Prompt generation method and device, computer equipment and readable storage medium

Abstract

The invention relates to the technical field of machine learning, and provides a prompt generation method and device, computer equipment and a readable storage medium, and the method comprises the steps: obtaining business data, rule data and historical processing records, and carrying out the preprocessing, and forming a standardized data set; obtaining a rule matrix corresponding to the standardized data set in real time by adopting increment extraction; performing dimension reduction and distribution equalization on the high-dimensional features corresponding to the rule matrix by using correlation analysis, a feature selection algorithm and a data balance technology; through loss distribution fitting, cross validation and hyper-parameter optimization, constructing a field association map of association business field semantics, rule logic and historical operation habits corresponding to the rule matrix; and when new service information is accessed, generating processing prompt information according to the field association map and the service information. According to the method, the intelligent level of a complex business process is fundamentally improved, and information processing pain points in the fields of finance, medical health, old-age care and the like are effectively handled.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a prompt generation method, apparatus, computer device, and readable storage medium. Background Technology

[0002] In sectors requiring complex information processing, such as financial services, healthcare, and elderly care, call center staff often face challenges such as cumbersome business processes, high difficulty in data entry and verification, and low efficiency in risk assessment. Taking the financial sector as an example, customer account opening and loan applications involve the entry of information across dozens of fields. These fields have complex semantics and diverse verification rules (such as identity information compliance and asset verification logic). Manual data entry is prone to errors due to rule memory bias or operational negligence, resulting in extremely high subsequent verification costs. In the healthcare and elderly care sectors, patient health record creation and elderly care service processing require the integration of heterogeneous information from multiple sources, including medical records, insurance terms, and service agreements. Traditional processes rely on manual field-by-field rule comparison, which is not only time-consuming but may also lead to the omission of key risk points (such as conflicts between coverage and medical history, and the compatibility of service terms with health status) due to insufficient data correlation analysis.

[0003] In existing technologies, some auxiliary tools can only provide simple keyword searches or fixed template filling, failing to achieve deep correlation between business field semantics, rule logic, and historical operating habits, and even more so, struggling to dynamically output targeted prompts based on process progress. For example, customer service systems in the financial sector lack real-time correlation analysis of "customer asset type - risk level - product suitability," and in healthcare and elderly care scenarios, it is difficult to automatically identify potential contradictions between "medical history records - insurance liability - claims rules," forcing agents to frequently rely on external knowledge bases or human experience for judgment, thus limiting both decision-making efficiency and accuracy.

[0004] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0005] This application provides a prompt generation method, apparatus, computer device and readable storage medium, which aims to solve the core defects of the prior art: (1) the sample selection relies on experience or random strategies and does not design a quantitative evaluation system for the professional terminology characteristics of insurance texts; (2) the feature extraction model is single and it is difficult to take into account both the global semantics and local fine-grained features of the text; (3) the error correction lacks business rule guidance and the quality of generated samples is low; (4) the data feedback mechanism is out of touch with the dynamic changes in the domain and the model has poor adaptability.

[0006] Firstly, this application provides a prompt generation method, including: Acquire business data, rule data, and historical processing records; preprocess the business data, rule data, and historical processing records to form a standardized dataset; Incremental extraction is used to obtain the rule matrix corresponding to the standardized dataset in real time; correlation analysis, feature selection algorithms and data balancing techniques are used to perform dimensionality reduction and distribution equalization on the high-dimensional features corresponding to the rule matrix. By fitting the loss distribution, cross-validation, and hyperparameter optimization, a field association graph of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix is ​​constructed. When new business information is accessed, processing prompts are generated based on the field association graph and the business information.

[0007] In some embodiments, the preprocessing of the business data, rule data, and historical processing records to form a standardized dataset includes: cleaning the business data, rule data, and historical processing records to remove duplicate records, correct abnormal records, and fill or remove missing records; and using encoding technology to normalize the format of heterogeneous data, eliminating naming, type, and format conflicts, and unifying the semantics and structure of data fields.

[0008] In some embodiments, the step of incrementally extracting the rule matrix corresponding to the standardized dataset in real time includes: monitoring the update status of the data source and obtaining the latest rule data in real time through a preset incremental extraction mechanism; parsing the obtained rule data into a structured rule matrix and generating a logical expression corresponding to the rule matrix, wherein the logical expression is used to characterize the verification logic and risk assessment logic of the business rules.

[0009] In some embodiments, the step of constructing a field association graph corresponding to the rule matrix, which includes semantic association of business fields, rule logic, and historical operation habits, through loss distribution fitting, cross-validation, and hyperparameter optimization, includes: fitting the loss distribution using the moment estimation method based on loss data in historical processing records and obtaining parameter estimation results; evaluating the predictive performance of the initial model using cross-validation and bootstrap methods to identify overfitting or underfitting problems; adjusting the model hyperparameters using a grid search combined with random search, and selecting, transforming, and extracting new features from the original features to construct a field association graph containing semantic association of business fields, rule logic mapping, and historical operation habit features.

[0010] In some embodiments, generating processing prompts based on the field association graph and business information includes: during the information entry process, automatically completing missing business information, generating customized entry templates, and recording user operation habits based on the semantic relationships of the fields in the field association graph; performing real-time compliance verification on the entered business information based on the verification logic of the rule matrix, marking key fields that trigger risk rules, and dynamically updating processing prompts in conjunction with the current status of the business processing flow.

[0011] In some embodiments, before constructing the field association map of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization, the method further includes: performing context association processing on the text information in the business data using the Transformer architecture; and fusing multiple types of business information such as text, numbers, and images through a multimodal learning model to achieve joint feature extraction of heterogeneous data.

[0012] In some embodiments, the acquisition of business data, rule data, and historical processing records includes: acquiring business data and historical processing records from internal business systems, application programming interfaces, and databases; and acquiring supplementary business data, rule data, and industry characteristic data from external partner data sources, industry statistical information platforms, and publicly available academic research data.

[0013] Secondly, this application provides a prompt generation device, comprising: The data acquisition unit is used to acquire business data, rule data, and historical processing records; and to preprocess the business data, rule data, and historical processing records to form a standardized dataset. The matrix acquisition unit is used to acquire the rule matrix corresponding to the standardized dataset in real time using incremental extraction; and to perform dimensionality reduction and distribution equalization processing on the high-dimensional features corresponding to the rule matrix using correlation analysis, feature selection algorithms and data balancing techniques. The graph generation unit is used to construct a field association graph of the semantics of the associated business fields, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization. The prompt generation unit is used to generate processing prompt information based on the field association graph and business information when new business information is accessed.

[0014] Thirdly, this application also provides a computer device, comprising: Memory and processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, when executing the computer program, implement the steps of the prompt generation method as described in the first aspect above.

[0015] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the prompt generation method described in the first aspect above.

