Big model data governance method based on cross-domain semantic drift

Abstract

The present disclosure provides a large model data governance method based on cross-domain semantic drift, which comprises constructing a cross-industry approval responsibility chain structure based on process actions, responsibility roles and transmission relationships in cross-industry approval data, generating a cross-industry approval responsibility chain graph and a process role position mapping set; responsibility position analysis and stage matching are performed on the process words in the newly added training data, and a process role semantic deviation set is identified; the process words are marked with responsibility chain attachment, split by industry, and isolated by stage, and a responsibility chain attachment semantic matrix is generated; the newly added approval data is subjected to real-time responsibility chain positioning, attribution verification and order correction, and a dynamic responsibility governance result is generated; the approval data after governance is subjected to responsibility chain consistency review, process role final state verification and reasoning stability detection, and the cross-industry approval semantic governance result is output, realizing stable governance of the responsibility chain of cross-industry approval semantics.

Description

Technical Field

[0001] This disclosure relates to the fields of data processing and artificial intelligence technology, and in particular to a method for governing large model data based on cross-domain semantic drift. Background Technology

[0002] In joint training scenarios involving multiple industry approval systems, such as large enterprise resource planning (ERP) approval systems and hospital consultation approval systems, large models often face the technical flaw of semantic inversion inheritance of process roles. Specifically, while process terms like "rejection," "review," and "confirmation" may appear structurally consistent across different industry approval chains, their actual responsible parties and semantic stages are completely different. For example, in a large enterprise ERP approval system, "confirmation" might represent the completion of financial settlement, belonging to the final state semantics; while in a hospital consultation approval system, "confirmation" might represent the risk requiring further observation, belonging to the intermediate state semantics. Traditional data governance solutions can establish industry semantic boundaries, but these boundaries are mainly constructed around static semantic attribution, making it difficult to handle the problem of role reversal for the same process term in different responsibility chains. When large models are jointly trained during cross-industry process reasoning, they can easily inherit the incorrect final state semantics from the industrial approval chain into the medical consultation chain, causing the model to prematurely infer the intermediate confirmation state as a completed state in the medical scenario, leading to confused reasoning logic and misjudgment of approval results. This phenomenon is particularly prominent in cross-domain joint training of multi-stage approval flows, seriously affecting the reliability and accuracy of large models in actual business.

[0003] Therefore, there is an urgent need for a large model data governance method based on cross-domain semantic drift to solve the problem of reversed inheritance of process role semantics caused by the different role positions of the same process semantics in different responsibility chains in the joint training scenario of multi-stage approval flow, and to achieve stable governance of cross-industry approval semantic responsibility chains. Summary of the Invention

[0004] In view of this, in order to solve the problems brought about by the existing technology, this application provides a large model data governance method based on cross-domain semantic drift.

[0005] Firstly, this disclosure provides a method for governing large model data based on cross-domain semantic drift, the method comprising: S1. Based on the process actions, responsible roles and transmission relationships in cross-industry approval data, construct a cross-industry approval responsibility chain structure, and generate a cross-industry approval responsibility chain map and a set of process role location mappings. S2. Based on the cross-industry approval responsibility chain graph and process role position mapping set, the responsibility position is parsed and the stage is matched for the process words in the newly added training data to identify the process role semantic offset set. S3. Perform responsibility chain dependency marking, industry affiliation splitting and stage isolation processing on the process words in the process role semantic offset set to generate a responsibility chain dependency semantic matrix. S4. Based on the responsibility chain dependency semantic matrix, perform real-time responsibility chain positioning, attribution verification and order correction on the newly added approval data to generate dynamic responsibility governance results. S5. Based on the dynamic responsibility governance results, perform responsibility chain consistency verification, process role final state verification, and reasoning stability detection on the governed approval data, and output cross-industry approval semantic governance results.

[0006] Optionally, S1 includes: Acquire cross-industry approval data and divide the cross-industry approval data into a set of semantically coherent industry approval process segments according to the approval items; Each node in the industry approval process segment set is marked with a responsibility role, forming a responsibility role marked process set; Based on the set of responsibility role marking processes, the responsibility transfer relationship between adjacent nodes under the same approval item is associated to generate a set of responsibility transfer chains; Based on the set of responsibility transfer chains, the responsibility positions of the same process words in different approval chains are mapped and compared, and the degree of overlap of common responsibility attributes is counted to generate the cross-industry approval responsibility chain map and the process role position mapping set.

[0007] Optionally, S2 includes: Based on the aforementioned process role location mapping set, the responsibility location of process terms in the newly added approval data is parsed to generate parsing results; Based on the analysis results and the industry standard stage sequence in the cross-industry approval responsibility chain diagram, the responsibility position sequence of the newly added approval process is matched with the industry standard responsibility chain stage to obtain the semantic stage matching result of the newly added process. When the stage matching result is a partial match, the responsibility advancement direction of the same process words in different approval systems is compared to obtain the result of the difference in responsibility advancement direction. Based on the results of the differences in the direction of responsibility advancement, a role offset propagation chain is established for the process semantics where there is a trend of responsibility position reversal; Based on the role offset propagation chain, risk statistics are performed on the inheritance direction change status of process roles in different approval systems to obtain role inversion risk scores, and the process role semantic offset set is generated based on the role inversion risk scores.

