Audit policy updating method and device, equipment and medium

Abstract

The application provides a method and device for auditing strategy updating, equipment and medium. The method comprises the following steps: obtaining business auditing content, an abnormal auditing track related to the business auditing content, and at least one auditing skill; based on the business auditing content, the abnormal auditing track, and the at least one auditing skill, a first target auditing node related to an abnormal auditing label is located in at least one auditing node, and node optimization description content for the target auditing node is output; the weight of each auditing skill affecting the target auditing node is evaluated to obtain the skill influence weight of each auditing skill; based on the skill influence weight of each auditing skill and the node optimization description content, an auditing strategy skill library is updated to obtain an auditing strategy updating skill library for reconfiguring an auditing model. The embodiments of the application can effectively improve the effectiveness of the strategy updating of the auditing strategy, and improve the accuracy of the content auditing of the auditing model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for updating an audit strategy. Background Technology

[0002] When existing moderation agents driven by large language models (LMs) conduct content review on relevant business review content, they perform multi-step logical reasoning on the business review content based on the standard operating procedure (SOP) indicated by the review strategy to obtain the content review trajectory for the business review content.

[0003] However, the inventors discovered in practice that when the review tags for business-reviewed content carried in the content review trajectory are abnormal (i.e., the large language model erred in reviewing the business-reviewed content), updating the review strategy often involves summarizing the overall trajectory anomalies and generating a new review standard operating procedure (SOP) based on the summary results. This new SOP then replaces the old one to update the review strategy. This method of updating the review strategy causes the abnormal signals from the content review trajectory to spread throughout the entire review strategy, reducing the effectiveness of the update and making it difficult to guarantee the accuracy of content review when facing other business-reviewed content. Therefore, improving the effectiveness of updating the review strategy, and thus improving the accuracy of the review model in content review, is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for updating review strategies, which can improve the effectiveness of updating review strategies and thus improve the accuracy of content review by the review model.

[0005] This application provides a method for updating an audit strategy, including: Obtain the business review content, the abnormal review trajectory related to the business review content, and at least one review skill; the abnormal review trajectory refers to the review trajectory that carries an abnormal review tag and contains at least one review node, and the abnormal review tag is obtained by the review model calling at least one review skill to review the business review content; at least one review skill refers to the skill in the review strategy skill library configured for the review model; Based on the business review content, abnormal review trajectory and at least one review skill, locate the first optimizable target review node related to the abnormal review tag in at least one review node, and output the node optimization description content for the target review node; Based on the business review content, abnormal review trajectory, at least one review skill, and target review node, the weight of each review skill used to influence the target review node is evaluated to obtain the skill influence weight of each review skill. Based on the skill influence weight and node optimization description of each review skill, the review strategy skill library is updated to obtain a review strategy updated skill library for reconfiguration to the review model.

[0006] One embodiment of this application provides an audit strategy update device, including: The audit trajectory acquisition module is used to acquire business audit content, abnormal audit trajectories related to the business audit content, and at least one audit skill. An abnormal audit trajectory refers to an audit trajectory that carries an abnormal audit tag and contains at least one audit node. The abnormal audit tag is obtained by the audit model calling at least one audit skill to audit the business audit content. At least one audit skill refers to a skill in the audit strategy skill library configured for the audit model. The optimized node location module is used to locate the first optimizable target audit node related to the abnormal audit tag in at least one audit node based on the business audit content, abnormal audit trajectory and at least one audit skill, and output the node optimization description content for the target audit node. The skill weight calculation module is used to evaluate the weight of each review skill that affects the target review node based on the business review content, abnormal review trajectory, at least one review skill, and the target review node, so as to obtain the skill influence weight of each review skill. The review strategy update module is used to update the review strategy skill library based on the skill influence weight and node optimization description of each review skill, and obtain the review strategy update skill library for reconfiguration to the review model.

[0007] One aspect of this application provides a computer device, including a memory and a processor. The memory is connected to the processor, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the method provided in one aspect of this application.

[0008] One aspect of this application provides a computer-readable storage medium storing a computer program adapted to be loaded and executed by a processor, so that a computer device having a processor performs the method provided in one aspect of this application.

[0009] According to one aspect of this application, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the method provided in the above aspect.

[0010] This application embodiment can acquire business review content, abnormal review trajectories related to the business review content, and at least one review skill. Here, an abnormal review trajectory refers to a review trajectory carrying an abnormal review tag and containing at least one review node. The abnormal review tag is obtained by the review model calling at least one review skill to review the business review content, and the at least one review skill refers to a skill in the review strategy skill library configured for the review model. Further, this application embodiment can, based on the acquired business review content, abnormal review trajectory, and at least one review skill, locate the first optimizable target review node related to the abnormal review tag among at least one review node, and output node optimization description content for the target review node. Further, based on the business review content, abnormal review trajectory, at least one review skill, and the target review node, a weight evaluation can be performed on each review skill used to influence the target review node to obtain the skill influence weight of each review skill for the target review node. Finally, based on the obtained skill influence weight of each review skill and the node optimization description content for the target review node, the review strategy skill library can be updated to obtain a review strategy update skill library for reconfiguration to the review model.

[0011] It is understood that, in this embodiment of the application, after anomaly localization of the acquired abnormal review trajectory and obtaining the first optimizable target review node among at least one review node contained in the abnormal review trajectory, the review strategy skill base can be updated with skills strongly correlated with the target review node by optimizing the description content of nodes related to the target review node and the skill influence weight of each review skill for the target review node. This can effectively prevent the abnormal signals of the abnormal review trajectory from being indiscriminately spread to all skills in the review strategy skill base, thereby improving the accuracy and effectiveness of updating the review strategy skill base. Furthermore, when this embodiment of the application reconfigures the obtained review strategy update skill base to the review model to update the review strategy of the review model, it can effectively improve the effectiveness of updating the review strategy, thereby improving the accuracy of the review model in content review. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application; Figure 2 This is a schematic diagram of a data interaction scenario provided in an embodiment of this application; Figure 3 This is a schematic diagram of an audit strategy update method provided in an embodiment of this application; Figure 4 This is a schematic diagram of a scenario for obtaining an abnormal audit trajectory provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating a scenario for acquiring auditing skills, provided in an embodiment of this application. Figure 6 This is a schematic diagram of a scenario for locating abnormal nodes provided in an embodiment of this application; Figure 7 This is a schematic diagram illustrating a scenario for obtaining skill influence weights according to an embodiment of this application; Figure 8 This is a schematic diagram illustrating a scenario where an audit strategy is updated through a revision process, as provided in an embodiment of this application. Figure 9 This is a schematic diagram of a scenario for testing an updated skill library, provided in an embodiment of this application. Figure 10 This is a schematic diagram of another audit strategy update method provided in an embodiment of this application; Figure 11 This is a schematic diagram illustrating a scenario for constructing an audit strategy skill library, provided in an embodiment of this application. Figure 12 This is a schematic diagram illustrating a scenario where an audit strategy is updated through a generation process, as provided in an embodiment of this application. Figure 13 This is a schematic diagram of the structure of an audit strategy update device provided in an embodiment of this application; Figure 14 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0014] 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, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0015] The following describes some custom terms used in the embodiments of this application: I. Review Strategy Skill Library: This is a data set that stores the review skills that the review model can call when performing content review. It is understood that, in this embodiment of the application, when calling the review model to perform content review on business review content, at least one review skill matching the business review content can be retrieved from multiple candidate skills included in the review strategy skill library. The review model can then complete the content review of the business review content by calling at least one review skill retrieved here. In this application, the business review content refers to content provided by users that is to be reviewed. For example, in business scenarios where certain users (e.g., user U1) can publish content themselves (e.g., business channel scenarios, video playback scenarios, etc.), when a user publishes relevant content (e.g., text content, image content, video content, etc.) in such a business scenario, the present application embodiment can identify the content published by the user as business review content and conduct content review on the business review content (i.e., determine whether the business review content contains illegal content indicated by the business scenario). If the review result (the content review label for the business review content) indicates that the business review content is legal content, the business review content can be displayed in the business scenario, so that users other than the publisher of the business review content (e.g., user U2) can also view the business review content in the business scenario.

[0016] It's understandable that the audit strategy skill library can also be called a Standard Operating Procedure (SOP) library. Here, the SOP library is a modular knowledge set composed of natural language text that can be read by large models. Each item in the audit strategy skill library corresponds to a complete candidate skill, or a standard operating procedure (SOP), such as a QR code parsing SOP or a slang metaphor decoding SOP. During the audit phase of business audit content, these candidate skills are dynamically retrieved to find at least one audit skill that matches the business audit content. This retrieved audit skill is then injected into the audit model's context window to guide the audit model to perform multi-step audits of the business audit content according to the preset audit logic indicated by the audit skill. It's also understandable that the candidate skills in the audit strategy skill library are not included in the model weights themselves, but rather are external soft modules of the audit model, supporting independent loading and flexible replacement, essentially a parameterized large model context constraint template. Optionally, the candidate skills in the review strategy skill library of this application embodiment can be stored in text form. This allows reviewers to directly read, audit, or even intervene in any candidate skill to adapt to new violation methods when dealing with rapid variations in new practices in actual business scenarios. It is understood that modifying a text rule (candidate skill) takes less time than fine-tuning the model, allowing for a faster response to the emergence of new violation methods and improving the accuracy of content review for business review.

[0017] II. Abnormal Review Trajectory: When the review model is invoked to review business review content, the resulting content review trajectory, consisting of at least one review node and carrying an abnormal review tag, is obtained. It can be understood that when the review model reviews business review content, it executes at least one review step. After completing the content review, it determines at least one review node based on the step execution actions (e.g., tool invocation, evidence summary, etc.) and step execution results (e.g., tool invocation results, evidence summary results, etc.). Each review node at this point includes two parts: an action and an observation. The action corresponds to the node execution action of the review node (i.e., the step execution action), and the observation corresponds to the node execution result of the review node (i.e., the step execution result).

[0018] It is understood that, after obtaining the content review trajectory for the business review content, this embodiment of the application can perform anomaly judgment on the content review tags carried in the content review trajectory (for example, it can manually review whether the content review tags match the legality of the business review content). If it is determined that the content review tag is an abnormal review tag (that is, the content review tag does not match the legality of the business review content; for example, when the business review content is illegal, the content review tag indicates that the business review content is legal, or when the business review content is legal, the content review tag indicates that the business review content is illegal), then this embodiment of the application can determine the content review trajectory carrying the abnormal review tag as an abnormal review trajectory.

[0019] III. Target Review Node: The first optimizable node related to the abnormal review tag among at least one review node included in the abnormal review trajectory. It is understood that in this embodiment, the target review node can be the first operable error node (Earliest Accountable Fault Step) in the content review trajectory. Specifically, when the review model performs multi-step reasoning on the business review content (e.g., step 1 extracts image and text information, then step 2 determines product attributes, and then step 3 determines violations; each step corresponds to a review node in the final generated content review trajectory) and ultimately produces a misjudgment conclusion (i.e., the content review tag carried by the content review trajectory is an abnormal review tag), the abnormal location unit provided in this embodiment can reverse-track the abnormal review trajectory (i.e., work backward from the final result), verifying the reasoning state of each intermediate link (each review node in the abnormal review trajectory) one by one, attempting to locate the first step (review node) where a logical jump, rule misuse, or information omission occurs. Correcting this step can correct the final review result. At this time, the review node found by the abnormal location unit can be determined as the first operable error node (i.e., the target review node).

[0020] It is understandable that the target review node here is not necessarily the first step (review node) in the content review trajectory that has an error. This is because when the review model reviews business review content, there may be flaws in the execution of some review steps. However, as long as the error in the step does not affect the final judgment (review result, i.e., content review label), this application embodiment will not identify the error step as the target review node that needs to be corrected. This helps to quickly and accurately locate the key steps in the trajectory when faced with an erroneous content review trajectory in actual business scenarios.

[0021] IV. Skill Influence Weight: This is a weight value used to describe the degree of correlation between a review skill and the target review node. It is understood that in the application embodiment, the review model can use at least one review skill when reviewing business review content, and each review skill corresponds to a candidate skill in the review strategy skill library. In this application embodiment, after determining the target review node in the abnormal review trajectory, the skill influence weight of each of these at least one review skills for the target review node can be calculated (i.e., the degree of abnormal correlation or responsibility correlation in causing the target review node to become abnormal). Therefore, the skill influence weight here can also be called the responsibility weight. It is understood that, after determining the skill influence weight of each audit skill, this application embodiment can determine how to update the audit strategy skill library based on the obtained skill influence weight. For example, this audit embodiment can enter the skill revision process, revise the audit skill with the largest skill influence weight to obtain the revised sub-audit skill, and then replace the audit strategy skill library with the revised sub-audit skill. Alternatively, it can enter the skill generation process, summarize the faults of the target audit node to generate a new audit skill (i.e., the summary sub-audit skill), and add the summary sub-audit skill to the audit strategy skill library.

[0022] It is understood that the embodiments of this application update the review strategy skill base by locating the target review node and calculating the skill influence weight of each review skill for the target review node, thereby enabling step-level attribution of abnormal review trajectories. Specifically, step-level attribution can compress the failure signal of abnormal review trajectories to the first operable error node and its associated review skill, making minimal edits (i.e., skill revision or skill generation) only to the rules (review skills) corresponding to that node. This refined attribution method can effectively improve the stability of updating the review strategy skill base.

[0023] V. Review Strategy Update Skill Library: The skill library obtained after updating the review strategy skill library is used to reconfigure it for the review model. It is understood that, in this embodiment of the application, after updating the review skill library (e.g., revising skills or adding newly generated skills), the resulting updated skill library (e.g., the first updated skill library) can undergo sandbox verification and qualification. Specifically, updating the review strategy skill library often leads to a risk of reduced accuracy when reviewing business content other than the current business review content. Therefore, after obtaining candidate new strategies (i.e., updating the skill library), if it is necessary to deploy the candidate new strategies (i.e., reconfigure the updated skill library for the review model), sandbox verification and qualification of the updated skill library is required. In other words, the sandbox verification and qualification at this time acts as a soft switch for deploying the updated skill library.

[0024] Specifically, when verifying the updated skill library through sandbox validation, a Golden Benchmark test set containing various representative cases can be retrieved first. Under this test set, the original strategy library (review strategy skill library) and the candidate library containing the new strategy (updated skill library) will each run an independent double-blind test. In this embodiment, the difference in key quantitative indicators such as the overall accuracy and recall of the two strategy libraries can be calculated based on the data output from the double-blind test. When the obtained performance difference indicates that the new strategy has accurately corrected the target misjudged cases (i.e., the updated review label obtained from the content review of business review content is a non-abnormal label), and indicates that the overall performance indicator of the new strategy has not decreased compared to the old strategy, then the updated skill library at this time can be determined as the review strategy update skill library, and the review strategy update skill library can be reconfigured to the review model. Optionally, if the obtained performance difference indicates that the overall performance indicator of the new strategy has decreased compared to the old strategy, then the new strategy will not be launched. In other words, the strategy update request for the review model will be rejected at this time, and the review model will still use the original review strategy skill library. Understandably, by using a sandbox verification admission mechanism, each iteration of the audit strategy can be limited to the scope of Pareto improvement, thereby increasing the reliability of updating the audit strategy.

