Insurance combination quantitative evaluation method and system and computer program product

Abstract

The invention provides an insurance combination quantitative evaluation method and system based on multi-task learning and an attention mechanism and a computer program product. The method and the system are used for quantitatively evaluating the matching degree of an insurance applicant and an insurance combination composed of a variable number of insurance products on a plurality of insurance business indexes. Inputting the insurance applicant feature vector into an insurance applicant tower to generate an insurance applicant feature embedding vector, and inputting each product feature vector of the insurance combination into a shared product tower to generate a combined product feature embedding vector; inputting a cross-tower attention aggregation layer by taking the insurer feature embedded vector as a query vector and taking the combined product feature embedded vector as a key vector and a value vector, generating a combined assessment context vector representing the interaction relationship between the insurer and the to-be-assessed combination, inputting the combined assessment context vector into a plurality of task-specific assessment headers prepared in parallel, and carrying out task-specific assessment on the to-be-assessed combination. And outputting the quantitative score of each business index. The accurate and flexible multi-dimensional quantitative evaluation of the matching degree of the insurer-insurance combination, especially the variable-length multi-entity combination, is realized.

Description

Technical Field

[0001] This invention relates to the field of financial technology, specifically to a method, system, and computer program product for quantitative evaluation of insurance portfolios. Specifically, it relates to a multi-task quantitative evaluation technology for insurance portfolios based on multi-task learning and attention mechanisms, used to provide decision support for insurance brokers, and an insurance interpretation generation technology that integrates the quantitative evaluation results. Background Technology

[0002] In the insurance brokerage field, recommending insurance portfolios to clients is a core and complex decision-making process that heavily relies on the decision-maker's professional experience and cognitive judgment. Given the highly complex interactions between client characteristics and multiple insurance product portfolios, and the specific gains and losses of different options under different business objectives, decisions are easily influenced by subjective factors such as empiricism, recency effects, and confirmation bias. Brokers cannot make precise, data-driven trade-offs, and even experienced experts often experience significant discrepancies between their instantaneous judgments and the actual business results. Existing technologies often simply provide brokers with keyword comparisons or search tools for reference, failing to effectively manage this computationally intensive task. There is an urgent need for a human-computer interaction method that enhances cognition and provides decision support for accurately assessing the suitability of specific insured individuals for specific insurance portfolios.

[0003] Previous recommendation tools have failed to effectively address the complex interactions between policyholders and insurance portfolios. For example, in deep learning-based recommendation systems, a common and fundamental paradigm involves concatenating features from different sources (e.g., user and item features) into a fixed-dimensional feature vector, which is then input into a deep neural network for end-to-end learning (Heng-Tze Cheng, et al. Wide & Deep Learning for Recommendmender Systems, Proceedings of the 1st Workshop on Deep Learning for Recommendmender Systems. 2016. DOI: 10.1145 / 2988450.2988454.). This paradigm suffers from the following shortcomings when handling multi-metric evaluation tasks involving a specific insurer and a variable number of insurance portfolios.

[0004] Forcibly concatenating the unstructured interactions between policyholders and dynamically generated insurance products into a flat, fixed-format input vector makes it difficult to effectively capture the deep and complex interactions between policyholders and products within a group, as well as between products themselves. For example, it cannot effectively learn the synergistic enhancement effect between the feature "the policyholder has minor children" and "the portfolio includes both high-value life insurance and critical illness insurance." The input dimensions are fixed at the beginning of training, making it impossible to flexibly respond to the business needs of dynamically changing portfolio numbers in reality, and it lacks technical scalability. This fragmented modeling approach, which trains a single task independently for each indicator, ignores the non-linear synergistic / counteracting effects between different indicators (e.g., high profit margins may lead to low conversion rates), and the prediction results are often contradictory, failing to provide brokers with a unified and self-consistent reference view.

[0005] On the other hand, some insurance recommendation schemes proposed in the past use the recommendation action (recommend or not recommend) as the model output, which lacks credibility and readability. Summary of the Invention

[0006] This invention was developed in view of the above problems, with the aim of providing a method, system, and computer program product for quantitative evaluation of insurance portfolios, so as to realize accurate, flexible, and multi-dimensional quantitative evaluation of the matching degree of the policyholder-insurance portfolio, especially the variable-length multi-entity portfolio.

[0007] This invention provides a quantitative evaluation method for insurance portfolios based on multi-task learning and attention mechanisms. This method quantifies the matching degree between an insured and an insurance portfolio consisting of a variable number of insurance products across multiple insurance business indicators. The method includes: inputting the insured's feature vector into an insured tower constructed using a self-attention mechanism-based neural network for encoding to generate an insured feature embedding vector; inputting the product feature vectors of each component product in the insurance portfolio into a product tower constructed using a fully connected neural network for encoding to generate the product feature embedding vector for that component product; concatenating the product feature embedding vectors of each component product to form a combined product feature embedding vector; using the insured's feature embedding vector as a query vector and the combined product feature embedding vector as a key and value vector, inputting them into a cross-tower attention aggregation layer constructed using an attention mechanism-based neural network to generate a combined evaluation context vector representing the interaction between the insured and the portfolio to be evaluated; and inputting the combined evaluation context vector into multiple task-specific evaluation heads constructed using fully connected neural networks in parallel configuration to output the quantitative scores of each insurance business indicator.

[0008] A quantitative evaluation system for insurance portfolios based on multi-task learning and attention mechanisms is also provided. The system comprises: a processor; a memory; and computer program instructions stored in the memory, wherein the computer program instructions are executed by the processor to perform the method.

[0009] A computer program product is also provided, characterized in that it includes a computer-readable storage medium on which computer program instructions are stored, the computer program instructions executing the method when run by a computer.

[0010] This invention, through a unique dual-tower encoding and cross-tower attention aggregation architecture, can handle variable-length insurance portfolios and quantification of policyholders in multi-task evaluation, thereby improving the reliability and usability of computer-aided decision-making. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating the neural network architecture for implementing the insurance portfolio quantitative evaluation method of the present invention.

[0012] Figure 2 This is a flowchart illustrating the insurance portfolio quantitative evaluation method of the present invention.

[0013] Figure 3 This is a schematic diagram illustrating the neural network architecture that implements the interpretation generation method of the present invention.

[0014] Figure 4 This is a flowchart illustrating the explanation and generation method of the present invention.

[0015] Figure 5 This is a schematic block diagram illustrating the insurance portfolio quantitative evaluation system of the present invention. Detailed Implementation

[0016] The specific embodiments of this invention are provided to enable those skilled in the art to fully understand the invention and are intended to effectively support the claims of this invention. This section also involves the explanation or definition of the features and terms of the technical solutions of this invention. Those skilled in the art should understand the scope of the claims of this invention in conjunction with the explanation of specific features and terms in the specification, and should not arbitrarily misinterpret the true meaning represented by the features and terms of this invention.

[0017] It must be understood that, in any work involving the definition of the scope of the claims of this invention, the specific embodiments provided by this invention can fully support the generalization of equivalent or superior suitable solutions derived therefrom, and the scope of the claims should never be construed as not exceeding the examples themselves. Similarly, it must be understood that, with the development of technology and the emergence of new scenarios, the technical solutions provided by the embodiments of this application will also be equally applicable to similar technical problems without any inventive effort by those skilled in the art, and the scope of protection of this invention should not be limited to the examples described in the specific embodiments of this invention.

[0018] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and drawings of this application are intended to cover a non-exclusive inclusion, so that a process, method, or system that comprises a series of elements is not necessarily limited to those elements, but may include other elements not expressly listed or inherent to and understood in such processes, methods, or systems. Hereinafter, specific embodiments of the invention will be described in detail with reference to the accompanying drawings.

[0019] Therefore, this invention provides a quantitative evaluation technique for insurance portfolios, which is suitable for the efficient and stable quantification of complex, high-dimensional, and unstructured insurance information.

[0020] The insurance portfolio involved in this invention comprises multiple insurance products, which are insurance products sold by licensed insurance institutions. The number of these multiple insurance products is variable, and the portfolio specifically refers to a combination of multiple insurance products selected by an insurance broker based on professional experience to match a specific policyholder. Each insurance product can be a single insurance policy or an insurance package containing multiple sub-insurance policies. The manufacturers of the component products in the portfolio can be the same or different, and the insurance package can be a product from one insurance company or a joint package product from multiple insurance companies.

[0021] This invention addresses the complex and insufficiently resolved evaluation task of multiple, interrelated performance metrics (conversion rate, profit margin, etc.) within the unstructured relationships of multiple structured objects (insurance portfolios), particularly the evaluation problem of a single object versus a variable-length set of products. This invention proposes a novel dual-tower encoding-cross-tower attention aggregation-multi-task architecture.

[0022] like Figure 1 As shown, the architecture includes an insured person tower 110, a dual-channel product tower 120, a cross-tower attention aggregation layer 200, and a multi-task quantification layer 300.

[0023] The policyholder tower 110 is a neural network encoder based on a self-attention mechanism. It is configured to receive the policyholder feature vector, calculate the correlation strength between any two feature elements in the vector through its internal self-attention mechanism, and perform information fusion and weighted reconstruction on the original feature vector based on this correlation strength, and finally output a context-aware policyholder feature embedding vector.

[0024] A neural network encoder is, for example, a standard Transformer encoder stack consisting of N encoder layers connected in sequence. Each encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer, and all of them employ residual connections and layer normalization.

[0025] The dual-channel product tower 120 includes a shared product tower 121 and an optional package tower 122, which is activated when a package product is detected.

[0026] Both product tower 121 and package tower 122 are composed of one or more fully connected neural networks. When package tower 122 identifies a package product, it receives the overall feature vector of the package product, encodes it, and generates an overall feature embedding vector for the package product. Product tower 121 receives the feature vectors of the component products in the package product and the product feature vectors of each sub-product in the package product separately, encodes them, and generates feature embedding vectors for each product. The feature embedding vectors of each product and the overall feature embedding vector are concatenated to form the feature embedding vector of the package product.