[0008] Optionally, the step of matching the responsibility position sequence of the newly added approval process with the stages of the industry standard responsibility chain to obtain the semantic stage matching result of the newly added process includes: The sequence of responsibility positions for process terms in the newly added approval process is compared with the industry standard stage sequence in the cross-industry approval responsibility chain diagram; The matching status is determined based on a preset threshold, and the matching status includes good stage matching, partial matching, and abnormal stage matching.

[0009] Optionally, S3 includes: Extract process words from the process role semantic offset set, associate them with the role offset propagation chain, calculate the responsibility chain dependency strength of each process word with respect to the role offset propagation chain, and attach and bind the process words to the role offset propagation chain according to the responsibility chain dependency strength to generate a dependency tag set. Based on the aforementioned dependency tag set, the industry affiliation status of the same process term in different responsibility stages is split, the necessity degree of industry affiliation splitting is calculated, and the industry affiliation splitting result is generated. Based on the industry affiliation decomposition results, a final state inheritance restriction region is established for the process semantics at the final state responsibility position; Based on the final state inheritance restriction region, a stage isolation region is established for the process semantics in the intermediate stage, and a process stage isolation result is generated. The dependency mark set, industry affiliation splitting results, final state inheritance restriction region, and process stage isolation results are summarized to generate the responsibility chain dependency semantic matrix.

[0010] Optionally, the step of splitting the industry affiliation status of the same process term in different responsibility stages based on the attachment tag set, calculating the necessity degree of industry affiliation splitting, and generating industry affiliation splitting results includes: Calculate the industry affiliation necessity for each process term; Based on whether the necessity of industry classification exceeds a preset threshold, determine whether to perform industry classification for the process term; When performing the split, the process term is used as the main index, and its responsibility chain attachment mark in different industries is split into at least one of the following categories according to the responsibility stage: final confirmation, intermediate confirmation, backtracking and review, observation confirmation, or archiving completion. A clear industry affiliation label and stage affiliation label are set for each type of split result, and the process term is retained as the minimum shared attribute of approval action term.

[0011] Optionally, S4 includes: For each batch of newly added approval data entering the training system, the extracted process words are compared with the responsibility chain dependency semantic matrix, real-time responsibility chain positioning is performed, and real-time responsibility chain positioning results are generated. Based on the real-time responsibility chain location results, the responsibility stage attribution of newly added process words is verified, process words with questionable attribution are marked, and responsibility stage attribution verification results are generated. Based on the responsibility stage attribution verification results, the semantics of processes with role inversion trends are dynamically adjusted, the industry inheritance scope is adjusted, and the industry inheritance scope adjustment results are generated. Based on the industry inheritance scope adjustment results, the consistency verification of the responsibility advancement order in the cross-industry approval reasoning results is performed, and the responsibility advancement order verification results are generated. Based on the results of the responsibility advancement sequence verification, the approval semantics that show abnormal responsibility sequence are re-executed in the isolation correction stage to generate the dynamic responsibility governance results.

[0012] Optionally, S5 includes: Based on the dynamic responsibility governance results, the responsibility chain consistency of the governed approval data is reviewed, and a responsibility chain consistency score is calculated. Based on the chain of responsibility consistency score, the final state verification of the process role attribution status in different approval systems is performed, and the credibility of the final state of the process role is calculated. Based on the credibility of the final state of the process roles, the stability of the cross-industry approval reasoning results is detected, and an approval reasoning stability score is calculated. Based on the stability score of the approval reasoning, the abnormal responsibility propagation path of the reasoning chain with the role inversion inheritance trend is backtracked, the backtracking strength is calculated, and the cross-industry approval semantic governance results are output.

[0013] In a second aspect, this disclosure provides an electronic device including a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the method of the first aspect described above.

[0014] Thirdly, this disclosure provides a computer storage medium storing a computer program that, when executed, implements the method described in the first aspect.

[0015] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following advantages: By constructing a cross-industry approval responsibility chain structure, identifying semantic offsets in process roles, isolating semantic dependencies in the execution responsibility chain, and implementing dynamic governance, the problem of semantic inversion inheritance caused by the reversed role positions of the same process terms in different responsibility chains is effectively solved. Specifically, a responsibility chain graph and position mapping are established, risks are quantified and an isolation matrix is ​​generated, and new data is located, verified, and anomaly corrected in real time, ultimately outputting stable governance results. These methods enable the same process terms to be accurately bound to their respective real responsibility stages in different industries, avoiding the erroneous inheritance of industrial final-state semantics to medical intermediate-state semantics, thereby significantly improving the inference reliability and approval accuracy of the large model in cross-industry joint training of approval processes. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0017] Figure 1 A flowchart of a large model data governance method based on cross-domain semantic drift provided in an embodiment of this disclosure is shown. Figure 2 A flowchart of process role semantic offset identification and risk scoring provided in an embodiment of this disclosure is shown.

