Insurance clause adaptive detection method and device, storage medium and electronic equipment

Abstract

The invention discloses an insurance clause adaptive detection method and device, a storage medium and electronic equipment. Relates to the technical field of insurance clause detection, can be applied to the field of financial science and technology businesses, and comprises the following steps: vectorizing to-be-detected insurance product clauses to obtain query clause vectors; querying a preset knowledge vector library based on the query clause vector to obtain a plurality of supervision clause vectors corresponding to the query clause vector, wherein the similarity of the supervision clause vectors meets a preset condition; based on the query term vector and each supervision term vector, performing compliance analysis by adopting a preset large language model to obtain a compliance analysis matrix; and performing risk detection on the compliance analysis matrix by adopting a preset classifier and a rule engine to obtain a detection result. According to the invention, the accuracy of insurance clause risk identification can be improved.

Description

Technical Field

[0001] This invention relates to the field of insurance clause detection technology, and can be applied to the financial technology business field. In particular, it relates to an adaptive detection method, device, storage medium, and electronic device for insurance clauses. Background Technology

[0002] Currently, insurance companies generally use manual or semi-automated methods for managing policy terms and conducting compliance inspections. The process mainly includes: legal personnel regularly reading new regulatory documents issued by regulatory agencies, manually comparing internal product terms with the latest regulatory requirements, revising non-compliant terms item by item, and submitting for approval. This method has significant shortcomings: manual tracking is time-consuming when regulatory policies change frequently, leading to the risk of delayed updates to product terms, sales scripts, and promotional materials. Existing systems can only rely on keyword searches or manual review, failing to understand semantic differences in terms and easily missing hidden violations. They cannot automate the inspection and risk classification of terms, rates, and disclaimers, resulting in high-risk issues often only being discovered after regulatory inspections. Traditional RAG (Retrieval-Augmented Generation) is used for text generation and knowledge-based question answering, but has not yet been applied to real-time regulatory tracking and corporate compliance comparison. The adaptive detection method for insurance terms proposed in this application can be applied to fintech business platforms that support functions such as shopping, social networking, interactive games, and resource transfer, and have functions such as applying for loans, credit cards, or purchasing insurance and wealth management products.

[0003] Therefore, there is an urgent need for an adaptive detection method for insurance clauses that has the ability to monitor autonomously, update automatically, and compare compliance, so as to achieve real-time matching and risk warning between insurance clauses and regulatory requirements. Summary of the Invention

[0004] In view of this, the present invention provides an adaptive detection method, device, storage medium and electronic device for insurance terms, the main purpose of which is to solve the problem of low efficiency caused by enterprises relying on manual labor when formulating insurance product terms.

[0005] To address the aforementioned issues, this application provides an adaptive detection method for insurance clauses, comprising: The terms and conditions of the insurance product to be tested are vectorized to obtain the query terms vector. Based on the query clause vector, a preset knowledge vector base is queried to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector. Based on the query clause vector and each of the regulatory clause vectors, a pre-set large language model is used to perform compliance analysis to obtain a compliance analysis matrix; The compliance analysis matrix is ​​subjected to risk detection using a preset classifier and rule engine to obtain the detection results.

[0006] Optionally, before querying the preset knowledge vector base based on the query term vector, the method further includes: constructing the preset knowledge vector base that is updated in real time; The construction of the preset knowledge vector base that is updated in real time specifically includes: The external regulatory clauses collected in real time through external interfaces and the enterprise regulatory clauses obtained through real-time monitoring are preprocessed to obtain structured clause data. The structured clause data is vectorized using the Embedding model to obtain a standardized knowledge unit stream; Based on the knowledge unit stream, historical knowledge data stored in the preset knowledge vector library is retrieved to obtain target historical knowledge data with the highest similarity to the same topic as the knowledge unit stream; Calculate the first similarity between the knowledge unit stream and the target historical knowledge data; When the first similarity is less than a preset threshold, the historical knowledge data is replaced based on the knowledge unit stream to update the preset knowledge vector library; When the first similarity is greater than a preset threshold, the current version number of the knowledge unit stream is identified and stored in the target storage area with the same theme as the target historical knowledge data.

[0007] Optionally, the compliance analysis based on the query clause vector and each of the regulatory clause vectors, using a preset large language model, to obtain a compliance analysis matrix, specifically includes: The query clause vector, each regulatory clause vector, and the structured task analysis requirements are concatenated to obtain the compliance analysis input statement; The preset large language model is used to perform multi-dimensional structured judgment on the compliance analysis input statement to obtain judgment results corresponding to different dimensions, and the judgment results carry evidence location identifiers; The compliance analysis matrix is ​​obtained by aligning and scoring the judgment results, the query clause vectors, and the regulatory clause vectors. The dimensions mentioned include consistency of rates, compliance of coverage scope, adequacy of disclaimer notices, and compliance of promotional language.

[0008] Optionally, the alignment scoring based on each of the judgment results, the query clause vector, and the regulatory clause vector to obtain the compliance analysis matrix specifically includes: A second similarity score is obtained by performing similarity calculation on the query clause vector and each of the regulatory clause vectors. Based on the risk label carried by the judgment result, a preset label scoring rule is queried to obtain the penalty coefficient; The structural consistency degree is calculated based on the mandatory regulatory conditions carried by the judgment result and the regulatory conditions actually satisfied by the query clause vector. The compliance analysis matrix is ​​obtained by performing alignment scoring based on the second similarity, the penalty coefficient, and the structural consistency.