[0027] When multiple package products exist, package tower 122 can also be shared. Product tower 121 and package tower 122 share parameters for their different input feature vectors.

[0028] The attention aggregation layer 200 sequentially includes a multi-head cross-attention layer, a first addition and normalization layer, a feedforward neural network layer, a second addition and normalization layer, and a global attention pooling layer. The multi-head cross-attention layer includes multiple attention heads in parallel. Each attention head receives the policyholder feature embedding vector as the query vector Query, and the combined product feature embedding vector as the key vector Key and value vector Valve, and calculates the attention between the two. The first addition and normalization layer, the feedforward neural network layer, and the second addition and normalization layer are used to fuse and enhance the interactive features. The global attention pooling layer compresses the fused features into a fixed-length combined evaluation context vector.

[0029] The multi-task quantization layer 300 includes multiple task-specific evaluation heads 3001, 3002, ..., 300 in parallel. n Each evaluation head includes one or more fully connected neural networks to receive the same combined evaluation context vector and output quantized scores using an activation function matched to the task. The multi-task evaluation layer 300 also includes a normalization processing layer 310 to normalize and map the quantized scores output by multiple evaluation heads, and output a set of standardized multi-task quantized scores.

[0030] The following details the quantitative evaluation method for insurance portfolios implemented based on the aforementioned architecture. This method is used to quantitatively evaluate the matching degree between an insured and an insurance portfolio consisting of a variable number of insurance products across multiple insurance business indicators.

[0031] like Figure 2As shown, in step S100, the policyholder feature vector is input into the policyholder tower 110, which is composed of a neural network based on a self-attention mechanism, and encoded to generate the policyholder feature embedding vector. The product feature vector of each component product in the insurance combination is input into the same product tower 121, which is composed of a fully connected neural network, and encoded to generate the product feature embedding vector of that component product. The product feature embedding vectors of each component product are concatenated to form the combined product feature embedding vector.

[0032] The order in which the product feature embedding vectors of multiple component products are concatenated in the combined product feature embedding vector can be arbitrary.

[0033] In step S100, optionally, when a package product is identified, the overall feature vector of the package product is input into the package tower 122 composed of a fully connected neural network for encoding to generate the overall feature embedding vector of the package product, and the product feature vector of each sub-product in the package product is input into the same product tower 121 for encoding to generate the product feature embedding vector of that sub-product. The product feature embedding vectors of each component product and each sub-product in the package product and the overall feature embedding vector of the package product are concatenated to form the combined product feature embedding vector.

[0034] Further optionally, when multiple package products are identified, the overall feature vector of each package product is input into the same package tower 122 for encoding to generate the overall feature embedding vector of the package product, and the product feature vector of each sub-product within the package product is input into the same component tower 121 for encoding to generate the product feature embedding vector of the sub-product. The product feature embedding vectors of each component product and each sub-product within each package product, as well as the overall feature embedding vector of each package product, are concatenated to form a combined product feature embedding vector.

[0035] The order of the product feature embedding vectors of component products and sub-products within package products, as well as the overall feature embedding vector of the package product, in the combined product feature embedding vector is not limited. However, it is preferable to place the overall feature vector of the package belonging to the same package adjacent to the feature vectors of all its sub-products in the sequence. More preferably, the overall feature vector of the package is placed before the feature vectors of all its sub-products, thereby reducing the learning difficulty and improving computational efficiency and effectiveness. The order between different packages and between component products and packages is not limited.

[0036] By sharing a single product neural network for all product features and a single package neural network for all package features, a shared dual-channel encoding process is achieved. This not only improves parameter efficiency and system scalability but also ensures the comparability of product and package embedding vectors. At the same time, it effectively utilizes data distribution to improve the feature representation quality of less popular products.

[0037] The policyholder feature vector is generated by embedding positional information into atomic feature vectors representing the policyholder's basic information and insurance history information, and then concatenating them in a predetermined order. The product feature vector is generated by embedding positional information into atomic feature vectors representing the insurance product's basic information and business status information, and then concatenating them in a predetermined order. Preferably, atomic feature vectors of the same category are assigned to a continuous interval, thereby forming a grouped structure.

[0038] The atomic feature vector representing the policyholder's basic information is a feature vector that characterizes the policyholder's inherent attributes and long-term stable state. The atomic feature vector representing the policyholder's insurance history information is a feature vector derived from the policyholder's historical insurance behavior data and used to characterize their stable decision-making tendencies and long-term business interaction patterns.

[0039] The atomic feature vectors representing the basic information of insurance products are feature vectors that characterize the core design and inherent rules of the products. The atomic feature vectors representing the business situation information of insurance products are feature vectors that characterize the market performance and strategic orientation of the products.

[0040] The overall feature vector of the package product is generated by embedding atomic feature vectors representing the overall design, comprehensive value, and bundling strategy of the package with location information and then concatenating them in a predetermined order. When the product feature vector is the product feature vector of a sub-product in the package product, it also includes atomic feature vectors representing the integration of the individual attributes and combined roles of the sub-product.

[0041] It should be understood that the embedded location information includes mapping the numerical representation of the atomic features to a unified dimension through a projection layer and then adding the feature vector to the location embedding vector.

[0042] It should be noted that the policyholder feature vector, product feature vector, and overall feature vector are vectors with predefined dimensions. After vectorization, atomic features are represented as two types: scalar features and vector features. Missing attribute values ​​are standardized and filled in.

[0043] It should be understood that atomic features are the minimum or aggregated feature information used as input to the model. Their feature vectorization process includes: for continuous numerical features, max-min normalization or standard fractionalization can be used; for categorical features, one-hot encoding or embedding layers can be used to map them into low-dimensional vectors; for unstructured text or complex classification information, a pre-defined risk level mapping table can be used to quantize them into ordered values, or lightweight semantic models such as Sentence-BERT can be used to encode them into fixed-dimensional semantic vectors. The following provides examples illustrating each type of information and its feature engineering.

[0044] The basic information of the insured can be selected from demographic information, financial status and asset burden information, health and medical risk information, occupational and behavioral environmental risk information, and derivative and processing information.

[0045] The demographic information can be selected from basic identity information, family structure and burden information, and long-term social attribute information. Basic identity information includes, for example, age, gender, occupation, city of residence, and residential stability, which is expressed as the length of time residing in the current place of residence. Occupational information includes, for example, occupation category, specific position, employment information, length of service, and whether frequent business travel is required. Family structure and burden information includes, for example, marital status, number and age of children, and number of elderly people to support. Long-term social attribute information includes, for example, highest level of education and industry attribute. Industry attribute includes, for example, the civil service system, large enterprises, and professional and technical positions.

[0046] The financial status and asset-liability information can be selected from income level information, asset status information, liability information, financial planning information, and income structure and stability information. Income level information includes, for example, individual annual income and family annual income. Asset status includes, for example, the value of real estate owned, total financial assets, and vehicle value. Liability includes, for example, outstanding mortgage payments, outstanding auto loan payments, and credit liabilities. Financial planning includes, for example, the annual savings rate and investment preference risk level. Income structure and stability information includes, for example, income sources and their sustainability, such as length of service, company operating years, disposable income, and emergency funds.

[0047] Health and medical risk information can be selected from past medical history and current illness, physical function indicators, behavioral and environmental risk indicators, lifestyle indicators, family health genetic information, and medical examination and claims records. Past medical history and current illness may include, for example, a history of major disease diagnoses, chronic disease management, and surgical history. Physical function indicators may include, for example, basic physiological indicators such as height, weight, blood pressure, vision, and hearing. Behavioral and environmental risk indicators may include, for example, participation in high-risk sports, vehicle type, driving experience, and driving record information. Lifestyle indicators may include, for example, smoking history, alcohol consumption habits, and weekly exercise frequency. Family health genetic information may include, for example, a history of major illnesses in immediate family members. Medical examination and claims records may include, for example, recent abnormal medical examination results.

[0048] It should be understood that, unlike information such as financial status and health risks which change more rapidly, education level, industry occupational attributes, and job stability have a strong statistical correlation with long-term social attributes and are easily learned by models to have a statistical relationship with the characteristics of combined insurance.

[0049] The occupational and behavioral environmental risk characteristics can be selected from occupational and behavioral environmental risk information and living environment and commuting risk information. Details of occupational environment risks include, for example, whether the worker works in special environments such as high pressure, high altitude, high temperature, high noise, or chemical environments; frequency of domestic and international business trips and frequently visited regions; and whether the worker works night shifts, rotating shifts, or long-term high-intensity overtime work. Living environment and commuting risk information includes environmental data such as the air quality index and drinking water quality of the city of residence; main commuting modes; one-way commuting distance and time; and typical road conditions along the commuting route.

[0050] Derivative and processed information includes, for example, a family responsibility index, a risk preference level, and a life cycle stage label. The family responsibility index is a weighted calculation derived from characteristics such as the insured's marital status, number and age of children, and number of elderly parents to support, used to quantify their family's economic burden. The risk preference level is a category label characterizing the insured's risk tolerance and aversion tendencies, such as conservative, moderate, or aggressive, derived through a classification model or rule mapping, considering factors like investment preferences, occupational risk, and the depth of previous insurance coverage. Life cycle stage labels are category labels based on the insured's age, family structure, and main income sources (e.g., salary, business operations), segmented through decision trees or predefined rules, used to identify specific life stages, such as single life, family formation, family growth, and retirement planning.

[0051] It should be noted that the value sets of some unstructured or high-dimensional categorical health information are enormous and open, and using one-hot encoding would lead to dimensionality explosion. As a solution example, information such as the history of important disease diagnoses, chronic disease management, surgical history, abnormal physical examination indicators, and the history of major diseases in immediate family members can be mapped to pre-defined risk level scalars based on actuarial knowledge. For example, a quantified risk weight can be preset for each disease diagnosis, and a comprehensive risk score can be calculated for the set of abnormal physical examination indicators.