[0025] Understandably, the golden regression test set used in the sandbox validation admission mechanism acts as an admission gate in the strategy iteration process; it does not participate in parameter fitting or model training. Whenever a candidate strategy is generated, the system can complete the entire inference and judgment process on this set and provide a set of evaluation metrics that align with the online environment. The test data in this set consists of pre-selected and well-labeled review cases, covering various representative scenarios, typical variants of illegal content, and boundary samples of normal content that are easily misidentified. If a new strategy can identify illegal content in the test set but causes normal content to be blocked as well, the new strategy will not be deployed. The size of the test set can be constrained here to improve the efficiency of strategy testing.

[0026] Optionally, embodiments of this application can update the review strategy through a dynamic moderation adapter. Specifically, this dynamic moderation adapter can connect four steps—anomaly localization (locating the target review node in the abnormal review trajectory), responsibility tracing (calculating the skill influence weight of each review skill), targeted modification (updating the review strategy skill library), and sandbox verification—into an automated pipeline. One end of this pipeline receives abnormal review trajectories generated by manual error correction or customer feedback, and the other end outputs an optimized review strategy updated skill library that has undergone rigorous verification and can be hot-updated online. The parameters of the underlying large language model (review model) remain completely frozen throughout the process, and the iteration of review capabilities is only accomplished by adding, deleting, and modifying the text-based SOP library. In actual business scenarios, this pipeline is usually deployed as an offline task. Specifically, it can centrally collect misjudgment cases and complete strategy updates and verifications offline to achieve seamless switching to online review capabilities.

[0027] For further details, please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. Figure 1 As shown, this network architecture may include a server 10d and a user terminal cluster. The user terminal cluster may include one or more user terminals; the number of user terminals is not limited here. Figure 1As shown, the user terminal cluster can specifically include user terminal 10a, user terminal 10b, and user terminal 10c. Server 10d can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. User terminals 10a, 10b, and 10c can all include: smartphones, tablets, laptops, PDAs, mobile internet devices (MIDs), wearable devices (such as smartwatches and smart bracelets), smart voice interaction devices, smart home appliances (such as smart TVs), and in-vehicle devices, as well as other electronic devices with video / image playback capabilities. Figure 1 As shown, user terminals 10a, 10b, and 10c can each connect to server 10d via a network, so that each user terminal can interact with server 10d through the network connection.

[0028] It is understood that the server 10d in this embodiment can obtain business review content, an abnormal review trajectory related to the business review content, and at least one review skill. Here, the abnormal review trajectory refers to a review trajectory carrying an abnormal review tag and containing at least one review node. The abnormal review tag is obtained by the review model calling at least one review skill to review the business review content, and the at least one review skill refers to a skill in the review strategy skill library configured for the review model. Optionally, in this embodiment, when a user terminal (e.g., user terminal 10a) obtains business review content, it can call the review model to obtain at least one review skill related to the business review content from the review strategy skill library, and perform content review on the business review content based on the obtained at least one review skill to obtain an abnormal review trajectory carrying an abnormal review tag. Then, the business review content, the abnormal review trajectory, and the at least one review skill used are sent to the server 10d. Optionally, in another possible implementation, after receiving the business review content provided by the user, the user terminal 10a can send the business review content to the server 10d. The server 10d then invokes the review model to perform content review based on at least one review skill from the review strategy skill library, thereby obtaining an abnormal review trajectory. The specific method by which the server 10d obtains the abnormal review trajectory is not limited here.

[0029] Understandably, Server 10d can, based on the acquired business review content, abnormal review trajectory, and at least one review skill, locate the first optimizable target review node related to the abnormal review tag among at least one review node included in the abnormal review trajectory, and output node optimization description content for the target review node. Furthermore, Server 10d can, based on the business review content, abnormal review trajectory, at least one review skill, and the target review node, perform weight evaluation on each review skill used to influence the target review node, obtaining the skill influence weight of each review skill for the target review node. Finally, Server 10d can, based on the obtained skill influence weight of each review skill and the node optimization description content for the target review node, update the review strategy skill library, obtaining an updated review strategy skill library for reconfiguration to the review model.

[0030] It is understood that, after obtaining an abnormal review trajectory, this embodiment of the application can perform anomaly localization based on the abnormal review trajectory, obtain the first optimizable target review node among at least one review node included in the abnormal review trajectory, and the node optimization description content for the target review node, and calculate the skill influence weight of each review skill used in the content review process for the target review node. Therefore, when updating the review strategy skill library through the node optimization description content and the skill influence weight of each review skill, the accuracy and effectiveness of updating the review strategy skill library can be improved. Furthermore, when this embodiment of the application reconfigures the obtained review strategy update skill library to the review model to update the review strategy of the review model, it can effectively improve the effectiveness of strategy updates for the review strategy, thereby improving the accuracy of the review model in content review.

[0031] It is understood that in the specific implementation of this application, business data of users, enterprises, institutions and other entities may be involved (e.g., information such as business review content published by users). When the above embodiments of this application are applied to specific products or technologies, permission or consent from users, enterprises, institutions and other entities is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0032] For further details, please see Figure 2 , Figure 2 This is a schematic diagram illustrating a data interaction scenario provided in an embodiment of this application. For example... Figure 2As shown, it can be understood that in this embodiment of the application, the content review unit in the provided review strategy update system can review the business review content (business review content 21a). Specifically, at least one review skill (i.e., review skill 22a, including skill 221a, skill 221b, ..., skill 221n) related to the business review content 21a can be obtained from the review strategy skill library (skill library 20a) configured for the review model (review model 23a). The review model 23a then reviews the business review content 21a based on the obtained review skill 22a, resulting in an abnormal review trajectory (review trajectory 24a) carrying an abnormal review tag (review tag 242a). Further, the content review unit can send the obtained review trajectory 24a, review skill 22a, and business review content 21a to the strategy update unit, which then updates the review strategy.

[0033] Specifically, the strategy update unit can, based on the acquired business review content 21a, review trajectory 24a, and review skills 22a, locate the first optimizable abnormal node related to the abnormal review label (review label 242a) among at least one review node (including node 241a, node 241b, and node 241c) included in the review trajectory 24a, i.e., the target review node (i.e., node 241a), and output node optimization description content (node ​​optimization description content 26a) for node 241a. Further, in this embodiment, based on the business review content 21a, review trajectory 24a, review skills 22a, and the target review node (node ​​241a), a weight evaluation can be performed on each review skill used to influence the target review node to obtain the skill influence weight of each review skill for the target review node (i.e., influence weight 25a, including the weight 251a of skill 221a, the weight 251b of skill 221b, ..., the weight 251n of skill 221n). Finally, based on the obtained influence weights 25a and the node optimization description content 26a for the target review node, the review strategy skill library (skill library 20a) can be updated to obtain a review strategy update skill library for reconfiguration to the review model 23a. Optionally, the strategy update skill library can be launched on the relevant content review platform, so that the review model 23a on the content review platform can subsequently perform content review based on the strategy update skill library.

[0034] It is understood that, after obtaining an abnormal review trajectory, this embodiment of the application can perform anomaly localization based on the abnormal review trajectory, obtain the first optimizable target review node among at least one review node included in the abnormal review trajectory, and the node optimization description content for the target review node, and calculate the skill influence weight of each review skill used in the content review process for the target review node. Therefore, when updating the review strategy skill library through the node optimization description content and the skill influence weight of each review skill, the accuracy and effectiveness of updating the review strategy skill library can be improved. Furthermore, when this embodiment of the application reconfigures the obtained review strategy update skill library to the review model to update the review strategy of the review model, it can effectively improve the effectiveness of strategy updates for the review strategy, thereby improving the accuracy of the review model in content review.

[0035] For further details, please see Figure 3 , Figure 3 This is a schematic diagram of an audit strategy update method provided in an embodiment of this application. Figure 3 As shown, it is understandable that Figure 3 The method for updating the audit policy shown can be executed by the server; for example, the server here could be something like... Figure 1 Server 10d is shown. (As shown) Figure 3 As shown, the audit strategy update method may include at least the following: Figure 3 Steps S101 to S104 in the process.

[0036] Step S101: Obtain the business review content and the abnormal review trajectory related to the business review content, as well as at least one review skill; It is understood that the embodiments of this application can first obtain the business review content, the abnormal review trajectory related to the business review content, and at least one review skill. It is understood that the abnormal review trajectory here refers to a review trajectory carrying an abnormal review tag and containing at least one review node. The abnormal review tag is obtained by the review model calling at least one review skill to review the business review content, and at least one review skill refers to a skill in the review strategy skill library configured for the review model. It is understood that the review model here is an intelligent model (e.g., a large language model) used for content review of business review content. When the review model reviews the business review content, it can first search for at least one review skill related to the current business review content in the review strategy skill library configured for the review model, thereby obtaining the content review trajectory. At this time, at least one review step executed by the review model during the content review process corresponds to at least one review node contained in the content review trajectory, and the content review trajectory also carries the review result (content review tag) of the content review of the business review content.

[0037] Optionally, in this application embodiment, the content review trajectory carrying the content review tag as an abnormal tag can be called an abnormal review trajectory. After obtaining the abnormal review trajectory, the corresponding business review content and at least one review skill can be obtained together, so as to use the review strategy dynamic adapter provided in this application embodiment to update the review strategy of the review model (i.e., update the review strategy skill library).

[0038] Optionally, in one possible implementation, before obtaining the business review content and the abnormal review trajectory and at least one review skill related to the business review content, this embodiment of the application can perform content review on the business review content through a review model. Specifically, when obtaining the review model used to review the business review content, this embodiment of the application can search for at least one review skill related to the business review content from the review strategy skill library configured for the review model. Optionally, the review model can call a relevant skill retrieval tool to search for at least one review skill related to the business review content from the review strategy skill library. It can be understood that the review strategy skill library here is a modular skill collection, which can be denoted as K. This skill collection stores multiple skills (which can be called candidate skills). Each skill s_i∈K can be a text-based review standard operating procedure, including a title, applicable conditions, and behavioral guidelines, etc.

[0039] Specifically, the details of a skill in the audit strategy skill set can be seen as follows: Skill Identifier: S-ILLEGAL-GOODS-ALIAS-01; Skill Name: SOP for Identifying Substandard Product Names and Off-Site Traffic Generation; Version: v1.3; Applicable Content Modal: Text, Image, QR Code; Applicable Risk Category: Substandard Product Transactions or Traffic Generation; Parameters: Optical Character Recognition (OCR) minimum confidence level ocr_min_conf=0.75; alias semantic similarity threshold alias_sim_threshold=0.82; minimum number of independent evidences required to establish a risky conclusion min_evidence_count=2; QR code parsing switch qr_decode=true; dictionary version alias_dict_version=2026-08.

[0040] Pre-verification conditions: The input must contain at least one of the following: body text, title, or image; if the image contains text areas, OCR must be completed first; if a QR code exists, QR code parsing must be completed, or "cannot be parsed" must be recorded; violations cannot be directly determined based on a single ordinary word.

[0041] Execution steps: 1. Normalize characters in the title, body text, and OCR text, handling character splitting, homophones, pinyin abbreviations, and symbol insertion; 2. Call the pseudonym mapping table to determine whether candidate words form a semantic mapping with the controlled goods; 3. Extract transaction or traffic-driving evidence, including price, quantity, contact information, QR code, and suggestive transaction verbs; 4. Perform reverse verification to exclude negative scenarios such as news discussions, popular science, law enforcement reports, and normal product name similarities; 5. Output high risk only when both "controlled object evidence" and "transaction / traffic-driving evidence" are valid, and the number of independent evidence is not less than min_evidence_count; otherwise, output low risk or transfer to manual review.

[0042] Understandably, a skill in the audit strategy skill library contains specific usage parameters, pre-verification conditions, and skill execution steps, enabling the audit model to execute the audit steps indicated by the skill when a skill is invoked, thereby achieving partial audit of the business audit content.

[0043] It is understood that the at least one review skill obtained from the review strategy library in this application embodiment is a skill that matches the content information of the business review content. Therefore, the review model can further call at least one review skill to perform content review on the business review content, obtaining at least one review node, and determining a content review tag for the business review content based on the at least one review node. Each of the at least one review nodes here corresponds to a review step executed by the review model when performing content review, and one review step corresponds to one review skill. The content review tag here is the content review result for the business review content output by the last review node in the at least one review node, which can indicate whether the current business review content is legal or illegal under the review of the review model. Further, this application embodiment can construct a content review trajectory for the business review content through the at least one review node and the content review tag, and determine the content review trajectory as an abnormal review trajectory when the content review tag (e.g., the review pass tag or review block tag output by the review model) is determined to be an abnormal review tag. The review pass tag here is a tag indicating that the business review content is legal content, and the review block tag is a tag indicating that the business review content is illegal content.

[0044] Optionally, the content review trajectory can be tagged using a different large language model than the review model, or through manual review, to determine whether the content review tags are abnormal. For example, when the content review tag is the "PASS" tag (review passed tag), which indicates that the business review content is legal, if the tag review determines that the business review content is illegal, then the content review tag is an abnormal review tag. When the content review tag is the "BLOCK" tag (review blocked tag), which indicates that the business review content is illegal, if the tag review determines that the business review content is legal, then the content review tag is an abnormal review tag. It is understandable that if the content review tag in the content review trajectory is determined to be an abnormal review tag, it can be determined that an error occurred during the current review process of the business review content by the review model. Therefore, the current content review trajectory can be considered an abnormal review trajectory with an error (abnormality).

[0045] Furthermore, after obtaining the abnormal review trajectory, this embodiment of the application can perform subsequent policy updates to the review strategy of the review model. In other words, the abnormal review trajectory of this embodiment can be understood as a signal source for updating the review strategy skill base. It is understood that the policy update procedure for the review model in this embodiment is performed under the premise that the underlying large language model (review model) is frozen (model parameters do not change, i.e., gradient updates). That is, by updating the review strategy skill base configured for the review model, the review strategy of the review model is updated, thereby improving the review capability of the review model. Compared to directly training the review model, updating the review strategy skill base of the review model can effectively reduce the update time of the review strategy and improve the policy update speed of the review strategies used by the review model.

