Hybrid collaborative approval method and system based on process engine and large model

Abstract

The application relates to a hybrid collaborative approval method and system based on a process engine and a large model, and relates to the technical field of intelligent approval, and comprises the following steps: obtaining an approval request, extracting approval content and approval metadata; based on a rule cluster, the approval metadata is subjected to rule-based screening and discrimination, and the approval request is divided into a first and a second alternative request set; the approval content corresponding to the first alternative request set is subjected to abnormal analysis, and a third alternative request set is extracted; the first, second and third alternative request sets are subjected to automatic routing approval, and the automatic routing approval result and the approval request are sent to a user terminal for verification decision, and an approval result is output according to the verification decision result. The application solves the problems that the traditional approval system cannot effectively process unstructured complex tasks, resulting in low complex business approval efficiency, inaccurate risk identification and insufficient intelligent level.

Description

Technical Field

[0001] This application relates to the field of intelligent approval, and in particular to a hybrid collaborative approval method and system based on process engines and large models. Background Technology

[0002] As business complexity increases and the proportion of unstructured approvals grows, the efficiency, accuracy of risk identification, and intelligence level of complex business approvals have become key requirements for supporting compliant operations and efficient decision-making.

[0003] Currently, traditional approval systems mostly rely on fixed process nodes to handle rule-based tasks, which cannot effectively cope with the in-depth review needs of unstructured and complex tasks. They are not only difficult to accurately uncover potential compliance risks and abnormal information in such tasks, resulting in inaccurate risk identification; but also cannot flexibly adapt to complex scenarios due to fixed processes, which slows down the approval process and reduces approval efficiency. At the same time, they lack intelligent collaborative processing mechanisms, resulting in insufficient overall intelligence in the approval process, which ultimately affects the efficiency of business progress and increases compliance decision-making risks. Summary of the Invention

[0004] This application provides a hybrid collaborative approval method and system based on a process engine and a large model, which improves the current situation of low efficiency, inaccurate risk identification, and insufficient intelligence level of traditional approval systems due to their inability to effectively handle unstructured and complex tasks.

[0005] The embodiments of this application disclose the following technical solutions:

[0006] In a first aspect, embodiments of this application provide a hybrid collaborative approval method based on a process engine and a large model, the method comprising:

[0007] Obtain the approval request and extract the approval content and approval metadata;

[0008] Based on a predetermined set of rules, the approval metadata is subjected to rule-based screening and discrimination to divide the approval request into a first set of alternative requests and a second set of alternative requests.

[0009] Perform a biased analysis based on data feature dimensions on the approval content corresponding to the first candidate request set, and extract a third candidate request set based on the biased analysis results;

[0010] Automatic routing approval is performed based on the first, second, and third alternative request sets, and the automatic routing approval results and approval requests are sent to the user terminal for verification and decision-making. The approval result is then output based on the collected verification and decision-making results.

[0011] Secondly, embodiments of this application provide a hybrid collaborative approval system based on a process engine and a large model, the system comprising:

[0012] The request information extraction module is used to obtain approval requests and extract approval content and approval metadata;

[0013] The metadata screening and classification module is used to perform rule-based screening and discrimination on the approval metadata based on a predetermined set of rules, and to divide the approval request into a first candidate request set and a second candidate request set.

[0014] The content anomaly analysis module is used to perform anomaly analysis on the approval content corresponding to the first candidate request set based on data feature dimensions, and extract the third candidate request set based on the anomaly analysis results.

[0015] The routing approval output module is used to automatically approve routes based on the first alternative request set, the second alternative request set, and the third alternative request set, and send the automatic routing approval result and approval request to the user terminal for verification and decision-making, and output the approval result based on the collected verification and decision-making results.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] This application proposes a hybrid collaborative approval method and system based on a process engine and a large model. By completing the approval information extraction, metadata screening and classification, content anomaly analysis, intelligent routing approval, and manual verification decision-making in steps, it achieves differentiated and efficient processing and precise risk control for different types of approval tasks. First, approval requests and corresponding data are received from the business system. A pre-defined rule engine is used to separate structured approval metadata from unstructured approval content. Then, based on a rule set including structural feature rules and keyword matching rules, the approval metadata is screened, with requests conforming to the rules assigned to the first candidate request set and those not conforming to the second candidate request set. Next, data feature dimension analysis is performed on the approval content of the first candidate request set, extracting statistical features, entropy features, and embedded vector features to construct multi-dimensional comparative features. The deviation is calculated and compared with the multi-dimensional reference features of historical regular approval content, and thresholds are used to filter out abnormal requests, forming a third candidate request set. Then, the second and third candidate request sets are merged and sent to a pre-trained large model to generate a decision support report with confidence. The difference between the first and third candidate sets is used as the process request set and input into the process engine for standardized approval, while simultaneously monitoring for approval anomalies and triggering supplementary approval from the large model. Finally, the automatic approval results are integrated with the original requests and sent to the user terminal for verification, and the final approval result is output based on the feedback.

[0018] The technical solution of this application solves the problems of low approval efficiency, inaccurate risk identification, and insufficient intelligence level caused by the inability of traditional approval systems to effectively handle unstructured complex tasks. It avoids the efficiency loss caused by mixing rule-based and complex tasks, as well as the risk omissions and misjudgments caused by relying solely on manual or single systems. At the same time, it further ensures the reliability of approval results through manual verification, thereby improving the efficiency, quality, and intelligence level of complex business approval. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the hybrid collaborative approval method based on a process engine and a large model provided in this application embodiment;

[0021] Figure 2 This is a schematic diagram of the structure of a hybrid collaborative approval system based on a process engine and a large model, provided in an embodiment of this application.

[0022] The components represented by each number in the attached diagram are explained below:

[0023] Request information extraction module 01, metadata screening and classification module 02, content anomaly analysis module 03, and route approval output module 04. Detailed Implementation

[0024] This application provides a hybrid collaborative approval method and system based on a process engine and a large model, which is used to solve the technical problems of traditional approval systems in the prior art, which are unable to effectively handle unstructured and complex tasks, resulting in low efficiency, inaccurate risk identification, and insufficient intelligence in complex business approval.

[0025] 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 skilled in the art without creative effort are within the scope of protection of this application.

[0026] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0028] Example 1, as shown in the appendix Figure 1 As shown, this application provides a hybrid collaborative approval method based on a process engine and a large model, the method comprising the following steps:

[0029] S110: Obtain the approval request, extract the approval content and approval metadata;

[0030] In this embodiment of the application, in the scenario of initiating a hybrid collaborative approval process, in order to clarify the processing objects of subsequent rule screening, abnormality analysis and routing approval, it is necessary to first receive approval-related data from the business system and complete the separation and extraction of structured and unstructured information to ensure the orderly progress of the entire approval process and the accuracy of analysis.

[0031] Specifically, the process begins by receiving the initiated approval request from the corresponding business system, and simultaneously retrieving all approval request data directly associated with that request. The scope of the retrieved approval request data must cover all information carriers related to the approval matter to ensure that subsequent information extraction comprehensively reflects the complete content of the approval matter.

[0032] Furthermore, based on a pre-configured rule engine, the received approval requests and approval request data are systematically parsed. During the parsing process, the focus is on identifying and extracting structured data fields that meet the preset format requirements. These fields are then identified as approval metadata, providing basic information support for subsequent rule-based rapid screening.