[0009] Optionally, the step of using a preset classifier and rule engine to perform risk detection on the compliance analysis matrix to obtain detection results specifically includes: The compliance analysis matrix of the same insurance product clauses of the same insurance product to be tested is weighted and scored to obtain the clause risk score value of the same insurance product clauses; The risk scores of each clause of the insurance product to be tested are aggregated to obtain the product risk score. Based on the product risk score, a preset classifier is used to classify the products to obtain risk levels. Based on the risk level query preset rule engine, a risk management strategy for the risk level is obtained.

[0010] Optionally, the compliance analysis matrix of different dimensions of the same insurance product clause of the insurance product to be tested is weighted and scored to obtain a clause risk score value of the same insurance product clause, specifically including: The risk score of the insurance product terms is obtained by calculating based on the second similarity, the penalty coefficient, the structural consistency, the predetermined first weight coefficient, the predetermined second weight coefficient, and the predetermined third weight coefficient.

[0011] Optionally, the method further includes: Based on the report templates in the risk level query report generation engine, a target report template corresponding to the risk level is obtained; A compliance risk report for the insurance product under test is generated based on the target report template, including the risk level and the risk management strategy corresponding to the risk level.

[0012] To address the aforementioned issues, this application provides an adaptive detection device for insurance terms, comprising: The vectorization module is used to vectorize the terms of the insurance product to be tested, and obtain the query terms vector. The query module is used to query a preset knowledge vector base based on the query clause vector to obtain several regulatory clause vectors that meet preset similarity conditions and correspond to the query clause vector. The analysis module is used to perform compliance analysis based on the query clause vector and each of the regulatory clause vectors using a preset large language model to obtain a compliance analysis matrix; The detection module is used to perform risk detection on the compliance analysis matrix using a preset classifier and rule engine to obtain the detection results.

[0013] To address the aforementioned problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned adaptive detection method for insurance terms.

[0014] To address the aforementioned problems, this application provides an electronic device, comprising at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the aforementioned adaptive detection method for insurance terms.

[0015] The beneficial effects of this application are as follows: By constructing an adaptive compliance inspection system that integrates the RAG large model, dynamic knowledge base, and intelligent agent technology, this invention achieves 24 / 7 real-time semantic comparison and risk warning of insurance terms and regulatory policies. This system transforms traditional manual compliance review into a fully automated intelligent process, significantly improving monitoring real-time performance and shortening the regulatory response cycle from "monthly" to "hourly"; enhancing the accuracy of risk identification by discovering hidden violations through multi-dimensional semantic analysis; drastically reducing compliance costs with an automation rate exceeding 90%; and constructing a self-evolving closed loop of "monitoring-identification-feedback-optimization," enabling insurance institutions to achieve continuous compliance and proactive risk prevention in a dynamically updated regulatory environment.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This illustration shows an application environment diagram of an adaptive detection method for insurance terms provided in an embodiment of this application; Figure 2 A flowchart illustrating an adaptive detection method for insurance terms provided in an embodiment of this application is shown. Figure 3 A flowchart illustrating an adaptive detection method for insurance terms according to another embodiment of this application is shown; Figure 4 A structural block diagram of an insurance clause adaptive detection device according to another embodiment of this application is shown; Figure 5 A schematic diagram of the structure of a computer device according to an embodiment of this application is shown; Figure 6 Another structural schematic diagram of a computer device according to one embodiment of this application is shown. Detailed Implementation

[0018] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0019] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0020] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0021] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0022] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0023] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0024] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0025] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0026] It should also be noted that the user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

[0027] The adaptive detection method for insurance terms provided in this application can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can vectorize the terms of the insurance product to be tested to obtain a query term vector; based on the query term vector, it queries a preset knowledge vector base to obtain several regulatory term vectors that meet preset similarity conditions and correspond to the query term vector; based on the query term vector and each of the regulatory term vectors, it performs compliance analysis using a preset large language model to obtain a compliance analysis matrix; and uses a preset classifier and rule engine to perform risk detection on the compliance analysis matrix to obtain the detection results. This application can achieve 24 / 7 automatic monitoring of regulatory and term updates, shortening the regulatory response cycle from "monthly" to "hourly"; automatic pre-checking before terms, rates, or wording go live reduces violations by more than 80%; multi-dimensional semantic comparison combined with regulatory citation makes risk reports traceable and interpretable; it can be adapted to different business lines such as health insurance, auto insurance, and life insurance, and supports multilingual compliance detection; the system report can be directly used as compliance proof material for internal audits and external regulatory reviews.

[0028] This application provides an adaptive detection method for insurance clauses, such as... Figure 2 As shown, it includes: Step S101: Vectorize the terms of the insurance product to be tested to obtain the query term vector; In this step, the terms of the insurance product to be tested undergo text cleaning and standardization to obtain preprocessed terms data. The preprocessed terms data is then segmented into text blocks and semantic units to obtain split data. Finally, the split data is used to generate vectors using a specialized embedding model for insurance regulations, resulting in the query terms vector. The embedding model can be the FinReg-BERT model.