[0046] The review trajectory (e.g., content review trajectory, abnormal review trajectory) obtained in the embodiments of this application can be understood as a sequence containing actions and observations, as shown in formula (1):

[0047] It is understandable that in formula (1), the audit trajectory It contains at least one review node (i.e., T nodes), and each review node corresponds to a specific review step performed by the review model during content review. For example, where... This refers to the action at step t (i.e., the action executed at step t corresponding to the t-th review step, or the action executed at step t of the review node). The observation returned by the environment (i.e., the execution result of the t-th review step obtained by the review model, which can also be considered as the node execution result of the t-th review node). For example, the business review content (content package q) received by the review model may include: title "Selling idle Bingtang"; body text "Those who understand, see the picture, large quantity"; image img_01. At this time, the embodiment of this application can perform content parsing on the business review content to generate a task description (d_q) for the review task of the business review content: "Review product posts containing text and images, focusing on identifying homophones, transaction intentions, and off-site traffic." Then, at least one review skill matching the task description of the current business review content can be found in the review strategy skill library. For example, the following four skills can be found: "Image and Text OCR Cleaning SOP", "Identification SOP of Obscure and Illegal Product Nicknames", "QR Code Traffic Recognition SOP", and "Risk Aggregation Skill SOP". The at least one review skill found is injected into the model context of the review model so that the review model can call these four review skills to perform content review of the current business review content.

[0048] For example, the content review trajectory generated by the review model at this time can be as follows: t1: Action a1=OCR(image=img_01, model=ocr_v3); Observation o1={text:“300 / copy, scan to chat”, confidence:0.94}. t2: Action a2=DecodeQR(image=img_01); Observation o2={url:“wx-contact: / / …”, type:“external contact information”}. t3: ​​Action a3=NormalizeAndMapAlias(term="Bingtang", dict=2026-08); Observation o3={normalized:“Bingtang”, matched_entity:null, similarity:0.61}. t4: Action a4=AggregateRisk(evidence=[o1,o2,o3]); Observation o4={risk_label:“PASS”, reason:“No controlled object identified”}.

[0049] It is understandable that each review node in the content review trajectory at this point (e.g., t1, t2, t3, t4) contains two parts: action and observation. The action can be a tool call or step judgment indicated by the review skills executed by the review model, while the observation is the step execution result returned by the tool call or the review environment. Specifically, the review skill corresponding to t1 could be "Image and Text OCR Cleaning SOP", t2 could be "QR Code Traffic Recognition SOP", t3 could be "Identification of Obscure and Inappropriate Product Names SOP", and t4 could be "Risk Aggregation Skill SOP". It is also understandable that the "PASS" indicated by the final review node t4 is the content review label indicated by the current content review trajectory. If the correct label given by the label review is BLOCK, then the current content review trajectory τ={(a1,o1)…(a4,o4)} can be determined as an abnormal review trajectory.

[0050] For further details, please see Figure 4 , Figure 4 This is a schematic diagram illustrating a scenario for obtaining an abnormal audit trajectory, provided in an embodiment of this application. For example... Figure 4 As shown in the embodiment of this application, when an audit model for auditing business audit content (business audit content 31a) is obtained, at least one audit skill (i.e., audit skill 32a, including skill 321a, skill 321b, ..., skill 321n) related to business audit content 31a can be searched from the audit strategy skill library (skill library 30a) configured for the audit model. Further, the audit model with frozen parameters can call audit skill 32a to perform content auditing on business audit content 31a, obtaining at least one audit node (including node 341a, node 341b, node 341c), and determining a content audit tag (audit tag 343a) for business audit content 31a based on the obtained at least one audit node. Then, a content audit trajectory (audit trajectory 34a) for business audit content 31a can be constructed using these at least one audit node and audit tag 343a. Understandably, if review tag 343a is an abnormal review tag, then review trajectory 34a can be identified as an abnormal review trajectory, and skill library 30a can be updated based on this abnormal review trajectory. If review tag 343a is not an abnormal review tag, then review trajectory 34a can be identified as a non-abnormal review trajectory, and this non-abnormal review trajectory can be archived.

[0051] In one possible implementation, the audit strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill. Here, a skill encoding vector is obtained by vector encoding a candidate skill using an encoding model. It is understood that the encoding model can be a pre-trained embedding model used for encoding processing; no further limitations are imposed. It is understood that when searching for at least one audit skill related to business audit content from the audit strategy skill base configured for the audit model, this embodiment can first perform structured processing on the business audit content to obtain structured content. This structured processing can be understood as the parsing unit provided in this embodiment performing a structured transformation of the audit intent into the business audit content, resulting in a structured audit task (structured content). Optionally, the parsing unit can be deployed on the audit model. The extracted structured content includes fixed slots such as task actions, task objects, and task constraints. The task action indicates what the current audit task should do, the object indicates what is being audited, and the constraints indicate which risks to focus on and which evidence must be output.

[0052] For example, if the business review content is: title "Selling idle Bingtang (a type of fruit)," body text "Those who understand, see the picture; large quantity available," and image img_01, then the structured content obtained by structuring the business review content can be: action=MODERATE (execute review); object={content_id:C123, content_type:image and text product post, text:“Selling idle Bingtang…”, images:[img_01]}; constraints={policy_version:P-2026-08, required_risk_types:[prohibited items, external traffic], language:zh-CN, decision_labels:[PASS,REVIEW,BLOCK],evidence_required:true, latency_budget_ms:1500}; where action is the task action of the current task, object is the task object, and constraints are the task constraints. It is understood that after obtaining this intent representation (structured content), the embodiments of this application can use the structured content to initiate a search of the review strategy skill base and find executable skill entries (i.e. review skills) that are semantically appropriate to the current task.

[0053] Specifically, the structured content can be encoded using an encoding model to obtain a content encoding vector. Then, among the multiple skill encoding vectors included in the review strategy skill base, at least one target skill encoding vector similar to the content encoding vector can be found. Optionally, cosine similarity calculation can be used to quickly find a target skill encoding vector similar to the content encoding vector from multiple skill encoding vectors. Specifically, the cosine similarity calculation process can be shown in formula (2):

[0054] As shown in formula (2), The structured content represents the current business review content, while s represents candidate skills in the review strategy skill library. This application embodiment can encode the structured content into a content encoding vector. Skill encoding vectors whose cosine similarity reaches a similarity threshold (e.g., 0.45, without specific limitation here) The top-k_ret ​​skill encoding vectors with the highest cosine similarity (e.g., k_ret ​​can be 10) are selected as target skill encoding vectors. Then, the candidate skills corresponding to each of these target skill encoding vectors are identified as target candidate skills, resulting in at least one target candidate skill. It can be understood that this at least one target candidate skill is the top-k_ret ​​candidate skill retrieved from the review strategy skill library that is similar to the business review content.

[0055] Optionally, embodiments of this application can recall at least one target candidate skill from the review strategy library using Approximate Nearest Neighbor Search (ANNS). ANNS is an algorithmic paradigm designed to address the computational bottleneck of exact nearest neighbor search in large-scale, high-dimensional data. ANNS sacrifices a certain degree of precision tolerance (i.e., allows returned points not to be globally optimal, but only probabilistically or geographically close) to achieve exponential performance improvements in retrieval speed, memory usage, and computational resources, thereby achieving millisecond-level real-time similarity matching in massive datasets (such as trillions of vectors). In other words, embodiments of this application can use Approximate Nearest Neighbor Search to find at least one target skill encoding vector similar to the content encoding vector from multiple skill encoding vectors contained in the review strategy library, and then determine the candidate skill corresponding to this at least one target candidate skill as the target candidate skill. Furthermore, embodiments of this application can determine at least one review skill related to the business review content based on the at least one target candidate skill obtained.

[0056] In one possible implementation, when determining at least one review skill related to business review content based on at least one target candidate skill, this embodiment of the application can perform content concatenation processing with the skill content of each of the at least one target candidate skill to obtain concatenated content for each target candidate skill. Then, a semantic analysis model can be invoked to calculate the semantic similarity of the concatenated content for each target candidate skill to obtain the semantic similarity of each target candidate skill to the business review content. Furthermore, when a target candidate skill whose semantic similarity to the business review content reaches a similarity threshold is determined as a review skill, at least one review skill related to the business review content is obtained. In other words, after obtaining at least one target candidate skill, this embodiment of the application can concatenate the intent text (structured content) with the skill description (skill content) of each target candidate skill, and then input the concatenated content (the concatenated content of each target candidate skill) into a semantic analysis model (e.g., a cross-encoder) for deep semantic interaction scoring (semantic similarity calculation), and finally select the Top-K skills (at least one review skill) according to the scores.

[0057] It is understood that, in obtaining at least one review skill related to business review content, the embodiments of this application can use a two-stage link of vector recall and cross-encoder reordering. In the vector recall stage, the pre-trained language model (encoding model) encodes the user intent (structured content of business review content) and skill description (candidate skills in the review strategy skill library) into fixed-length vectors respectively. Then, the candidate skill set (i.e., at least one target candidate skill) is quickly delineated by approximate nearest neighbor search or cosine similarity. Then, the reordering stage can be entered, where the intent text is combined with the skill description of the target candidate skill, and the semantic analysis model (e.g., cross-encoder) performs deep semantic interaction scoring (semantic similarity calculation), and finally, the Top-K skills (at least one review skill) are selected according to the score. The specific process of obtaining this at least one review skill can be shown in formula (3):

[0058] As shown in formula (3), For structured content, s represents candidate skills in the review strategy skill base. In this embodiment, at least one target candidate skill can be recalled from the review strategy skill base, and then a semantic analysis model (i.e., ...) can be used to... Semantic similarity calculation and ranking are performed on at least one target candidate skill and structured content to obtain at least one review skill. It is understood that, through a two-stage process of vector recall and reordering, this application embodiment can quickly and accurately find at least one review skill for content review of business review content from multiple candidate skills contained in the review strategy skill library. Optionally, this application embodiment can set a timeout budget (e.g., 200ms) for the skill retrieval process. When the time taken to recall at least one review skill is close to the threshold, the system can skip the reordering stage and directly determine the recall result (target candidate skill) as the at least one review skill to be used, so as to ensure the stability of the delay in responding to business review content.

[0059] For further details, please see Figure 5 , Figure 5 This is a schematic diagram illustrating a scenario for acquiring auditing skills, as provided in an embodiment of this application. Figure 5 As shown, this embodiment of the application can perform structured processing on the business review content (business review content 35a) to obtain structured content (structured content 35b), and call the encoding model to encode the structured content 35b to obtain the content encoding vector (content encoding vector 36a). Further, in the review strategy skill base (skill base 37a), which includes multiple skill encoding vectors (including skill encoding vectors 371a, 371b, 371c, 371d, ..., 371n), at least one target skill encoding vector (including skill encoding vectors 371b, 371c, and 371d) similar to the content encoding vector can be searched. It is understood that each skill encoding vector in skill base 37a corresponds to a candidate skill (including candidate skill 372a, candidate skill 372b, candidate skill 372c, candidate skill 372d, ..., candidate skill 372n). In this embodiment of the application, the candidate skill corresponding to each target skill encoding vector in the obtained at least one target skill encoding vector can be determined as a target candidate skill to obtain at least one target candidate skill (including candidate skill 372b, candidate skill 372c, and candidate skill 372d).

[0060] Furthermore, in this embodiment of the application, the structured content 35b can be concatenated with the skill content of each of the at least one target candidate skill to obtain concatenated content for each target candidate skill. Then, a semantic analysis model can be invoked to calculate the semantic similarity of the concatenated content for each target candidate skill, obtaining the semantic similarity of each target candidate skill to the business review content 35a, such as... Figure 5As shown, the semantic similarity between candidate skill 372b and business review content 35a is 0.5, the semantic similarity between candidate skill 372c and business review content 35a is 0.7, and the semantic similarity between candidate skill 372d and business review content 35a is 0.8. Furthermore, at least one target candidate skill whose semantic similarity to the business review content reaches a similarity threshold (the top K target candidate skills in semantic similarity, where K is 2) can be identified as a review skill, thus obtaining at least one review skill (including candidate skill 372c and candidate skill 372d) related to the business review content.

[0061] Step S102: Based on the business review content, abnormal review trajectory and at least one review skill, locate the first optimizable target review node related to the abnormal review tag in at least one review node, and output the node optimization description content for the target review node. It is understood that, after obtaining the business review content, the abnormal review trajectory, and at least one review skill, this application embodiment can locate the first optimizable target review node related to the abnormal review tag in at least one review node based on the business review content, the abnormal review trajectory, and at least one review skill, and output the node optimization description content for the target review node. In other words, after receiving the abnormal review trajectory, the dynamic adapter of the review strategy provided in this application embodiment can trace back along the multi-step reasoning link (at least one review node included in the abnormal review trajectory) through the abnormal location unit (Localizer). The core objective of this backtracking process is to find the first operable abnormal node (i.e., the target review node) in the link. The operability here indicates that the abnormality in the target review node is not a simple semantic deviation, but a reasoning defect with a clear space for correction. For example, there is a misunderstanding of the applicable conditions of a certain review standard operating procedure (SOP), a omitting of necessary verification steps, or a logical jump when processing the mapping of variants of illegal words. It is understood that the positioning unit in this application embodiment will eventually output a specific step number to indicate the target review node, as well as an improvement suggestion for the target review node (i.e., node optimization description content).

[0062] In one possible implementation, when this embodiment locates the first optimizable target review node related to the abnormal review tag within at least one review node based on business review content, abnormal review trajectory, and at least one review skill, and outputs node optimization description content for the target review node, it can first obtain an anomaly analysis model for anomaly analysis. Then, based on the anomaly location template indicated by the anomaly analysis model, the abnormal review trajectory, business review content, and at least one review skill, prompt words are generated to obtain anomaly location prompt words. It is understood that the anomaly analysis model here can be a trained large language model; optionally, the anomaly analysis model and the review model can be the same model or different models, which is not limited here.

[0063] Furthermore, in this embodiment, anomaly location prompts can be input into anomaly analysis models. Under the constraints of these prompts, the model performs backtracking analysis on at least one review node in the abnormal review trajectory, obtaining backtracking analysis results for the abnormal review trajectory. Based on the backtracking analysis evidence present in these results, the first optimizable target review node related to the abnormal review tag can be located within at least one review node. For example, in the abnormal review trajectory shown below: t1: Action a1=OCR(image=img_01, model=ocr_v3); Observation o1={text:“300 / copy, scan to chat”, confidence:0.94}. t2: Action a2=DecodeQR(image=img_01); Observation o2={url:“wx-contact: / / …”, type:“External contact information”}. t3: Action a3 = NormalizeAndMapAlias(term = "Bingtang", dict = 2026-08); Observation o3 = {normalized: "Bingtang", matched_entity: null, similarity: 0.61}. t4: Action a4 = AggregateRisk(evidence = [o1, o2, o3]); Observation o4 = {risk_label: "PASS", reason: "No controlled object identified"}. When performing anomaly localization, the anomaly analysis model can trace back from t4 to determine that the text extraction in t1 was correct, the QR code recognition in t2 was correct, but t3 failed to recognize "Bingtang" as a homophone variant of the target's name, resulting in t4 lacking object evidence. If t3 is corrected, combined with price and QR code evidence, the final conclusion will be changed from PASS to BLOCK. Therefore, the target review node at this point is t3.