[0033] Meanwhile, relying on the preset rule engine, content that does not conform to the structured format and exists in the form of natural text is filtered out from the approval request data. This content is extracted as approval content, reserving processing objects for subsequent abnormal analysis and risk identification of unstructured information.

[0034] This step obtains complete approval data from the business system and uses a rules engine to separate and extract approval metadata from approval content, providing a clear data foundation for subsequent approval processing based on different scenarios.

[0035] Step S110 in the method provided in this application embodiment includes:

[0036] Receive approval requests and corresponding approval request data from the business system;

[0037] Based on a preset rule engine, the approval request and the approval request data are parsed, and structured data fields are extracted as the approval metadata.

[0038] Based on a preset rule engine, the unstructured text content in the approval request data is extracted as the approval content.

[0039] In this embodiment of the application, in order to avoid deviations or reduced efficiency in subsequent process processing due to the chaotic splitting of approval information, it is necessary to first obtain complete approval data from the business system, and then use a rule engine to separate and extract structured and unstructured information to ensure the orderly progress and processing accuracy of the entire hybrid collaborative approval process.

[0040] Specifically, the system first receives the initiated approval request from the corresponding business system, and at the same time obtains all approval request data directly associated with the approval request.

[0041] The scope of the business system needs to be determined based on the actual approval scenario, covering the company's internal financial approval system, project management system, contract management system, etc., to ensure that all information related to the approval matters can be obtained, and to avoid incomplete subsequent analysis due to missing data.

[0042] Meanwhile, during the data reception process, it is necessary to ensure the integrity and security of data transmission. By using a preset data verification mechanism, potential packet loss and code errors during transmission can be detected to ensure that the received approval request data can accurately reflect the actual situation of the approval matters and provide a reliable basis for subsequent parsing and extraction.

[0043] Furthermore, based on a pre-defined rule engine, the acquired approval requests and approval request data are systematically parsed to extract structured data fields as approval metadata. The pre-defined rules of the rule engine need to be formulated in conjunction with the structured information requirements of common approval scenarios to clarify the types and format requirements of the fields to be extracted.

[0044] For example, in the financial reimbursement approval scenario, it is necessary to extract fields such as the name of the person requesting reimbursement, department, reimbursement amount, reimbursement reason code, number of invoices, and reimbursement cycle; in the project initiation approval scenario, it is necessary to extract fields such as project name, initiating department, project budget, project cycle, and person in charge information.

[0045] Among them, the above fields must conform to the preset structured format, such as the date field following the "year-month-day" format, the amount field retaining two decimal places, and the code field adopting the preset encoding rules, so as to ensure that the extracted approval metadata has a unified format standard and can be directly used for subsequent rule set matching and discrimination.

[0046] Meanwhile, during the parsing and extraction process, the rule engine will automatically verify the completeness and compliance of the fields. If key structured fields are found to be missing or have incorrect formats, a data correction prompt will be triggered. The extraction of approval metadata will be completed only after the data has been supplemented and corrected, so as to avoid invalid data from entering the subsequent process.

[0047] Furthermore, relying on the pre-defined rule engine, unstructured text content is extracted from the approval request data as the approval content. The extraction scope of unstructured text content must cover information in the approval request data that cannot be presented in a fixed field format. The rule engine will use pre-defined text recognition logic to filter out and integrate this type of information.

[0048] For example, in a contract approval scenario, it is necessary to extract the contract terms, supplementary agreement content, risk warnings, etc.; in a project plan approval scenario, it is necessary to extract the project feasibility analysis report, market research conclusions, technical solution descriptions, etc.

[0049] During the extraction process, the rule engine automatically filters out redundant text that is irrelevant to the core content of the approval, such as format specifiers, duplicate header information, and irrelevant automatically generated system notes, ensuring that the extracted approval content focuses on the core information of the approval matter.

[0050] Meanwhile, for approval request data containing multiple unstructured text files, the rule engine will classify and integrate the extracted text according to the file type and content relevance. For example, unstructured text related to contracts will be classified into one category, and unstructured text related to attachment descriptions will be classified into another category, providing a clear text foundation for subsequent abnormal analysis of unstructured content.

[0051] Furthermore, after extracting the approval metadata and approval content, the rule engine will cross-validate the two types of information to ensure that the extracted structured fields and unstructured text can corroborate each other and there are no logical contradictions.

[0052] For example, in the financial expense reimbursement approval process, if the reimbursement reason code in the approval metadata is "travel expenses," it is necessary to verify whether the unstructured reimbursement description text contains descriptions related to travel, such as travel destination and purpose. If a logical contradiction is found, such as the code being "travel expenses" but the text containing no travel-related information, a secondary parsing verification will be triggered to check for extraction errors and ensure that the final extracted approval metadata is consistent with the approval content, laying the foundation for reliable processing in subsequent procedures.

[0053] S120: Based on a predetermined set of rules, the approval metadata is subjected to rule-based screening and discrimination to divide the approval request into a first set of alternative requests and a second set of alternative requests.

[0054] In this embodiment of the application, in order to quickly distinguish between rule-based approval tasks and unstructured complex approval tasks, it is necessary to systematically screen and judge the approval metadata based on a predetermined set of rules in order to achieve accurate classification of approval requests.

[0055] Specifically, the first step is to configure a rule set that includes structural feature rules and keyword matching rules. Among them, the structural feature rules need to be formulated based on the structured data characteristics of rule-based tasks in common approval scenarios to clarify the format, numerical range, field relationships, and other conditions that the approval metadata must meet.

[0056] In addition, keyword matching rules need to sort out the core information identifiers that frequently appear in rule-based approval tasks and form a pre-set keyword list to ensure that the rule set can fully cover the identification dimensions of rule-based approval requests.

[0057] Furthermore, the extracted approval metadata is matched with structural feature rules to identify approval requests that meet the rule requirements and store them in the first candidate request set.

[0058] During the matching process, it is necessary to verify whether each field of the approval metadata meets the preset conditions of the structural feature rules. If all key fields meet the rule requirements, the approval request is determined to be a rule-based task and is included in the first candidate request set. If there are fields that do not meet the rule conditions, they are temporarily excluded from the first candidate request set and proceed to the subsequent keyword matching stage.

[0059] Furthermore, after completing the structural feature rule matching, the approval metadata is further matched with the keyword matching rules to identify approval requests containing preset keywords, which are also stored in the first candidate request set.

[0060] During matching, keyword retrieval is performed on text fields in the approval metadata. If a preset keyword is found and the frequency and position of the keyword meet the rule requirements, the approval request is determined to have rule-based task characteristics and is included in the first candidate request set. If no preset keyword is found, or the occurrence of the keyword does not meet the rule requirements, the temporary state of the approval request is maintained.

[0061] Finally, after completing the dual screening using structural feature rules and keyword matching rules, the approval requests that were not included in the first candidate request set are output as the second candidate request set.

[0062] This step, through dual rule screening, effectively distinguishes between rule-based approval requests and unstructured complex approval requests, ensuring that different types of approval requests can be processed in an appropriate manner, thereby improving the efficiency and accuracy of the overall approval process.

[0063] Step S120 in the method provided in this application embodiment includes:

[0064] Configure the rule set that includes structural feature rules and keyword matching rules;

[0065] The approval metadata is matched with the structural feature rules to identify eligible approval requests and store them in the first candidate request set.

[0066] The approval metadata is matched with the keyword matching rules to identify approval requests containing preset keywords and stored in the first candidate request set.