[0052] The policyholder's insurance history information can be selected from the number of insurance companies, product diversity, average annual total premium, coverage depth trend, percentage of the company with the largest premium contribution, brand switching frequency, historical complaint rate, historical claims rate, historical underwriting information, category rejection and preference information, decision-making model information, etc.

[0053] Product diversity can be expressed as the number of primary category insurance products.

[0054] Characterizing the coverage depth trend includes obtaining the sum assured and incorporation year of all policies over a past period, and calculating the average sum assured (AvgAmount) of all policies over the past year. year4 and the average sum assured (AvgAmount) of all policies in the previous three years year1-3 Trend = AvgAmountyear4 / AvgAmount year1-3 .

[0055] The characterization of the company contributing the largest premium includes obtaining the underwriting companies and premiums for all policies over the past specified period, calculating the total premiums paid to each insurance company, determining the maximum premium (MaxPremium) and the total premiums over the past specified period (TotalPremium), and calculating the percentage = MaxPremium / TotalPremium.

[0056] Characterizing brand switching frequency involves sorting all policies within a specified time period by their effective date, listing the underwriting company sequence, and counting the number of changes in the sequence (ChangeCount). Frequency = ChangeCount / Year.

[0057] Category exclusion and preference information includes, for example, insurance type holding intensity. Given a predefined set of insurance type categories C = {c1, c2, ..., cn}, a comprehensive holding intensity score is calculated for each insurance type ci, using the formula: Intensity i =Whether currently held × (1 + log(1 + number of historical policies)) × (1 + log(1 + total holding time)), concatenate the strength scores of all insurance types in a fixed order to generate a [strength] score. c1 Intensity c2 Intensity cn The vector of ].

[0058] Category rejection and preference information also includes information on insurance product interaction behavior. A vector is created with a length equal to the number of insurance products, n. If a certain insurance product has recently had a surrender or policy lapse record, the corresponding position is 1; otherwise, it is 0. The final result is a [surrender / vacancy] vector. c1 Cancel insurance c2 ..., Insurance cancellation cn Multi-hot encoded vectors.

[0059] Decision-making model characteristics include, for example, policy concentration, which involves counting the number of different insurance companies involved in the policyholder's historical policies, calculating the total premium paid to each insurance company, determining the premium MaxPremium of the company with the highest premium paid, and calculating the total premium TotalPremium. The policy concentration is MaxPremium / TotalPremium.

[0060] Decision-making model features include, for example, the average insurance period, obtaining the first contact date and policy effective date of the policyholder's historical policies, calculating the insurance period for each policy as effective date - first contact date, and calculating the average insurance period for all policies.

[0061] The historical complaint rate is calculated as the ratio of the number of valid complaints filed by the policyholder within a certain period to the total number of inactive policies held by the policyholder during that period. The historical claims rate is calculated as the ratio of the total number of claims filed by the policyholder within a certain period to the total number of inactive policies held by the policyholder during that period.

[0062] Historical underwriting information can be represented by a multi-dimensional vector. The proportion of standard-risk policies in historical policies is calculated out of the total number of applications, and normalized to obtain the underwriting approval rate. A three-dimensional multi-hot encoded vector is created: if there is a history of premium-increased underwriting, the first dimension is 1; if there is a history of delayed underwriting, the second dimension is 1; if there is a history of rejection, the third dimension is 1; otherwise, the corresponding dimension is 0, resulting in premium / delay / rejection indicators. Each historical policy's underwriting conclusion is assigned a numerical value, for example, standard-risk = 0, premium-increased = 1, delayed = 2, rejection = 3. Then, the average of all historical policies is calculated and normalized to obtain the average underwriting conclusion strength. Finally, the above scalar and encoded vectors are concatenated to form the final feature vector representing historical underwriting information.

[0063] This invention does not process the original behavioral sequence of the policyholder's insurance history data. Instead, it compresses the historical information into a small number of low-dimensional static statistical features with clear business significance through feature engineering driven by business logic.

[0064] Basic information about insurance products can be selected from product identification and classification information, protection attribute information, financial attribute information, target customer group attribute information, etc.

[0065] Product identification and classification information includes, for example, product name, underwriting company, product ID, primary category, secondary category, and product tags, such as savings, protection, consumption type, and investment type.

[0066] Coverage attributes include, for example, core coverage liabilities and exclusion information. Core coverage liabilities are characterized by encoding based on a pre-defined set of coverage liabilities, and the sum insured can be binned and normalized. Exclusion clauses can be generated by extracting keywords from a pre-defined keyword library from the exclusion clause text to form an exclusion clause vector.

[0067] Financial attributes include, for example, standard premiums, payment periods, cash value tables (e.g., 1000 in year 1, 3500 in year 2), and internal rate of return (IRR), which can be obtained through product rate tables and actuarial data. The cash value can be extracted from the cash values ​​of the first N years, forming a sequence, which is then normalized to form a cash value curve vector. Optionally, the standardized premium value, the normalized payment period value, the normalized IRR value, and the cash value curve vector can be further concatenated to form a financial attribute feature vector.

[0068] Target customer attribute information, such as the target customer age range [min]age ,max age [Target customer income level, target customer family stage, such as established families, retirement planning, can further target the minimum] age ,max age The target customer group attribute feature vector is formed by splicing together the single hot vector of income level and the multi hot vector of family stage.

[0069] Business status information for insurance products can be selected from market performance characteristics, sales and service characteristics, strength and service characteristics, and corporate strategic characteristics.

[0070] Market performance characteristics include historical conversion rates, market share, and customer satisfaction. Furthermore, the processed conversion rate, market share, and satisfaction scores can be concatenated to generate a market performance feature vector.

[0071] Sales and service characteristics include commission rates, underwriting approval rates, average claims settlement time, and service value-added features such as expedited processing channels and telephone doctors. Furthermore, the processed vectors of commission rates, underwriting approval rates, claims processing time, and service value-added features can be concatenated to generate a sales and service feature vector.

[0072] Strength and service characteristics include solvency adequacy ratio, claims reputation rating, and service quality. Solvency adequacy ratio data can be collected from all insurance companies, and the mean and standard deviation can be calculated to convert each company's raw values ​​into standardized scores. Service quality includes underwriting friendliness, customer service response speed, and value-added services. Underwriting friendliness can be expressed as whether health disclosures are required, while value-added services include the availability of expedited access for critical illnesses and advance payment of medical expenses.

[0073] Correspondingly, the conversion rate assessment task will learn the relationship between providing services such as medical expense prepayment and conversion drivers, the profit margin / price task will learn the relationship between high solvency ratio, good reputation and brand premium, and the customer risk task will learn the relationship between lenient underwriting and high claims risk.

[0074] Company strategic characteristics include, for example, strategic product identifiers, product lifecycle stages, and target profit margins. Furthermore, the processed strategic product identifiers, lifecycle one-hot vectors, and normalized profit margin values ​​can be concatenated to generate a company strategic feature vector.

[0075] The overall feature vector of a package product can be selected from package identifier and structural features, comprehensive value features, and bundling strategy features.

[0076] Package identifiers and structures include, for example, package ID, package type and package structure. Package types include, for example, family sharing package, couple mutual insurance package, and car-family combination package. Package structures include, for example, the number of sub-products included in the package and package tiers such as basic version and premium version.

[0077] Comprehensive value features may include price and discount features, such as the total premium of the package, the discount rate compared to the total price of purchasing sub-products separately, whether there are premium reduction or exemption clauses, and whether installment payment discounts are offered. They may also include service synergy features, such as whether a unified family policy concierge service is provided, whether joint underwriting is available, and whether unified value-added services are included across products.

[0078] Bundling strategy features include, for example, sales-oriented features, such as whether the package is designed as a lead generation product, a profit product, or a strategic defense product, as well as synergy features, such as those selected from comprehensive family protection, package bundling strength, and mandatory bundling, encoded into fixed-length vectors using a lightweight sentence encoder or a bag-of-words model.

[0079] The coverage features of sub-products can be redistributed to each sub-product according to certain rules based on the discount rate of the overall coverage of the package, forming new discounted coverage features.

[0080] The overall feature vector of the package and the product feature vectors of sub-products can share combined strategic features to identify package affiliation, such as sharing package identification IDs, like B-Family-001, and can also share a list of cooperating product IDs, such as [P-Health Insurance-001, P-Accident Insurance-005]. By introducing association identifiers, feature vectors generated by different encoding channels can be implicitly associated in the attention mechanism.

[0081] In step S200, the policyholder feature embedding vector is used as the query vector Query, and the combined product feature embedding vector is used as the key vector Key and value vector Value. The input is a cross-tower attention aggregation layer 200 composed of a neural network based on an attention mechanism to generate a combined evaluation context vector representing the interaction relationship between the policyholder and the combination to be evaluated.

[0082] The multi-head attention layer of the neural network based on the attention mechanism calculates the correlation between the query vector and each key vector to obtain the attention weight. The attention weight is used to perform a weighted summation on the corresponding value vector to generate a context vector representing the interaction relationship between the policyholder and the portfolio to be evaluated. The context vector is further processed by an addition and normalization layer, a feedforward neural network layer, a second addition and normalization layer, and a global attention pooling layer to obtain the portfolio evaluation context vector.

[0083] The value of an insurance portfolio is not simply the sum of the values ​​of its individual products; there are complex interactions between the products. For example, the presence of a particular product can significantly enhance the value or suitability of another product. Treating the entire portfolio as a whole in attention calculations allows us to capture the complex synergistic and antagonistic relationships between the products, enabling structured modeling of the aforementioned unstructured business problems.

[0084] In step S300, the combined evaluation context vector is input into multiple task-specific evaluation heads 3001-300, each configured in parallel and composed of fully connected neural networks. n Each task indicator will generate a quantitative score.

[0085] The generation of quantitative scores for each task indicator also includes each evaluation head using an activation function that matches the task based on the mathematical characteristics of the predicted task, and outputting a quantitative score.