[0064] Furthermore, the anomaly analysis model can identify the node action indicated by the target review node as an abnormal action, and generate an anomaly description and action optimization content for the abnormal action. Then, based on the anomaly description and action optimization content, it outputs an optimized node description for the target review node. In other words, the optimized node description output by the anomaly analysis model includes a description of the abnormal behavior (abnormal action) in the target review node, as well as corrective suggestions for the abnormal action. Specifically, the anomaly localization process performed by the anomaly localization unit in this embodiment can be shown in formula (4):

[0065] In formula (4), the anomaly localization unit receives the failed trajectory. The business audit content q and the injected skill subset (at least one audit skill) Then, an actionable fault step (i.e., the target audit node) can be output. Understandably, errors in the target review node need to be clearly correctable. This could include misinterpreting the applicable conditions of a review rule, omitting necessary verification steps, or logical jumps in the mapping of violating word variants. Simultaneously, the anomaly localization unit can also output optimized descriptions of the target review node. It is understandable that here This can be a natural language text describing the failed behavior and improvement suggestions for the target review node. Taking the example of a review agent (i.e., the review model) misjudging a normal commercial advertisement as illegal content, the anomaly localization unit will find that the image and text extraction in step 1 is correct. The real error lies in step 2, the determination of the illegal product attributes—step 4 of a certain "specific illegal product determination skill" has a logical jump. Therefore, the review node corresponding to step 2 can be identified as the target review node.

[0066] It is understood that, by locating the target audit node of the abnormal audit trajectory in this embodiment of the application, and then revising the audit strategy skill base based on the target audit node, the update of the audit strategy skill base can be narrowed down to precise modifications related to the target audit node, while modifications unrelated to the target audit node can be made to the audit strategy skill base. This can effectively improve the accuracy of updating the audit strategy.

[0067] For further details, please see Figure 6 , Figure 6 This is a schematic diagram illustrating a scenario for abnormal node localization provided in an embodiment of this application. For example... Figure 6As shown, this embodiment of the application can obtain an anomaly analysis model (anomaly analysis model 44a) for anomaly analysis. Based on the anomaly location template indicated by the anomaly analysis model 44a, the anomaly review trajectory (anomaly review trajectory 42a), the business review content (business review content 40a), and at least one review skill (i.e., review skill 41a), prompt words are generated to obtain anomaly location prompt words (anomaly location prompt words 43a). Then, the anomaly location prompt words 43a can be input into the anomaly analysis model 44a. Under the constraints of the anomaly location prompt words 43a, the anomaly analysis model 44a performs node backtracking analysis on at least one review node (including nodes 421a, 421b, and 421c) in the anomaly review trajectory 42a. Based on the backtracking analysis evidence present in the obtained backtracking analysis results, the first optimizable target review node (i.e., node 421b) related to the anomaly review label (review label 422a) is located among at least one review node. Furthermore, the anomaly analysis model 44a can identify the node execution action 451a indicated by the target audit node (including node execution action 451a and node execution result 452a) as an abnormal action, and generate anomaly description content (anomaly description content 461a) and action optimization content (action optimization content 462a) for the abnormal action, and output node optimization description content (node ​​optimization description content 46a) for the target audit node based on the anomaly description content 461a and action optimization content 462a.

[0068] Step S103: Based on the business review content, abnormal review trajectory, at least one review skill, and target review node, perform a weight evaluation on each review skill used to influence the target review node to obtain the skill influence weight of each review skill. It is understood that, after determining the target audit node in the abnormal audit trajectory, the responsibility tracing unit in the dynamic adapter of the audit strategy provided in this application embodiment can perform a weight evaluation on each audit skill that affects the target audit node based on the business audit content, the abnormal audit trajectory, at least one audit skill, and the target audit node, to obtain the skill influence weight of each audit skill. It is understood that the skill influence weight here can be used to describe the abnormal correlation of each audit skill with respect to the target audit node (or the degree of responsibility for causing the target audit node to be abnormal). For the specific process of the responsibility tracing unit calculating the skill influence weight, please refer to formula (5):

[0069] As shown in formula (5), the responsibility traceability unit (Link) can be based on the current business audit content. Abnormal review trajectory and at least one auditing skill and target review nodes Calculate each audit skill (e.g., audit skill) The skill affects the weight (e.g., weight). It is understandable that the acquired skills influence the weighting. The larger the value, the higher the skill level. The greater the responsibility for anomalies in the target audit node, the heavier the liability.

[0070] In one possible implementation, this embodiment of the application, when weighting each review skill that influences the target review node based on business review content, abnormal review trajectory, at least one review skill, and the target review node, to obtain the skill influence weight of each review skill, firstly, an anomaly analysis model for anomaly analysis can be obtained. Then, prompt words are generated using the weight evaluation template indicated by the anomaly analysis model, the abnormal review trajectory, business review content, at least one review skill, and the target review node, resulting in weight evaluation prompt words. Next, the weight evaluation prompt words can be input into the anomaly analysis model. The anomaly analysis model then obtains a weight evaluation strategy based on the prompt words and evaluates the influence weight of each of the at least one review skill on the target review node according to the weight evaluation strategy. The evaluated influence weight of each review skill is then used as the skill influence weight of each review skill. Optionally, the weight evaluation template indicated by the anomaly analysis model can be the same fault analysis prompt template as the anomaly location template used for anomaly location. The two can be distinguished only in terms of output format. In this way, when the anomaly analysis model is called multiple times for anomaly location and weight evaluation, the delay caused by calling the base model (anomaly analysis model) multiple times can be effectively reduced, and the update speed of the strategy update in this application embodiment can be improved.

[0071] It is understood that the weight evaluation strategy indicated by the anomaly analysis model here is a strategy for determining how to assign a weight to the skill influence of each review skill, or in other words, the basis for weight judgment. For example, it may include: whether the skill text directly misleads the judgment of the target review node, whether the skill text omits key triggering conditions, and whether the skill text provides operational guidance that conflicts with the context. It is understood that by calculating the skill influence weight of each review skill for the target review node, this application embodiment can identify at least one review skill with a high degree of correlation with the target review node. Subsequently, when updating the review strategy skill library, skills with a high degree of correlation with the target review node in the review strategy skill library can be updated, effectively improving the accuracy of review strategy updates.

[0072] For further details, please see Figure 7 , Figure 7 This is a schematic diagram illustrating a scenario for obtaining skill influence weights according to an embodiment of this application. For example... Figure 7 As shown, this application embodiment can obtain an anomaly analysis model (anomaly analysis model 44a) for anomaly analysis. The weight evaluation prompt word (weight evaluation prompt word 43b) is generated by using the weight evaluation template, anomaly review trajectory (anomaly review trajectory 42a), business review content (business review content 40a), at least one review skill (i.e., review skill 41a) and target review node indicated by the anomaly analysis model 44a. Then, the weight assessment prompt 43b can be input into the anomaly analysis model 44a. The anomaly analysis model 44a obtains the weight assessment strategy for weight assessment based on the weight assessment prompt 43b, and assesses the influence weight of each of the at least one review skill (i.e., review skill 41a, including skill 411a, skill 411b, ..., skill 411n) on the target review node according to the weight assessment strategy. Then, the skill influence weight of each review skill (i.e., influence weight 48a, including the weight 481a corresponding to skill 411a, the weight 481b corresponding to skill 411b, ..., the weight 481n corresponding to skill 411n) can be obtained.

[0073] Step S104: Based on the skill influence weight and node optimization description of each review skill, update the review strategy skill library to obtain the review strategy update skill library for reconfiguration to the review model; It is understood that, after outputting the skill influence weight of each review skill, the responsibility tracing unit in this application embodiment can output a corresponding operation type based on the skill influence weight of each review skill. REVISE indicates that the node anomaly of the target review node is mainly related to at least one of the currently used review skills (referred to as the influencing review skill). A skill revision process will then be implemented to revise the influencing review skill and replace it with a candidate skill from the review strategy skill library. GENERATE indicates that at least one current review skill is not highly correlated with the node anomaly of the target review node. In other words, the existing review strategy skill library cannot cover the new types of violations involved in the current business review content. Therefore, a skill generation process can be implemented, specifically adding new skills to the review strategy skill library to resolve the node anomaly of the target review node.

[0074] Optionally, the accountability unit can attribute the node anomaly of the target audit node based on the maximum influence weight among the skill influence weights of each audit skill obtained so far, and determine whether to enter the revision process or the generation process. For example, if the maximum influence weight does not reach the weight threshold (e.g., 0.7), it is considered that the correlation between each audit skill and the node anomaly of the target audit node is not high, and the process enters the generation process; if the maximum influence weight reaches the weight threshold (e.g., 0.7), it is considered that the audit skill corresponding to the maximum influence weight has a high correlation with the node anomaly of the target audit node, and the process enters the revision process. Specifically, the embodiments of this application are based on the operation type output by the accountability unit. For the subsequent process of updating the audit strategy library, please refer to formula (6):

[0075] It is understandable that in formula (6), This is the skill set for review strategies prior to the update. For the target review node, The operation type output by the responsibility traceability unit. Optimize the node descriptions output by the anomaly localization unit. It is understandable that when... When set to =REVISE, Modify can rewrite the skill with the highest weight (the one that has the greatest impact on skill weight). And replace the skills in the audit strategy skill set. When When =GENERATE is used, Modify can be obtained from Extract reusable behavior patterns from the context (i.e., the node content corresponding to the target review node), synthesize a new general skill, and add it to K. The new skill generated here contains clear triggering conditions and execution steps.

[0076] In other words, after obtaining at least one audit skill actually invoked by the current audit model during the inference phase and the target audit node output by the anomaly localization unit, the responsibility tracing unit in this embodiment does not make indiscriminate modifications to the entire audit skill library, nor does it modify whichever skill is used. Instead, it calculates the responsibility weight (skill influence weight) of each audit skill for the target audit node and sorts them according to the obtained responsibility weights for each sub-skill. It is understandable that the responsibility tracing unit bases its weight estimation on several factors, including whether the skill text directly misleads the judgment of that step, whether key triggering conditions are omitted, and whether operational instructions conflict with the context. It is also understandable that the obtained responsibility weights can be used to attribute errors to the target audit node, attributing the error to a specific audit skill or pointing to a missing skill gap.

[0077] Specifically, if an error is attributed to a flaw in an existing skill, the process proceeds to revision; if the existing skill set cannot cover the new types of violations encountered, the process proceeds to generation. In subsequent updates to the review strategy, the underlying large language model (review model) remains frozen (parameters are not optimized or modified), and only review strategy skills that use text for content storage are updated. In the revision process, erroneous skills can be minimally edited: supplementing negative examples that trigger misjudgments, or adding pre-verification conditions to the step descriptions, while the core logic of the rule remains unchanged. In the generation process, a new general skill can be summarized based on the characteristics of new violation methods, clarifying its triggering conditions and execution steps.

[0078] Understandably, this method of accountability and strategy update effectively transforms a general error in the review strategy into a deviation in the guidance of a specific skill, thereby reducing the scope of modifications to the review strategy and improving the accuracy and efficiency of strategy updates. By keeping the review model frozen throughout its operation without parameter fine-tuning, all updates to the review strategy fall within the text-based review strategy skill library: revision paths undergo minimal editing, generation paths add new rules, and the weights of the underlying large language model remain unchanged. This effectively prevents catastrophic forgetting caused by frequent fine-tuning, avoiding misjudging correctly handled cases due to fixing individual bad cases, thus preventing a decrease in overall accuracy. Simultaneously, computational costs are significantly reduced; adding, deleting, modifying, and querying the skill library can be completed solely through text generation and vector retrieval, effectively increasing the speed of review strategy updates.

[0079] In one possible implementation, when updating the review strategy skill library based on the skill influence weight and node optimization description content of each review skill to obtain the review strategy update skill library for reconfiguration to the review model, this application embodiment first obtains the maximum influence weight among the skill influence weights of each review skill, and performs weight analysis on the maximum influence weight to obtain the weight analysis result. If the weight analysis result indicates that the maximum influence weight has reached the influence weight threshold, then the revision process can proceed. Specifically, the influencing review skill corresponding to the maximum influence weight can be determined among at least one review skill; in other words, the review skill with the largest skill influence weight among at least one review skill can be determined as the influencing review skill.

[0080] Furthermore, this embodiment of the application can obtain the skill revision template indicated by the skill revision model, and generate prompt words by using the skill revision template, business review content, influencing review skills, target review nodes, and node optimization descriptions to obtain skill revision prompt words. Then, the skill revision prompt words can be input into the skill revision model, which will revise the influencing review skills based on the prompt words to obtain revised review skills. Optionally, the revision strategy used by the skill revision model when revising skills can be an addition-only strategy. Specifically, it can include: retaining the correct main logic in the original skill, only supplementing negative samples that trigger misjudgments, or adding pre-verification conditions in the step description. No further examples or limitations are provided for the revision strategies used here. It is understood that this embodiment of the application can update the review strategy skill library using the obtained revised review skills to obtain an updated review strategy skill library for reconfiguration to the review model. It is understood that this embodiment of the application can improve the accuracy and effectiveness of updating review strategies by revising the influencing review skills related to the target review node in at least one review node and updating the review strategy skill library with the obtained revised review skills.

[0081] In one possible implementation, the audit strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill. A skill encoding vector is obtained by vector encoding a candidate skill using an encoding model. In this embodiment, when updating the audit strategy skill base by revising audit skills to obtain an updated audit strategy skill base for reconfiguration to the audit model, redundancy control can first be implemented on the revised audit skills. Specifically, the encoding model can be called to encode the revised audit skills to obtain revised encoding vectors. The revised encoding vectors are then compared with the multiple skill encoding vectors included in the audit strategy skill base to calculate the first vector similarity of each skill encoding vector relative to the audit encoding vector. Optionally, the vector similarity calculation performed here can be a cosine similarity calculation, and the specific calculation process can be shown in formula (7):

[0082] As shown in formula (7), For current revision review skills, To review candidate skills in the strategy library, the coding model The revision review skills and candidate skills here can be encoded to obtain revision code vectors. and skill encoding vector Then, the cosine similarity between the two can be calculated as shown in formula (7) to obtain the first vector similarity between the skill encoding vector and the audit encoding vector. .