[0029] Step S102: Based on the query clause vector, query a preset knowledge vector base to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector; In this step, the process involves querying a preset knowledge vector base based on the query clause vector to obtain several regulatory clause vectors corresponding to the query clause vector whose similarity meets preset conditions. Specifically, cosine similarity calculation is performed on the query clause vector and the regulatory clause vectors pre-stored in the preset knowledge vector base to obtain various similarities. First, the dot product of the query clause vector and the regulatory clause vector is calculated to obtain the dot product value. Second, the first modulus corresponding to the query clause vector and the second modulus corresponding to the regulatory clause vector are calculated. Finally, cosine similarity calculation is performed based on the dot product value, the first modulus, and the second modulus to obtain various similarities. The dot product value is divided by the product of the first modulus and the second modulus. Based on each similarity, each regulatory clause vector is filtered to obtain several regulatory clause vectors with a similarity greater than a first preset threshold. The preset condition can be that the similarity between the query clause vector and the regulatory clause vector is greater than the first preset threshold, and the first preset threshold can be set according to actual needs.

[0030] Step S103: Based on the query clause vector and each of the regulatory clause vectors, a preset large language model is used to perform compliance analysis to obtain a compliance analysis matrix; In this step, the query clause vector, the regulatory clause vectors, and the structured task analysis requirements are concatenated to obtain the compliance analysis input statement. The pre-defined large language model is used to perform multi-dimensional structured judgment on the compliance analysis input statement, obtaining judgment results corresponding to different dimensions. These judgment results carry evidence location identifiers. Alignment scoring is performed based on the judgment results, the query clause vector, and the regulatory clause vectors to obtain the compliance analysis matrix. The dimensions include rate consistency, coverage compliance, sufficiency of disclaimer notifications, and compliance of promotional language. The pre-defined large language model can be an autoregressive large language model based on the Transformer architecture. Its basic capabilities include: contextual semantic modeling (joint understanding of multiple sentences and paragraphs of text), instruction following, and structured output capabilities (JSON / Schema constraint output). The model can be a general-purpose large model (such as Qwen, Deepseek, etc.) or a fine-tuned model specifically for the insurance field.

[0031] Step S104: Use a preset classifier and rule engine to perform risk detection on the compliance analysis matrix and obtain the detection results.

[0032] In this step, the compliance analysis matrix of different dimensions of the same insurance product clause of the insurance product to be tested is weighted and scored to obtain the clause risk score value of the same insurance product clause; the risk score values ​​of each clause of the insurance product to be tested are aggregated to obtain the product risk score value; the product risk score value is classified using a preset classifier to obtain the risk level; and the risk level is queried by a preset rule engine to obtain the risk handling strategy for the risk level.

[0033] This application enables 24 / 7 automatic monitoring of regulatory and clause updates, shortening the regulatory response cycle from "monthly" to "hourly"; it automatically pre-checks clauses, rates, or wording before they go live, reducing violations by over 80%; multi-dimensional semantic comparison combined with regulatory citations makes risk reports traceable and interpretable; it is adaptable to different business lines such as health insurance, auto insurance, and life insurance, and supports multilingual compliance testing; system reports can be directly used as compliance documentation for internal audits and external regulatory reviews.

[0034] Another embodiment of this application provides a different adaptive detection method for insurance terms, such as... Figure 3 As shown, it includes: Step S201: Vectorize the terms of the insurance product to be tested to obtain the query term vector; In this step, the terms of the insurance product to be tested undergo text cleaning and standardization to obtain preprocessed terms data. The preprocessed terms data is then segmented into text blocks and semantic units to obtain split data. Finally, the split data is used to generate vectors using a specialized embedding model for insurance regulations, resulting in the query terms vector. The embedding model can be the FinReg-BERT model.

[0035] Step S202: Construct a preset knowledge vector base that is updated in real time; In this step, the external regulatory clauses collected in real time through external interfaces and the enterprise regulatory clauses obtained through real-time monitoring are preprocessed to obtain structured clause data. Specifically, regulatory agency documents, such as announcements from XX regulatory agencies and documents from local branches, are automatically crawled through external interfaces such as network APIs, RSS subscriptions, and web crawling modules. Internal document monitoring includes monitoring company clauses and rate folders (PDF, DOCX, Markdown, etc.) and detecting new version generation. Data extraction is performed using OCR recognition and structured parsing methods to obtain structured clause data. The structured clause data is then vectorized to obtain clause vectors. The mathematical expression of the clause vectors is as follows:

[0036] The structured clause data is vectorized using an embedding model to obtain a standardized knowledge stream. A dense vector is generated using an insurance regulation-specific embedding model, such as the FinReg-BERT model, to obtain the standardized knowledge stream. The data science expression is as follows:

[0037] in, represent t The flow of knowledge points in units at any given moment; Represents the source of knowledge units; This represents the type of knowledge unit; Represents the version of a knowledge unit; Vectors representing knowledge units; The timestamp represents the knowledge unit; based on the knowledge unit stream, historical knowledge data pre-stored in a preset knowledge vector library is retrieved to obtain the target historical knowledge data with the highest similarity to the same topic as the knowledge unit stream; for example, the knowledge unit stream of the insurance product terms to be detected is: "clause_id: CL-2025-001-EX12; product_id: P-LIFE-A; clause_type: Disclaimer clause; clause_text: "Medical expenses incurred due to pre-existing conditions will not be reimbursed (including outpatient and inpatient expenses)." The query text can be obtained directly from the clause_text of the knowledge unit flow, or by concatenating the business context. The query text is: Q_text = clause_text + product type / insurance type + effective date / version number; the query clause vector can be represented as:

[0038] Calculate the first similarity between the knowledge unit stream and the target historical knowledge data; when the first similarity is less than a preset threshold, replace the historical knowledge data based on the knowledge unit stream to update the preset knowledge vector base; the preset threshold can be 0.9, and the preset threshold can be set according to actual needs; the mathematical expression of the adaptive incremental update algorithm is as follows:

[0039] in, express t The increment of knowledge unit flow at any given moment; express t The flow of knowledge units at any given moment; express t The knowledge unit stream at time -1; when the first similarity is greater than a preset threshold, the current version number of the knowledge unit stream is identified and stored in a target storage area with the same theme as the target historical knowledge data, so as to construct the preset knowledge vector library. The application uses both new and old clause vectors for differential comparison analysis. The preset knowledge vector library of this application can be dynamically stored using Milvus or Faiss. Furthermore, the knowledge data in the preset knowledge vector base is filtered based on a preset time window, retaining the knowledge vectors closest to the current time for a predetermined duration, thereby updating the preset knowledge vector base. This achieves "manual synchronization" of the knowledge base, providing version management and differential tracking capabilities, and supporting the analysis of clause evolution trajectories. The preset knowledge vector base stores the latest set of legal clause vectors, including metadata such as document number, clause number, effective date, and version.

[0040] Step S203: Based on the query clause vector, query a preset knowledge vector base to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector; In this step, the query clause vector is used to query a preset knowledge vector base to obtain several regulatory clause vectors corresponding to the query clause vector whose similarity meets preset conditions. Specifically, the preset conditions can be that the similarity between the query clause vector and the regulatory clause vector is greater than a first preset threshold, resulting in several regulatory clause vectors. The first preset threshold can be set according to actual needs. A Top-K recall regulatory item list, including text and evidence location markers, is shown in the following example: "reg_id: REG-2024-32-12; reg_title: "XX Supervision and Administration Document No. 32 of 2024"; reg_clause: "Article 12: Disclaimer clauses must clearly list and provide a definition of 'pre-existing conditions', and must not broaden the scope of disclaimers with general statements." effective_date: 2024-05-01; Similarity: 0.87.

[0041] Step S204: Concatenate the query clause vector, each regulatory clause vector, and the structured task analysis requirements to obtain the compliance analysis input statement; In the specific implementation process, this step involves concatenating the query clause vector, the vectors of each regulatory clause, and the structured task analysis requirements to obtain the compliance analysis input statement, which is then used to perform compliance analysis on the compliance analysis input statement using a large model.

[0042] Step S205: Use the preset large language model to perform multi-dimensional structured judgment on the compliance analysis input statement to obtain judgment results corresponding to different dimensions, and the judgment results carry evidence location identifiers; In the specific implementation of this step, the dimensions mentioned include rate consistency, compliance of coverage scope, adequacy of disclaimer disclosure, and compliance of promotional language; for example, the judgment result for the adequacy of disclaimer disclosure is as follows: { "clause_id":"CL-2025-001-EX12", "dimension_results": [ { "dimension" "exclusion_disclosure", "reg_ref": "REG-2024-32-12", "finding": "If the terms mention 'pre-existing conditions' but fail to provide a definition or explanation, it may be deemed as insufficient disclosure." "evidence": { "clause_span": "Medical expenses incurred due to pre-existing conditions will not be reimbursed." "reg_span": "Prompts should be provided for the definition of 'pre-existing conditions'". }, "risk_tag": "potential_noncompliance", "suggestion": "Add details regarding the definition and highlighting method of pre-existing conditions (bold / highlighted)." } ] }".

[0043] The compliance judgment is broken down into calculable and traceable dimensional conclusions, and evidence location markers are provided; these markers include clause fragments (clause_span) and regulatory fragments (reg_span). The preset large language model can be an autoregressive large language model based on the Transformer architecture, whose basic capabilities include: contextual semantic modeling capability (joint understanding of multiple sentences and paragraphs of text), instruction following capability, and structured output capability (JSON / Schema constraint output); the model can be a general large model (such as Qwen, Deepseek, etc.) or a fine-tuned model specifically for the insurance field. This application enables it to have multi-dimensional structured compliance judgment capabilities through the following three mechanisms, specifically, using RAG. The input structure is enhanced by incorporating facts and norms, specifically by concatenating the query clause vector, the regulatory clause vectors, and the structured task analysis requirements as input. The judgment task is decomposed using dimensional constraints, employing dimension-specific judgment instruction templates to output judgment results from independent dimensions such as the sufficiency of disclaimer prompts, consistency of coverage scope, compliance of rate descriptions, and compliance of promotional language. This approach constrains the inference space of the large language model within specific dimensions and slot sets, preventing generalized responses. Furthermore, predefined structured output constraints are provided for each dimension, explicitly requiring the model to output preset field data. This ensures that the model output is computable, traceable, and can be incorporated into subsequent scoring modules, thereby achieving the large language model's multi-dimensional structured compliance judgment capability.