[0067] Approval requests that are not stored in the first alternative request set are output as the second alternative request set.

[0068] In this embodiment of the application, in order to avoid inefficiency or judgment bias caused by processing all approval requests in the same way, it is necessary to first build a set of rules that are adapted to the approval scenario, and complete the classification of approval requests through dual rule matching, so as to ensure that the corresponding processing path can be matched for different types of requests in the future.

[0069] Specifically, the first step is to configure a rule set that includes structural feature rules and keyword matching rules. The configuration of structural feature rules needs to be determined in conjunction with the structured data characteristics of various rule-based approval tasks to clarify the field formats, numerical ranges, and inter-field relationships that the approval metadata must meet.

[0070] For example, in the financial reimbursement approval scenario, rules are set such as the department field must be a value in the preset department list, the reimbursement amount field must be a positive number with two decimal places, and the reimbursement period field must be within the current year; in the project initiation approval scenario, rules are set such as the project budget field must be within the department's annual budget range, and the project period field must be greater than 30 days and less than 365 days.

[0071] Furthermore, configuring keyword matching rules requires identifying frequently occurring core information identifiers in rule-based approval tasks and creating a pre-defined keyword list. Examples include routine office expenses and travel allowances in financial reimbursement, and annual routine projects and small-scale purchases in project approval.

[0072] At the same time, it is necessary to clarify the matching requirements of keywords, such as requiring them to appear at least once in the approval reason field and to match the department field, so as to ensure that the rule set can fully cover the identification dimensions of rule-based approval requests.

[0073] In addition, the configuration of rule clusters needs to be updated regularly in light of changes in actual approval scenarios. For example, when adding a new approval type, the corresponding structural features and keywords should be added, and when adjusting business standards, the field rules should be updated to avoid classification deviations caused by rule lag.

[0074] After completing the rule cluster configuration, the approval metadata is further matched with the structural feature rules to identify the matching approval requests and store them in the first alternative request set.

[0075] Specifically, during the matching process, each key field of the approval metadata needs to be verified one by one to confirm whether it meets the preset requirements of the structural feature rules. If all key structured fields in the metadata of a certain approval request meet the corresponding rules and there are no logical contradictions between the fields, then the request is determined to be a rule-based task and is directly stored in the first candidate request set. Conversely, if any key field does not meet the rules or there is a logical conflict between the fields, it is temporarily excluded from the first candidate request set and proceeds to the subsequent keyword matching stage for further evaluation.

[0076] Furthermore, after completing the structural feature rule matching, the approval metadata is further matched with the keyword matching rules to identify approval requests containing preset keywords and store them in the first candidate request set.

[0077] Specifically, during the matching process, keyword searches are performed on text fields in the approval metadata, and the results are combined with preset matching requirements to determine whether the conditions are met. For example, in a project approval scenario, if the "Project Description" field contains the keyword "Quarterly Regular Material Procurement," and the keyword appears at least once and is consistent with the "Procurement" attribute of the "Project Type" field, then the request is determined to have regular task characteristics and is stored in the first candidate request set.

[0078] Conversely, if no preset keywords are found in the text fields, or if the found keywords conflict with other field attributes, the request will remain in a temporary state. Meanwhile, to avoid duplicate classification, before storing a matching request in the first candidate request set, it must be verified whether the request has already been entered into the set through structural feature rules. If it already exists, it will not be stored again, ensuring the uniqueness of data within the first candidate request set.

[0079] After completing the dual screening using structural feature rules and keyword matching rules, approval requests not included in the first candidate request set are output as the second candidate request set. These requests typically do not meet the characteristics of rule-based tasks and may contain unstructured, complex information that cannot be quickly identified using fixed rules.

[0080] Ultimately, through the above rule set configuration and dual matching steps, the accurate classification of approval requests was achieved. The first set of alternative requests can then be processed through the workflow engine to execute a standardized approval process, while the second set of alternative requests triggers the large model to conduct in-depth content review and risk identification. This lays the foundation for the efficient advancement of hybrid collaborative approval and also provides practical data reference for the subsequent optimization of the rule set.

[0081] S130: Perform a biased analysis on the approval content corresponding to the first candidate request set based on data feature dimensions, and extract a third candidate request set based on the biased analysis results;

[0082] In this embodiment of the application, in the scenario where the first set of candidate requests has been determined through rule screening, in order to accurately identify the abnormal items in the first set of candidate requests, it is necessary to conduct anomaly analysis on the approval content from the data feature dimension and extract abnormal requests, so as to ensure that only rule-based tasks without abnormalities are allowed to enter the traditional approval node, thereby improving the risk control capability and result reliability of the approval process.

[0083] Specifically, the approval content in the first candidate request set is first analyzed for data feature dimensions to obtain multi-dimensional comparative features including statistical features, entropy features and embedded vector features.

[0084] Furthermore, historical routine approval content is obtained, and processed using the same data feature dimension analysis method as the approval content of the first alternative request set, to extract multi-dimensional reference features.

[0085] Among them, the selection of historical routine approval content should cover rule-based tasks that have been successfully approved without any abnormalities within a certain period in the past, so as to ensure that the multi-dimensional reference features can truly reflect the feature benchmark of routine rule-based approval content, and it is necessary to ensure that the dimensions of the multi-dimensional reference features and the multi-dimensional comparison features correspond one-to-one, so as to avoid the failure of subsequent deviation calculation due to dimension mismatch.

[0086] Furthermore, after extracting the multidimensional contrast features and multidimensional reference features, the multidimensional feature deviation is calculated based on the two types of features, and the calculation result is compared with the preset multidimensional confidence deviation threshold.

[0087] During the calculation process, the deviation value needs to be calculated separately for each feature dimension to ensure accurate identification of anomalies in each dimension. During comparison, if the deviation value of any feature dimension is greater than or equal to the corresponding multidimensional confidence deviation threshold, it is determined that the approval content is abnormal, and this multidimensional feature deviation is added to the abnormality analysis results.

[0088] Finally, based on the results of the anomaly analysis, the corresponding alternative requests are extracted from the first alternative request set and output as the third alternative request set.

[0089] This step, through the analysis of abnormalities in data feature dimensions, enables the precise screening of anomalies in rule-based tasks. This allows the third alternative request set to enter the deep processing stage of the large model together with the second alternative request set, effectively avoiding approval errors caused by abnormal requests passing through traditional approval nodes, and further improving the classification and processing mechanism of approval tasks.

[0090] Step S130 in the method provided in this application embodiment includes:

[0091] Data feature dimension analysis is performed on the approval content in the first candidate request set to obtain multi-dimensional comparison features, wherein the multi-dimensional comparison features include statistical features, entropy features and embedding vector features;

[0092] Obtain historical routine approval content and perform corresponding data feature dimension analysis to extract multidimensional reference features, wherein the dimensions of the multidimensional reference features correspond one-to-one with those of the multidimensional comparison features;

[0093] The multidimensional feature deviation is calculated based on the multidimensional reference feature and the multidimensional comparison feature, and compared with the preset multidimensional confidence deviation threshold. Any multidimensional feature deviation whose dimension is greater than or equal to the multidimensional confidence deviation threshold is added to the anomaly analysis result.

[0094] Based on the anomaly analysis results, the corresponding alternative requests are extracted from the first alternative request set and output as the third alternative request set.

[0095] In this embodiment of the application, in order to avoid the risk of abnormal requests entering the traditional approval node due to relying solely on rule screening, it is necessary to screen out abnormal items in rule-based tasks through in-depth analysis of data feature dimensions and comparison with historical routine content, so as to improve the approval task classification system.