[0083] Furthermore, embodiments of this application can search among the obtained multiple first vector similarities to find those that reach a similarity threshold (which can be denoted as...). For example, a first target similarity of 0.95 can be used to obtain the first search result. Optionally, if the first search result indicates that there is a first target similarity among multiple first vector similarities, it means that there are candidate skills in the review strategy skill base that have a similarity to the current revised review skills exceeding the threshold. To avoid redundant updates to the review strategy skill base leading to storage expansion, affecting the retrieval accuracy of the review model when performing review skills retrieval, and lengthening the context window during inference due to redundant text skills, the revised review skills can be discarded here. In other words, the revised review skills are not used to update the review strategy skill base to ensure the conciseness of the review strategy skill base.

[0084] Optionally, if the first search result indicates that there is no first target similarity among multiple first vector similarities, then the candidate skill in the review strategy skill library that matches the influencing review skill can be replaced with the revised review skill to obtain the first updated skill library. It is understood that the candidate skill that matches the influencing review skill here is the candidate skill in the review strategy skill library that has the same skill identifier as the influencing review skill. It is understood that in this embodiment of the application, when there is no candidate skill in the review strategy skill library similar to the revised review skill, the review strategy skill library is updated by revising the review skill. This is equivalent to configuring a deduplication gate (noise reduction filtering based on semantic similarity) for the review strategy skill library, which can effectively avoid redundant updates to the review strategy skill library, thereby improving the update efficiency of review strategy updates.

[0085] It is understood that, after obtaining the first updated skill library, this embodiment of the application can test the first updated skill library through a sandbox verification admission method. Specifically, the first updated skill library can be tested through the audit strategy skill library, and the first updated skill library that passes the test is determined as the audit strategy update skill library for reconfiguration to the audit model. In other words, this embodiment of the application will launch the first updated skills that have passed the sandbox verification admission, that is, reconfigure them to the audit model, so that the audit model can use the audit strategy update skill library here to audit newly received business content.

[0086] For further details, please see Figure 8 , Figure 8 This is a schematic diagram illustrating a scenario where an audit strategy is updated through a revision process, as provided in an embodiment of this application. Figure 8As shown, in this embodiment of the application, after calculating the skill influence weight (i.e., influence weight 51a, including weight 511a corresponding to skill 501a, weight 511b corresponding to skill 501b, ..., weight 511n corresponding to skill 501n) of each audit skill (i.e., audit skill 50a, including skill 501a, skill 501b, ..., skill 501n) for the target audit node, the maximum influence weight (i.e., weight 511b) of each audit skill's skill influence weight can be obtained, and weight analysis can be performed on the maximum influence weight to obtain the weight analysis result. If the weight analysis result indicates that the maximum influence weight has not reached the influence weight threshold, the embodiment of the application can enter the generation process, add skills to the current audit strategy skill library (skill library 52a) through the newly generated skills to obtain an updated skill library 53a, and perform sandbox admission testing on the obtained updated skill library 53a to determine the updated skill library 53a that passes the test as an audit strategy updated skill library that can be reconfigured to the audit model.

[0087] Optionally, if the weight analysis results indicate that the maximum impact weight has not reached the impact weight threshold, the impact review skill (i.e., skill 501b) corresponding to the maximum impact weight can be determined among at least one review skill. Then, the skill revision model can revise skill 501b based on the target review node and the node optimization description content to obtain the revised review skill (revised review skill 54a). Further, in this embodiment, the encoding model can be called to encode the revised review skill 54a to obtain a revised encoding vector (revised encoding vector 55a). Further, the skill library 52a can be searched for a skill encoding vector similar to the revised encoding vector 55a. If one exists, the current test update is stopped. If no similar skill encoding vector exists in the skill library 52a, the candidate skill matching the impact review skill (skill 501b) in the skill library 52a can be replaced with the revised review skill 54a to obtain the first updated skill library (updated skill library 53b). Furthermore, a sandbox access test can be conducted on the updated skill library 53b, that is, the updated skill library 53b can be tested using skill library 52a, and the updated skill library 53b that passes the test can be identified as the updated skill library for reconfiguration to the audit model.

[0088] In one possible implementation, this embodiment of the application, when testing a first updated skill library using an audit strategy skill library and determining the first updated skill library that passes the test as the audit strategy update skill library for reconfiguration to the audit model, can perform content auditing on business audit content using the first updated skill library to obtain updated audit tags. When the updated audit tag is a non-abnormal audit tag, the test sample content indicated by the skill library testing strategy is obtained. Specifically, before deploying the first updated skill library, it is first necessary to ensure that the first updated skill library can fix the target Bad Case (i.e., business audit content) (i.e., it can accurately audit the business audit content), and then obtain the test sample content for skill library testing. This test sample content can be content from the golden regression test set, including various representative cases (various types of sample audit cases).

[0089] Furthermore, in this embodiment, the review model can perform content review on the test sample content using the review strategy skill library to obtain a first test review result, and then perform content review on the test sample content using the first updated skill library to obtain a second test review result. It is understood that the specific process of the review model performing content review using the review strategy skill library or the first updated skill library is similar to the specific process of the review model reviewing business review content using the review strategy skill library, and will not be elaborated further here. This can also be understood as a process of performing skill library testing (e.g., double-blind testing) on ​​the review strategy skill library and the first updated skill library respectively using the obtained test sample content, thereby obtaining a first test review result for the review strategy skill library and a second test review result for the first updated skill library.

[0090] Furthermore, in this embodiment, the obtained first test audit result can be used to analyze the second test audit result to obtain a test analysis result. It is understood that if the test analysis result indicates that the second test audit result meets the test pass conditions indicated by the skill library testing strategy, then this embodiment can determine the first updated skill library as the audit strategy update skill library. Optionally, the test result analysis of the second test audit result using the first test audit result can be understood as a process of comparing the two results. If the second test audit result is better than the first test audit result (e.g., when both accuracy and recall are higher), it can be considered that the second test audit result meets the test pass conditions indicated by the skill library testing strategy. Therefore, when the first updated skill library that has passed the test is determined as the audit strategy update skill library that can be reconfigured to the audit model, the effectiveness of the audit strategy update skill library can be effectively ensured, thereby ensuring the effectiveness and reliability of the entire strategy update process.

[0091] In one possible implementation, the test sample content used in this embodiment includes sample content under multiple test types. This embodiment analyzes the test results of the second test review result using the first test review result. To obtain the test analysis results, it first obtains the first result index of the first test review result for each of the multiple test types, and then obtains the test weights for each test type indicated by the skill library testing strategy. Further, this embodiment can perform a weighted summation of the obtained first result indexes using the test weights for each test type to obtain a first weighted result index. This first weighted result index can be understood as the overall macro-average F1 score (Macro-F1) corresponding to the first test review result. The F1 score is the harmonic mean determined by comprehensively considering accuracy and recall. When calculating the overall macro-average F1 score of the first test review result, this embodiment can obtain the first result index of the first test review result for each test type (i.e., the F1 score under each test type), and then perform a weighted average of all obtained F1 scores (i.e., the first result indexes) using the test weights for each test type to obtain the first weighted result index for the first test review result.

[0092] Similarly, in this embodiment, the second test review result can be obtained as a second result indicator for each test type. The obtained second result indicators are then weighted and summed using the test weights for each test type to obtain a second weighted result indicator. This second weighted result indicator can be understood as the overall macro average F1 score corresponding to the second test review result. Furthermore, if the obtained second weighted result indicator is greater than or equal to the first weighted result indicator, the second test review result can be determined as meeting the test pass conditions indicated by the skill library testing strategy and thus be considered a test analysis result. In other words, the first updated skill library can then be determined as the skill library updated under the review strategy.

[0093] It is understood that, in the embodiments of this application, when obtaining the updated skill library (e.g., the first updated skill library) obtained by updating the review strategy skill library (e.g., skill replacement), the test sample content (denoted as) in the golden regression test set can be obtained. Then, the original policy library K and the candidate library with the new policy added are used respectively. right Perform a double-blind test and then compare the performance feedback metrics of the two. For the specific comparison process, please refer to formula (8):

[0094] As shown in formula (8), M() is an evaluation metric that compares the test results obtained by reviewing the test sample content using the updated skill library and the review strategy skill library. For example, it can be accuracy, recall, or Macro-F1. It is understandable that when Δ is greater than 0 (if multiple evaluation metrics are used simultaneously, the differences between multiple metrics must all be greater than 0), the second test review result can be considered superior to the first test review result, and thus the skill update library can be updated accordingly. If an update is not approved, it will be merged into the online effective library (i.e., confirmed as part of the skill library for review policy updates); otherwise, the update will be rejected, and K will remain unchanged. This is determined by checking if Δ is greater than 0. Whether to deploy a strategy is essentially a soft switch for incremental performance differential. Under this mechanism, the new strategy needs to prove that it can fix the target bad case without causing overall metric degradation on the regression test set (sample test content). Even if the new strategy performs perfectly on the target case, if it introduces even a slight degradation on other cases, Δ will become negative, leading to the blocking of updates to K. This method can effectively block unstable or overgeneralized strategy changes, avoid online degradation issues caused by strategy updates, and thus ensure the reliability of the strategy update library.

[0095] For further details, please see Figure 9 , Figure 9 This is a schematic diagram illustrating a scenario for testing an updated skill library, as provided in an embodiment of this application. Figure 9 As shown, in this embodiment of the application, after updating the review strategy library (skill library 56a) through a revision process or a generation process, a corresponding updated skill library 56b can be obtained (e.g., a first updated skill library, a second updated skill library). It is understood that in this embodiment of the application, the updated skill library 56b can first be tested using business review content to determine whether the review model can effectively review the business review content using the updated skill library 56b. If it can (i.e., the obtained updated review label is a non-abnormal review label), then the updated skill library 56b can be tested using skill library 56a (sandbox access test).

[0096] Specifically, the review model 59a can review the test sample content through skill library 56a to obtain a first test review result (test review result 57a), and the review model 59a can then review the test sample content through updated skill library 56b to obtain a second test review result (test review result 57b). Further, test result analysis can be performed on the obtained test review results 57a and 57b. Specifically, result metrics 58a (e.g., precision, recall, or a first weighted result metric) and result metrics 58b (e.g., precision, recall, or a second weighted result metric) of test review result 57a can be obtained. Further, the result metrics 58a and 58b can be compared. If result metric 58a is not greater than result metric 58b, then test review result 57b is considered to meet the test pass conditions indicated by the skill library testing strategy. In this case, the review strategy can be updated, and the updated skill library 56b is determined as the review strategy update skill library to be reconfigured for the review model 59a. If result indicator 58a is greater than result indicator 58b, then the test audit result 57b is considered not to meet the test pass conditions indicated by the skill library test strategy. In this case, skill library 56b can be discarded and updated directly without updating the audit strategy.

[0097] It is understood that the embodiments of this application can use an anomaly localization unit to locate anomalies in the abnormal review trajectory, determine the target review node in the abnormal review trajectory, and then use a responsibility backtracking unit to output the node optimization description content related to the target review node and the skill influence weight of each review skill for the target review node. Based on the node optimization description content and the skill influence weight of each review skill, new skills can be generated or old skills can be revised. Redundancy control is then performed on the revised and generated review skills obtained through redundancy control. The revised and generated review skills update the review strategy skill library, resulting in an updated review strategy skill library. This effectively prevents the indiscriminate spread of abnormal signals from abnormal review trajectories to all skills in the review strategy skill library, thereby improving the accuracy and effectiveness of updating the review strategy skill library. Simultaneously, the reliability of the review strategy skill library can be effectively ensured through redundancy control and sandbox control access mechanisms.

[0098] For further details, please see Figure 10 , Figure 10 This is a schematic diagram of another audit strategy update method provided in an embodiment of this application. For example... Figure 10 As shown, it is understandable that Figure 10 The method for updating the audit policy shown can be executed by the server; for example, the server here could be something like... Figure 1 Server 10d is shown. (As shown) Figure 10 As shown, the audit strategy update method may include at least the following: Figure 10 Steps S201 to S209 in the process.

[0099] Step S201: When an audit model for auditing business audit content is obtained, at least one audit skill related to the business audit content is searched from the audit strategy skill library configured for the audit model. It is understood that the review model in this application embodiment can review business review content. Specifically, it can first search for at least one review skill related to the business review content from the review strategy skill library configured for the review model, and call this at least one review skill to review the business review content.

[0100] Optionally, in the initialization phase of the review strategy skill library of the review model in this embodiment, it can be obtained by summarizing successful content review trajectories. Specifically, the initial sample content indicated by the skill generation strategy can be obtained, and the review model can be called to review the initial sample content to obtain the sample review trajectory. It is understood that at this time, the review model does not need any external skill library when reviewing the initial sample content. The underlying LLM (review model) first performs a round of review reasoning on the task set (initial sample content) without injecting any skills or experience. Further, when the sample review label for the initial sample content carried by the sample review trajectory is a non-abnormal review label (i.e., the sample review trajectory can be regarded as a successful review trajectory), the trajectory skills can be summarized according to the skill generation strategy to obtain at least one summarized trajectory skill. Optionally, the trajectory skills summary of the sample review trajectory according to the skill generation strategy can be one or more of the following methods: manual rules, historical review specifications, or LLM induction, to summarize the successful experience in the sample review trajectory into initial skills that can be reused (i.e., summarized trajectory skills).

[0101] Furthermore, each summary trajectory skill in at least one summary trajectory skill can be identified as a candidate skill, and a review strategy skill library can be constructed using the obtained candidate skills. This review strategy skill library is then configured to the review model. It is understood that the review strategy skill library obtained at this point can be repeatedly called by the review model for content review, or iteratively updated when abnormal review trajectories are obtained. All subsequent content review and strategy updates are performed with the support of this review strategy skill library. Failure trajectories continuously input into the review strategy dynamic adapter for a closed-loop update process of anomaly attribution, strategy modification, and strategy verification, driving the iterative evolution of the review strategy skill library. Optionally, in this embodiment, when iteratively updating the skill library for the same batch of failure samples, multiple rounds of iterative updates can be performed using these failure samples. When the number of iterations or several consecutive rounds of no effective updates occur, the strategy update process based on the current batch of failure samples is terminated, and a new iterative update of the review strategy skill library is triggered again when a new failure sample arrives.

[0102] For further details, please see Figure 11 , Figure 11 This is a schematic diagram illustrating a scenario for constructing an audit strategy skill library, as provided in an embodiment of this application. Figure 11 As shown, this embodiment of the application can obtain the initial sample content indicated by the skill generation strategy (i.e., initial sample content 60a, including sample content 601a, sample content 601b, sample content 601c, and sample content 601d), call the review model to perform content review on the initial sample content 60a, and obtain the sample review trajectory (i.e., sample review trajectory 61a, including review trajectory 611a corresponding to sample content 601a, review trajectory 611b corresponding to sample content 601b, review trajectory 611c corresponding to sample content 601c, and review trajectory 611d corresponding to sample content 601d). Then, non-abnormal trajectories in the sample review trajectory 61a can be obtained (carrying sample review tags that are non-abnormal review tags, including review trajectories 611a, 611b, and 611c). Furthermore, this embodiment of the application can summarize the trajectory skills of the non-abnormal trajectories according to the skill generation strategy to obtain at least one summarized trajectory skill (including skill 621a, skill 621b, ..., skill 621n). Furthermore, each summary trajectory skill in at least one summary trajectory skill can be identified as a candidate skill, and an audit strategy skill library (skill library 63a) can be constructed using the obtained at least one candidate skill, and the skill library 63a can be configured to the audit model.