[0044] Step S206: Perform similarity calculation on the query clause vector and each of the regulatory clause vectors to obtain a second similarity. In the specific implementation process of this step, the similarity calculation is performed on the query clause vector and each of the regulatory clause vectors. The mathematical expression is as follows:

[0045] in, Represents the second similarity; Represents the vector of query terms; Represents the vector of regulatory provisions.

[0046] Step S207: Based on the risk label carried by the judgment result, query the preset label scoring rules to obtain the penalty coefficient; In the specific implementation process of this step, based on the risk label carried by the judgment result, a preset label scoring rule is queried to obtain the penalty coefficient; specifically, based on the risk label, a risk strategy configuration table is queried to obtain the penalty coefficient corresponding to the risk label; the risk strategy configuration table is shown in Table 1:

[0047] For example, in step S205, when the result corresponding to the risk tag "risk_tag" is "potential_noncompliance", the penalty coefficient obtained by querying the risk strategy configuration based on potential_noncompliance is 0.3.

[0048] Step S208: Calculate the structural consistency degree based on the mandatory regulatory conditions carried by the determination result and the regulatory conditions actually satisfied by the query clause vector; In this step, the structural consistency degree is calculated based on the mandatory regulatory conditions carried by the judgment result and the regulatory conditions actually satisfied by the query clause vector. Specifically, the mandatory regulatory conditions are the regulatory requirements slot conditions that must be met; for example, [Mandatory Definition = Pre-existing Condition Definition], [Mandatory Reminder = Significant Reminder]; the regulatory conditions actually satisfied by the query clause vector are the actual structural slots of the query clause vector, for example, [Exclusion Object = Pre-existing Condition], [Scope = Outpatient / Inpatient], [Reminder = None]. Because the "Mandatory Definition / Mandatory Reminder" slots are missing, the structural consistency degree is calculated accordingly. .

[0049] Step S209: Perform alignment scoring based on the second similarity, the penalty coefficient, and the structural consistency to obtain the compliance analysis matrix; In this step, calculations are performed based on the second similarity, the penalty coefficient, the structural consistency, the predetermined first weight coefficient, the predetermined second weight coefficient, and the predetermined third weight coefficient to obtain a clause risk score for the same insurance product clause. A compliance analysis matrix is ​​then generated based on the clause risk score and the judgment result. An example of a compliance risk matrix is ​​shown below. { "product_id": "P-LIFE-A", "clause_id": "CL-2025-001-EX12", "matched_regs": ["REG-2024-32-12"], "dimensions": { "exclusion_disclosure": { "align_score": 0.56, "risk_level": "medium", "violation_type": "Insufficient warning / Missing definition", "evidence_link": { "clause_span": "...pre-existing conditions...", "reg_span": "...A prompt is needed regarding the definition of pre-existing conditions..." }, "action": "Suggest rectification and trigger manual review", } }, "overall_risk": "medium", }".

[0050] Step S210: Weight the compliance analysis matrix of different dimensions of the same insurance product clause of the insurance product to be tested to obtain the clause risk score value of the same insurance product clause; In the specific implementation process of this step, the mathematical formula for calculating the risk score of the clause is as follows:

[0051] in, Refers to the first i Clause in d Alignment scores across compliance dimensions This refers to the importance weight of that compliance dimension (e.g., disclaimers > promotional language). Refers to the risk weight of violations (by) (mapping) D This refers to the number of compliance dimensions; the more "misaligned" (lower Align), the more "important" the dimension, and the more "serious the type of violation," the higher the risk score.

[0052] Step S211: Aggregate the risk scores of each clause of the insurance product to be tested to obtain the product risk score; In this step, the risk scores of each clause of the insurance product to be inspected are aggregated to obtain the product risk score. Specifically, if the inspection target is an insurance product, the system further calculates the product-level risk, as shown in the following mathematical expression:

[0053] in, N The product contains the number of clauses. The average value is used as the product risk score for the insurance product under test. In practice, the maximum value, weighted average, and priority of regulatory sensitive clauses can also be used to calculate the product risk score. This application supports calculation using multiple aggregation strategies.

[0054] Step S212: Based on the product risk score, classify the products using a preset classifier to obtain the risk level; In this step, the product risk score is classified using a preset classifier to obtain a risk level. Specifically, the current product risk score is determined by querying the risk level intervals of the preset classifier, and the risk level interval into which the current product risk score falls is determined as the target risk partition interval. The risk level corresponding to the target risk partition interval is determined as the risk level corresponding to the terms of the insurance product to be tested. The classification rules of the preset classifier are shown in Table 2.

[0055] The boundaries between each risk zone can be set according to actual needs.