[0096] Specifically, the data feature dimension analysis is first performed on the approval content in the first set of candidate requests to obtain multi-dimensional comparative features including statistical features, entropy features and embedded vector features, so as to comprehensively capture the feature differences of the approval content.

[0097] The method provided in this application embodiment performs data feature dimension analysis on the approval content in the first candidate request set to obtain multi-dimensional comparison features, wherein the multi-dimensional comparison features include statistical features, entropy features, and embedding vector features, including:

[0098] Based on a pre-set list of phrases, the approval content is statistically analyzed to obtain statistical features, wherein the statistical features include at least word frequency distribution information;

[0099] Based on the word frequency distribution information, calculate the approximate information entropy corresponding to the approval content;

[0100] Based on the attribute tags of the phrase list, the word frequency distribution information is deredundant and filtered, and the deredundant information entropy corresponding to the approval content is calculated based on the deredundant filtering results. The ratio of the deredundant information entropy to the approximate information entropy is then output as the deredundant coefficient.

[0101] Activate the pre-built embedding vector model to map the approval content into embedding vector features;

[0102] The deredundant information entropy, the approximate information entropy, and the deredundancy coefficient are used as the entropy features, and the statistical features and the embedded vector features are combined to output the multidimensional comparison features.

[0103] In this embodiment of the application, in order to comprehensively capture the characteristics of the approval content of the first candidate request set from multiple dimensions and provide quantifiable analytical basis for subsequent abnormal comparison with historical routine approval content, it is necessary to construct a complete multi-dimensional comparison feature system through multi-step feature extraction and integration to improve the accuracy and reliability of subsequent abnormal analysis.

[0104] Specifically, the approval content is first statistically analyzed based on a pre-defined list of terms to obtain statistical features that include at least word frequency distribution information. This pre-defined list of terms needs to be constructed based on the actual needs of the approval scenario, covering frequently occurring core compliance terms, invalid clichés, and business-related expressions in the approval process. The terms in the list are categorized into core terms, invalid clichés, and business expressions, and attribute tags are added to ensure that the statistical analysis can specifically capture key textual information.

[0105] During the statistical process, it is necessary to perform word segmentation and matching on the text in the approval content one by one, record the number of times and frequency of each phrase in the list, and form word frequency distribution information. For example, in the financial reimbursement approval content, the frequency of core terms such as travel expenses and office supplies, as well as the number of times invalid clichés such as "this is to request" appear, are counted. The word frequency distribution can intuitively reflect the textual composition characteristics of the approval content.

[0106] Furthermore, after obtaining word frequency distribution information, the approximate information entropy corresponding to the approved content is calculated based on this information. Specifically, the method of "approximate substitution probability for word or phrase occurrence frequency" is used to substitute into the information entropy formula.

[0107] For example, if the frequency of "travel expenses" is 0.2, the frequency of "office supplies" is 0.15, the frequency of "this application is hereby submitted" is 0.1, and the total frequency of the remaining phrases is 0.55, then the frequency of each phrase is calculated using the formula: "frequency × log2". (1 / 频率) The sum of these terms yields an approximate information entropy.

[0108] Among them, the value of approximate information entropy can quantify the information complexity of the approval content. The higher the value, the more dispersed the distribution of phrases and the more diverse the information types in the approval content. The lower the value, the more focused the content is on a few phrases and the information types are relatively simple, thus making a preliminary judgment on the information characteristics of the approval content.

[0109] Furthermore, the word frequency distribution information is deredundanted based on the attribute tags of the word list, eliminating the word frequency data corresponding to invalid cliché phrases and retaining only the word frequencies corresponding to core terms, business expressions, and other valid information. Then, the deredundant information entropy is calculated based on the deredundant word frequency distribution information. The calculation method is the same as that of approximate information entropy, but only the frequency of valid word phrases is calculated.

[0110] Meanwhile, the ratio of the deredundant information entropy to the approximate information entropy is used as the deredundant coefficient (with the deredundant information entropy as the numerator and the approximate information entropy as the denominator). If the deredundant information entropy is 1.8 and the approximate information entropy is 2.5, then the deredundant coefficient is 0.72 (1.8 / 2.5=0.72).

[0111] Among them, the deredundancy information entropy can reflect the complexity of the effective information in the approval content, while the deredundancy coefficient can reflect the proportion of effective information in the overall information. The higher the deredundancy coefficient, the greater the density of effective information, and the lower the deredundancy coefficient, the higher the proportion of invalid platitudes. These two indicators together constitute the entropy feature to reflect the effective information feature of the approval content.

[0112] Furthermore, after calculating the entropy features, the pre-built embedding vector model is activated to map the approval content into embedding vector features. This embedding vector model needs to be pre-trained with a large amount of text data from the approval domain to ensure accurate capture of the semantic information of the approval text. The vector output by the model must have a uniform dimension, and each dimension in the vector must correspond to a specific semantic feature. For example, one dimension might reflect semantics related to the approval type, while another dimension might reflect semantics related to the amount range.

[0113] During the mapping process, the complete approval content text is input into the embedding vector model. The model converts the text into a fixed-dimensional embedding vector through semantic encoding, so that the semantic information of the approval content is transformed into a computable and comparable vector form, solving the problem that text semantics are difficult to quantify directly.

[0114] Finally, the deduplication information entropy, approximate information entropy, and deduplication coefficient together constitute the entropy feature, which is then integrated with statistical features (word frequency distribution information) and embedded vector features to form a complete multidimensional comparative feature. During integration, it is necessary to ensure that each feature dimension is clear and the data format is consistent.

[0115] For example, word frequency distribution information is converted into vector form, matched with the dimension of embedded vector features, and then combined with numerical data of entropy features to form a multidimensional array. This allows the multidimensional comparison features to reflect the textual composition (statistical features) and effective information density (entropy features) of the approval content, as well as the semantic connotation (embedded vector features), providing comprehensive feature support for subsequent multidimensional reference feature comparison with historical routine approval content and calculation of feature deviation.

[0116] Furthermore, after obtaining the multidimensional comparison features, it is necessary to obtain the historical routine approval content and extract the multidimensional reference features that correspond one-to-one with the multidimensional comparison feature dimensions using the same data feature dimension analysis method as the approval content of the first alternative request set.

[0117] Among them, the selection of historical routine approval content should follow the principle of no abnormalities and full coverage, that is, to screen the rule-based task content that has been successfully approved in the past 1-3 years and has no compliance risk record, and to cover content of different approval types and different business cycles, so as to ensure that it can reflect the true characteristic benchmark of routine rule-based approval content.

[0118] For example, in the area of ​​financial reimbursement, it is necessary to include information on travel reimbursements and office procurement reimbursements within the past 12 months that have no abnormal records; in the area of ​​project approval, it is necessary to include compliant information such as the establishment of small-amount routine projects and the approval of projects within the annual budget within the past 2 years.

[0119] When extracting multidimensional reference features, the analysis process for multidimensional comparative features must be strictly reused. That is, statistical features are obtained by statistically analyzing word frequency distribution information based on the same pre-set word list, and approximate information entropy, deduplicated information entropy, and deduplication coefficient are calculated according to the same formula to form entropy features. The embedded vector features are then mapped through the same pre-built embedding vector model, ultimately forming multidimensional reference features that are completely consistent with the dimensions of the multidimensional comparative features, ensuring the effectiveness of subsequent bias calculations.