[0103] Step S202: The review model calls at least one review skill to review the business review content, obtains at least one review node, and determines the content review tag for the business review content based on at least one review node; It is understood that the embodiments of this application can use the review model to call at least one retrieved review skill to perform content review on business review content, obtain at least one review node, and determine content review tags for business review content based on at least one review node. Specifically, steps can be generated for each of the at least one review skill to obtain review steps for each review skill, and then a content review link for business review content can be constructed through the review steps corresponding to each review skill. In other words, the content review link at this time contains at least one review step. The content review link here can be understood as a task execution plan obtained by breaking down the review task corresponding to business review content into steps through at least one review skill. Each review step corresponds to a review skill and has corresponding preconditions and expected effects. For example, the obtained content review link can be as follows: Step 1, Extract Image and Text Evidence: Prerequisite is the existence of an image. Use OCR cleaning skills; expected result is to obtain image text with confidence. Step 2, Analyze Traffic Carrier: Prerequisite is the detection of a QR code area. Use QR code traffic generation skills; expected result is to obtain link types. Step 3, Identify Hidden Objects: Prerequisite is the acquisition of standardized text. Use alias recognition skills; expected result is to output candidate controlled entities and matching evidence. Step 4, Comprehensive Judgment: Prerequisite is that the aforementioned evidence has been generated or is clearly unavailable. Use risk aggregation skills; expected result is to output PASS / REVIEW / BLOCK and evidence chains.

[0104] It is understood that, in the embodiments of this application, when the review model performs content review on business review content based on the content review chain, it can output the review node corresponding to each review step in the content review chain. Then, the review node corresponding to each review step can be determined as at least one review node in the content review trajectory, and a content review tag for the business review content can be determined based on at least one review node.

[0105] In one possible implementation, when the review model performs content review on business content based on a content review chain, and outputs the review node corresponding to each review step in at least one review step included in the content review chain, it can select one review step from the at least one review step included in the content review chain as the target review step based on the review status of the review model (i.e., execute the review steps one by one according to the content review chain; here, the review step to be executed can be determined as the target review step), and determine the review skill corresponding to the target review step as the target review skill. The review status here includes the link execution progress of the review model in the content review chain, and also includes the model context (i.e., review evidence) required to execute the review task at the current link node (target review step).

[0106] Specifically, when the audit evidence indicated by the audit status meets the skill usage conditions indicated by the target audit skill, the target audit skill can be invoked to execute the step execution action indicated by the target audit step, and the step execution result for the target audit step can be obtained.

[0107] Specifically, during runtime, the execution engine of the review model can extract the required slot values ​​from the model context by calling the current target review skill. These slot values ​​can be understood as structured evidence fields in the task state of the review model. Different skills require different fields when used. During task execution, the execution engine reads the skill input conditions declared in the skill declaration and then retrieves the corresponding required field data from the task state. In essence, if the required slot values ​​for the target review skill can be filled at this point, the current review state can be considered to meet the skill usage conditions indicated by the target review skill. The task state of the review model can be maintained by the review task state manager, which records the current task execution progress, filled slots, skill call history, and strategy configuration. Optionally, if there is a vacancy in the target review skill slot, the current review process can be suspended. In this embodiment, the evidence missing handling strategy can be triggered. When the required evidence field is missing, the execution engine of the review model can supplement the evidence by triggering tool retry, entering the alternative evidence chain, or performing manual review based on the failed branch declared by the current target review skill. After the evidence is supplemented, the execution can continue from the current target review step.

[0108] It is understood that, in the embodiments of this application, when it is determined that the execution result of the step matches the expected result of the skill indicated by the target audit skill, the step execution action of the audit step is determined as the node execution action, and the obtained step execution result is determined as the node execution result. Therefore, an audit node for the target audit step can be constructed based on the node execution action and the node execution result. It is also understood that, when each audit step in at least one audit step is selected as the target audit step (i.e., each audit step is executed through the audit model), the audit node corresponding to each audit step can be obtained.

[0109] Optionally, if a skill returns a failure signal during the execution of a step in the review model (i.e., the step execution result does not match the expected result indicated by the target review skill), the review model's state manager can perform fallback processing based on a preset degradation strategy. For example, in this embodiment, different parameter combinations under the same skill can be tried first according to the controlled retry parameters and failure branches pre-declared by the target review skill. For instance, when the target review skill is an OCR skill, the skill parameters may include primary_model=ocr_v3, fallback_model=ocr_v2, min_confidence=0.75, and rotation_angles=[0,90,180,270]. It is understood that when the review model first calls the OCR skill, if the confidence level of the result is only 0.42, it can first perform rotation correction on the image in the business review content according to the fallback_policy pre-declared by the skill, and then use fallback_model to retry (i.e., re-perform character recognition). If the process still fails, the embodiments of this application can fall back to the alternative skill path (a pre-declared alternative evidence acquisition path; for example, if the image resolution in the business review content is low, causing OCR failure, the alternative path could be to first perform an image enhancement SOP and then call the OCR skill). It is understood that if the alternative path also fails, the review model can generate review tags and return the generated trajectory to the reviewer for manual review.

[0110] It is understandable that the reasoning process performed by the review model can be applied to relevant content review scenarios, such as: veiled references to prohibited products in product posts; external traffic redirection carried by images or QR codes; word splitting, homophony, and symbol confusion in sensitive text; and scenarios where false positives are likely to occur in normal commercial advertisements, news reports, and popular science content, etc. Further examples and elaborations are omitted here. Optionally, when the review model in this embodiment of the application determines that the currently invoked review skill cannot effectively review the business review content, it can trigger an incremental search to retrieve new review skills from the review strategy library and add the retrieved review skills to the current model context, so that subsequent review processes can call the newly acquired review skills for reasoning. It is understandable that, in order to control reasoning costs, this embodiment of the application can limit the skill invocation depth, that is, a single task is allowed to execute a maximum of N consecutive logical steps (i.e., call N review skills). After exceeding the depth, the current review task will be forcibly terminated, and partial results will be returned to manual review. In this way, the response speed of the review model to review tasks can be effectively improved.

[0111] Step S203: If the content review tag is an abnormal review tag obtained by reviewing the review pass tag or review block tag output by the review model, then the content review trajectory carrying the abnormal review tag is determined as an abnormal review trajectory. It is understood that, in this embodiment of the application, after obtaining at least one review node and content review tag through model inference by the review model, a content review trajectory for business review content can be constructed using at least one review node and content review tag. The review nodes can be sorted based on the execution order of their actions to determine the content review trajectory. Furthermore, when a content review tag is determined to be an abnormal review tag, the content review trajectory can be identified as an abnormal review trajectory. Optionally, in this embodiment of the application, the sampling temperature of the review model during content review can be set to 0. Here, the sampling temperature is a decoding parameter in the model generation stage. Setting the sampling temperature to 0 (i.e., using a greedy decoding method) indicates that the highest probability output should be selected as much as possible to reduce the randomness of repeated runs, thereby improving the output stability and reproducibility of the review model. For the specific processes of steps S201 and S203, please refer to the above. Figure 3 The description of step S101 in the corresponding embodiments will not be repeated here.

[0112] Step S204: Based on the business review content, abnormal review trajectory and at least one review skill, locate the first optimizable target review node related to the abnormal review tag in at least one review node, and output the node optimization description content for the target review node. Step S205: Based on the business review content, abnormal review trajectory, at least one review skill, and target review node, perform a weight evaluation on each review skill used to influence the target review node to obtain the skill influence weight of each review skill. It is understood that, after obtaining the business review content, the abnormal review trajectory, and at least one review skill, the embodiments of this application can, based on the business review content, the abnormal review trajectory, and at least one review skill, locate the first optimizable target review node related to the abnormal review tag among at least one review node, and output the node optimization description content for the target review node. In other words, after receiving the abnormal review trajectory, the dynamic adapter of the review strategy provided in the embodiments of this application can trace back along the multi-step inference link (at least one review node included in the abnormal review trajectory) through the abnormal location unit (Localizer). The core objective of this backtracking process is to find the first operable abnormal node (i.e., the target review node) in the link.

[0113] Furthermore, after determining the target audit node in the abnormal audit trajectory, the responsibility tracing unit in the dynamic adapter of the audit strategy provided in this application embodiment can perform weight evaluation on each audit skill that affects the target audit node based on the business audit content, the abnormal audit trajectory, at least one audit skill, and the target audit node, thereby obtaining the skill influence weight of each audit skill. It is understood that by performing weight evaluation through the responsibility tracing unit, the intermediate states (i.e., audit nodes in the abnormal audit trajectory) in the multi-step inference chain generated by the audit model can be explicitly determined, and responsibility weights can be calculated. This provides a systematic traceability path for auditing the behavior of the audit agent. The responsibility tracing unit outputs a weight value for each invoked audit skill. When it is necessary to review the behavior of the audit agent in a certain task, these weight values ​​and the corresponding skill invocation records together constitute a traceable link, allowing tracing back from the final conclusion to the specific decision node. In specific business scenarios, this traceable link can serve as an automated tool for internal auditing, effectively improving audit efficiency and speed compared to the manual process of reviewing multi-turn dialogue logs.

[0114] Step S206: Obtain the maximum influence weight among the skill influence weights of each audit skill, and perform weight analysis on the maximum influence weight to obtain the weight analysis results; Step S207: If the weight analysis result indicates that the maximum influence weight has not reached the influence weight threshold, then obtain the skill generation template indicated by the skill generation model, and generate prompt words by using the skill generation template, business review content, target review node and node optimization description content to obtain skill generation prompt words; Step S208: Input the skill generation prompts into the skill generation model, and the skill generation model generates the missing skills for the target review node based on the skill generation prompts to obtain the generated review skills; It is understood that, in this embodiment of the application, after determining the skill influence weight of each review skill through the responsibility tracing unit, a weight analysis can be performed on the maximum influence weight among the obtained skill influence weights to obtain the weight analysis result. If the weight analysis result indicates that the maximum influence weight has not reached the influence weight threshold, it indicates that at least one review skill used for content review has a low correlation with the target review node. It is understood that, in the responsibility tracing stage of this embodiment of the application, the failure of abnormal review trajectories will be treated in two categories: existing skill defects (Skill_Wrong) and skill omissions (Skill_Missing). It is understood that, when a skill defect occurs, a skill revision process can be initiated, the specific process of which can be referred to the above. Figure 3 Step S104 in the corresponding embodiment.

[0115] Optionally, if this application embodiment identifies an erroneous node (target review node) and confirms that the current violation pattern exceeds the existing coverage of the skill library (at which point the maximum impact weight has not reached the impact weight threshold), it can be determined that a skill is missing. In this case, the skill generation process can be initiated, using the observational evidence exposed in the failure trajectory to generate a new general SOP (generate review skill). The new SOP will list the triggering conditions, execution steps, and verification points one by one, and then write them into the current skill library.

[0116] Specifically, in this embodiment of the application, when entering the generation process, a skill generation model can be invoked to generate skills. First, the skill generation template indicated by the skill generation model can be obtained. Then, prompt words are generated using the skill generation template, business review content, target review node, and node optimization description, resulting in skill generation prompt words. These prompt words are then input into the skill generation model, which generates the missing skills for the target review node based on the prompt words, thus generating the review skill.

[0117] Step S209: By generating an audit skills update audit strategy skills library, an audit strategy update skills library is obtained for reconfiguration to the audit model; It is understood that, after obtaining the generated review skills, this application embodiment can update the review strategy skill library by generating review skills, resulting in an updated review strategy skill library for reconfiguration to the review model. Optionally, the review skills obtained and generated in this application embodiment are stored in natural language, which can effectively improve the readability of the review skills and facilitate manual review and verification. It is understood that, regardless of whether the generated skills are revisions or additions, the original structured form of the skills is well preserved, including titles, applicable conditions, execution steps, and negative sample descriptions. This makes the review strategy skill library a white-box format, presenting reviewers with interpretable text rules that are easy to read and modify, greatly improving the efficiency of manual review and revision. It is understood that the embodiments of this application can explicitly control the redundant growth of the skill library. In the adaptive iteration of the review strategy, whether it is a newly generated or revised skill, it needs to go through a vector similarity calculation before being written into the review strategy skill library. It compares the skills with the existing candidate skills in the library one by one. Once the semantic repetition exceeds the threshold, the skill to be written can be directly discarded and will no longer enter the subsequent verification and update, so as to improve the reliability of updating the review strategy.

[0118] In one possible implementation, the review strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill. A skill encoding vector is obtained by vector encoding a candidate skill using an encoding model. In this embodiment, when updating the review strategy skill base by generating review skills to obtain a review strategy update skill base for reconfiguration to the review model, redundancy control can be implemented for the generated review skills. Specifically, the encoding model can be invoked to encode the generated review skills, resulting in generated encoding vectors. The generated encoding vectors are then compared with the multiple skill encoding vectors included in the review strategy skill base to calculate vector similarity, obtaining a second vector similarity degree for each skill encoding vector relative to the review encoding vector. Further, a second target similarity degree that reaches a similarity threshold can be searched among the obtained second vector similarities to obtain a second search result.

[0119] It is understandable that if the second search result indicates that there is no second target similarity among multiple second vector similarities, it means that there are no candidate skills in the current review strategy library that are highly similar to the generated review skills. Therefore, the generated review skills can be added to the review strategy skill library to obtain the second updated skill library. Furthermore, in this embodiment, the second updated skill library can be tested using the review strategy skill library, and the second updated skill library that passes the test is determined as the review strategy updated skill library for reconfiguration to the review model. It is understood that the skill library test performed on the second updated skill library using the review strategy skill library is for sandbox testing and admission verification of the second updated skill library; the specific process is the same as described above. Figure 3 The step S104 in the corresponding embodiment, which involves testing the first updated skill library using the audit strategy skill library, is similar and will not be described again here. It is understood that, after obtaining the generated audit skills, this application embodiment effectively improves the simplicity and reliability of the final audit strategy updated skill library by implementing redundancy control on the generated audit skills and conducting sandbox testing admission on the subsequently obtained updated skill library.