[0086] The number of assessment heads and prediction targets can be configured according to actual business needs. These can include quantitative assessment heads for conversion rates, renewal rates, profit margins, and claims risk. Optionally, they can also include at least one of the following: quantitative assessment heads for adequate risk coverage, protection leverage ratio, potential complaint risk, short-term silent risk, and complaint tendency risk. Introducing multi-task indicators into the assessment can enhance the information richness and decision support capabilities of the assessment model output.

[0087] Building upon this, the outputs of the attention heads are concatenated into a wider, more information-rich super vector. This vector contains the policyholder's assessment of the combination from multiple perspectives. This super vector undergoes deep processing through a shared DNN and is then distributed to the output layers of various simpler tasks. The advantage of this approach is that all tasks have access to complete information from all angles; for example, the profit margin task can also utilize information extracted from the protection perspective.

[0088] In step S300, the quantified scores of each task indicator are further input into the normalization processing layer for normalization mapping, and a set of standardized multi-task quantified scores are output.

[0089] This normalization process maps the quantified scores, which have different dimensions, output by each task-specific evaluation head to a unified and comparable scoring system.

[0090] As an example of the normalization mapping, the quantized score sequence of each task indicator is linearly scaled to generate a floating-point vector constrained in the interval [0,1], where the normalization formula for each quantized score is as follows.

[0091]

[0092] in Quantitative score for a certain assessment head Normalized value of a certain rating score X is the score sequence [x1, x2, ..., xn] ε is a very small positive number, for example, 0.00001. By using a preset minimum positive constant ε, the normalized value range is ensured to be a closed interval [ε, 1-ε], thus maintaining the distinguishability of the values ​​and the stability of the calculation.

[0093] The backbone neural network consists of multiple neural networks that form the policyholder tower 110, the dual-channel product tower 120, the cross-tower attention aggregation layer 200, and the multi-task quantization layer 300. It is trained end-to-end, and the training samples are derived from multi-source heterogeneous historical data of the insurance business system. Each sample is a complete policyholder-product combination pair, which includes a structured policyholder feature vector, a variable number of product feature vectors, and corresponding multi-dimensional task actuarial labels.

[0094] The training samples specifically include real policyholder characteristic information and product characteristic information of insurance products in different combinations, along with their corresponding conversion rate scores, renewal rate scores, profit margin scores, and claims risk scores. Optionally, they may also include risk coverage adequacy scores, protection leverage ratio scores, potential complaint risk scores, short-term silent risk scores, and complaint tendency risk scores.

[0095] Optionally, the training set also includes a certain proportion of package samples containing overall feature information of the package product and product feature information of the sub-products within the package product. These can include different types of packages, different discount strategies such as direct price reduction, free insurance coverage, bundled services, and various combinations of sub-products. The sample proportion is preferably no less than 5% to 10%, but can also be increased to 15% to 30%.

[0096] It should be understood that the training set should contain a large number of regular positive / negative samples, including combinations of successful transactions and obviously failed combinations. To improve the model's ability to evaluate complex and edge cases, the training set can also include a certain proportion of key, small-scale difficult / edge samples. These samples originate from real business history cases where decision-making uncertainty is high or the final result deviates from initial expectations. Examples include cases where a broker recommended a client who initially showed interest but ultimately did not close the deal, cases where the client accepted the recommendation but the renewal rate was very low, and cases where different senior brokers had differing opinions.

[0097] The construction of training labels for the conversion rate quantification evaluation head in each evaluation head is, for example, extracting data records from sales logs. The label assignment is based on whether the policyholder has completed the insurance payment process for the presented combination within a preset behavioral time window. If the payment is completed, the conversion rate label of this data record is assigned a value of 1; if the payment is not completed, the label is assigned a value of 0. The learning objective of the model is to fit the probability from the feature to the binary event.

[0098] The construction of training labels for the renewal rate assessment head can rely on policy lifecycle data. For any policy in the sample that has completed its first coverage period, it is queried whether the premium for the next period was successfully paid within the grace period after the end of the period. If the payment is successful, the system records that the renewal behavior has occurred and assigns a renewal label 1 to the data record of the policy at the renewal decision time point; if the payment is not made or the contract is explicitly terminated, a renewal label 0 is assigned. The learning objective of the model is the probability of the binary event of the policyholder's retention decision after completing the first insurance period.

[0099] The training label for the profit margin quantitative assessment head is a continuous value. Its construction can be calculated based on policy-level financial data. For a policy that has ended or has sufficient sample data, the total earned premiums in a complete cycle are obtained from the financial system. The total claims amount, channel commissions and other direct costs incurred in the same period are deducted, as well as the allocated operating costs. The specific value is calculated by the formula (total earned premiums - total cost expenditures) / total earned premiums.

[0100] The training label for the claims risk quantification assessment head is a continuous numerical value, such as the loss rate. In its construction, for each sample policy, after determining that it has passed the fixed observation period, the total amount of claims paid for it during the observation period is collected from the claims database and divided by the total earned premiums of the policy during the same period. The calculated ratio is the loss rate, which serves as the claims risk label for that data record.

[0101] The training label for the claims risk quantification assessment head can also be the claims risk index. Its construction is based on an actuarial model that comprehensively considers multiple factors such as whether a claim event occurred in the policy during the observation period, the number of claims, the distribution of claim amount, and the claim time pattern. For example, the sample policies are risk-stratified and divided into different risk levels according to their claims characteristics. The basic risk coefficient of each risk level is calculated based on a statistical model. Combined with the specific claims records of the policies, including zero claims records, a continuous risk index is obtained through weighted calculation.

[0102] The training labels for the adequacy assessment head are constructed through actuarial models. For example, based on core parameters such as the policyholder's annual income, total outstanding debt, number of dependent family members, and education and living costs, the actuaries calculate the theoretical minimum adequacy amount for the policyholder in a specific protection dimension, such as death benefit or critical illness benefit, using a standardized model predefined by the actuaries. This theoretical amount is used as the benchmark for risk assessment. The total amount of insurance in the protection dimension is compared with the theoretical minimum adequacy amount, and the ratio is the protection adequacy label. This ratio is a continuous value. When the ratio is greater than 1, it indicates that the coverage of the combination exceeds the theoretical minimum requirement; when it is equal to 1, it indicates full coverage; and when it is less than 1, it indicates that there is a protection gap.

[0103] The construction of training labels for the potential complaint risk quantification prediction head is based on the correlation analysis between customer service data and policy characteristics. For example, sample policy data and their corresponding customer service records are collected, and policies that have caused policyholder complaints or claims disputes are extracted as positive samples. The label is defined as the probability that the combination will cause a complaint or dispute in a specific period of time in the future, which is a continuous value between 0 and 1. In its specific implementation, the sample data is divided into time windows, and it is statistically analyzed whether each insurance product combination has a complaint behavior within a specific period after it takes effect. Combinations with complaints are marked as high-risk samples, and combinations without complaints are marked as low-risk samples.

[0104] The training labels for the leverage ratio quantitative assessment head can be obtained through financial calculations. For example, in its construction, the total sum insured and total premium data of all products in the insurance portfolio are obtained. The total sum insured refers to the arithmetic sum of the coverage of each insurance product, and the total premium refers to the total annual premium that the policyholder needs to pay for the portfolio. The leverage ratio label is the ratio of the total sum insured to the total premium, which is a unitless continuous value.

[0105] The construction of training labels for the customer risk coverage quantification assessment head is based on a standardized risk exposure model, such as a standard risk exposure classification system defined by the actuarial team. This system should cover core risk categories such as death risk, health and medical risk, accidental injury risk, property loss risk, and legal liability risk. Each risk category has specific coverage items and their corresponding weight coefficients. When calculating the labels, the coverage provided by the insurance portfolio is compared with the standard risk exposure model item by item. The ratio of covered risk items to total risk items is calculated, and the weight coefficients of each risk item are considered to finally obtain a weighted coverage score.

[0106] The Short-Term Silence Risk Quantification Header and the Complaint Propensity Risk Quantification Header are used for proactive identification and assessment of potential service risks. The Short-Term Silence Risk Header assesses the probability that an insured will no longer respond to any contact from the broker or explicitly state that they will not renew the policy within a specific period after purchasing insurance. Its training labels are sourced from follow-up records in the customer relationship management system and the response status of renewal notices. If a customer remains silent / explicitly does not renew the policy within the first year, the sample label is marked as 1, representing high risk; otherwise, it is 0. The Complaint Propensity Risk Header assesses the probability that an insured will file a complaint or dispute during or after the purchase process.

[0107] In a preferred embodiment, the plurality of task-specific evaluation heads include at least a conversion rate quantification evaluation head, a profit margin quantification evaluation head, a renewal rate quantification evaluation head, and a claims risk quantification evaluation head. During the end-to-end training, a prediction consistency index is calculated based on the quantified scores of conversion rate, profit margin, renewal rate, and claims risk to quantify the relative deviation and linkage anomalies among the four task evaluations. This index serves as the global control signal for the differentiated learning rate of the backbone network, guiding the gradient allocation strategy in backpropagation and achieving adaptive optimization of model parameters driven by prediction consistency.

[0108] The predictive consistency index is a function Ψ with the quantified scores of the four tasks as variables. It is constructed such that the factor value decreases when the quantified score of claims risk deviates positively from the level of synergy among the other three quantified scores. The formula is as follows:

[0109] The numerator is the output of a two-layer perceptron network that is constructed and trained together with the backbone network. The task is to calculate the value consensus scalar that represents the level of inherent consistency and reliability between task predictions from the loss signals of market tasks (conversion rate, profit margin, renewal rate).