[0016] This invention provides a prompt generation method that, through multi-source data standardization processing, field association graph construction, and dynamic prompt mechanism, breaks through the limitations of traditional auxiliary tools and provides a universal intelligent decision support solution for complex business scenarios.

[0017] By acquiring business data, rule data, and historical processing records from internal and external data sources, abnormal data is cleaned and formatted through encoding techniques to form a standardized dataset. Incremental extraction is used to update rule data in real time, transforming it into a structured rule matrix. Correlation analysis, feature selection, and data balancing techniques are employed to reduce feature dimensionality and balance data distribution. The Transformer architecture is used to process text information, combining multimodal learning to fuse text, numbers, and images. Through loss distribution fitting and model optimization, a field association graph is constructed that relates business field semantics, rule logic, and user habits. During business information entry, information is automatically completed based on the field association graph, customized templates are generated, and user habits are memorized. Simultaneously, the entered content is verified in real time according to the rule matrix, and risk fields are dynamically marked and processing prompts are updated.

[0018] By cleaning, normalizing, and incrementally extracting multi-source data, this solution addresses the challenges of merging heterogeneous data in fields such as finance and healthcare, forming a unified and standardized data foundation that adapts to the diverse data needs of complex business scenarios. Field association graphs integrate business field semantics, rule logic, and historical operational habits, enabling automatic information completion and customized template generation, significantly reducing manual data entry time and lowering the error rate caused by field omissions or rule misjudgments. Based on real-time rule validation and model risk assessment, key fields are dynamically marked and combined with process progress update prompts, helping agents quickly locate core issues and improving decision-making efficiency and accuracy. This is particularly suitable for risk-sensitive scenarios such as financial risk control and medical insurance compliance review. After stripping away specific domain terminology, a general technical solution covering "data integration - model building - real-time assistance" is formed, which can be flexibly migrated to multiple scenarios such as financial customer service, medical record creation, and elderly care services, solving common problems in complex information processing and possessing broad application value.

[0019] Through the above technical solutions, this method fundamentally improves the intelligence level of complex business processes, realizes the upgrade of decision-making mode from "human experience-driven" to "data and model-driven", and effectively addresses the pain points of information processing in fields such as finance, medical health and elderly care.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic flowchart illustrating the steps of a prompt generation method provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the application of a prompt generation method provided in this application in the insurance field; Figure 3 This is a schematic diagram of the structure of a prompt generation device provided in an embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0027] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0030] In sectors requiring complex information processing, such as financial services, healthcare, and elderly care, call center staff often face challenges such as cumbersome business processes, high difficulty in data entry and verification, and low efficiency in risk assessment. Taking the financial sector as an example, customer account opening and loan applications involve the entry of information across dozens of fields. These fields have complex semantics and diverse verification rules (such as identity information compliance and asset verification logic). Manual data entry is prone to errors due to rule memory bias or operational negligence, resulting in extremely high subsequent verification costs. In the healthcare and elderly care sectors, patient health record creation and elderly care service processing require the integration of heterogeneous information from multiple sources, including medical records, insurance terms, and service agreements. Traditional processes rely on manual field-by-field rule comparison, which is not only time-consuming but may also lead to the omission of key risk points (such as conflicts between coverage and medical history, and the compatibility of service terms with health status) due to insufficient data correlation analysis.

[0031] In existing technologies, some auxiliary tools can only provide simple keyword searches or fixed template filling, failing to achieve deep correlation between business field semantics, rule logic, and historical operating habits, and even more so, struggling to dynamically output targeted prompts based on process progress. For example, customer service systems in the financial sector lack real-time correlation analysis of "customer asset type - risk level - product suitability," and in healthcare and elderly care scenarios, it is difficult to automatically identify potential contradictions between "medical history records - insurance liability - claims rules," forcing agents to frequently rely on external knowledge bases or human experience for judgment, thus limiting both decision-making efficiency and accuracy.

[0032] Therefore, a method is urgently needed to solve at least one of the above problems.

[0033] To resolve the above issues, please refer to [link / reference]. Figure 1 , Figure 1 This is a schematic flowchart illustrating a prompt generation method according to an embodiment of this application. The prompt generation method can be implemented by a computer device, which can be deployed on a single server or a server cluster. It can also be deployed on a handheld terminal, laptop, wearable device, or robot, etc.

[0034] It should be noted that the acquisition of any information mentioned in the provided methods is in compliance with relevant regulations and is carried out with the user's consent, and will not infringe on the user's privacy or violate relevant laws and regulations.

[0035] To solve the above problem, please refer to Figure 1 Specifically, such as Figure 1 As shown, the provided prompt generation method includes steps S101 to S104. Details are as follows: Step S101. Obtain business data, rule data, and historical processing records; preprocess the business data, rule data, and historical processing records to form a standardized dataset.

[0036] Specifically, business data (such as basic customer information and business processing content), rule data (such as business verification rules and risk assessment logic), and historical processing records (such as historical data entry and historical decision results) are collected from multiple sources. The data is then preprocessed, including cleaning and format normalization, to eliminate semantic and structural differences between heterogeneous data and form a unified standard dataset.

[0037] In financial scenarios, data acquisition involves obtaining business data such as customer identity information, asset type, and risk level requirements from the bank's core system (account opening data), credit reporting platform (credit records), and loan application system (asset proof, income statements); extracting identity information compliance rules (such as ID card validity verification) and asset proof logic rules (such as the ratio between income statements and loan amounts) from regulatory compliance documents; and obtaining the completeness of previously entered fields, approval results, and risk marking information from historical loan approval records.

[0038] Preprocessing involves cleaning up duplicate customer information (such as multiple duplicate account opening records for the same customer), filling in missing income verification fields (derived through historical averages or rules), performing One-Hot encoding on the "Asset Type" field (e.g., converting "Real Estate", "Deposit", and "Stocks" into numerical vectors), and standardizing the date format (e.g., converting "2024 / 01" and "January 2024" into "2024-01").

[0039] In the context of healthcare and elderly care, data acquisition involves obtaining patient medical history (such as past illnesses and surgical records), coverage (such as drug reimbursement catalogs), and service terms (such as health indicators corresponding to nursing levels) from hospital HIS systems (electronic medical records, diagnostic reports), medical insurance platforms (coverage terms), and elderly care institution management systems (service agreements, health monitoring data); extracting coverage responsibility rules (such as reimbursement ratios for chronic disease medications) and service matching rules (such as the correspondence between disability levels and nursing service types) from medical insurance policy documents; and obtaining the completeness of fields in past health records, coverage plan matching results, and risk warning records from historical records.

[0040] Preprocessing involves cleaning contradictory diagnostic records (such as repeated diagnoses of the same patient at different times), removing invalid medical histories (such as irrelevant symptoms older than 10 years), uniformly encoding "symptom names" (such as using the ICD-10 standard), and converting unstructured medical record text into structured fields (such as extracting keywords such as "hypertension" and "diabetes" through NLP).