[0120] For further details, please see Figure 12 , Figure 12 This is a schematic diagram illustrating a scenario where an audit strategy is updated through a generation process, as provided in an embodiment of this application. Figure 12 As shown, in this embodiment of the application, after calculating the skill influence weight (i.e., influence weight 71a, including the weight 711a corresponding to skill 701a, the weight 711b corresponding to skill 701b, ..., the weight 711n corresponding to skill 701n) of each review skill for the target review node, the maximum influence weight (i.e., weight 711b) of each review skill's skill influence weight can be obtained, and weight analysis can be performed on the maximum influence weight to obtain the weight analysis result. If the weight analysis results indicate that the maximum impact weight has reached the impact weight threshold, the application embodiment can enter the revision process, revise the skill 701b corresponding to the weight 711b, and replace the current audit strategy skill library (skill library 72a) with the revised skill to obtain the updated skill library 73a. The updated skill library 73a is then subjected to sandbox admission testing to determine the updated skill library 73a that passes the test as the audit strategy update skill library that can be reconfigured to the audit model.

[0121] Optionally, if the weight analysis results indicate that the maximum influence weight has reached the influence weight threshold, the generation process can proceed. Missing skills are generated through the target review node and node optimization content to obtain the generated review skill 74a. Further, this embodiment can call an encoding model to encode the generated review skill 74a to obtain a generated encoding vector (generated encoding vector 75a). Further, it can search in the skill library 72a for a skill encoding vector similar to the generated encoding vector 75a. If it exists, the current test update is stopped. If no similar skill encoding vector exists in the skill library 72a, the generated review skill 74a can be added to the skill library 72a to obtain a second updated skill library (updated skill library 73b). Further, a sandbox admission test can be performed on the updated skill library 73b, i.e., a skill library test is performed on the updated skill library 73b using the skill library 72a, and the updated skill library 73b that passes the test is determined as the review strategy update skill library for reconfiguration to the review model.

[0122] It is understandable that after obtaining the abnormal review trajectory, the embodiments of this application will enter a processing closed loop consisting of abnormal attribution, skill modification or generation, and access verification (review strategy iterative update process). The goal of this closed loop can be expressed in a training-independent optimization form, as shown in formula (9):

[0123] In formula (9), q is the task distribution (business review content), Run(q,K) represents the execution of task q with the support of the review strategy skill library K, and success() is the success metric defined by the benchmark test. The entire optimization process does not touch the model weights, but improves the expected success rate of content review by modifying K. It can be understood that the embodiments of this application update the review strategy through the review strategy update process of anomaly attribution, skill modification or generation, and access verification, which can effectively improve the reliability and accuracy of updating the review strategy.

[0124] For further details, please see Figure 13 , Figure 13 This is a schematic diagram of the structure of an audit strategy update device provided in an embodiment of this application. Figure 13 As shown, the audit policy update device 1 can be a computer program (including program code) running on a computer device. For example, the audit policy update device 1 can be an application software. The audit policy update device 1 can be used to execute the functions provided in the embodiments of this application. Figure 3 and Figure 11 The corresponding steps in the method.

[0125] The device 1 includes: a review trajectory acquisition module 11, an optimized node positioning module 12, a skill weight calculation module 13, and a review strategy update module 14. The audit trajectory acquisition module 11 is used to acquire business audit content and abnormal audit trajectories related to the business audit content and at least one audit skill; abnormal audit trajectory refers to an audit trajectory that carries an abnormal audit tag and contains at least one audit node. The abnormal audit tag is obtained by the audit model calling at least one audit skill to audit the business audit content; at least one audit skill refers to a skill in the audit strategy skill library configured for the audit model. The node positioning optimization module 12 is used to locate the first optimizable target audit node related to the abnormal audit tag in at least one audit node based on the business audit content, abnormal audit trajectory and at least one audit skill, and output the node optimization description content for the target audit node. The skill weight calculation module 13 is used to evaluate the weight of each audit skill that affects the target audit node based on the business audit content, abnormal audit trajectory, at least one audit skill, and target audit node, and obtain the skill influence weight of each audit skill. The review strategy update module 14 is used to update the review strategy skill library based on the skill influence weight and node optimization description content of each review skill, and obtain the review strategy update skill library for reconfiguration to the review model.

[0126] Before performing the steps of acquiring business audit content, abnormal audit trajectories related to the business audit content, and at least one audit skill, the audit trajectory acquisition module 11 is also used for: When an audit model for auditing business audit content is obtained, at least one audit skill related to the business audit content is searched from the audit strategy skill library configured for the audit model. The review model invokes at least one review skill to review the business review content, resulting in at least one review node, and determines the content review tag for the business review content based on at least one review node; If the content review label is an abnormal review label obtained by reviewing the review pass label or review block label output by the review model, then the content review trajectory carrying the abnormal review label is determined as an abnormal review trajectory.

[0127] The review strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by vector encoding a candidate skill through an encoding model. When performing the step of searching for at least one audit skill related to the business audit content from the audit strategy skill library configured for the audit model, the audit trajectory acquisition module 11 is specifically used for: The business review content is structured to obtain structured content, and the encoding model is called to encode the structured content to obtain the content encoding vector; Among the multiple skill coding vectors included in the review strategy skill base, at least one target skill coding vector similar to the content coding vector is found, and the candidate skills corresponding to each of the found target skill coding vectors are determined as at least one target candidate skill; Based on at least one target candidate skill, identify at least one audit skill that is relevant to the business audit content.

[0128] Specifically, when performing the step of determining at least one audit skill related to the business audit content based on at least one target candidate skill, the audit trajectory acquisition module 11 is used for: The structured content is concatenated with the skill content of each of the at least one target candidate skill to obtain the concatenated content for each target candidate skill. The semantic analysis model is invoked to calculate the semantic similarity of the concatenated content for each target candidate skill, thereby obtaining the semantic similarity of each target candidate skill to the business review content. Among at least one target candidate skill, obtain the target candidate skill whose semantic similarity to the business review content reaches a similarity threshold, and determine the obtained target candidate skill as at least one review skill related to the business review content.

[0129] Specifically, when executing the steps of calling at least one review skill by the review model to review the business review content, obtaining at least one review node, and determining the content review tag for the business review content based on at least one review node, the review trajectory acquisition module 11 is used for: For each of the at least one review skills, generate steps to obtain review steps for each review skill, and construct a content review link for business review content through the review steps corresponding to each review skill; When the review model performs content review on business review content based on the content review chain, it outputs the review node corresponding to each review step in at least one review step included in the content review chain. Each review step corresponds to at least one review node, and content review tags for business review content are determined based on at least one review node.

[0130] Specifically, when the review trajectory acquisition module 11 executes the steps of the review model performing content review on business review content based on the content review chain, and outputs the review node corresponding to each review step in at least one review step included in the content review chain, it is used for: When the review model reviews business content based on the content review chain, it selects one review step from at least one review step included in the content review chain as the target review step based on the review status of the review model, and determines the review skill corresponding to the target review step as the target review skill. When the audit evidence indicated by the audit status meets the skill usage conditions indicated by the target audit skill, the target audit skill is invoked to execute the step execution action indicated by the target audit step, and the step execution result for the target audit step is obtained; When the result of a step execution matches the expected result of the skill indicated by the target audit skill, the step execution action is determined as a node execution action, and the step execution result is determined as a node execution result. Based on the node execution action and the node execution result, an audit node for the target audit step is constructed. When each audit step in at least one audit step is selected as the target audit step, the audit node corresponding to each audit step is obtained.

[0131] For details on the implementation process of the audit trajectory acquisition module 11, please refer to the above. Figure 3 The description of step S101 in the corresponding embodiments will not be repeated here.

[0132] Specifically, when the optimization node positioning module 12 performs the step of locating the first optimizable target audit node related to the abnormal audit tag in at least one audit node based on business audit content, abnormal audit trajectory, and at least one audit skill, and outputting node optimization description content for the target audit node, it is used for: Obtain the anomaly analysis model for anomaly analysis, and generate prompt words based on the anomaly location template, anomaly review trajectory, business review content, and at least one review skill indicated by the anomaly analysis model; Input the anomaly location prompt words into the anomaly analysis model. Under the constraints of the anomaly location prompt words, the anomaly analysis model performs node backtracking analysis on at least one review node in the anomaly review trajectory, and obtains the backtracking analysis results for the anomaly review trajectory. Based on the backtracking analysis evidence in the backtracking analysis results, the first optimizable target audit node related to the abnormal audit tag is located in at least one audit node; The anomaly analysis model identifies the node action indicated by the target review node as an anomalous action, generates anomaly description and action optimization content for the anomalous action, and outputs node optimization description content for the target review node based on the anomaly description and action optimization content.

[0133] For details on the implementation process of optimizing node positioning module 12, please refer to the above. Figure 3 The description of step S102 in the corresponding embodiments will not be repeated here.

[0134] Specifically, when the skill weight calculation module 13 performs the step of evaluating the weight of each audit skill that affects the target audit node based on the business audit content, abnormal audit trajectory, at least one audit skill, and the target audit node, it is used to obtain the skill influence weight of each audit skill. Obtain the anomaly analysis model for anomaly analysis, and generate prompt words by using the weight evaluation template, anomaly review trajectory, business review content, and at least one review skill and target review node indicated by the anomaly analysis model; Input the weight assessment prompts into the anomaly analysis model. The anomaly analysis model obtains the weight assessment strategy for weight assessment based on the weight assessment prompts, and assesses the influence weight of each of the at least one review skills on the target review node according to the weight assessment strategy. The influence weight of each review skill obtained is used as the skill influence weight of each review skill.

[0135] For details on the implementation process of skill weight calculation module 13, please refer to the above. Figure 3 The description of step S103 in the corresponding embodiments will not be repeated here.

[0136] Specifically, when executing the step of updating the audit strategy skill library based on the skill influence weight and node optimization description of each audit skill to obtain an audit strategy update skill library for reconfiguration to the audit model, the audit strategy update module 14 is used for: Obtain the maximum influence weight among the skill influence weights for each review skill, and perform weight analysis on the maximum influence weight to obtain the weight analysis results; If the weight analysis results indicate that the maximum impact weight has reached the impact weight threshold, then the impact review skill corresponding to the maximum impact weight is determined in at least one review skill. Obtain the skill revision template indicated by the skill revision model, and generate prompt words by combining the skill revision template, business review content, skills affecting the review, target review nodes, and node optimization descriptions. Input skill revision prompts into the skill revision model. The skill revision model then revises the skills affecting the review based on the skill revision prompts, resulting in revised review skills. The revised review skills are then used to update the review strategy skill library, which is used to reconfigure the review strategy updated skill library for the review model.

[0137] The review strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by vector encoding a candidate skill through an encoding model. When executing the step of updating the audit strategy skill library by revising audit skills to obtain an updated audit strategy skill library for reconfiguration to the audit model, the audit strategy update module 14 is specifically used for: The coding model is called to encode the revision review skills to obtain revision coding vectors. The revision coding vectors are then compared with multiple skill coding vectors included in the review strategy skill library to calculate the vector similarity, and the first vector similarity of each skill coding vector with respect to the review coding vector is obtained. Among the multiple first vector similarities obtained, the first target similarity that reaches the similarity threshold is searched to obtain the first search result; If the first search result indicates that there is no first target similarity among multiple first vector similarities, then the candidate skills in the review strategy skill library that match the impact review skills are replaced with revised review skills to obtain the first updated skill library; The first updated skill library is tested using the audit strategy skill library, and the first updated skill library that passes the test is identified as the audit strategy update skill library to be reconfigured to the audit model.

[0138] Specifically, when executing the step of performing a skill library test on the first updated skill library using the audit strategy skill library and determining the first updated skill library that passes the test as the audit strategy update skill library to be reconfigured to the audit model, the audit strategy update module 14 is used for: The business review content is reviewed by updating the skill library first, and the updated review tag is obtained. When the updated review tag is a non-abnormal review tag, the test sample content indicated by the skill library test strategy is obtained. The test sample content is reviewed by the review strategy skill library to obtain the first test review result, and the test sample content is reviewed by the first update skill library to obtain the second test review result. The test results are analyzed based on the results of the first test review, and the test analysis results are obtained. When the test analysis results indicate that the second test audit results meet the test pass conditions indicated by the skill library test strategy, the first updated skill library will be identified as the audit strategy updated skill library.

[0139] The test sample content includes sample content under multiple test types; When the audit strategy update module 14 performs the step of analyzing the test results of the second test audit based on the first test audit result to obtain the test analysis results, it is specifically used for: Obtain the first result metric for each of the multiple test types in the first test audit result, and obtain the test weight for each test type as indicated by the skills library test strategy; The first result indicator is obtained by weighting and summing the results based on the test weights of each test type to obtain the first weighted result indicator. The second test review result is then obtained by weighting and summing the results based on the test weights of each test type to obtain the second weighted result indicator. If the second weighted result indicator is greater than or equal to the first weighted result indicator, then the second test audit result is determined as the test analysis result if it meets the test pass conditions indicated by the skills library test strategy.

[0140] The audit strategy update module 14 is also used for: If the weight analysis results indicate that the maximum impact weight has not reached the impact weight threshold, then the skill generation template indicated by the skill generation model is obtained. The prompt words are generated by using the skill generation template, business review content, target review node and node optimization description content to obtain the skill generation prompt words. Input the skill generation prompts into the skill generation model, and the skill generation model will generate the missing skills for the target review node based on the skill generation prompts to obtain the generated review skills; By generating an audit skills update audit strategy skills library, an audit strategy update skills library is obtained for reconfiguration to the audit model.

[0141] The review strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by vector encoding a candidate skill through an encoding model. When executing the step of generating an audit strategy update skill library to obtain an audit strategy update skill library for reconfiguration to the audit model, the audit strategy update module 14 is specifically used for: The encoding model is called to encode the generated review skills to obtain the generated encoding vector. The generated encoding vector is then compared with the multiple skill encoding vectors included in the review strategy skill library to calculate the vector similarity, and the second vector similarity of each skill encoding vector with respect to the review encoding vector is obtained. Among the multiple second vector similarities obtained, the second target similarity that reaches the similarity threshold is searched to obtain the second search result; If the second search result indicates that there is no second target similarity among multiple second vector similarities, then the generated review skill is added to the review strategy skill library to obtain the second updated skill library; The second updated skill library is tested using the audit strategy skill library, and the second updated skill library that passes the test is identified as the audit strategy update skill library to be reconfigured to the audit model.

[0142] For details on the implementation process of the audit strategy update module 14, please refer to the above. Figure 3 The description of step S104 in the corresponding embodiments will not be repeated here.

[0143] The device 1 also includes an audit strategy generation module 15, which is specifically used for: Obtain the initial sample content indicated by the skill generation strategy, call the review model to review the initial sample content, and obtain the sample review trajectory; When the sample review label for the initial sample content carried by the sample review trajectory is a non-abnormal review label, the trajectory skills are summarized according to the skill generation strategy to obtain at least one summarized trajectory skill. Each summary trajectory skill in at least one summary trajectory skill is identified as a candidate skill, and an audit strategy skill library is constructed using the obtained candidate skills. The audit strategy skill library is then configured to the audit model.