[0120] Furthermore, after completing the extraction of multidimensional reference features, the multidimensional feature deviation is calculated based on the two types of features and compared with the preset multidimensional confidence deviation threshold to locate the difference between the current approval content and the historical routine approval content, providing a quantitative basis for determining whether the approval content is abnormal.

[0121] The multidimensional confidence bias threshold is determined based on the distribution pattern of multidimensional reference features of historical routine approval content and the deviation data of past abnormal approval cases. During the setting process, it is necessary to first statistically analyze the numerical distribution range of each dimension in the multidimensional reference features, calculate the standard deviation and mean of each dimension, and use "mean + 1.5 times standard deviation" as the initial threshold reference.

[0122] Furthermore, by combining past approval cases that were deemed abnormal, we analyze the deviation values ​​of these cases in each dimension. If the core term frequency deviation of most abnormal cases exceeds 0.15 and the redundancy coefficient deviation exceeds 0.2, then the thresholds of the corresponding dimensions need to be adjusted to 0.15 and 0.2 to ensure that the multidimensional confidence deviation threshold can effectively distinguish between normal and abnormal deviations.

[0123] For example, the mean of the multidimensional reference feature of travel expense term frequency in the statistical features is 0.2 and the standard deviation is 0.03. The initial threshold is 0.2 + 1.5 × 0.03 = 0.245. However, if in the past approvals where the deviation of travel expense term frequency exceeded 0.245, 90% were normal fluctuations, while approvals with a deviation exceeding 0.26 were abnormal, then the threshold of this dimension needs to be adjusted to 0.26 to improve the accuracy of anomaly detection.

[0124] When calculating the multidimensional feature bias, it is necessary to proceed one by one along the dimension. Specifically, for statistical features, the core term frequency in the word frequency distribution information is selected as the key dimension. The absolute difference between the core term frequency in the current approval content and the average term frequency of the corresponding term in the multidimensional reference features is calculated. If the average term frequency of "within budget" in the multidimensional reference features is 0.18, and the term frequency of "within budget" in the current approval is 0.05, then the bias of this dimension is 0.13 (0.18-0.05=0.13).

[0125] In addition, for the entropy feature, the absolute difference between the current deredundancy information entropy, the deredundancy coefficient and the mean of the corresponding index in the multidimensional reference feature is calculated respectively. If the mean of the deredundancy coefficient in the multidimensional reference feature is 0.78 and the current deredundancy coefficient is 0.52, then the deviation of this dimension is 0.26 (0.78-0.52=0.26).

[0126] The method provided in this application embodiment, which calculates the multidimensional feature deviation based on the multidimensional reference feature and the multidimensional comparison feature, and compares it with a preset multidimensional confidence deviation threshold, further includes:

[0127] Based on the aforementioned multidimensional reference features, a historical routine approval vector baseline is constructed;

[0128] Compare the embedded vector features with the historical regular approval vector baseline, and calculate the mean cosine similarity and nearest neighbor Mahalanobis distance accordingly.

[0129] The average cosine similarity and the nearest neighbor Mahalanobis distance are weighted and fused together, and the weighted fusion result is added to the multidimensional feature bias.

[0130] Specifically, a baseline of historical routine approval vectors is first constructed based on multidimensional reference features. The multidimensional reference features contain a large number of embedding vector features of historical routine approval content. When constructing the baseline, these embedding vectors need to be classified and clustered according to the approval type. For example, the routine embedding vectors of financial reimbursement can be clustered into 3-5 subclasses using the K-means clustering algorithm, with each subclass corresponding to a typical financial reimbursement scenario.

[0131] Furthermore, the mean vector of all embedded vectors in each subclass is calculated, and the obtained mean vector is used as the baseline of the historical regular approval vector under that approval type, so as to form a baseline set covering different approval types and different typical scenarios.

[0132] For example, in the financial reimbursement approval type, the mean vector of the regular embedding vector for the travel reimbursement scenario is calculated as the regular baseline for travel reimbursement; the mean vector of the regular embedding vector for the office procurement reimbursement scenario is calculated as the regular baseline for office procurement reimbursement, so as to ensure that the baseline can accurately reflect the semantic features under different regular scenarios.

[0133] After constructing a baseline of historical routine approval vectors, the embedded vector features of the current approval content are further compared with the corresponding baseline set, and the mean cosine similarity and nearest neighbor Mahalanobis distance are calculated to more accurately analyze the semantic differences between the current approval content and the historical routine approval content.

[0134] Among them, mean cosine similarity can reflect the overall similarity level between the current vector and all regular baseline vectors, while nearest neighbor Mahalanobis distance can focus on the distribution deviation between the current vector and the most similar regular scene. The combination of the two can avoid the one-sided judgment of semantic differences by a single indicator.

[0135] Specifically, when calculating the mean cosine similarity, the cosine similarity between the current embedded vector and all regular baseline vectors of the same approval type in the baseline set needs to be calculated one by one (the value ranges from 0 to 1, and the closer the value is to 1, the more similar the semantics are), and then the arithmetic mean of all similarities is taken as the mean cosine similarity.

[0136] For example, if the current approval is for travel reimbursement, the cosine similarity of the corresponding embedding vector with the regular baseline of travel reimbursement is 0.85 and the cosine similarity with the regular baseline of office procurement reimbursement is 0.32. The average cosine similarity is (0.85+0.32) / 2=0.585. This value can initially reflect the overall similarity between the current vector and the regular semantics.

[0137] In addition, when calculating the nearest neighbor Mahalanobis distance, it is necessary to first determine the regular baseline vector with the highest cosine similarity to the current embedding vector (i.e., the nearest neighbor baseline vector), and then calculate the Mahalanobis distance between the current vector and the nearest neighbor baseline vector based on the covariance matrix of the subclass to which the nearest neighbor baseline vector belongs (the smaller the value, the smaller the deviation of the current vector in the data distribution of that subclass).

[0138] For example, the nearest neighbor baseline vector of the current travel reimbursement embedding vector is the regular travel reimbursement baseline. The covariance matrix of the subclass to which this baseline belongs has been calculated and determined through historical data. After substituting it into the Mahalanobis distance formula, the distance value is 1.2. This value can reflect the deviation of the current vector from the distribution of the most similar regular scenario data.

[0139] Furthermore, after calculating the two indicators, the mean cosine similarity and the nearest neighbor Mahalanobis distance are weighted and fused, and the fusion result is added to the multidimensional feature bias to integrate the judgment results of overall semantic similarity and local data distribution deviation, forming a single and comprehensive embedding vector dimension deviation quantification value.

[0140] Specifically, before fusion, the two indicators need to be normalized. For example, the mean cosine similarity is converted to "1-mean cosine similarity" (the larger the value after conversion, the greater the semantic difference, which is consistent with the trend of Mahalanobis distance). The nearest neighbor Mahalanobis distance is scaled to the 0-1 range according to "Malanobis distance / historical maximum Mahalanobis distance".

[0141] Meanwhile, preset weights are set according to the importance of the two indicators. Generally, semantic similarity has a higher priority than data distribution deviation. Therefore, the weight of "1-mean cosine similarity" can be set to 0.6 and the weight of normalized nearest neighbor Mahalanobis distance can be set to 0.4.