[0041] By addressing the issues of fragmented data sources and inconsistent formats in financial and healthcare scenarios (such as asset verification documents in finance and unstructured medical records in healthcare), a standardized dataset that can be processed uniformly is created, laying the foundation for subsequent rule parsing and model training. By cleaning outliers and filling in missing values, the impact of manual data entry errors on subsequent analysis is reduced. For example, this avoids misjudgments of loan amounts due to missing income statements in financial scenarios, or mismatches in insurance plans due to omitted medical history in healthcare scenarios.

[0042] Step S102. Incremental extraction is used to obtain the rule matrix corresponding to the standardized dataset in real time; correlation analysis, feature selection algorithm and data balancing technology are used to perform dimensionality reduction and distribution equalization on the high-dimensional features corresponding to the rule matrix.

[0043] Specifically, incremental extraction technology is used to obtain updated rule data in real time and parse it into a structured rule matrix (such as "field A + condition B → verification result C"). For high-dimensional features in the rule matrix (such as dozens of business fields and their associated rules), correlation analysis is used to filter key features, and L1 regularization, PCA and other algorithms are combined to reduce dimensionality. Oversampling / undersampling is used to balance the data distribution.

[0044] In financial scenarios, incremental extraction and rule matrix are achieved by monitoring new regulatory rules (such as adjustments to loan limit caps) and internal bank policies (such as updates to risk level classifications) in real time. New rules are incrementally obtained through API interfaces and parsed into a rule matrix (e.g., the logical expression corresponding to "age < 25 years old and no fixed income → loan limit ≤ 50,000 yuan": (Age < 25) ∧ (Income = none) → Limit ≤ 5W).

[0045] Feature processing involves performing correlation analysis on 50+ fields such as "age", "income", and "asset type" to remove fields with a correlation to loan approval results below a threshold (e.g., <0.1) (e.g., irrelevant hobbies fields); PCA is used to project high-dimensional features into a 20-dimensional subspace to reduce the computational complexity of the model; and oversampling is performed on "high-risk customers" who account for less than 5% of historical approval records to prevent the model from biasing towards the "low-risk" classification due to data imbalance.

[0046] In the healthcare and elderly care scenarios, incremental extraction and rule matrices are parsed into rule matrices by real-time synchronization of updates to the medical insurance reimbursement catalog (such as adding targeted drugs to the reimbursement scope) and adjustments to the service standards of elderly care institutions (such as changes in disability level assessment indicators). For example, the logical expression for "diagnosed with diabetes and medication not listed in the catalog → reimbursement rate reduced by 30%" is: (Disease = Diabetes) ∧ (Drug (Table of Contents) → Reimbursement -30%).

[0047] Feature processing involves L1 regularization on 80+ fields such as “symptoms”, “medication”, and “coverage period” to filter out 20 core fields that significantly affect the matching results of the coverage plan (such as removing fields related to rare diseases); SMOTE oversampling is performed on “patients with multiple symptoms” samples, which account for a very small proportion in the historical records, to balance the distribution of the dataset and avoid the model ignoring the adaptability verification of complex health conditions.

[0048] Incremental extraction technology ensures that updates to financial regulatory rules and medical insurance policies are synchronized to the system in real time. This includes timely responses to adjustments in loan limit caps or additions to the medical insurance reimbursement list, avoiding compliance risks caused by manual oversights of rule changes. Dimensionality reduction and data balancing techniques address the computational burden of high-dimensional features and sample imbalance. For example, in financial scenarios, this reduces the interference of redundant fields on risk assessment; in medical scenarios, it enhances the model's ability to identify complex health conditions, preventing mismatches in insurance schemes due to the neglect of "long-tail samples."

[0049] Step S103. Through loss distribution fitting, cross-validation and hyperparameter optimization, construct a field association graph of the semantics of the associated business fields, rule logic and historical operation habits corresponding to the rule matrix.

[0050] Specifically, based on the Transformer architecture, text-based business data (such as rule descriptions and medical history texts) is processed. Through multimodal learning, heterogeneous data such as text, numbers, and images (such as ID images in finance and image reports in medicine) are fused. Combined with loss distribution fitting (such as the probability of financial risk loss and the probability of medical security gap), cross-validation, and hyperparameter optimization, a field association graph is constructed that includes semantic associations of fields (such as the mapping between "asset type" and "risk level"), rule logic mappings (such as the derivation relationship of "age + income → loan amount"), and historical habit features (such as the input templates frequently used by agents).

[0051] In financial scenarios, multimodal processing uses Transformer to parse text fields in loan applications (such as "occupation description"), combines OCR technology to identify key information in ID card and asset certificate images (such as ID card validity period and property certificate number), and uses cross-modal attention mechanism to associate the semantic relationship between "occupation" and "income stability" (such as "civil servant" corresponding to the "high stability" label).

[0052] Model training and optimization are achieved by using moment estimation to fit the loss distribution (such as the curve of default probability changing with loan amount) based on historical loan loss data. The accuracy of the model's prediction of "asset type-risk level-product suitability" is evaluated by 10-fold cross-validation. Grid search is used to optimize hyperparameters such as the number of Transformer layers and the number of attention heads. Finally, the relationship between graph nodes (such as "customer age 28 years old" and "no house or car") and edges (such as "→low-risk product recommendation") is constructed.

[0053] In the medical, health and elderly care scenarios, multimodal processing uses Transformer to process medical record text (such as "Hypertension Grade III, combined with coronary heart disease"), combines image recognition to analyze lesion features in CT images, and uses a multimodal fusion model to associate the logic of "severity of disease" with "limitation of insurance liability" (such as "cancer diagnosis" → "increased reimbursement amount for targeted drugs").

[0054] Model training and optimization are based on historical insurance claim data to fit the distribution of coverage gaps (such as the probability of out-of-pocket expenses for chronic disease patients due to medications not listed in the drug catalog). The self-help method is used to review and predict the model's omission of "medical history record-insurance clause conflict" issues. After adjusting the hyperparameters, the relationship between graph nodes (such as "5-year history of diabetes" and "insurance plan A") and edges (such as "→need to supplement medication list verification") is constructed.

[0055] Breaking away from the limitations of traditional keyword matching tools, this tool achieves deep semantic and rule-based associations between fields. For example, in financial scenarios, it automatically derives the risk level corresponding to "freelance work + high liquidity assets," and in medical scenarios, it identifies the implicit conflict between "history of kidney disease medication + dialysis treatment" and "medical insurance reimbursement restrictions," reducing reliance on human experience. By learning from historical processing records of frequently used agent operations (such as "quick entry of corporate customer templates" commonly used in finance and "chronic disease record-keeping templates" commonly used in healthcare), it generates customized entry templates, improving the processing efficiency of repetitive scenarios.