[0056] Step S213: Based on the risk level query preset rule engine, obtain the risk management strategy for the risk level; In the specific implementation process of this step, a risk management strategy for the risk level is obtained based on the preset rule engine for querying the risk level. Specifically, the preset rule engine representing the risk level and the risk management strategy is shown in Table 3:

[0057] Step S214: Based on the report templates in the risk level query report generation engine, obtain the target report template corresponding to the risk level; In the specific implementation process of this step, the report template includes a summary of risk clauses, a description of risk type, references to corresponding regulatory provisions, and improvement suggestions. It uses a templated Prompt to generate standardized output, integrates with the enterprise messaging channel, and realizes a hierarchical mechanism of "instant push for high risks and batch daily reports for medium and low risks".

[0058] Step S215: Generate a compliance risk report for the insurance product to be tested, including the risk level and the risk management strategy corresponding to the risk level, based on the target report template.

[0059] In the specific implementation process of this step, a compliance risk report for the insurance product to be tested is generated based on the target report template, including the risk level and the risk handling strategy corresponding to the risk level. An example report is as follows: Output 1: Clause-level Risk Report: { "clause_id": "CL-2025-001-EX12", "risk_score": 0.62, "risk_level": "high", "recommended_action": "Manual review + clause revision"}; Output 2: Product-level risk report: (available for management / regulatory purposes) { "product_id": "P-LIFE-A", "overall_risk_score": 0.47, "overall_risk_level": "medium", "high_risk_clauses": ["CL-2025-001-EX12"], "summary": "There is a systemic risk that the disclaimer clauses are not adequately presented." }

[0060] This invention constructs an adaptive compliance inspection system that integrates a RAG large-scale model, a dynamic knowledge base, and intelligent agent technology, enabling 24 / 7 real-time semantic comparison and risk warning of insurance terms and regulatory policies. This system transforms traditional manual compliance review into a fully automated intelligent process, significantly improving monitoring real-time performance and shortening the regulatory response cycle from "monthly" to "hourly"; enhancing the accuracy of risk identification by uncovering hidden violations through multi-dimensional semantic analysis; drastically reducing compliance costs with an automation rate exceeding 90%; and constructing a self-evolving closed loop of "monitoring-identification-feedback-optimization," enabling insurance institutions to achieve continuous compliance and proactive risk prevention in a dynamically updated regulatory environment.

[0061] Another embodiment of this application provides an adaptive detection device for insurance terms, such as... Figure 4 As shown: Vectorization module 1 is used to vectorize the terms of the insurance product to be tested to obtain the query terms vector. Query module 2 is used to query a preset knowledge vector base based on the query clause vector to obtain several regulatory clause vectors that meet preset similarity conditions and correspond to the query clause vector; Analysis module 3 is used to perform compliance analysis based on the query clause vector and each of the regulatory clause vectors using a preset large language model to obtain a compliance analysis matrix; The detection module 4 is used to perform risk detection on the compliance analysis matrix using a preset classifier and rule engine to obtain the detection results.

[0062] In specific implementation, the device further includes a construction module, which is specifically used to construct the preset knowledge vector library that is updated in real time. The construction of the preset knowledge vector library includes: preprocessing external regulatory clauses collected in real time through an external interface and enterprise regulatory clauses obtained through real-time monitoring to obtain structured clause data; vectorizing the structured clause data using an embedding model to obtain a standardized knowledge unit stream; retrieving historical knowledge data pre-stored in the preset knowledge vector library based on the knowledge unit stream to obtain target historical knowledge data with the highest similarity to the same topic as the knowledge unit stream; calculating a first similarity between the knowledge unit stream and the target historical knowledge data; when the first similarity is less than a preset threshold, replacing the historical knowledge data based on the knowledge unit stream to update the preset knowledge vector library; when the first similarity is greater than the preset threshold, identifying the current version number of the knowledge unit stream and storing it in a target storage area with the same topic as the target historical knowledge data.

[0063] In the specific implementation process, the analysis module 3 is specifically used to: concatenate the query clause vector, each of the regulatory clause vectors, and the structured task analysis requirements to obtain a compliance analysis input statement; use the preset large language model to perform multi-dimensional structured judgment on the compliance analysis input statement to obtain judgment results corresponding to different dimensions, and the judgment results carry evidence location identifiers; perform alignment scoring based on each of the judgment results, the query clause vector, and each of the regulatory clause vectors to obtain the compliance analysis matrix; wherein, each of the dimensions includes rate consistency, compliance of coverage scope, sufficiency of disclaimer notice, and compliance of promotional language.

[0064] In the specific implementation process, the analysis module 3 is also used to: perform similarity calculation processing on the query clause vector and each of the regulatory clause vectors to obtain a second similarity; query the preset label scoring rules based on the risk label carried by the judgment result to obtain the penalty coefficient; calculate the structural consistency degree based on the mandatory regulatory conditions carried by the judgment result and the regulatory conditions actually satisfied by the query clause vector; and perform alignment scoring based on the second similarity, the penalty coefficient and the structural consistency degree to obtain the compliance analysis matrix.