[0142] For example, after normalization, the "1-mean cosine similarity" is 0.415 and the nearest neighbor Mahalanobis distance is 0.3. Then the weighted fusion result is 0.415×0.6+0.3×0.4=0.369. This result is the feature deviation value of the embedded vector dimension. It is included together with the statistical feature deviation and the entropy feature deviation into the multidimensional feature deviation set to form a complete deviation analysis system.

[0143] Furthermore, after calculating the deviation for each dimension, the deviation value for each dimension is compared with the corresponding multidimensional confidence deviation threshold: if the deviation value for a certain dimension is greater than or equal to the threshold for that dimension, it is determined that the dimension is skewed, and the deviation information for that dimension needs to be added to the skew analysis results.

[0144] For example, the core term frequency deviation threshold is 0.15, and the current deviation of 0.13 does not reach the threshold and is therefore not included; the redundancy coefficient deviation threshold is 0.2, and the current deviation of 0.26 ≥ 0.2, so it is included in the results; the embedding vector deviation threshold is 0.38, and the current deviation of 0.42 ≥ 0.38, so it is also included in the results. Ultimately, all dimensional deviations reaching the threshold will be integrated into the anomaly analysis results, providing a clear basis for the subsequent extraction of the third candidate request set.

[0145] Furthermore, based on the obtained anomaly analysis results, the corresponding alternative requests are extracted from the first alternative request set and output as the third alternative request set.

[0146] The extraction process requires establishing a mapping between the first set of candidate requests and the results of the anomaly analysis, assigning a unique identifier to each approval request to ensure accurate matching of its corresponding deviation judgment results across various dimensions. This involves determining whether each request exhibits anomalies such as redundancy coefficient deviation exceeding the threshold, embedding vector deviation exceeding the threshold, or core term frequency deviation exceeding the threshold.

[0147] Specifically, each approval request in the first candidate request set needs to be checked individually. If the anomaly analysis result of a request contains at least one dimensional bias exceeding the threshold, the request is determined to be anomaly and needs to be included in the third candidate request set. For example, for a financial reimbursement request, the anomaly analysis result shows a redundancy coefficient bias of 0.26 (exceeding the threshold of 0.2) and an embedding vector bias of 0.42 (exceeding the threshold of 0.38). Although the core term frequency bias of 0.13 does not exceed the threshold, it still needs to be extracted into the third candidate request set because there are two biases exceeding the threshold.

[0148] In addition, the third set of candidate requests in the final output needs to be classified and labeled. Based on the main abnormality dimensions marked in the abnormality analysis results, a corresponding abnormality type label is added to each request so that content review and risk identification can be carried out in a targeted manner during subsequent large model processing, thereby improving the processing efficiency of complex tasks.

[0149] Ultimately, the resulting third set of alternative requests, along with the second set of alternative requests, must enter the subsequent large-scale model deep processing flow to achieve comprehensive control over approval risks.

[0150] S140: Perform automatic routing approval based on the first alternative request set, the second alternative request set, and the third alternative request set, and send the automatic routing approval result and approval request to the user terminal for verification decision, and output the approval result based on the collected verification decision result.

[0151] In this embodiment of the application, in order to achieve differentiated and efficient processing of different types of approval requests, it is necessary to match the corresponding approval path based on the features of the three types of request sets, and at the same time combine manual verification to ensure the reliability of the approval results, so as to improve the efficiency, accuracy and compliance of the overall approval process.

[0152] Specifically, the three types of request sets are first routed and assigned: the third alternative request set and the second alternative request set are merged and sent together to the pre-trained large model. The large model conducts in-depth review and risk identification of the unstructured content, analyzes the compliance points and potential risk points in the content, and finally generates a structured decision support report with confidence level labels.

[0153] At the same time, the difference between the first alternative request set and the third alternative request set is calculated, and the difference set is determined as the process request set. These requests are all regular tasks without anomalies, and are directly input into the preset process engine. The process engine automatically flows to the corresponding approval position to perform standardized approval operations according to the pre-configured traditional approval node sequence.

[0154] Furthermore, during the process of the workflow engine executing the workflow approval, the approval progress and status need to be monitored in real time. When an abnormal approval event that exceeds the expected range of the workflow engine is identified, the corresponding approval request is immediately retrieved from the traditional approval node sequence and transferred to the pre-trained large model for supplementary approval. The large model combines the abnormal situation with the approval content to provide supplementary review opinions and form a supplementary approval result.

[0155] Furthermore, after completing the above-mentioned automatic routing approval, the decision support report generated by the large model and the supplementary approval results of the process engine are integrated into the automatic routing approval result, which is sent to the user terminal along with the original approval request. The approver then verifies and decides on the automatic approval result to confirm whether to accept the automatic review conclusion and whether further adjustments to the approval opinion are needed.

[0156] Finally, after receiving the verification decision results from the user terminal, the final decision on whether to approve or reject the application is made based on the results, and a complete approval result is output to ensure that the entire approval process balances automation efficiency with the retention of key human control links.

[0157] Step S140 in the method provided in this application embodiment includes:

[0158] The third alternative request set is merged with the second alternative request set and sent to a pre-trained large model for content review and risk identification, generating a structured decision support report with confidence level labels;

[0159] The difference between the first alternative request set and the third alternative request set is determined as the process request set, which is then input into a preset process engine and transferred to a preset traditional approval node sequence for process-oriented approval.

[0160] The process engine monitors the execution of the process approval in real time. When an abnormal approval event that is not within the expected range of the process engine is identified, the corresponding approval request is transferred from the traditional approval node sequence to the pre-trained large model for supplementary approval.

[0161] The output of the supplementary approval results and the decision support report constitutes the automatic routing approval results.

[0162] In this embodiment of the application, in order to enable regular tasks without anomalies to be approved quickly through a standardized process, while allowing complex or abnormal tasks to be reviewed in depth to avoid risks, and at the same time to ensure the reliability of the approval results through manual verification, it is necessary to construct an approval process of "automatic routing + dynamic supplementary review + manual verification" based on the feature matching of the three types of request sets, so as to improve the efficiency and accuracy of the overall approval.

[0163] Specifically, the third and second alternative request sets are first merged, and both types of requests are sent together to a pre-trained large model for content review and risk identification.

[0164] The third set of alternative requests consists of anomalous and regular tasks identified through anomaly analysis in the first set of alternative requests, while the second set consists of unstructured and complex tasks initially screened. Both types of requests require in-depth processing to uncover potential risks.

[0165] Furthermore, the pre-trained large model needs to be trained on a large amount of text data in the approval domain to be capable of identifying compliance loopholes, logical contradictions, and missing key information in unstructured content. During the review process, the large model will perform semantic parsing and risk scanning on each approval content in the request. For example, in contract approval requests, it will identify risk points such as missing breach of contract clauses or payment cycles that exceed the industry norm.

[0166] In addition, the project budget approval request verifies whether the budget items match the project requirements and whether large expenditures have a compliant basis. After the review is completed, the big data model generates a structured decision support report with confidence level markers, which clearly marks the identified risk points, risk levels, and confidence levels, such as "Risk point: The contract does not stipulate a warranty period; Risk level: Medium; Confidence level: 92%", providing a clear basis for subsequent manual verification.

[0167] Simultaneously, the difference between the first and third candidate request sets is calculated, and this difference set is determined as the process request set. These requests are all rule-based tasks that have been screened by rules and confirmed to be anomaly-free through anomaly analysis; they do not require in-depth review and can be directly input into the preset process engine.