[0110] Where m=[L c , L p , Lr] T It is the market value vector, L c , L p Lr are the quantitative scores of the rate of return, profit margin, and renewal rate, respectively. z = m⊕[GeometricMean(m)] is the concatenation of the m vector and its geometric mean. GeometricMean(m) = (Lr / Lr) c × L p × Lr) / 3; The enhanced feature z is fed into a single hidden layer of the two-layer perceptron network, and then processed by the weights W of the hidden layer. h Bias b h And the tanh activation function performs nonlinear transformations and interactions on the input features; W o , b o These are the weights and biases of the output layer σ of the two-layer perceptron network, where σ is the sigmoid function and the output is compressed to (0,1).

[0111] The denominator is the risk deviation penalty term. Where v = [L] c , L p , L r , L claim ] T It is a vector of full-task predicted values ​​for conversion rate, profit margin, renewal rate, and claims risk, L claimIt is a quantitative score of claims risk; μ is the L2 norm of the vector v after standardization, which is used to quantify the multidimensional dispersion of each predicted value relative to the overall mean, where μ v , σ v ε represents the mean and standard deviation of vector v, where 1 represents a vector of all 1s and ε is a minimal positive constant. Cov(m,v) is the covariance between the market vector m and the total vector v, used to capture the correlation between market indicators and overall forecasts. exp(-β·Cov(m,v)) constitutes a dynamic adjustment factor based on covariance, which reduces the penalty when market indicators are negatively correlated with risk prediction. λ and β are preset scaling coefficients and covariance influence coefficients.

[0112] Here, the prediction consistency index Ψ is a global state scalar. It extracts nonlinear consensus features from market value indicators through a two-layer perceptron network. A dynamic penalty mechanism based on the L2 norm of the prediction vector and cross-task covariance is introduced into the denominator, which can keenly capture and quantify the relative deviation and linkage anomalies between multi-task predictions.

[0113] Furthermore, in each training iteration, the loss function of the policyholder feature embedding network, product feature embedding network, package feature embedding network, market task network, and risk task network is dynamically reconstructed using the prediction consistency index Ψ. Its general formula is:

[0114] Where: L is the loss function value of the current iteration, t represents the t-th iteration, Φ is the dynamic loss function mapping relationship, Ψ(t) is the prediction consistency index of the current iteration, Θ is the network parameter, and D represents the training data.

[0115] The consistency of forward prediction guides the gradient allocation strategy in backpropagation, which in turn guides the model parameters to evolve in a direction that improves prediction consistency. When Ψ indicates a decrease in prediction consistency, the weight of market-related task losses is automatically reduced, while the weight of risk-related task losses is increased. This guides the model parameters to optimize towards prioritizing the correction of risk predictions during backpropagation. This method achieves a closed loop from global state awareness to loss function reconstruction, enabling the model to make intelligent trade-offs and priority decisions under complex conflicts. Through iterative optimization, the model ultimately masters the ability to dynamically trade off multiple business objectives. The converged model after iterative optimization outputs quantitative evaluation results with higher business rationality and global balance.

[0116] Specifically, in each training iteration, the loss function of the policyholder feature embedding network, product feature embedding network, and package feature embedding network is dynamically reconstructed using the prediction consistency index Ψ. The formula for the loss function is as follows:

[0117] This formula is essentially a smooth ramp function implemented using the Sigmoid function, where η is the base learning rate. eff The effective learning rate is defined by ε, which is the minimum learning rate coefficient. K controls the steepness of the slope. For policyholder feature embedding networks, product feature embedding networks, and package feature embedding networks, specific steepness K values ​​can be used respectively. a , K b , K c Ψ0 is the learning rate switching center point, and δ is the numerical stability constant. When Ψ is very high, the denominator Sigmoid output is close to 1, and the learning rate is normal; when Ψ is very low, the Sigmoid output is close to 0, and the learning rate drops to η·ε.

[0118] In each training iteration, the loss function of the cross-tower attention network is dynamically reconstructed using the prediction consistency index Ψ. The formula for the loss function is as follows:

[0119] Here, γ is the inhibition strength coefficient, used to control the severity of inhibition. When Ψ is close to 1, the integral value is close to 0, the exponential term is 1, and the learning rate is normal; when Ψ decreases, the integral value increases rapidly, causing the exponential term to decrease sharply, and the learning rate is strongly suppressed.

[0120] In each training iteration, the loss function of the market task network is dynamically reconstructed using the prediction consistency index Ψ. The formula for the loss function is as follows:

[0121] Here, α is the market task inhibition coefficient, used to control the decay intensity of the market task learning rate. When Ψ=1, the learning rate is η; as Ψ decreases, the α(1-Ψ) term in the denominator will be amplified, causing the learning rate to decrease in a non-linear manner.

[0122] In each training iteration, the loss function of the claims risk task network is dynamically reconstructed using the prediction consistency index Ψ. The formula for the loss function is as follows:

[0123] Among them, erf -1 Let Ψ be the inverse error function, and ζ be the risk task gain coefficient, used to control the gain strength of the inverse error function. The larger the value of ζ, the more significant the improvement in the risk task learning rate under low consistency. When Ψ=1, the error rate... -1(1)→∞, pruning is required; when Ψ is low (e.g., Ψ<0.99), the inverse error function will give a positive value, thus achieving superlinear growth of the learning rate.

[0124] The prediction consistency index Ψ serves as a global control signal for the learning rate. By quantifying the collaborative state among multi-task predictions in real time, it generates a scalar signal characterizing the global state of the system. This scalar signal is then injected into the loss function construction process of each network component as a dynamic adjustment benchmark, dynamically adjusting the effective learning rate of each network component. This intelligent allocation of dynamic weights enables the model to autonomously weigh the optimization priorities of different tasks during training, achieving optimization guidance based on this global awareness. This gives the gradient backpropagation process system-level state awareness, driving the model parameters to converge towards the overall optimal solution. Thus, a spiral dynamic adjustment feedback mechanism is constructed, moving from the global to the local, and then from the local feedback back to the global.

[0125] To this end, this invention provides insurance brokers with a unified quantitative assessment tool. Its core capability lies in outputting multi-dimensional, unbiased, and objective assessment scores for any policyholder-insurance product combination. It is not only suitable for the rapid assessment and verification of conventional combinations, but its unique value is also reflected in its decision support for complex and challenging combinations: by revealing the quantitative performance of the combination in different dimensions such as market demand, profitability, and long-term risk in parallel, the system can accurately locate the root cause of decision-making dilemmas and transform vague empirical trade-offs into clear multi-objective optimization problems, thereby supporting brokers to make optimal decisions under the constraints of strategic objectives.

[0126] In a preferred embodiment, a premium optimization head 300, composed of one or more fully connected neural networks, is configured in parallel with multiple task evaluation heads. m In step S300, the combined evaluation context vector is input into the premium optimization head, and an optimized premium value is generated. The optimized premium value does not participate in the normalization mapping.

[0127] The training of the fully connected neural network of the premium optimization head participates in the end-to-end training. Based on available information such as the optimized premium, the market average premium, and the current iteration's claim risk quantification score (claim risk index), reinforcement learning optimization is performed through a reward function R.

[0128] The reward function R is:

[0129] Where Rrisk is the risk-matching reward, and its formula is:

[0130] P optimized To optimize the premium output by the premium optimization head, I riskThe claims risk index is output by the claims risk quantification assessment head, where k is a positive hyperparameter representing the company's risk pricing strategy, used to convert the risk index into a benchmark premium. When optimizing the premium P... optimized Approaching the risk benchmark k·I risk At this point, the reward value reaches its maximum, and any deviation in any direction will cause the reward value to decrease, thereby encouraging a positive correlation between premiums and risk levels and reflecting the fairness of pricing.

[0131] Where R profit For profit-driven incentive items, the formula is:

[0132] C expected The expected total cost is calculated as C. expected =I risk ·A total +C fixed A total For the total sum insured, C fixed This is a fixed cost. This encourages the model to pursue higher profits while satisfying other constraints.

[0133] Where R penalty_high The market competitiveness penalty item is formulated as follows:

[0134] P market For the market average premium, τ high This is a hyperparameter greater than 1, defining the maximum acceptable premium cap (e.g., 1.2). This penalty only takes effect when the optimized premium exceeds the market premium cap, and the penalty strength is proportional to the extent of the excess, preventing the model from setting prices that are out of line with the market in pursuit of high profits.

[0135] Where R penalty_low The formula for the base price guarantee penalty is as follows:

[0136] τ low is a hyperparameter greater than or equal to 1, representing the minimum cost markup rate (e.g., 1.05). This parameter takes effect when the optimized premium is lower than the cost floor, ensuring the commercial sustainability of pricing and preventing the model from incurring losses in pursuit of risk matching or to avoid high penalties.

[0137] Where λ, μ, ν are positive hyperparameters used to balance the relative importance of various rewards and punishments.

[0138] This invention is based on a multi-objective optimization reward function. The training process aims to maximize the overall reward and updates the model parameters using a gradient ascent algorithm. During backpropagation, the parameters of the premium optimization head are adjusted according to the reward function to optimize the gradient direction of the premium, enabling the model to gradually learn a pricing strategy that balances risk matching, profit objectives, and market competitiveness. This significantly improves the mathematical stability and business rationality of the model output.

[0139] The training of the optimization head's neural network can participate in the training of the backbone network in different ways to ensure synchronous convergence. For example, the backbone network can be fully trained to convergence first, its weights frozen, and then the optimization head can be trained using a validation-level method; the learning rate of the backbone network can be set much smaller than that of the optimization head; the optimization head can be trained for a period of time with the backbone network fixed, and then the backbone network can be released at a very low frequency for fine-tuning, alternating between these methods; strong weight regularization can be applied to the backbone network, and changes in its output feature distribution can be closely monitored, but these methods are not limited to these.

[0140] By combining the risk indicators quantified by the risk assessment head, an optimized premium that matches the risk and takes into account the profit is automatically calculated. The final assessment score is provided to the broker as output, enabling the invention to provide combined intelligent pricing while completing quantitative assessment, and to achieve human-computer interaction that is closer to the essence of business decision-making.

[0141] In a preferred embodiment, the present invention further includes an insurance interpretation generation step for generating insurance interpretation text for the policyholder.