[0065] In the specific implementation process, the detection module 4 is specifically used to: perform weighted scoring on the compliance analysis matrix of different dimensions of the same insurance product clause of the insurance product to be tested, to obtain the clause risk score value of the same insurance product clause; perform risk aggregation on the risk score values ​​of each clause of the insurance product to be tested, to obtain the product risk score value; classify the product risk score value using a preset classifier to obtain the risk level; and query the preset rule engine based on the risk level to obtain the risk handling strategy for the risk level.

[0066] In the specific implementation process, the detection module 4 is also used to perform calculations based on the second similarity, the penalty coefficient, the structural consistency, the predetermined first weight coefficient, the predetermined second weight coefficient, and the predetermined third weight coefficient to obtain the clause risk score value of the same insurance product clause.

[0067] In specific implementation, the device also includes a report generation module, which is specifically used to query each report template in the report generation engine based on the risk level to obtain a target report template corresponding to the risk level; and generate a compliance risk report of the insurance product to be tested, including the risk level and the risk handling strategy corresponding to the risk level, based on the target report template.

[0068] This invention constructs an adaptive compliance inspection system that integrates a RAG large-scale model, a dynamic knowledge base, and intelligent agent technology, enabling 24 / 7 real-time semantic comparison and risk warning of insurance terms and regulatory policies. This system transforms traditional manual compliance review into a fully automated intelligent process, significantly improving monitoring real-time performance and shortening the regulatory response cycle from "monthly" to "hourly"; enhancing the accuracy of risk identification by uncovering hidden violations through multi-dimensional semantic analysis; drastically reducing compliance costs with an automation rate exceeding 90%; and constructing a self-evolving closed loop of "monitoring-identification-feedback-optimization," enabling insurance institutions to achieve continuous compliance and proactive risk prevention in a dynamically updated regulatory environment.

[0069] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps: Step 1: Vectorize the terms and conditions of the insurance product to be tested to obtain the query terms vector; Step 2: Based on the query clause vector, query a preset knowledge vector base to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector; Step 3: Based on the query clause vector and each regulatory clause vector, a pre-set large language model is used to perform compliance analysis to obtain a compliance analysis matrix; Step 4: Use a preset classifier and rule engine to perform risk detection on the compliance analysis matrix and obtain the detection results.

[0070] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0071] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0072] The specific implementation process of the above method steps can be found in the embodiment of the above-mentioned adaptive detection method for any insurance clause, which will not be repeated here.

[0073] This invention constructs an adaptive compliance inspection system that integrates a RAG large-scale model, a dynamic knowledge base, and intelligent agent technology, enabling 24 / 7 real-time semantic comparison and risk warning of insurance terms and regulatory policies. This system transforms traditional manual compliance review into a fully automated intelligent process, significantly improving monitoring real-time performance and shortening the regulatory response cycle from "monthly" to "hourly"; enhancing the accuracy of risk identification by uncovering hidden violations through multi-dimensional semantic analysis; drastically reducing compliance costs with an automation rate exceeding 90%; and constructing a self-evolving closed loop of "monitoring-identification-feedback-optimization," enabling insurance institutions to achieve continuous compliance and proactive risk prevention in a dynamically updated regulatory environment.

[0074] Another embodiment of this application provides an electronic device, which can be a server. The electronic device includes a processor, memory, network interface, and database connected via a system bus. Its internal structure diagram is shown below. Figure 5 As shown, the processor of this electronic device provides computing and control capabilities. The memory of this electronic device includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of this electronic device is used to communicate with external clients via a network connection. When the program of this electronic device is executed by the processor, it implements the functions or steps of a server-side adaptive detection method for insurance terms.

[0075] In one embodiment, an electronic device is provided, which may be a client. The electronic device includes a processor, memory, network interface, display screen, and input device connected via a system bus. Its internal structure diagram is shown below. Figure 6 As shown, the processor of this electronic device provides computing and control capabilities. The memory of this electronic device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of this electronic device is used to communicate with an external server via a network connection. When the program of this electronic device is executed by the processor, it implements the functions or steps on the client side of an adaptive detection method for insurance terms.

[0076] Another embodiment of this application provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, performs the following method steps: Step 1: Vectorize the terms and conditions of the insurance product to be tested to obtain the query terms vector; Step 2: Based on the query clause vector, query a preset knowledge vector base to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector; Step 3: Based on the query clause vector and each regulatory clause vector, a pre-set large language model is used to perform compliance analysis to obtain a compliance analysis matrix; Step 4: Use a preset classifier and rule engine to perform risk detection on the compliance analysis matrix and obtain the detection results.

[0077] The specific implementation process of the above method steps can be found in the embodiment of the above-mentioned adaptive detection method for any insurance clause, which will not be repeated here.

[0078] This invention constructs an adaptive compliance inspection system that integrates a RAG large-scale model, a dynamic knowledge base, and intelligent agent technology, enabling 24 / 7 real-time semantic comparison and risk warning of insurance terms and regulatory policies. This system transforms traditional manual compliance review into a fully automated intelligent process, significantly improving monitoring real-time performance and shortening the regulatory response cycle from "monthly" to "hourly"; enhancing the accuracy of risk identification by uncovering hidden violations through multi-dimensional semantic analysis; drastically reducing compliance costs with an automation rate exceeding 90%; and constructing a self-evolving closed loop of "monitoring-identification-feedback-optimization," enabling insurance institutions to achieve continuous compliance and proactive risk prevention in a dynamically updated regulatory environment.