[0056] Step S104. When new business information is accessed, a processing prompt message is generated based on the field association graph and the business information.

[0057] Specifically, when new business information is accessed, information is automatically completed, rules are validated in real time, and dynamic prompts are achieved based on the field association graph: Input assistance: missing information is automatically filled or high-frequency templates are recommended based on the association relationship of fields in the graph; Risk prompts: fields that trigger risk rules are highlighted based on the validation results of the rule matrix, and the prompt content is dynamically updated in combination with the current business process node (such as data entry in progress, underwriting in progress).

[0058] In financial scenarios (loan application entry), the autocomplete feature automatically completes the "default income range" (derived from historical data of similar professions) when a customer agent enters "customer age 25, occupation: programmer" and recommends a "quick entry template for internet professionals" (including high-frequency fields such as "housing provident fund contribution ratio" and "employment contract term"). Real-time verification and prompts are also implemented. When "loan purpose = home purchase" and "customer already owns 2 properties" is entered, the rule matrix triggers "purchase restriction policy verification." The system highlights the "number of properties" field and prompts: "According to current policy, a down payment ratio of 50% is required for a third property loan. Please confirm the customer's qualifications." As the process enters the "underwriting stage," the prompt updates to: "Bank statements for the previous 12 months and scanned copies of property ownership certificates are required."

[0059] In the healthcare and elderly care scenario (elderly care insurance registration), automatic completion is achieved when "elderly person is 75 years old and diagnosed with Alzheimer's disease" is entered. The system automatically completes "default care level = special level" and recommends "cognitive impairment elderly registration template" (including high-frequency fields such as "guardian information", "emergency contact person", and "common medication list"). Real-time verification and prompts are provided when "the insurance plan includes 'inpatient medical reimbursement'" and "medical history includes 'recurrent mental illness'" is entered. The rule matrix triggers "insurance liability conflict verification". The system highlights the "insurance plan" field and prompts: "The reimbursement rate for mental illness treatment under this plan is only 40%, which needs to be clearly explained to the client". When the process enters the "plan confirmation stage", the prompt is updated to: "It is recommended to supplement 'psychiatric follow-up records in the past 6 months' to assess the suitability of the insurance".

[0060] Automatic completion and template recommendations reduce repetitive data entry by over 50%. For example, in financial scenarios, loan application entry time is reduced from 30 minutes to 15 minutes, and in medical scenarios, health record creation time is reduced from 20 minutes to 10 minutes. Real-time rule validation and dynamic prompts reduce human error rates by over 70%. For example, this avoids irregular loan disbursements due to ignoring purchase restrictions in financial scenarios, or claims disputes due to conflicting insurance terms in medical scenarios, improving business compliance and customer satisfaction. The prompts are dynamically updated according to the process progress (e.g., from "field validation in the data entry stage" to "document supplementation in the underwriting stage"), helping agents handle complex business according to standardized processes. This is especially suitable for beginners to quickly get started, reducing reliance on the experience of senior employees.

[0061] In some embodiments, the preprocessing of the business data, rule data, and historical processing records to form a standardized dataset includes: cleaning the business data, rule data, and historical processing records to remove duplicate records, correct abnormal records, and fill or remove missing records; and using encoding technology to normalize the format of heterogeneous data, eliminating naming, type, and format conflicts, and unifying the semantics and structure of data fields.

[0062] Preprocessing of business data, rule data, and historical processing records includes data cleaning (removing duplicate records, correcting abnormal records, and filling / removing missing records) and encoding normalization (eliminating naming, type, and format conflicts of heterogeneous data through encoding techniques, and unifying semantics and structure).

[0063] In financial scenarios, data cleaning includes: removing duplicate loan application records by customer ID; correcting abnormal "income amount" fields (e.g., correcting negative values ​​to 0); for missing "occupation type" fields, records with a percentage less than 5% are removed, while those with a percentage greater than or equal to 5% are filled with "freelance" (a high-frequency value). Encoding normalization involves unifying heterogeneous representations of the "asset type" field (e.g., "real estate" and "immovable property" to "fixed assets") using One-Hot encoding to convert "risk level" (low / medium / high) into a numerical vector; and extracting key indicators (e.g., number of overdue payments, debt ratio) from unstructured "credit report text" using NLP and converting them into structured fields.

[0064] In healthcare settings, data cleaning involves: removing duplicate records from patient health records using medical record numbers; correcting abnormal "body temperature values" (e.g., >42℃ is corrected to the measured maximum of 41℃); for missing "allergy history" fields, if a field is required, it is marked as "no allergy history," and if not required, the field is removed. Encoding normalization involves mapping "disease names" to ICD-10 codes (e.g., "hypertension" corresponds to I10), using Label Encoding to convert "nursing level" (special / level 1 / level 2) into numerical values; and extracting key information such as "disease" and "onset time" from unstructured "diagnosis descriptions" using entity recognition and filling them into structured fields.

[0065] By cleaning and filling in data, we can reduce the interference of duplicate, abnormal, and missing data on subsequent analysis (such as avoiding misjudgments of loan amounts due to incorrect income data in financial scenarios, or medication risks caused by missing allergy history in medical scenarios). Unifying data formats and semantics enables direct correlation and analysis between regulatory rules data and historical approval data in finance, and between medical record data and medical insurance terms data in healthcare, reducing the cost of cross-system data integration.

[0066] In some embodiments, the step of incrementally extracting the rule matrix corresponding to the standardized dataset in real time includes: monitoring the update status of the data source and obtaining the latest rule data in real time through a preset incremental extraction mechanism; parsing the obtained rule data into a structured rule matrix and generating a logical expression corresponding to the rule matrix, wherein the logical expression is used to characterize the verification logic and risk assessment logic of the business rules.

[0067] By monitoring data source updates in real time through a preset incremental extraction mechanism, the latest rule data is obtained and parsed into a structured rule matrix to generate logical expressions (such as "field A + condition B → verification result C"), which represent the verification logic and risk assessment logic of business rules.

[0068] In financial scenarios, the incremental extraction mechanism monitors updates to the "Personal Loan Overdue Record Rules" in the central bank's credit reporting system in real time via API interfaces, or notifications from banks regarding adjustments to the "Loan Limit for Small and Micro Enterprises." Rule parsing and matrix construction involve parsing a new rule, "2025 Small and Micro Enterprise Loan Limit = Annual Turnover × 0.3 (original rule was × 0.25)," into a rule matrix (field: annual turnover; condition: none; logical expression: limit = turnover × 0.3), and updating the corresponding entries in the rule matrix for subsequent loan limit verification.