[0142] An insurance explanatory text is an explanatory text provided to policyholders in insurance practice to explain key information such as the coverage, applicability, risks, and value of an insurance package, aiming to assist and guide policyholders in making decisions.

[0143] Therefore, such as Figure 3 As shown, the architecture of this invention further includes an interpretation generation layer 400 and a bias generator 500. The interpretation generation layer 400 is composed of a natural language generation model based on a self-attention mechanism, such as a Transformer with a decoder frame, whose masking operation can be deployed on a server. In this embodiment, a model based on the GPT-3.5 architecture is used, and supervised fine-tuning is performed using an insurance instruction dataset. The GPT vocabulary is used as is.

[0144] The bias generator 410 is used to generate a bias vector with the same dimension as the vocabulary size of the language generation model, which is added to the logistic value vector of the language generation model based on the quantization score and strategic weight. It includes a first, second and third fully connected layer composed of a fully connected neural network. The first fully connected layer is used to capture the complex, non-linear interaction between the quantization score and strategic weight to generate hidden layer feature vectors. The second fully connected layer is used to generate a projection vector with the same dimension as the vocabulary size of the language generation model based on the hidden layer feature vector. The third fully connected layer is used to generate a gating vector for selectively amplifying or suppressing the bias based on the hidden layer feature vector.

[0145] Specifically, such as Figure 4 As shown, in step S401, for the policyholder feature vector, product feature vector, and optionally, each atomic feature vector in the overall feature vector is embedded into the corresponding keyword vector in the pre-built feature keyword library to obtain atomic semantic vectors. The atomic semantic vectors are then multiplied by the global decision coefficient k to obtain the gain atomic semantic vector.

[0146] Wherein, the global decision coefficient k=v quant ×w strategic T .

[0147] Where v quant The quantitative score vector, composed of the quantitative scores from each assessment head in step S300, represents the objective matching degree of the insurance portfolio in multiple dimensions.

[0148] w strategic This is the strategic weight vector for the quantitative tasks of each evaluation head, where strategy represents the importance attached to business metrics.

[0149] Where v quant and w strategic All elements should remain consistent according to the order of the assessment tasks.

[0150] Strategic weight vectors, pre-configured for each assessment head to reflect the company's strategy, are used to inject the strategic preferences of the insurance portfolio recommender into the interpretation generation process.

[0151] As an example, the quantified scores are normalized by linearly scaling the sequence of quantified scores for each task indicator to generate a floating-point vector constrained to the [0,1] interval. Strategic weights can be predefined as floating-point numbers within the [0,1] interval.

[0152] In step S401, simultaneously, the quantization score vector v is... quant and strategic weight vector w strategic The input consists of a bias generator network composed of three fully connected networks, and the output is a bias vector B with the same dimension as the vocabulary size of the language generation model.

[0153] The formula for generating B is: B = (W2· tanh( W1·z + b1) + b2) ⊙ Softplus( W3· tanh( W1· z + b1)+ b3).

[0154] in

[0155] Z is a d+1 dimensional vector, where d is the total number of evaluation heads, and z represents a d+1 dimensional fusion vector composed of the product of the quantization score vector and the strategic weight vector and the superposition of the magnitude of the product vectors.

[0156] Through element-wise multiplication v quant ⊙w strategic This approach amplifies the impact of scores on metrics prioritized by business strategy while weakening the influence of less important metrics, thereby integrating quantitative results with strategic weights. The magnitude of the fused vector is concatenated with the vector itself, preserving both the vector's directionality and the strength of the aggregation. The magnitude serves as a scalar, providing a global strength signal.

[0157] tanh(W1·z + b1) represents the output of the first fully connected layer, which takes the fused vector z as input to weights W1 and bias b1 and uses the hyperbolic tangent tanh activation function. This output is the hidden feature vector of the first fully connected layer. This layer captures the complex, non-linear interaction between quantization scores and strategic weights, for example, learning that "a stronger bias is needed when high profit margins align with strategic priorities."

[0158] W2·tanh(W1·z + b1) + b2 represents the input of the hidden layer feature vector output by the first fully connected layer into the output of the second fully connected layer with weights W2 and biases b2. This output is a projection vector with the same dimension as the vocabulary size of the language generation model, with each dimension corresponding to a word in the vocabulary, thereby projecting the hidden layer features onto the high-dimensional vocabulary space.

[0159] Softplus( W3· tanh( W1· z + b1) + b3) means that the hidden layer feature vector output from the first fully connected layer is input into the weights W3 and the bias b3, and the output of the third fully connected layer is activated by the Softplus activation function. This output is a gating vector. Softplus ensures that the gating value is always positive, acting as a magnifying glass and / or filter.

[0160] This means that the final bias applied to each word in the vocabulary is determined by the initial projection bias vector and a dynamically calculated gating factor vector, allowing the model to learn to selectively amplify or suppress the bias in different regions of the vocabulary under different evaluation scenarios.

[0161] In step S402, the gain atomic semantic vector is serialized, and the input is a natural language generation model based on a self-attention mechanism, which calculates the initial logical value vector L at its output layer. original The bias vector B is compared with the initial logic value vector L. original Perform vector addition to generate a corrected logical value vector L. modified Based on the modified logical value vector L modified Calculate the probability distribution of the next word to ultimately generate the insurance explanation text.

[0162] The serialization process includes concatenating the gain atomic semantic vectors in a certain order and embedding positional information. The certain order of concatenation includes placing features belonging to the policyholder, the product, and the overall package in the same segment, preferably with the overall package features and sub-product features being adjacent to each other to prevent contamination.

[0163] The dimension of this logical value vector is the same as the vocabulary size of the language generation model, and each element of the vector represents the probability that the model considers the corresponding word in the vocabulary to be the next word.

[0164] The vector addition operation is represented as follows: ∈ [1, V], (L modified = (L original +

[0165] This means that for a word index i from 1 to V, the i-th element of the new vector is equal to the sum of the i-th element of the original vector and the i-th element of the bias vector. That is, the bias vector is added to the original logical value vector element by element to generate the corrected logical value vector.

[0166] The training of the language generation model and the bias generation network are performed simultaneously, and their total loss function is: L total = L LM + λ · L reg Where L LM It is the loss of the language generation model, used to ensure that the generated text is fluent and reasonable.

[0167] L regIt is a component regularization loss term used to control bias behavior, and its loss function is:

[0168] Where B is the bias vector, ||B| ² represents the square of the L2 norm of vector B, used for regularization to prevent excessively large bias values ​​that could lead to choppy generated text. δ is a preset minimum similarity threshold, a real constant greater than 0.

[0169] Represent vectors B and v quant • Cosine similarity of M.

[0170] v quant • M represents a target direction vector mapped from the quantized score vector through a learnable matrix M. The initial value of M is random or initialized according to a predetermined strategy and updated during training using backpropagation and gradient descent algorithms to minimize the loss function. Its purpose is to learn how to map the quantized score vector v... quant It is mapped to a semantic direction related to the vocabulary space.

[0171] The purpose of this loss term is to force the direction of the bias vector B to be as close as possible to the target direction v. quant • M remains consistent. That is, the loss term does not directly require the model to generate a specific word, but requires that the bias it generates be directionally related to the current quantitative evaluation result. For example, it forces the bias generation network to learn to map "high profit margin" to a semantic direction related to "value".

[0172] In the entire network iteration process of this preferred embodiment architecture, the backbone neural network of the quantization evaluation architecture is first trained, and after it converges, the parameters of the quantization evaluation architecture network are frozen. Then, the interpretation generation network is trained using a validation sample set.

[0173] In explaining each forward propagation of the generative network training, the input text features are evaluated, and the output v of the network architecture is quantized. quant and the pre-booked w strategy The bias vector B is calculated using the bias generation network according to the above formula. The bias vector B is then added element-wise to the Logits output by the language generation model, and the loss L is calculated. total This allows for the simultaneous updating of the parameters of the bias generator network through backpropagation. , , , , , The bias generation network is iteratively trained to produce biases that both preserve text quality and are semantically aligned with the quantization evaluation results.

[0174] The training samples for the explanatory text should contain explanatory text labels with different recommendation strengths, as shown in the example below.

[0175] Tag 1: Weak Recommendation Strength - Neutral Explanatory Type

[0176] Mr. Wang, this insurance package provides you with core protection against death and critical illness. Based on your age and family situation, the coverage of this package is appropriate for your current stage of life. Please carefully read the terms and conditions regarding coverage and exclusions, taking into account your budget.

[0177] Tag 2: Moderate Recommendation Strength - Suggestion Guidance

[0178] Mr. Wang, this combination of life insurance and critical illness insurance is a perfect match for your needs. It provides your family with solid risk protection, ensuring that your family's finances are not severely impacted in the event of an accident or illness. This is a market-proven and reliable protection plan, and we recommend that you pay close attention to it.

[0179] Tag 3: Strong Recommendation Strength - Active Recommendation Type

[0180] Mr. Wang, we strongly recommend this insurance package to you. It precisely meets your core protection needs as the breadwinner of your family. This package not only provides substantial life and health protection, but the underlying products also feature stable risk management and efficient service. Building this safety net for you and your family is a wise and responsible decision.

[0181] Therefore, based on the quantitative evaluation results, this embodiment injects a single, aggregated scalar coefficient into the input, enabling the model to learn the correlation between the input scale and the output text style. The model adjusts the tone and recommendation strength of the generated text according to the volume of the input. At the output, a high-dimensional, differential bias vector is injected to guide the probability distribution of lexical terms during the generation process. This generates insurance explanation text for policyholders that is aligned with the quantitative evaluation results and has a guiding nature, thereby improving the transparency of decision support and the depth of human-computer interaction.

[0182] Application Example 1

[0183] As a typical application, this invention can be applied to family-type policyholders. The feature system and assessment tasks can be arranged with reference to the following example.