[0079] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. Those skilled in the art can make various modifications or equivalent substitutions to this application within the scope and nature of this application, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. An adaptive detection method for insurance clauses, characterized in that, include: The terms and conditions of the insurance product to be tested are vectorized to obtain the query terms vector. Based on the query clause vector, a preset knowledge vector base is queried to obtain several regulatory clause vectors that meet the preset similarity conditions and correspond to the query clause vector. Based on the query clause vector and each of the regulatory clause vectors, a pre-set large language model is used to perform compliance analysis to obtain a compliance analysis matrix; The compliance analysis matrix is ​​subjected to risk detection using a preset classifier and rule engine to obtain the detection results.

2. The method as described in claim 1, characterized in that, Before querying the preset knowledge vector base based on the query term vector, the method further includes: constructing the preset knowledge vector base that is updated in real time; The construction of the preset knowledge vector base that is updated in real time specifically includes: The external regulatory clauses collected in real time through external interfaces and the enterprise regulatory clauses obtained through real-time monitoring are preprocessed to obtain structured clause data. The structured clause data is vectorized using the Embedding model to obtain a standardized knowledge unit stream; Based on the knowledge unit stream, historical knowledge data stored in the preset knowledge vector library is retrieved to obtain target historical knowledge data with the highest similarity to the same topic as the knowledge unit stream; Calculate the first similarity between the knowledge unit stream and the target historical knowledge data; When the first similarity is less than a preset threshold, the historical knowledge data is replaced based on the knowledge unit stream to update the preset knowledge vector library; When the first similarity is greater than a preset threshold, the current version number of the knowledge unit stream is identified and stored in the target storage area with the same theme as the target historical knowledge data.

3. The method as described in claim 1, characterized in that, The compliance analysis, based on the query clause vector and each regulatory clause vector, employs a pre-defined large language model to obtain a compliance analysis matrix, specifically including: The query clause vector, each regulatory clause vector, and the structured task analysis requirements are concatenated to obtain the compliance analysis input statement; The preset large language model is used to perform multi-dimensional structured judgment on the compliance analysis input statement to obtain judgment results corresponding to different dimensions, and the judgment results carry evidence location identifiers; The compliance analysis matrix is ​​obtained by aligning and scoring the judgment results, the query clause vectors, and the regulatory clause vectors. The dimensions mentioned include consistency of rates, compliance of coverage scope, adequacy of disclaimer notices, and compliance of promotional language.

4. The method as described in claim 3, characterized in that, The compliance analysis matrix is ​​obtained by aligning and scoring based on the judgment results, the query clause vector, and the regulatory clause vector, specifically including: A second similarity score is obtained by performing similarity calculation on the query clause vector and each of the regulatory clause vectors. Based on the risk label carried by the judgment result, a preset label scoring rule is queried to obtain the penalty coefficient; The structural consistency degree is calculated based on the mandatory regulatory conditions carried by the judgment result and the regulatory conditions actually satisfied by the query clause vector. The compliance analysis matrix is ​​obtained by performing alignment scoring based on the second similarity, the penalty coefficient, and the structural consistency.

5. The method as described in claim 1, characterized in that, The process of using a preset classifier and rule engine to perform risk detection on the compliance analysis matrix and obtaining detection results specifically includes: The compliance analysis matrix of the same insurance product clauses of the same insurance product to be tested is weighted and scored to obtain the clause risk score value of the same insurance product clauses; The risk scores of each clause of the insurance product to be tested are aggregated to obtain the product risk score. Based on the product risk score, a preset classifier is used to classify the products to obtain risk levels. Based on the risk level query preset rule engine, a risk management strategy for the risk level is obtained.

6. The method as described in claim 5, characterized in that, The compliance analysis matrix of the same insurance product clause for the insurance product under test is weighted and scored to obtain a clause risk score for the same insurance product clause, specifically including: The risk score of the insurance product terms is obtained by calculating based on the second similarity, the penalty coefficient, the structural consistency, the predetermined first weight coefficient, the predetermined second weight coefficient, and the predetermined third weight coefficient.

7. The method as described in claim 4, characterized in that, The method further includes: Based on the report templates in the risk level query report generation engine, a target report template corresponding to the risk level is obtained; A compliance risk report for the insurance product under test is generated based on the target report template, including the risk level and the risk management strategy corresponding to the risk level.

8. An adaptive detection device for insurance terms, characterized in that, include: The vectorization module is used to vectorize the terms of the insurance product to be tested, and obtain the query terms vector. The query module is used to query a preset knowledge vector base based on the query clause vector to obtain several regulatory clause vectors that meet preset similarity conditions and correspond to the query clause vector. The analysis module is used to perform compliance analysis based on the query clause vector and each of the regulatory clause vectors using a preset large language model to obtain a compliance analysis matrix; The detection module is used to perform risk detection on the compliance analysis matrix using a preset classifier and rule engine to obtain the detection results.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the adaptive detection method for insurance terms as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the insurance clause adaptive detection method according to any one of claims 1-7.

Images