[0168] The process engine needs to pre-configure the traditional approval node sequences corresponding to different approval types. For example, the node sequence for financial reimbursement approval is "department review - financial preliminary review - financial final review", and the node sequence for project initiation approval is "department head approval - technical department evaluation - supervisor approval". The process engine will automatically assign requests to the corresponding node sequences according to the approval type of the process request set, and flow them to the respective approval positions in a preset order to execute standardized approval operations.

[0169] For example, travel expense reimbursement requests can be automatically pushed to the review position of the department to which the person seeking reimbursement belongs. After the department approves the request, it will be transferred to the finance department for initial review. There is no need for manual intervention in the allocation of nodes, which greatly improves the approval efficiency of rule-based tasks.

[0170] In addition, during the process of the workflow engine executing the workflow approval, it is necessary to monitor the execution status and progress of the approval in real time, including the processing time of each node, approval opinions, data interaction, etc. When an abnormal approval event that is not within the expected scope of the workflow engine is identified, the corresponding approval request is immediately retrieved from the traditional approval node sequence and transferred to the pre-trained large model for supplementary approval.

[0171] Among these, abnormal approval events include, but are not limited to, a node's processing time exceeding a preset threshold, logical conflicts in approval opinions from different nodes, and the discovery of new unstructured information during the approval process.

[0172] For example, in an office procurement reimbursement process, a request is judged at the initial financial review stage to have an invoice information that does not match the procurement list. This exception exceeds the scope of the automatic approval / rejection logic preset by the process engine and needs to be transferred to the big model for supplementary approval. The big model will re-verify the relationship between the invoice and the procurement list, analyze the reasons for the discrepancy, and generate a supplementary approval result, clearly suggesting rejection and requiring the supplementation of a correct invoice or confirmation that the list entry is incorrect, and agreeing to continue the approval process, etc.

[0173] Furthermore, after completing the above-mentioned automatic routing approval and supplementary approval, the decision support report generated by the large model and the supplementary approval results are integrated into the automatic routing approval result, and then the result is sent to the user terminal (i.e. the operator's terminal) along with the original approval request.

[0174] Specifically, the approver will combine the automated approval results with the original request content to make a verification decision. For example, they will check whether the risk points in the decision support report are accurate and whether the supplementary approval results are reasonable. If the risk point of missing warranty period in the contract identified by the large model is deemed to be true, and the applicant is required to supplement the clause, the approver will submit a verification opinion on the terminal, rejecting the application and noting the modification requirement. In addition, if the approval decision that the checklist entry in the supplementary approval result is correct is deemed to be correct, the approver will submit an approval opinion.

[0175] Finally, after collecting the verification decision results from all user terminals, the final approval result (approved or rejected) for each approval request is determined based on the results, and a complete approval record is generated, including the automatic approval process, manual verification opinions, and final conclusions, to ensure that the entire approval process is traceable and its compliance is verifiable.

[0176] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects:

[0177] This application proposes a hybrid collaborative approval method based on a process engine and a large model. First, it receives approval requests and corresponding data from the business system. Then, relying on a pre-defined rule engine, it parses and separates structured approval metadata from unstructured approval content, while cross-validation ensures that the two types of information are logically consistent. Next, it configures a rule set containing structural feature rules and keyword matching rules to perform a double screening of the approval metadata. Requests that meet the rules are assigned to the first candidate request set, while those that do not are assigned to the second candidate request set. Subsequently, it conducts data feature dimension analysis on the approval content of the first candidate request set, extracting statistical features, entropy features, and embedded vector features to construct multi-dimensional data structures. The system compares the features of the first and third candidate request sets, and simultaneously obtains historical routine approval content and extracts multi-dimensional reference features using the same method. By calculating the deviation between the two types of features and comparing it with a preset confidence threshold, requests with any deviation exceeding the threshold are classified into the third candidate request set. Then, the second and third candidate request sets are merged and sent to a pre-trained large model to generate a decision support report with confidence. The difference between the first and third candidate sets is used as the process request set and input into the process engine. Standardized approval is executed according to a preset node sequence, while approval anomalies are monitored and the large model is triggered to supplement the approval. Finally, the automatic approval result and the original request are sent to the user terminal for verification, and the final approval result is output based on the feedback.

[0178] The method provided in this application, through the technical solution of "information extraction - rule screening - abnormality analysis - intelligent routing - manual verification", solves the problems of low efficiency, inaccurate risk identification, and insufficient intelligence caused by the inability of traditional approval systems to effectively handle unstructured complex tasks and mixed processing of rule-based and abnormal tasks. It improves the efficiency, quality, and compliance of complex business approvals and provides suitable technical support for multiple scenarios such as financial reimbursement, project initiation, and contract approval.

[0179] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of the hybrid collaborative approval method based on process engine and large model provided in Embodiment 1, this application also provides a hybrid collaborative approval system based on process engine and large model, specifically including:

[0180] The request information extraction module 01 is used to obtain approval requests, extract approval content and approval metadata;

[0181] Metadata screening and classification module 02 is used to perform rule-based screening and discrimination on the approval metadata based on a predetermined set of rules, and divide the approval request into a first candidate request set and a second candidate request set;

[0182] The content anomaly analysis module 03 is used to perform anomaly analysis on the approval content corresponding to the first candidate request set based on data feature dimensions, and extract the third candidate request set based on the anomaly analysis results.

[0183] The routing approval output module 04 is used to automatically approve routes based on the first alternative request set, the second alternative request set, and the third alternative request set, and send the automatic routing approval result and approval request to the user terminal for verification and decision-making, and output the approval result based on the collected verification and decision-making results.

[0184] In one embodiment, the request information extraction module 01 is further configured to:

[0185] The system receives approval requests and corresponding approval request data from the business system; it parses the approval requests and approval request data based on a preset rule engine, extracts structured data fields as the approval metadata; and it extracts unstructured text content from the approval request data as the approval content based on the preset rule engine.

[0186] In one embodiment, the metadata screening and classification module 02 is also used for:

[0187] Configure the rule set containing structural feature rules and keyword matching rules; match the approval metadata with the structural feature rules to identify matching approval requests and store them in the first candidate request set; match the approval metadata with the keyword matching rules to identify approval requests containing preset keywords and store them in the first candidate request set; output the approval requests not stored in the first candidate request set as the second candidate request set.

[0188] In one embodiment, the content deviation analysis module 03 is also used for:

[0189] Data feature dimension analysis is performed on the approval content in the first candidate request set to obtain multi-dimensional comparison features, wherein the multi-dimensional comparison features include statistical features, entropy features, and embedding vector features; historical routine approval content is obtained, and corresponding data feature dimension analysis is performed to extract multi-dimensional reference features, wherein the dimensions of the multi-dimensional reference features and the multi-dimensional comparison features correspond one-to-one; multi-dimensional feature deviation is calculated based on the multi-dimensional reference features and the multi-dimensional comparison features, and compared with a preset multi-dimensional confidence deviation threshold; any multi-dimensional feature deviation whose dimension is greater than or equal to the multi-dimensional confidence deviation threshold is added to the anomaly analysis result; according to the anomaly analysis result, the corresponding candidate requests are extracted from the first candidate request set and output as the third candidate request set.