[0184] Family insurance often takes the form of product packages, which include multiple sub-products such as life insurance for the male head of the household, critical illness insurance for the female head of the household, and accident insurance for the children.

[0185] The policyholder's basic information also includes the basic information of the decision-maker in the family's finances, as well as family structure information, such as the number of family members, the number and age of minor children, and the number of parents living with the family.

[0186] Insurance coverage information may include, for example, total family premium expenditure, which can be calculated as the average annual total premium for all family members; family coverage breadth, which can be calculated as the number of different types of insurance the family already holds; and family risk preference, which can be comprehensively scored based on the family's coverage depth (e.g., the ratio of average sum assured to income) and insurance propensity (e.g., the ratio of family investment-linked insurance premiums to total family premiums). For example, the standardized coverage depth and investment propensity indicators can be analyzed using principal component analysis (PCA). The first principal component PC1 score is then mapped to the (0,1) interval using the Sigmoid function 1 / (1 + exp(-PC1)). A score close to 1 indicates high risk aversion, a score close to 0 indicates high risk aversion, and a score around 0.5 indicates risk neutrality.

[0187] As a preferred embodiment, the present invention also provides a lifecycle event detector for quantifying the likelihood of significant changes in a customer's family or financial structure based on policyholder characteristics. Specifically, it generates a beneficiary concentration change intensity index, a core relationship network stability index, and a policyholder lifecycle fluctuation index vector based on the policy's basic information change log, including policy number, change date, change type (such as beneficiary change, address change), previous content, and new content, as well as core policy information such as policy number, policyholder, insured, effective date, and sum insured.

[0188] The beneficiary concentration change intensity index is obtained by identifying beneficiary change operations that occurred in all policies within the past N months (e.g., 12 months) and clustering them by change date. The cluster with the largest number of policies that changed within the shortest time window (e.g., 30 days) is then used as the beneficiary concentration change intensity index. The larger the number and the higher the value, the stronger the volatility. A concentration of beneficiary changes in multiple policies within a short period is a strong signal of significant relationship changes.

[0189] The protection tier mutation index is calculated by arranging the effective dates and sum insured of all life insurance and critical illness insurance policies in the customer's history in chronological order, and calculating the rate of change of sum insured between two adjacent policies: (new sum insured - old sum insured) / old sum insured. The maximum absolute value among all the change rates is then taken as the protection tier mutation index. The larger the value, the more drastic the adjustment of the customer's protection needs has been. Dramatic changes in the depth of protection reflect a major shift in economic responsibility or mindset.

[0190] The core relationship network stability index is calculated by statistically analyzing the number of new insured persons appearing in a customer's policy over the past N months, and the number of insured persons who no longer appear during the same period. The sum of the number of new insured persons and the number of insured persons who have disappeared is the core relationship network stability index. A value greater than 0 indicates that the core protection relationship network is in flux. This index observes the addition and disappearance of the policyholder-insured relationship network in a customer's policy.

[0191] This combination of indices generates a policyholder life cycle volatility index vector, such as [beneficiary concentration change intensity = 3, protection level mutation degree = 0.85, core relationship network stability = 2]. This vector comprehensively represents the probability and intensity of a customer experiencing major life cycle events in the near future.

[0192] The lifecycle event detector is a leading early warning indicator. Its generated vector is used as the input of the policyholder's features to indicate to the model that the policyholder's background is changing drastically, and their traditional preferences may no longer be reliable, requiring more attention to their latest behavior.

[0193] This invention also generates a recent protection demand change characteristic to directly and quantitatively describe the recent changing trend of a customer's protection portfolio, including: defining a recent time window, such as the past 6 months, and a baseline window, such as the 18 months prior to the recent window; mapping all policies to core protection dimensions, such as life insurance, critical illness insurance, medical insurance, accident insurance, and savings / annuity, based on product master data; calculating the total sum insured for each protection dimension within the recent time window and the baseline time window; and calculating the demand change trend value for each dimension: Trend value = (Total sum insured in the recent window - Total sum insured in the baseline window) / The baseline window total sum assured, where a smoothing term can be added to the denominator to prevent division by zero, represents a standardized rate of change. Positive values ​​indicate increasing demand, while negative values ​​indicate decreasing demand. Combining these trend values ​​generates a demand trend vector, which represents the recent changes in protection demand. For example: [Life Insurance: +0.35, Critical Illness Insurance: +0.10, Medical Insurance: 0.00, Accident Insurance: -0.20, Savings / Annuity: +0.50]. This vector clearly depicts that the policyholder is significantly increasing their investment in life insurance and savings, while reducing their focus on accident insurance. This profile is highly instructive when recommending family-type or savings-type life insurance. Recent changes in protection demand are a descriptive indicator of the current situation. Inputting this as a policyholder characteristic tells the model: regardless of the reason, this customer is currently demonstrating through their actions that they need more A and less B.

[0194] By inputting the two feature vectors—the policyholder's life cycle fluctuation index vector and the recent protection need change vector—along with other features into the model, the model gains a dynamic insight. It can learn not only what the policyholder's past state was, but also what state it is becoming. This enables a truly forward-looking and accurate assessment. In scenarios involving sensitive information such as changes in marital and reproductive status, this invention's quantitative assessment can achieve the same assessment effect based on life cycle stages and changes in protection needs, even without accessing private data, thus avoiding the use of biased labels.

[0195] Application Example 2

[0196] As another typical application, this invention can be applied to corporate group policyholders, and the feature system and assessment tasks can be arranged with reference to the following example.

[0197] Basic information about corporate group policyholders can include, for example, industry category (which can be embedded using standard industry classification codes), company size (which can be mapped to discrete size levels such as micro-enterprises, medium-sized enterprises, and large enterprises based on the total number of employees and annual revenue), years of operation (calculated based on the current date minus the establishment date), and geographic location information (including the company's registered address or main operating address). It can also be linked to external data such as the average salary level and medical costs in the region.

[0198] This can also include employee risk profiles, such as calculating the proportion of employees in high-risk positions based on demographic data of all employees, calculating the company's expected morbidity rate, and using industry-standard expected morbidity / mortality rates as a benchmark to generate risk deviation characteristics. It can also include information on financial stability, such as weighted summation of data like debt-to-equity ratio, profit margin, and credit rating after standardization and dimensionless transformation to obtain a financial health score. Furthermore, it can include the aggressiveness of the benefits strategy, such as comparing the company's historical benefits expenditure with the industry average, and comparing per capita benefits expenditure with the industry benchmark to obtain a relative score. It can also include talent competition pressure; for example, companies in industries / regions with high turnover rates are more motivated to offer generous benefits, and this can be achieved using talent turnover rate data for the company's industry and region. It can also include information related to risk management investment and awareness, such as the proportion of employees in high-risk positions and the proportion and level of investment in dedicated risk control services; a lower proportion and higher investment indicate that the company is actively managing its own risks, a strong signal of a high-quality client. Finally, it can include industry risk level information, such as mapping industry classification codes to risk numerical levels, referencing external authoritative risk ratings, or calculating based on statistical data of historical industry claims to generate numerical features with ranking properties.

[0199] The insurance history information of corporate group policyholders can include historical claims experience, such as historical loss ratio (total claims amount / total premium), claims frequency (total claims number / total number of policies), and average claims frequency per person (total claims number / total number of insured persons). Historical claims data for enterprises is a statistical result of group behavior, supported by the law of large numbers, representing the stable risk level of the enterprise as a risk pool. It has high reliability and is a strong signal for future risk prediction. When the model processes corporate policyholders, historical claims experience characteristics will shift from a secondary reference to a dominant factor. It can also include employee participation intentions, for example, using historical participation rate as an agency indicator: participation rate = actual number of insured employees / total number of employees meeting the participation conditions.

[0200] Figure 5 This is a schematic block diagram illustrating the quantitative evaluation system according to the present invention. Figure 3As shown, according to an embodiment of the present invention, the quantitative evaluation system 70 includes: a processor 71, a memory 72, and computer program instructions 73 stored in the memory 72.

[0201] When the computer program instructions 73 are executed by the processor 71, they can implement the functions of each functional module of the quantitative evaluation method according to the embodiments of the present invention, and / or execute each step of the quantitative evaluation method according to the embodiments of the present invention.

[0202] Each module in the quantitative evaluation system according to the present invention can be implemented by a processor running computer program instructions stored in memory, or by computer instructions stored in a computer-readable storage medium of a computer program product according to the present invention being implemented by a computer running.

[0203] The computer-readable storage medium can be any combination of one or more computer-readable storage media. For example, one computer-readable storage medium contains computer-readable program code for randomly generating sequences of action instructions, and another computer-readable storage medium contains computer-readable program code for target recognition. Computer-readable storage media may include, for example, a memory card of a smartphone, a storage component of a tablet computer, a hard disk of a personal computer, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB memory, or any combination of the above storage media.

[0204] In any work relating to defining the scope of the claims of this invention, the specific embodiments provided are intended to adequately support appropriate equivalent or superior embodiments derived therefrom, and should in no way be construed as not extending beyond the examples themselves.