[0144] For details on the implementation process of the audit strategy generation module 15, please refer to the above. Figure 10 The description of step S201 in the corresponding embodiments will not be repeated here.

[0145] Further, please see Figure 14 , Figure 14 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 14 As shown, the computer device 1000 can be a user terminal, for example, the one described above. Figure 1 The user terminal 10a in the corresponding embodiment can also be a server, for example, as described above. Figure 1The server 10d in the corresponding embodiment will not be limited here. For ease of understanding, this application takes a computer device as an example of a server. The computer device 1000 may include: a processor 1001, a network interface 1004, and a memory 1005. In addition, the computer device 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The memory 1005 may also optionally be at least one storage device located remotely from the aforementioned processor 1001. Figure 14 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0146] The network interface 1004 in the computer device 1000 can also provide network communication functions, and the optional user interface 1003 can also include a display screen and a keyboard. Figure 14 In the computer device 1000 shown, the network interface 1004 provides network communication functionality; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the aforementioned... Figure 3 or Figure 10 The description of the audit strategy update method in the corresponding embodiment can also be executed as described above. Figure 13 The description of the audit strategy update device 1 in the corresponding embodiment will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0147] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium, which stores a computer program executed by the aforementioned audit policy update device 1. The computer program includes computer instructions, and when the processor executes the computer instructions, it can execute the aforementioned... Figure 3 or Figure 10 The description of the audit strategy update method in the corresponding embodiments is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments related to this application, please refer to the description of the method embodiments of this application.

[0148] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of a computer device reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the computer device to perform the aforementioned... Figure 3 or Figure 10 The description of the audit strategy update method in the corresponding embodiments is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program products or computer program embodiments involved in this application, please refer to the description of the method embodiments of this application.

[0149] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0150] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0151] The steps in the method of this application embodiment can be adjusted, combined, or deleted according to actual needs.

[0152] The modules in the device of this application embodiment can be merged, divided, and deleted according to actual needs.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0154] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for updating an audit strategy, characterized in that, The method includes: The system acquires business review content, abnormal review trajectories related to the business review content, and at least one review skill. The abnormal review trajectory refers to a review trajectory carrying an abnormal review tag and containing at least one review node. The abnormal review tag is obtained by the review model calling the at least one review skill to review the business review content. The at least one review skill refers to a skill in the review strategy skill library configured for the review model. Based on the business review content, the abnormal review trajectory, and the at least one review skill, the first optimizable target review node related to the abnormal review tag is located in the at least one review node, and the node optimization description content for the target review node is output. Based on the business review content, the abnormal review trajectory, the at least one review skill, and the target review node, a weight evaluation is performed on each review skill used to influence the target review node to obtain the skill influence weight of each review skill. Based on the skill influence weight of each review skill and the node optimization description, the review strategy skill library is updated to obtain a review strategy update skill library for reconfiguration to the review model.

2. The method according to claim 1, characterized in that, Before acquiring the business review content and the abnormal review trajectory and at least one review skill related to the business review content, the method further includes: When an audit model for auditing business audit content is obtained, at least one audit skill related to the business audit content is searched from the audit strategy skill library configured for the audit model. The review model invokes at least one review skill to review the business review content, obtaining at least one review node, and determines a content review tag for the business review content based on the at least one review node; A content review trajectory for the business review content is constructed using the at least one review node and the content review tag. If the content review tag is an abnormal review tag obtained by reviewing the review pass tag or review block tag output by the review model, then the content review trajectory carrying the abnormal review tag is determined as an abnormal review trajectory.

3. The method according to claim 2, characterized in that, The audit strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by performing vector encoding processing on a candidate skill through an encoding model; The step of searching for at least one audit skill related to the business audit content from the audit strategy skill library configured for the audit model includes: The business review content is structured to obtain structured content, and the encoding model is called to encode the structured content to obtain a content encoding vector; Among the multiple skill encoding vectors included in the review strategy skill base, at least one target skill encoding vector similar to the content encoding vector is searched, and the candidate skills corresponding to each of the searched target skill encoding vectors are determined as at least one target candidate skill; Based on the at least one target candidate skill, at least one audit skill related to the business audit content is determined.

4. The method according to claim 3, characterized in that, The step of determining at least one review skill related to the business review content based on the at least one target candidate skill includes: The structured content is concatenated with the skill content of each of the at least one target candidate skill to obtain concatenated content for each target candidate skill. The semantic analysis model is invoked to calculate the semantic similarity of the concatenated content of each target candidate skill, so as to obtain the semantic similarity of each target candidate skill to the business review content. Among the at least one target candidate skill, a target candidate skill whose semantic similarity to the business review content reaches a similarity threshold is obtained, and the obtained target candidate skill is determined as at least one review skill related to the business review content.

5. The method according to claim 2, characterized in that, The step of calling at least one review skill by the review model to perform content review on the business review content, obtaining at least one review node, and determining content review tags for the business review content based on the at least one review node, includes: For each of the at least one review skill, steps are generated to obtain review steps for each review skill, and a content review link for the business review content is constructed through the review steps corresponding to each review skill. When the review model performs content review on the business review content based on the content review chain, it outputs the review node corresponding to each review step in at least one review step included in the content review chain; The review node corresponding to each review step is determined as at least one review node, and a content review tag for the business review content is determined based on the at least one review node.

6. The method according to claim 5, characterized in that, When the review model performs content review on the business review content based on the content review chain, it outputs the review node corresponding to each review step in at least one review step included in the content review chain, including: When the review model performs content review on the business review content based on the content review link, it selects one review step from at least one review step included in the content review link as the target review step based on the review status of the review model, and determines the review skill corresponding to the target review step as the target review skill. When the audit evidence indicated by the audit status meets the skill usage conditions indicated by the target audit skill, the target audit skill is invoked to execute the step execution action indicated by the target audit step, and the step execution result for the target audit step is obtained; When the result of the step execution matches the expected result of the skill indicated by the target audit skill, the step execution action is determined as a node execution action, and the step execution result is determined as a node execution result. Based on the node execution action and the node execution result, an audit node for the target audit step is constructed. When each of the at least one audit step is selected as the target audit step, the audit node corresponding to each audit step is obtained.

7. The method according to claim 1, characterized in that, Based on the business review content, the abnormal review trajectory, and the at least one review skill, the first optimizable target review node related to the abnormal review tag is located among the at least one review node, and node optimization description content for the target review node is output, including: Obtain an anomaly analysis model for anomaly analysis, and generate prompt words by using the anomaly location template indicated by the anomaly analysis model, the anomaly review trajectory, the business review content, and the at least one review skill to obtain anomaly location prompt words; The abnormal location prompt is input into the abnormal analysis model. Under the constraint of the abnormal location prompt, the abnormal analysis model performs node backtracking analysis on at least one review node in the abnormal review trajectory to obtain the backtracking analysis results for the abnormal review trajectory. Based on the backtracking analysis evidence present in the backtracking analysis results, the first optimizable target audit node related to the abnormal audit tag is located in the at least one audit node; The anomaly analysis model identifies the node action indicated by the target review node as an anomalous action, generates an anomaly description and action optimization content for the anomalous action, and outputs an optimized node description for the target review node based on the anomaly description and action optimization content.

8. The method according to claim 1, characterized in that, Based on the business review content, the abnormal review trajectory, the at least one review skill, and the target review node, a weight evaluation is performed on each review skill that influences the target review node to obtain the skill influence weight of each review skill, including: Obtain an anomaly analysis model for anomaly analysis, and generate prompt words by using the weight evaluation template indicated by the anomaly analysis model, the anomaly review trajectory, the business review content, the at least one review skill, and the target review node; thus, obtain weight evaluation prompt words. The weight assessment prompt is input into the anomaly analysis model. The anomaly analysis model obtains a weight assessment strategy for weight assessment based on the weight assessment prompt and assesses the influence weight of each of the at least one review skills on the target review node according to the weight assessment strategy. The influence weight of each review skill obtained by assessment is used as the skill influence weight of each review skill.

9. The method according to claim 1, characterized in that, The review strategy skill library is updated based on the skill influence weight of each review skill and the node optimization description content to obtain a review strategy update skill library for reconfiguration to the review model, including: Obtain the maximum influence weight among the skill influence weights of each review skill, and perform weight analysis on the maximum influence weight to obtain the weight analysis results; If the weight analysis result indicates that the maximum influence weight has reached the influence weight threshold, then the influence review skill corresponding to the maximum influence weight is determined among the at least one review skill; Obtain the skill revision template indicated by the skill revision model, and generate prompt words by using the skill revision template, the business review content, the skills affecting the review, the target review node, and the node optimization description content to obtain skill revision prompt words; The skill revision prompt is input into the skill revision model, which then revises the influencing review skill based on the prompt to obtain the revised review skill. The revised review skill is then used to update the review strategy skill library, resulting in a review strategy updated skill library for reconfiguration to the review model.

10. The method according to claim 9, characterized in that, The audit strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by performing vector encoding processing on a candidate skill through an encoding model; The step of updating the review strategy skill library through the revised review skills to obtain a review strategy updated skill library for reconfiguration to the review model includes: The coding model is invoked to encode the revised review skill to obtain a revised coding vector. The revised coding vector is then compared with multiple skill coding vectors included in the review strategy skill library to calculate the vector similarity, thereby obtaining the first vector similarity of each skill coding vector among the multiple skill coding vectors with respect to the review coding vector. Among the multiple first vector similarities obtained, the first target similarity that reaches the similarity threshold is searched to obtain the first search result; If the first search result indicates that the first target similarity does not exist among the plurality of first vector similarities, then the candidate skill in the review strategy skill library that matches the influence review skill is replaced with the revised review skill to obtain the first updated skill library; The first updated skill library is tested using the audit strategy skill library, and the first updated skill library that passes the test is determined as the audit strategy update skill library to be reconfigured to the audit model.

11. The method according to claim 10, characterized in that, The step of testing the first updated skill library using the audit strategy skill library, and determining the first updated skill library that passes the test as the audit strategy update skill library for reconfiguration to the audit model, includes: The business review content is reviewed by the first updated skill library to obtain an updated review tag. When the updated review tag is a non-abnormal review tag, the test sample content indicated by the skill library test strategy is obtained. The test sample content is reviewed using the review strategy skill library to obtain a first test review result, and the test sample content is reviewed using the first update skill library to obtain a second test review result. The test results are analyzed based on the first test review results to obtain the test analysis results; When the test analysis result indicates that the second test review result meets the test pass conditions indicated by the skill library test strategy, the first updated skill library is determined as the skill library updated by the review strategy.

12. The solution according to claim 11, characterized in that, The test sample content includes sample content under multiple test types; The step of analyzing the second test audit result based on the first test audit result to obtain the test analysis result includes: Obtain the first result indicator of the first test review result for each of the multiple test types, and obtain the test weight for each test type indicated by the skill library test strategy; The first result indicator is obtained by weighting and summing the test weights of each test type to obtain the first weighted result indicator. The second result indicator of the second test review result is obtained by weighting and summing the test weights of each test type to obtain the second weighted result indicator. If the second weighted result indicator is greater than or equal to the first weighted result indicator, then the second test review result that meets the test pass conditions indicated by the skill library test strategy is determined as the test analysis result.

13. The method according to claim 9, characterized in that, The method further includes: If the weight analysis result indicates that the maximum influence weight has not reached the influence weight threshold, then the skill generation template indicated by the skill generation model is obtained, and prompt words are generated by using the skill generation template, the business review content, the target review node, and the node optimization description content to obtain skill generation prompt words; The skill generation prompts are input into the skill generation model, which then generates missing skills for the target review node based on the prompts, thus generating the review skills. The review strategy skill library is updated by generating review skills to obtain a review strategy update skill library for reconfiguration to the review model.

14. The method according to claim 13, characterized in that, The audit strategy skill base includes multiple candidate skills and a skill encoding vector corresponding to each candidate skill; a skill encoding vector is obtained by performing vector encoding processing on a candidate skill through an encoding model; The step of updating the review strategy skill library through the generated review skills to obtain a review strategy update skill library for reconfiguration to the review model includes: The encoding model is invoked to encode the generated review skill to obtain a generated encoding vector. The generated encoding vector is then compared with multiple skill encoding vectors included in the review strategy skill library to calculate the vector similarity, thereby obtaining the second vector similarity of each skill encoding vector with respect to the review encoding vector. Among the multiple second vector similarities obtained, the second target similarity that reaches the similarity threshold is searched to obtain the second search result; If the second search result indicates that the second target similarity does not exist among the plurality of second vector similarities, then the generated review skill is added to the review strategy skill library to obtain the second updated skill library; The second updated skill library is tested using the audit strategy skill library, and the second updated skill library that passes the test is determined as the audit strategy update skill library to be reconfigured to the audit model.

15. The method according to claim 1, characterized in that, The method further includes: Obtain the initial sample content indicated by the skill generation strategy, call the review model to perform content review on the initial sample content, and obtain the sample review trajectory; When the sample review label for the initial sample content carried by the sample review trajectory is a non-abnormal review label, the trajectory skill is summarized according to the skill generation strategy to obtain at least one summarized trajectory skill. Each of the at least one summary trajectory skills is identified as a candidate skill, and an audit strategy skill library is constructed using the obtained at least one candidate skill. The audit strategy skill library is then configured to the audit model.

16. An audit strategy update device, characterized in that, The device includes: The review trajectory acquisition module is used to acquire business review content, abnormal review trajectories related to the business review content, and at least one review skill; the abnormal review trajectory refers to a review trajectory carrying an abnormal review tag and containing at least one review node, the abnormal review tag is obtained by the review model calling the at least one review skill to review the business review content; the at least one review skill refers to a skill in the review strategy skill library configured for the review model; The optimization node positioning module is used to locate the first optimizable target audit node related to the abnormal audit tag in the at least one audit node based on the business audit content, the abnormal audit trajectory and the at least one audit skill, and output the node optimization description content for the target audit node; The skill weight calculation module is used to evaluate the weight of each review skill that affects the target review node based on the business review content, the abnormal review trajectory, the at least one review skill, and the target review node, so as to obtain the skill influence weight of each review skill. The review strategy update module is used to update the review strategy skill library based on the skill influence weight of each review skill and the node optimization description content, so as to obtain a review strategy update skill library for reconfiguration to the review model.

17. A computer device, characterized in that, Including memory and processor; The memory is connected to the processor, the memory is used to store computer programs, and the processor is used to invoke the computer programs so that the computer device performs the method according to any one of claims 1-15.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-15.

19. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium, the computer program being adapted to be read and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-15.

Images