[0190] Furthermore, the content deviation analysis module 03 also includes:

[0191] Based on a preset list of phrases, statistical analysis is performed on the approval content to obtain statistical features, wherein the statistical features include at least word frequency distribution information; based on the word frequency distribution information, the approximate information entropy corresponding to the approval content is calculated; based on the attribute tags of the phrase list, the word frequency distribution information is deredundant-filtered, and based on the deredundant-filtering results, the deredundant information entropy corresponding to the approval content is calculated, and the ratio of the deredundant information entropy to the approximate information entropy is output as the deredundancy coefficient; a pre-constructed embedding vector model is activated to map the approval content into embedding vector features; the deredundant information entropy, the approximate information entropy, and the deredundancy coefficient are used as the entropy features, and the statistical features and the embedding vector features are merged to output the multidimensional comparison features.

[0192] Furthermore, the content deviation analysis module 03 also includes:

[0193] Based on the multidimensional reference features, a historical routine approval vector baseline is constructed; the embedded vector features are compared with the historical routine approval vector baseline, and the mean cosine similarity and nearest neighbor Mahalanobis distance are calculated accordingly; the mean cosine similarity and nearest neighbor Mahalanobis distance are weighted and fused, and the weighted fusion result is added to the multidimensional feature deviation.

[0194] In one embodiment, the routing approval output module 04 is also used for:

[0195] The third alternative request set is merged with the second alternative request set and sent to a pre-trained large model for content review and risk identification, generating a structured decision support report with confidence level labels. The difference between the first and third alternative request sets is determined as the process request set, which is input into a preset process engine and flows to a preset traditional approval node sequence for process-oriented approval. The process engine monitors the execution of process-oriented approval in real time. When an approval anomaly event that does not fall within the expected range of the process engine is identified, the corresponding approval request is transferred from the traditional approval node sequence to the pre-trained large model for supplementary approval. The supplementary approval result and the decision support report are output as the automatic routing approval result.

[0196] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0197] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0198] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A hybrid collaborative approval method based on a process engine and a large model, characterized in that, The method comprises the following steps: Obtaining an approval request, extracting the approval content and approval metadata; Based on a predetermined rule cluster, the approval metadata is rule-based, screened and distinguished, and the approval request is divided into a first alternative request set and a second alternative request set; Based on the data feature dimension, the approval content corresponding to the first alternative request set is analyzed for abnormality, and a third alternative request set is extracted based on the abnormality analysis result; According to the first alternative request set, the second alternative request set and the third alternative request set, automatic routing approval is carried out, and the automatic routing approval result and the approval request are sent to the user terminal for verification decision, and the approval result is output according to the recovered verification decision result; Wherein, based on a predetermined rule cluster, the approval metadata is rule-based, screened and distinguished, and the approval request is divided into a first alternative request set and a second alternative request set, comprising: The rule cluster includes structure feature rules and keyword matching rules; The approval metadata is matched with the structure feature rules to identify the approval requests that meet the requirements and store them in the first alternative request set; The approval metadata is matched with the keyword matching rules to identify the approval requests containing the preset keywords and store them in the first alternative request set; The approval requests not stored in the first alternative request set are output as the second alternative request set Wherein, based on the data feature dimension, the approval content corresponding to the first alternative request set is analyzed for abnormality, and a third alternative request set is extracted based on the abnormality analysis result, comprising: Data feature dimension analysis is performed on the approval content in the first alternative request set to obtain multi-dimensional comparison features, wherein the multi-dimensional comparison features include statistical features, entropy features and embedded vector features; Historical regular approval content is obtained and corresponding data feature dimension analysis is performed to extract multi-dimensional reference features, wherein the multi-dimensional reference features correspond one-to-one with the multi-dimensional comparison features; Based on the multi-dimensional reference features and the multi-dimensional comparison features, multi-dimensional feature deviations are calculated and compared with a preset multi-dimensional confidence deviation threshold, and any feature deviation dimension greater than or equal to the multi-dimensional confidence deviation threshold is added to the abnormality analysis result; According to the abnormality analysis result, the corresponding alternative request is extracted from the first alternative request set and output as the third alternative request set; Wherein, according to the first alternative request set, the second alternative request set and the third alternative request set, automatic routing approval is carried out, comprising: Merge the third alternative request set and the second alternative request set, send to the pre-trained large model for content review and risk identification, and generate a structured decision support report with confidence label; The difference set of the first alternative request set and the third alternative request set is determined as a process request set, which is input to a preset process engine and transferred to a preset traditional approval node sequence for process approval; Through the process engine, the execution of the process approval is monitored in real time, and when an approval abnormal event that does not belong to the expected range of the process engine is identified, the corresponding approval request is transferred from the traditional approval node sequence to the pre-trained large model for supplementary approval; The output supplementary examination result and the decision support report are the automatic routing examination result.

2. The hybrid collaborative approval method based on a process engine and a large model according to claim 1, characterized in that, An examination request is obtained, and examination content and examination metadata are extracted, including: An examination request and corresponding examination request data are received from a business system; The examination request and the examination request data are parsed based on a preset rule engine, and structured data fields are extracted as the examination metadata; Unstructured text content in the examination request data is extracted as the examination content based on the preset rule engine.

3. The hybrid collaborative approval method based on a process engine and a large model according to claim 1, characterized in that, Data feature dimension analysis is performed on the examination content in the first alternative request set to obtain multi-dimensional comparison features, wherein the multi-dimensional comparison features include statistical features, entropy features, and embedding vector features, including: Statistical analysis is performed on the examination content based on a preset phrase list to obtain statistical features, wherein the statistical features at least include word frequency distribution information; Approximate information entropy corresponding to the examination content is calculated according to the word frequency distribution information; The word frequency distribution information is de-duplication filtered according to the attribute labels of the phrase list, and a de-duplication information entropy corresponding to the examination content is calculated based on the de-duplication filtering result, and a ratio of the de-duplication information entropy and the approximate information entropy is output as a de-duplication coefficient; A pre-constructed embedding vector model is activated to map the examination content to an embedding vector feature; The de-duplication information entropy, the approximate information entropy, and the de-duplication coefficient are taken as the entropy features, and the statistical features and the embedding vector features are merged to output the multi-dimensional comparison features.

4. The hybrid collaborative approval method based on a process engine and a large model according to claim 1, characterized in that, Multi-dimensional feature deviations are calculated based on the multi-dimensional reference features and the multi-dimensional comparison features, and compared with a preset multi-dimensional confidence deviation threshold, further including: A historical regular examination vector baseline is constructed based on the multi-dimensional reference features; The embedding vector feature is compared with the historical regular examination vector baseline to calculate an average cosine similarity and a nearest neighbor Mahalanobis distance; The average cosine similarity and the nearest neighbor Mahalanobis distance are weighted and fused, and the weighted fusion result is added to the multi-dimensional feature deviation.

5. A hybrid collaborative approval system based on a process engine and a large model, characterized in that, The system is used to perform the hybrid collaborative examination method based on the process engine and the large model according to any one of claims 1-4, and the system includes: A request information extraction module for obtaining an examination request and extracting examination content and examination metadata; A metadata screening and classification module for screening and classifying the examination metadata based on a predetermined rule cluster to divide the examination request into a first alternative request set and a second alternative request set; A content abnormality analysis module for performing data feature dimension-based abnormality analysis on the examination content corresponding to the first alternative request set and extracting a third alternative request set based on the abnormality analysis result; A routing examination output module for automatically routing examination according to the first alternative request set, the second alternative request set, and the third alternative request set, and sending the automatic routing examination result and the examination request to a user terminal for verification decision, and outputting an examination result according to the recovered verification decision result.

Images