Claims

1. A quantitative evaluation method for insurance portfolios based on multi-task learning and attention mechanisms, characterized in that, This method is used to quantitatively assess the fit between an insured person and an insurance portfolio consisting of a variable number of insurance products across multiple insurance business metrics. The method includes: The policyholder feature vector is input into the policyholder tower, which is composed of a neural network based on a self-attention mechanism, and then encoded to generate the policyholder feature embedding vector. The product feature vector of each component product in the insurance portfolio is input into the same product tower, which is composed of a fully connected neural network, and then encoded to generate the product feature embedding vector of that component product. The product feature embedding vectors of each component product are concatenated to form the combined product feature embedding vector. Optionally, when a package product is identified, the overall feature vector of the package product is input into a package tower composed of a fully connected neural network for encoding to generate an overall feature embedding vector of the package product. The product feature vector of each sub-product in the package product is input into the same product tower for encoding to generate the product feature embedding vector of that sub-product. The product feature embedding vectors of each component product and each sub-product in the package product, as well as the overall feature embedding vector of the package product, are concatenated to form the combined product feature embedding vector. Optionally, when multiple package products are identified, the overall feature vector of each package product is input into the same package tower for encoding to generate the overall feature embedding vector of the package product, and the product feature vector of each sub-product in the package product is input into the same product tower for encoding to generate the product feature embedding vector of the sub-product. The product feature embedding vectors of each component product and each sub-product in each package product, as well as the overall feature embedding vector of each package product, are concatenated to form the combined product feature embedding vector. Using the policyholder feature embedding vector as the query vector and the combined product feature embedding vector as the key vector and value vector, the input is a cross-tower attention aggregation layer composed of a neural network based on an attention mechanism to generate a combined evaluation context vector representing the interaction relationship between the policyholder and the combination to be evaluated. The combined evaluation context vector is input into multiple task-specific evaluation heads configured in parallel, each composed of fully connected neural networks, and the quantitative scores of each insurance business indicator are output respectively.

2. The quantitative evaluation method as described in claim 1, wherein, Also includes: The quantitative scores of each insurance business indicator are input into the normalization processing layer for normalization mapping, and a set of standardized multi-task quantitative scores are output.

3. The quantitative evaluation method as described in claim 1, wherein, The backbone neural network, composed of the neural networks that make up the policyholder tower, product tower, package tower, cross-tower attention aggregation layer, and multiple evaluation heads, is trained end-to-end. The multiple task-specific evaluation heads include at least a conversion rate evaluation head, a profit margin evaluation head, a renewal rate evaluation head, and a claims risk evaluation head. In each training iteration, a prediction consistency index Ψ is calculated based on the quantitative scores of conversion rate, profit margin, renewal rate and claims risk to represent the relative deviation and abnormal linkage among the four task evaluations. The loss function of each neural network is dynamically reconstructed using the prediction consistency index Ψ.

4. The quantitative evaluation method as described in claim 3, wherein, The formula for calculating the predictive consistency index Ψ is: In the formula, the numerator is the output of a two-layer perceptron network constructed and trained together with the backbone neural network. Where m is the market value vector, m=[L c , L p , Lr] T , L c , L p Lr are the quantitative scores for conversion rate, profit margin, and renewal rate, respectively. z is the concatenation of vector m and its geometric mean, z = m ⊕ [GeometricMean(m)] GeometricMean(m)=(L c × L p × Lr) / 3 , z is fed into a single hidden layer of a two-layer perceptron network, through the weights W of the hidden layer. h Bias b h And the tanh activation function performs nonlinear transformations and interactions on the input features. W o , b o These are the weights and biases of the output layer σ of the two-layer perceptron network. σ is the sigmoid function, and its output is compressed to (0,1); In the formula, the denominator is the risk deviation penalty term. Where v is the vector of predicted values ​​for the entire task, including conversion rate, profit margin, renewal rate, and claims risk. v=[L c , L p , L r , L claim ] T L claim It is a quantitative score of claims risk. It is the L2 norm of vector v after standardization. μ v , σ v Let v be the mean and standard deviation of vector v, 1 represents a vector of all 1s, and ε is a minimal positive constant. λ and β are the preset scaling factor and covariance influence factor, respectively.

5. The quantitative evaluation method as described in claim 4, wherein, In each training iteration The loss function for the neural network that dynamically reconstructs the policyholder tower, product tower, and package tower using the predictive consistency index Ψ is: Where η is the basic learning rate, η eff For the effective learning rate, ε is the minimum learning rate coefficient, K is the control slope steepness, Ψ0 is the learning rate switching center point, and δ is the numerical stability constant. The loss function for dynamically reconstructing the attention convergence layer of the neural network using the prediction consistency index Ψ is: Where γ is the inhibition strength coefficient; The loss function of the neural network that dynamically reconstructs the evaluation heads of conversion rate, profit margin, and renewal rate using the predictive consistency index Ψ is: Where α is the market task inhibition coefficient; The loss function of the neural network that dynamically reconstructs the claims risk assessment head using the predictive consistency index Ψ is: Among them erf -1 Let ζ be the inverse error function, and ζ be the risk task gain coefficient.

6. The quantitative evaluation method as described in claim 4, wherein, It also includes a premium optimization head composed of a fully connected neural network, which is configured in parallel with the multiple evaluation heads to input the combined evaluation context vector and generate an optimized premium value. The premium optimization head is trained in the end-to-end training process, using reinforcement learning optimization through a reward function R. The reward function R is: Where R risk The formula for the risk-matching reward item is as follows: P optimized To optimize the premium output by the premium optimization head, I risk The claim risk assessment head outputs a quantitative score for claim risk, where k is a positive hyperparameter. Where R profit For profit-driven incentive items, the formula is: C expected The expected total cost is calculated as C. expected =I risk ·A total +C fixed A total For the total sum insured, C fixed For fixed costs, Where R penalty_high The market competitiveness penalty item is formulated as follows: P market For the market average premium, τ high For a hyperparameter greater than 1, Where R penalty_low The formula for the penalty item for the minimum price guarantee is: τ low is a hyperparameter greater than or equal to 1, representing the minimum cost markup rate. Where λ, μ, ν are positive hyperparameters.

7. The quantitative evaluation method as described in claim 1, wherein, The policyholder feature vector is generated by embedding atomic feature vectors representing the policyholder's basic information and insurance history information with position information and then concatenating them in a predetermined order. The atomic feature vector representing the policyholder's basic information is a feature vector characterizing the policyholder's inherent attributes and long-term stable state. The atomic feature vector representing the policyholder's insurance history information is a feature vector derived from the policyholder's historical insurance behavior data, used to characterize their stable decision-making tendencies and long-term business interaction patterns; and / or The product feature vector is generated by embedding atomic feature vectors representing the basic information and business status information of the insurance product into position information and then concatenating them in a predetermined order. The atomic feature vectors representing the basic information of the insurance product are feature vectors characterizing the core design and inherent rules of the product, while the atomic feature vectors representing the business status information of the insurance product are feature vectors characterizing the product's market performance and strategic orientation; and / or The overall feature vector of the package product is generated by embedding atomic feature vectors representing the overall design, comprehensive value, and bundling strategy of the package into positional information and then concatenating them in a predetermined order. The product feature vectors of sub-products within the package product include atomic feature vectors representing the integration of individual attributes and combined roles of the sub-products; and / or The plurality of assessment heads also includes at least one of the following: a risk coverage adequacy quantification assessment head, a guarantee leverage ratio quantification assessment head, a potential complaint risk quantification assessment head, a short-term silent risk quantification assessment head, and a complaint tendency risk quantification assessment head.

8. The quantitative evaluation method as described in claim 1, wherein, The quantitative assessment method also includes an insurance explanation generation method for generating insurance explanation texts for policyholders. The method for generating this insurance interpretation includes: (1) Embed the corresponding keyword vectors from the pre-built feature keyword library into each atomic feature vector in the policyholder feature vector, product feature vector, and overall feature vector to obtain atomic semantic vectors. Multiply the atomic semantic vectors by the global decision coefficient k to obtain the gain atomic semantic vectors. Wherein, the global decision coefficient k=v quant ·w strategic T , Where v quant w is a quantized score vector composed of the quantized scores of each evaluation head. strategic The strategic weight vector corresponding to the quantization task of each evaluation head; Simultaneously, the quantized score vector and strategic weight vector are input into a bias generation network composed of three fully connected neural networks, which outputs a bias vector B with the same dimension as the vocabulary size of a natural language generation model composed of neural networks based on a self-attention mechanism. The formula for generating the bias vector B is: B = (W2· tanh( W1·z + b1) + b2) ⊙ Softplus( W3· tanh( W1· z + b1) +b3). in d represents the total number of evaluation heads, and z represents the d+1-dimensional fusion vector formed by the superposition of the multiplication vector of the quantized score vector and the strategic weight vector and the scalar of the magnitude of the multiplication vector. tanh(W1·z + b1) represents the hidden layer feature vector output by the first fully connected neural network with the fused vector z input to weights W1 and bias b1, and the hyperbolic tangent tanh activation function. W2·tanh(W1·z + b1) + b2 represents the projection vector of the second fully connected neural network with the same dimension as the vocabulary size of the natural language generation model, obtained by inputting the hidden layer feature vector into weights W2 and biases b2. Softplus(W3·tanh(W1·z + b1) + b3) represents a gated vector with the same dimension as the vocabulary size of the natural language generation model, which is output by the third fully connected neural network using the Softplus activation function after inputting the hidden layer feature vector with weights W3 and bias b3. (2) Serialize the gain atomic semantic vector and input it into the natural language generation model. Calculate the initial logical value vector with the same output dimension as the vocabulary size of the language generation model at its output layer. Perform vector addition operation between the bias vector B and the initial logical value vector to generate the corrected logical value vector. Calculate the lexical probability distribution based on the corrected logical value vector to finally generate the insurance explanation text. (3) The natural language generation model and the biased generation network are trained together, and the total loss function is: L total = L LM + λ · L reg Where L LM It is the loss of the language generation model. L reg This is the regularization loss term for the biased generator network, and its loss function is: Where δ is the preset minimum similarity threshold, which is a real constant greater than 0. Represent vectors B and v quant • Cosine similarity of M v quant M represents the target direction vector mapped from the quantized score vector through the learnable matrix M. The parameters of the bias-generated network are updated simultaneously during backpropagation. , , , , , M and the parameters of the natural language generation model.

9. A quantitative evaluation system for insurance portfolios based on multi-task learning and attention mechanisms, characterized in that, The system includes: processor; A memory; and computer program instructions stored in the memory, which, when executed by the processor, perform the method according to any one of claims 1 to 8.

10. A computer program product, characterized in that, Including computer-readable storage media, The computer-readable storage medium stores computer program instructions that, when executed by a computer, perform the method according to any one of claims 1 to 8.

Images