[0069] In healthcare scenarios, the incremental extraction mechanism uses RPA robots to periodically crawl relevant websites for updates to the "Drug Reimbursement Catalog" or notifications of adjustments to the "Disability Level Assessment Indicators" in elderly care institutions. Rule parsing and matrix construction involve parsing the new rule "Diabetes Complications Treatment Included in Outpatient Reimbursement in 2025" into a rule matrix (fields: disease type; condition: diabetes complications; logical expression: reimbursement scope = outpatient + inpatient), and generating the logical expression "(disease = diabetes complications) → reimbursement scope includes outpatient," for subsequent verification of the insurance plan.

[0070] By ensuring that dynamic updates to financial regulatory policies and medical insurance rules take effect immediately, compliance risks caused by rule lag can be avoided (such as over-lending due to failure to adjust loan limit caps in financial scenarios, and loss of customer rights due to failure to update reimbursement catalogs in medical scenarios). Unstructured rule documents are transformed into computable logical expressions, facilitating automatic system verification (such as limit calculation in finance and reimbursement scope matching in medical care), reducing the time and errors associated with manual rule interpretation.

[0071] In some embodiments, the step of constructing a field association graph corresponding to the rule matrix, which includes semantic association of business fields, rule logic, and historical operation habits, through loss distribution fitting, cross-validation, and hyperparameter optimization, includes: fitting the loss distribution using the moment estimation method based on loss data in historical processing records and obtaining parameter estimation results; evaluating the predictive performance of the initial model using cross-validation and bootstrap methods to identify overfitting or underfitting problems; adjusting the model hyperparameters using a grid search combined with random search, and selecting, transforming, and extracting new features from the original features to construct a field association graph containing semantic association of business fields, rule logic mapping, and historical operation habit features.

[0072] The model fits the loss distribution based on historical loss data (such as financial default losses and medical security gaps), evaluates the model performance through cross-validation (such as 10-fold cross-validation) and bootstrapping, and adjusts hyperparameters using grid search + random search to construct an association graph that includes field semantics, rule logic, and historical habits.

[0073] In financial scenarios, loss distribution fitting utilizes historical loan default data to fit a normal distribution of "loan amount - default probability" using moment estimation, obtaining the mean (500,000 yuan) and variance (200,000 yuan) to assess the default risk of new applications. Model optimization: 10-fold cross-validation is performed on the "customer asset type - risk level - product suitability" model. If overfitting is found (training set accuracy 95%, validation set 70%), the number of attention heads in the Transformer is reduced through random search. When constructing the graph, the historical operational habits of "high-liquidity customers prefer short-term wealth management" are used as edge weights to optimize the recommendation logic.

[0074] In the healthcare scenario, loss distribution fitting utilizes historical claims data to fit a gamma distribution of "proportion of out-of-pocket chronic disease medications - out-of-pocket expenses," obtaining shape parameters (2.3) and scale parameters (1500 yuan) to assess the gap risk of the coverage plan. Model optimization: The "medical history - coverage clause conflict" model is evaluated using a self-help method. If underfitting is found (missed detection rate of 30%), the number of training rounds is increased through grid search. When constructing the graph, the historical habit of "frequently supplementing guardian information for Alzheimer's patients" is used as a node association to generate customized input prompts.

[0075] By fitting loss distributions, financial risks and healthcare gaps are transformed into calculable probability values, improving the scientific rigor of risk assessment (e.g., quantifying "the default probability of customers without fixed income is 20% higher"). Cross-validation and hyperparameter optimization reduce model overfitting / underfitting issues (e.g., improving default prediction accuracy from 75% to 85% in financial scenarios and clause conflict identification rate from 60% to 80% in medical scenarios), ensuring the reliability of the graph association logic.

[0076] In some embodiments, generating processing prompts based on the field association graph and business information includes: during the information entry process, automatically completing missing business information, generating customized entry templates, and recording user operation habits based on the semantic relationships of the fields in the field association graph; performing real-time compliance verification on the entered business information based on the verification logic of the rule matrix, marking key fields that trigger risk rules, and dynamically updating processing prompts in conjunction with the current status of the business processing flow.

[0077] During information entry, missing information is automatically filled in based on field association graphs, customized templates are generated, and operating habits are recorded; compliance is verified in real time based on rule matrix, risk fields are marked, and prompts are dynamically updated in combination with process status.

[0078] In financial scenarios (loan application entry), autocomplete and template features include: when "customer occupation = teacher" is entered, the system automatically completes "default housing provident fund contribution ratio = 12%" (a historically high-frequency value for the same occupation) and generates a "quick entry template for public institution employees" (including high-frequency fields such as "professional title level" and "years of service"). Dynamic prompts include: when "loan term = 30 years" and "customer age = 55 years old" are entered, the rule matrix triggers a "retirement age limit" check (retirement age ≤ 60 years old, remaining repayment period ≤ 5 years), the system marks the "age" and "loan term" fields, and prompts: "Repayment period covers 5 years after retirement, proof of retirement income is required"; when the process enters the "review stage", the prompt is updated to: "Please confirm whether your post-retirement income meets the monthly payment requirements".

[0079] In healthcare scenarios (health record creation), automatic completion and templates include: when "Patient Diagnosis = Cerebral Infarction" is entered, the atlas automatically completes "Default Rehabilitation Plan = Limb Function Training" (a frequently used plan for the same condition in history) and generates a "Stroke Patient Record Creation Template" (including frequently used fields such as "Sequelae Assessment" and "Medication Contraindications"). Dynamic prompts include: when "Coverage Plan = Basic Version" is entered and "Medical History includes 'Requires Long-Term Medication'", the rule matrix triggers a "Coverage Scope Insufficient" check (the Basic Version does not include targeted drug reimbursement), the system marks the "Coverage Plan" field, and prompts: "The Basic Version reimbursement scope does not include long-term medication for chronic diseases; upgrading the plan is recommended"; when the process enters the "Plan Signing Stage", the prompt is updated to: "Please explain the proportion of out-of-pocket medications and the annual limit to the client."

[0080] By using autocomplete and template generation, manual data entry time can be reduced by 40%-60% (e.g., financial loan entry from 25 minutes to 12 minutes, medical record creation from 18 minutes to 8 minutes), thus reducing repetitive labor costs. Combined with dynamic adjustment of prompt granularity at process nodes (prompting field-level risks during data entry and process-level risks during review), it helps agents process business according to standardized paths, making it especially suitable for beginners to quickly master complex rules (such as financial compliance guidelines and medical insurance details).

[0081] In some embodiments, before constructing the field association map of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization, the method further includes: performing context association processing on the text information in the business data using the Transformer architecture; and fusing multiple types of business information such as text, numbers, and images through a multimodal learning model to achieve joint feature extraction of heterogeneous data.

[0082] Before constructing the field association graph, the Transformer architecture is used to process the contextual association of textual information (such as rule descriptions and medical history texts); a multimodal learning model is used to fuse heterogeneous data such as text, numbers, and images to extract joint features.

