and the environmental information are detected. In response to the instruction to fill in the detected missing information, context-aware filling processing is performed on the time-series data in the user information and / or the environmental information based on a preset time-series prediction model. For the discrete data in the user information, obtain the user profile data corresponding to the user, and infer and populate the data based on the user profile data and the profile dataset of the neighboring user groups. In response to the abnormal handling instruction for the detected abnormal information, a secondary confirmation process for the abnormal information by the user is triggered, wherein, upon receiving the user's operation instruction for the abnormal information, the abnormal information is updated based on the operation instruction; In response to the confirmation command that the detection is complete, a secondary derivation is performed based on the user information and environmental information after the detection confirmation to obtain user information and environmental information with derived characteristics.

6. The method according to claim 1, characterized in that, The process of mining potential correlations among keywords based on preset rule templates and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths includes: Based on a preset rule template, entities are extracted from the keywords, and entity relationships are established between the entities, wherein the entity relationships include at least one of causal relationships, triggering relationships, and collaborative relationships; Based on the relational graph neural network, potential topological relationships of the entity relationships are mined and relation completion processing is performed to obtain a preliminary risk graph in which the keywords are interconnected. The nodes in the risk graph include risk entity nodes, attribute nodes and environmental event nodes. The edges in the risk graph represent the relationships between the nodes. The edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

7. The method according to claim 1, characterized in that, The process of mining potential correlations among keywords based on preset rule templates and a relational graph neural network to obtain a preliminary risk map containing risk transmission paths includes: Based on the user information and the environmental information, the disease risk transmission path and the financial risk transmission path are determined. Based on the aforementioned disease risk transmission path and the aforementioned financial risk transmission path, the correlation strength, time decay factor, and environmental impact factor are determined. The dynamic weights are determined based on the correlation strength, the time decay factor, and the environmental impact factor. Based on the dynamic weights, preset rule templates, and relational graph neural networks, potential relationships among the keywords are mined to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information.

8. A device for generating dynamic insurance policies, characterized in that, The device includes: The acquisition module is used to acquire user information authorized by the user side and environmental information from the environment side in response to the generation instruction of generating the insurance policy. The user information includes at least one of user behavior logs, user physiological data and user financial information. The keyword extraction module is used to extract risky keywords from the user information and the environmental information based on a pre-trained language model. The graph construction module is used to mine the potential correlations in the keywords based on the preset rule template and the relation graph neural network to obtain a preliminary risk graph containing risk transmission paths. The nodes in the risk graph include the risk nodes corresponding to the risk keywords, and the edges in the risk graph are assigned a dynamic weight that changes with the user information and the environmental information. An encoding module is used to encode the risk map into a feature vector, wherein the topological structure information of the risk map is preserved during encoding; The mapping module is used to map the feature vector from risk representation to policy structure based on a preset condition generation model, so as to obtain the structured policy draft corresponding to the feature vector. The policy adjustment module is used to respond to the user's modification instruction on the policy draft and adjust and update the policy draft to obtain the target policy.

9. A device for generating dynamic insurance policies, characterized in that, The method includes a memory, a processor, and a dynamic policy generation program stored in the memory and executable on the processor. The processor executes the dynamic policy generation program to implement the steps of the dynamic policy generation method according to any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a program that implements a method for generating dynamic insurance policies, the program being executed by a processor to implement the steps of the method for generating dynamic insurance policies as described in any one of claims 1 to 7.

Images