[0083] In financial scenarios, Transformer text processing uses Transformer encoding on "customer occupation description" text (e.g., "senior engineer at an internet company") to capture the contextual relationship between "senior engineer" and "high income potential," generating semantic vectors. Multimodal fusion: The text semantic vectors, along with "salary slip data" and "work permit image OCR information," are input into a multimodal model (e.g., ViT-BERT) to extract joint features of "occupation-income-stability," which are used to construct the association edge between "occupation type → risk level."

[0084] In healthcare scenarios, Transformer text processing captures the medical semantic association between "ST-segment elevation" and "acute myocardial infarction" by encoding "medical record diagnostic descriptions" (such as "sudden chest pain, ST-segment elevation on ECG") using Transformer encoding, generating symptom vectors. Multimodal fusion inputs the text symptom vectors along with "ECG waveform digital data" and "CT image features" into a multimodal model (such as CLIP) to extract joint features of "symptom-examination results-symptom severity," which are used to construct the association edges of "medical history → insurance policy suitability."

[0085] By using Transformers to capture textual contextual relationships (e.g., in finance, "freelance" ≠ "no fixed income," and in medicine, "hypertension" needs to distinguish between "Level I" and "Level III"), the ambiguity problems of traditional keyword matching are avoided. Multimodal learning breaks down data type barriers (e.g., images, text, numbers), enhances the feature richness of the association graph (e.g., in finance, combining work permit images to verify the authenticity of an occupation, and in medicine, combining image data to accurately assess disease risk), and enhances the model's adaptability to complex scenarios.

[0086] In some embodiments, the acquisition of business data, rule data, and historical processing records includes: acquiring business data and historical processing records from internal business systems, application programming interfaces, and databases; and acquiring supplementary business data, rule data, and industry characteristic data from external partner data sources, industry statistical information platforms, and publicly available academic research data.

[0087] Data is acquired through multiple channels: internal data sources (business systems, APIs, databases) are used to obtain business data and historical records; external data sources (partners, industry platforms, academic data) are used to obtain supplementary data, rules, and industry characteristics.

[0088] In financial scenarios, internal data includes data obtained from the bank's core system (basic customer information), credit approval system (historical loan records), and customer service system (agent operation logs); and data is integrated with the credit reporting system (personal credit scores) and enterprise business registration information system (business status) via API. External data includes the latest compliance rules obtained from the official website of the State Financial Regulatory Commission, and characteristic data such as industry average loan interest rates and default rates obtained from third-party data companies to supplement the industry benchmark values ​​of the rule matrix.

[0089] In the healthcare context, internal data includes data obtained from hospital HIS systems (electronic medical records), medical insurance settlement systems (reimbursement records), and elderly care institution management systems (service records); and data is accessed via APIs through regional health information platforms (infectious disease history) and pharmaceutical company databases (drug instructions). External data includes reimbursement policy documents obtained from relevant websites and academic data such as "Chronic Disease Management Guidelines" from medical journals, supplementing the medical professional rules for the design of the insurance plan.

[0090] By combining internal and external data sources, the system covers information across the entire business chain (e.g., in financial scenarios, it includes both customer data and industry benchmark risk values; in medical scenarios, it includes both medical records and medical insurance policies and guidelines), avoiding the information silo problem caused by a single data source. The introduction of industry characteristics and external rules enables the system to adapt to business differences across regions and institutions (e.g., mortgage policies in different provinces in finance, and medical insurance reimbursement details in different cities in healthcare), improving the cross-institutional reusability of the technical solution.

[0091] In some embodiments, when applied to the insurance industry, the traditional insurance policy issuance process is extremely lengthy, especially in the inquiry and underwriting stages. The semantics of the fields are complex, the learning cost is high, the number of fields to be entered is numerous, and the validation rules are also complex. Manual entry has a high probability of error, and manual underwriting carries a very high risk. Currently, AI code plugins based on large models, such as Copilot, Fitten Code, and Tongyi Lingma, are gradually maturing. These plugins provide developers with functions such as explaining code snippets, automatically completing code, generating code based on prompts, and providing code analysis and optimization suggestions.

[0092] If we use AIGC to centrally train a large amount of insurance application data and build a general and detailed field factor graph, we can then conduct training in vertically segmented fields based on massive amounts of insurance application data and underwriting rules. After training, we can achieve automatic completion of data entry, one-click input of general templates, remembering user habits, and real-time rule verification during data entry. In the underwriting process, fields with high risk coefficients are highlighted, risk factors are intelligently analyzed, and prompts are generated for agents, which are updated in real time according to the progress of the process. In this way, we can create a highly professional agent who can make quick decisions.

[0093] For example, in actual insurance policy issuance, manual data entry used to be prone to errors due to complex fields and rules. New technologies can automatically complete and verify entry information, reducing error rates and improving underwriting accuracy and efficiency. Combined with... Figure 2The training process for the group insurance call center agent is as follows: 1. Collection and preprocessing of group insurance policy data and underwriting rules: a) Obtain group insurance policy information and underwriting records from multiple internal data sources such as PG and API, while extracting data sources from relevant upstream and downstream partners, industry statistics, customer information statistics, and publicly released research data in the group property insurance field. 2. Data collection and integration (ET operation): a) Extract customer information, group policy information, and group insurance product information, and extract verification rules and risk assessment rules from the underwriting rules. b) Clean the group insurance policy data, processing problematic values ​​such as duplicate values, abnormal records, and missing records to ensure quality and promote standardized group insurance policy data. c) Use One-Hot to solve naming, type, and format conflicts in policy information, input fields, and risk rules, as well as field normalization. For group underwriting records, a specific incremental extraction mechanism is adopted to obtain the latest underwriting record data in real time. Organize and parse group underwriting rules, and convert rule information into a rule matrix and corresponding expressions. d) Serializing and normalizing data scales the values ​​of different features to a similar range, ensuring a relatively balanced impact of each feature on the model. Furthermore, the large number of fields, including insurance information, customer information, indicator fields, and risk rule fields, may lead to an excessive number of features, potentially posing a significant risk to model performance. To address this, correlation analysis is used to assess the correlation between features, L1 regularization can regulate model sparsity and filter out unimportant features, while PCA rules can project the original features into a smaller subspace through linear transformation. e) Data sampling employs over-sampling and under-sampling techniques to balance the dataset, thereby addressing the problem of imbalanced field distribution.

[0094] 3. Model selection and training after data collection: a) Use the Transformer model to process text information in the insurance application data, thereby associating it with contextual information. b) The insurance application data contains various types of information such as text, numbers, and images. Multimodal learning is used to process various types of information for automatic data entry and automatic underwriting. c) In the modeling method for stochastic processes, loss distribution fitting and parameter estimation techniques were employed. By performing moment estimation on a set of historical insurance policy loss data and fitting it with historical insurance information data, the parameter estimation results of the loss distribution were obtained. d) Taking an actuarial loss model as an example, risk assessment and premium pricing for agricultural insurance subsidies were used, but poor predictive performance was found, requiring optimization. Through cross-validation and retrospective prediction using the bootstrap method on the original model, the problems of overfitting and underfitting were found. To improve the model, the original features were selected and transformed, and new features related to loss were extracted, such as the insured's insured object, occupational risk, and duration of the insured object. To find the optimal and suitable hyperparameter combination, grid search combined with random search methods was used to adjust the model's input parameters. Simultaneously, deep learning was used to optimize the model. The fusion of linear and nonlinear methods was an unavoidable step after optimization of multiple models, thus greatly improving the model's predictive performance.

[0095] As shown above, the evaluation and optimization of the actuarial loss model are crucial for measuring its predictive ability. The F1 score, which combines accuracy and recall, is an important standard for evaluating the overall performance of the model. Accuracy measures the model's complete coverage of positive examples, recall reflects the model's ability to correctly identify positive examples, and the AUC value is used to judge the ranking quality of the model's prediction results. e) By analyzing historical premium loss data and fitting it with the corresponding probability distribution, we predict future losses for various indicators. Based on the loss distribution and our company's risk control thresholds, we apply the expected loss method to define the premium range. When formulating pricing strategies for group insurance products such as group property insurance and agricultural insurance, we comprehensively consider domestic and international market demand, the competitive landscape of PICC, CPIC, and China Life, and our company's operational goals to ensure the competitiveness of our insurance products. More importantly, risk control requires the reasonable allocation of various reserves according to regulatory requirements and the company's actual situation to ensure the company's sound operation. We use premium actuarial forecasts to regularly assess the company's solvency and prepare risk warnings and risk plans. Furthermore, we constantly monitor and comply with the regulations of regulatory authorities such as the banking and insurance sectors to ensure the compliance of our group insurance business operations. We also coordinate reinsurance needs and select appropriate reinsurance methods. For different types of risks, we design corresponding reinsurance contract terms, clearly defining the rights and obligations of both parties. To reduce the operational risks of property insurance, we have established a comprehensive risk management system for group insurance, managing risks through risk identification, assessment, control, and reporting. We use models to fit and predict historical group insurance underwriting data and formulate reasonable premium determination strategies. Based on the characteristics of group insurance business and regulatory requirements, we use actuarial methods to extract reserves in a robust manner. For sustainable development, we need to design reasonable reinsurance schemes to achieve maximum risk diversification and transfer. However, as model complexity gradually increases, model interpretability decreases. Therefore, we must control model complexity and, while ensuring predictive accuracy, strengthen data governance and integration to improve data quality standardization and usability.

[0096] In the traditional insurance policy issuance process, dozens or even hundreds of fields might need to be manually reviewed one by one. This is not only time-consuming but also highly prone to errors. However, thanks to AIGC's training and optimization, the system can complete this task quickly and accurately, much like setting up an intelligent traffic control system at a busy intersection to efficiently direct traffic flow. For another example, the highlighting and intelligent analysis of risk coefficient fields gives agents the feeling of having a high-precision navigation system, allowing them to quickly identify key issues and make informed and accurate decisions.

[0097] Please see Figure 3 As shown, Figure 3This is a schematic diagram of the structure of the prompt generation device 200 provided in the embodiments of this application. The prompt generation device 200 is used to execute the steps of the prompt generation method shown in the above embodiments. The prompt generation device 200 can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0098] like Figure 3 As shown, the prompt generation device 200 includes: The data acquisition unit 201 is used to acquire business data, rule data, and historical processing records; and to preprocess the business data, rule data, and historical processing records to form a standardized dataset. The matrix acquisition unit 202 is used to acquire the rule matrix corresponding to the standardized dataset in real time by incremental extraction; and to perform dimensionality reduction and distribution equalization processing on the high-dimensional features corresponding to the rule matrix using correlation analysis, feature selection algorithms and data balancing techniques. The graph generation unit 203 is used to construct a field association graph of the semantics of the associated business fields, rule logic and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation and hyperparameter optimization. The prompt generation unit 204 is used to generate processing prompt information based on the field association graph and the business information when new business information is accessed.

[0099] In some embodiments, the preprocessing of the business data, rule data, and historical processing records to form a standardized dataset includes: cleaning the business data, rule data, and historical processing records to remove duplicate records, correct abnormal records, and fill or remove missing records; and using encoding technology to normalize the format of heterogeneous data, eliminating naming, type, and format conflicts, and unifying the semantics and structure of data fields.

[0100] In some embodiments, the step of incrementally extracting the rule matrix corresponding to the standardized dataset in real time includes: monitoring the update status of the data source and obtaining the latest rule data in real time through a preset incremental extraction mechanism; parsing the obtained rule data into a structured rule matrix and generating a logical expression corresponding to the rule matrix, wherein the logical expression is used to characterize the verification logic and risk assessment logic of the business rules.

[0101] In some embodiments, the step of constructing a field association graph corresponding to the rule matrix, which includes semantic association of business fields, rule logic, and historical operation habits, through loss distribution fitting, cross-validation, and hyperparameter optimization, includes: fitting the loss distribution using the moment estimation method based on loss data in historical processing records and obtaining parameter estimation results; evaluating the predictive performance of the initial model using cross-validation and bootstrap methods to identify overfitting or underfitting problems; adjusting the model hyperparameters using a grid search combined with random search, and selecting, transforming, and extracting new features from the original features to construct a field association graph containing semantic association of business fields, rule logic mapping, and historical operation habit features.

[0102] In some embodiments, generating processing prompts based on the field association graph and business information includes: during the information entry process, automatically completing missing business information, generating customized entry templates, and recording user operation habits based on the semantic relationships of the fields in the field association graph; performing real-time compliance verification on the entered business information based on the verification logic of the rule matrix, marking key fields that trigger risk rules, and dynamically updating processing prompts in conjunction with the current status of the business processing flow.

[0103] In some embodiments, before constructing the field association map of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization, the method further includes: performing context association processing on the text information in the business data using the Transformer architecture; and fusing multiple types of business information such as text, numbers, and images through a multimodal learning model to achieve joint feature extraction of heterogeneous data.

[0104] In some embodiments, the acquisition of business data, rule data, and historical processing records includes: acquiring business data and historical processing records from internal business systems, application programming interfaces, and databases; and acquiring supplementary business data, rule data, and industry characteristic data from external partner data sources, industry statistical information platforms, and publicly available academic research data.

[0105] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the above-described prompt generation device and its modules can be referred to the corresponding processes in the prompt generation method embodiments described above, and will not be repeated here.

[0106] The above-described prompt generation method can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the device shown.

[0107] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0108] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any prompt generation method.

[0109] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0110] Internal memory provides an environment for the execution of computer programs stored in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any prompt generation method.

[0111] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0112] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0113] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: Acquire business data, rule data, and historical processing records; preprocess the business data, rule data, and historical processing records to form a standardized dataset; Incremental extraction is used to obtain the rule matrix corresponding to the standardized dataset in real time; correlation analysis, feature selection algorithms and data balancing techniques are used to perform dimensionality reduction and distribution equalization on the high-dimensional features corresponding to the rule matrix. By fitting the loss distribution, cross-validation, and hyperparameter optimization, a field association graph of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix is ​​constructed. When new business information is accessed, processing prompts are generated based on the field association graph and the business information.

[0114] In some embodiments, the preprocessing of the business data, rule data, and historical processing records to form a standardized dataset includes: cleaning the business data, rule data, and historical processing records to remove duplicate records, correct abnormal records, and fill or remove missing records; and using encoding technology to normalize the format of heterogeneous data, eliminating naming, type, and format conflicts, and unifying the semantics and structure of data fields.

[0115] In some embodiments, the step of incrementally extracting the rule matrix corresponding to the standardized dataset in real time includes: monitoring the update status of the data source and obtaining the latest rule data in real time through a preset incremental extraction mechanism; parsing the obtained rule data into a structured rule matrix and generating a logical expression corresponding to the rule matrix, wherein the logical expression is used to characterize the verification logic and risk assessment logic of the business rules.

[0116] In some embodiments, the step of constructing a field association graph corresponding to the rule matrix, which includes semantic association of business fields, rule logic, and historical operation habits, through loss distribution fitting, cross-validation, and hyperparameter optimization, includes: fitting the loss distribution using the moment estimation method based on loss data in historical processing records and obtaining parameter estimation results; evaluating the predictive performance of the initial model using cross-validation and bootstrap methods to identify overfitting or underfitting problems; adjusting the model hyperparameters using a grid search combined with random search, and selecting, transforming, and extracting new features from the original features to construct a field association graph containing semantic association of business fields, rule logic mapping, and historical operation habit features.

[0117] In some embodiments, generating processing prompts based on the field association graph and business information includes: during the information entry process, automatically completing missing business information, generating customized entry templates, and recording user operation habits based on the semantic relationships of the fields in the field association graph; performing real-time compliance verification on the entered business information based on the verification logic of the rule matrix, marking key fields that trigger risk rules, and dynamically updating processing prompts in conjunction with the current status of the business processing flow.

[0118] In some embodiments, before constructing the field association map of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization, the method further includes: performing context association processing on the text information in the business data using the Transformer architecture; and fusing multiple types of business information such as text, numbers, and images through a multimodal learning model to achieve joint feature extraction of heterogeneous data.

[0119] In some embodiments, the acquisition of business data, rule data, and historical processing records includes: acquiring business data and historical processing records from internal business systems, application programming interfaces, and databases; and acquiring supplementary business data, rule data, and industry characteristic data from external partner data sources, industry statistical information platforms, and publicly available academic research data.

[0120] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method provided in any embodiment of this application.

[0121] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating prompts, characterized in that, include: Acquire business data, rule data, and historical processing records; preprocess the business data, rule data, and historical processing records to form a standardized dataset; Incremental extraction is used to obtain the rule matrix corresponding to the standardized dataset in real time; correlation analysis, feature selection algorithms and data balancing techniques are used to perform dimensionality reduction and distribution equalization on the high-dimensional features corresponding to the rule matrix. By fitting the loss distribution, cross-validation, and hyperparameter optimization, a field association graph of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix is ​​constructed. When new business information is accessed, processing prompts are generated based on the field association graph and the business information.

2. The method according to claim 1, characterized in that, The preprocessing of the business data, rule data, and historical processing records to form a standardized dataset includes: The business data, rule data, and historical processing records are cleaned to remove duplicate records, correct abnormal records, and fill or remove missing records. Encoding techniques are used to normalize the format of heterogeneous data, eliminating naming, type, and format conflicts, and unifying the semantics and structure of data fields.

3. The method according to claim 1, characterized in that, The step of incrementally extracting the rule matrix corresponding to the standardized dataset in real time includes: The system uses a pre-defined incremental extraction mechanism to monitor the update status of the data source in real time and obtain the latest rule data. The acquired rule data is parsed into a structured rule matrix, and a logical expression corresponding to the rule matrix is ​​generated. The logical expression is used to characterize the verification logic and risk assessment logic of the business rules.

4. The method according to claim 1, characterized in that, The process involves constructing a field association graph of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization. This includes: Based on the loss data in historical processing records, the moment estimation method is used to fit the loss distribution and obtain the parameter estimation results; The predictive performance of the initial model is evaluated using cross-validation and bootstrap methods to identify overfitting or underfitting issues. The model hyperparameters are adjusted by combining grid search with random search. The original features are selected, transformed, and new features are extracted to construct a field association graph that includes semantic associations of business fields, rule logic mappings, and features of historical operation habits.

5. The method according to claim 1, characterized in that, The step of generating processing prompt information based on the field association graph and business information includes: During the information entry process, based on the semantic relationship of the fields in the field association graph, the missing business information is automatically filled in, a customized entry template is generated, and user operation habits are recorded. Based on the verification logic of the rule matrix, the input business information is verified for compliance in real time, key fields that trigger risk rules are marked, and processing prompts are dynamically updated in combination with the current status of the business process.

6. The method according to claim 1, characterized in that, Before constructing the field association graph of the associated business field semantics, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization, the method further includes: The Transformer architecture is used to perform contextual processing on text information in business data. By integrating multiple types of business information such as text, numbers, and images through a multimodal learning model, joint feature extraction of heterogeneous data can be achieved.

7. The method according to claim 1, characterized in that, The acquisition of business data, rule data, and historical processing records includes: Obtain business data and historical processing records from internal business systems, application programming interfaces, and databases; We obtain supplementary business data, rule data, and industry characteristic data from external partner data sources, industry statistical information platforms, and publicly available academic research data.

8. A prompt generation device, characterized in that, include: The data acquisition unit is used to acquire business data, rule data, and historical processing records; and to preprocess the business data, rule data, and historical processing records to form a standardized dataset. The matrix acquisition unit is used to acquire the rule matrix corresponding to the standardized dataset in real time using incremental extraction; and to perform dimensionality reduction and distribution equalization processing on the high-dimensional features corresponding to the rule matrix using correlation analysis, feature selection algorithms and data balancing techniques. The graph generation unit is used to construct a field association graph of the semantics of the associated business fields, rule logic, and historical operation habits corresponding to the rule matrix through loss distribution fitting, cross-validation, and hyperparameter optimization. The prompt generation unit is used to generate processing prompt information based on the field association graph and business information when new business information is accessed.

9. A computer device, characterized in that, The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 7.

Images