[0018] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0019] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and should not be used to limit the scope of protection of the present disclosure.

[0020] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0022] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0023] Figure 1 A flowchart of the large model data governance method based on cross-domain semantic drift provided in this disclosure embodiment is shown below. Figure 1 As shown, the process may include the following steps: S1: Based on the process actions, responsible roles and transmission relationships in cross-industry approval data, construct a cross-industry approval responsibility chain structure, and generate a cross-industry approval responsibility chain map and a set of process role position mappings.

[0024] S1.1: Obtain cross-industry approval data and divide the cross-industry approval data into a set of semantically coherent industry approval process segments according to the approval items.

[0025] Acquire enterprise ERP approval data, hospital consultation approval data, and other industry approval data. This data typically includes approval record timestamps, approval document numbers, and approval operation texts. For ease of processing, all approval records are merged according to the approval document number, ensuring that all operation records under the same approval item are grouped into the same process unit. The operation records in each process unit are then sequentially arranged based on the approval record timestamps, forming an initial process sequence from initiation, review, verification, confirmation to archiving. Based on this, process action words, such as "submit," "reject," "verify," "confirm," and "archive," are extracted from the approval operation text and marked as process actions.

[0026] Furthermore, by combining the job information, department, and operation time of the approvers, operation records that occur consecutively and belong to the same approval item are merged into a set of industry approval process segments. Each segment in this set corresponds to a complete approval process, providing a structured data foundation for subsequent role assignment.

[0027] S1.2: Mark the nodes in each approval process segment of the industry approval process segment set with responsibility roles to form a responsibility role marked process set.

[0028] Based on the obtained set of industry approval process fragments, responsibility roles are assigned to approval nodes in each process fragment. The specific marking rules are as follows: the node that submits the approval request for the first time is marked as the initiating node; the node that performs business compliance checks is marked as the review node; the node that confirms the review results a second time is marked as the verification node; the node that determines whether the process enters the completion, observation, return, or termination status is marked as the confirmation node; and the node that creates an unmodifiable record is marked as the archiving node.

[0029] For the different meanings of the same process term across different industries, instead of directly labeling it according to its literal meaning, the labeling is determined by combining information about the responsible party, the type of preceding nodes, the type of succeeding nodes, and the node's processing result. For example, in an enterprise ERP system, if the "confirmation" node directly leads to the financial accounting process, then that node is marked as a final confirmation node; while in a hospital consultation system, if the "confirmation" node leads to an observation or supplementary examination stage, then that node is marked as an intermediate confirmation node. Each node is assigned a label reflecting its true responsibility role, thus forming a set of process terms labeled with responsibility roles. This set binds process terms to specific responsibility stages, laying the data foundation for subsequent identification of semantic inversion of process roles.

[0030] S1.3: Based on the set of responsibility role marking processes, associate the responsibility transfer relationship between adjacent nodes under the same approval item to generate a set of responsibility transfer chains.

[0031] After obtaining the set of responsibility role-marked processes, a correlation analysis is performed on the responsibility transfer relationships between adjacent nodes under the same approval item. The preceding and subsequent responsibility roles in each pair of adjacent nodes are identified, and the transfer type of the node pair is determined, such as initiation to review, review to verification, verification to confirmation, confirmation to archiving, or a rollback relationship like rejection, correction, or resubmission. Sequential progression relationships, rollback correction relationships, and termination archiving relationships are marked as different responsibility transfer types. Through these markings, a set of responsibility transfer chains is obtained, which depicts the ordered transfer paths between responsibility roles in each approval process.

[0032] Furthermore, potential role reversals that may occur during the transfer of responsibility are pre-marked. For example, if "confirmation" in one industry is followed by archiving, while "confirmation" in another industry still leads to review or observation, this difference is recorded as a potential difference in responsibility position.

[0033] S1.4: Based on the set of responsibility transfer chains, the responsibility positions of the same process words in different approval chains are mapped and compared, and the degree of overlap of common responsibility attributes is counted to generate the cross-industry approval responsibility chain map and the process role position mapping set.

[0034] Based on the obtained set of responsibility transfer chains, the responsibility positions of the same process words in different approval chains are mapped and compared. Using process words as indices, the occurrence positions of these words in enterprise ERP approval chains, hospital consultation approval chains, and other industry approval chains are extracted. The preceding responsibility role, subsequent responsibility role, and processing result corresponding to the process word in each industry are statistically analyzed. The responsibility stages of the same process word in different industries are compared, distinguishing them into final responsibility positions, intermediate responsibility positions, fallback responsibility positions, and observation responsibility positions. For process words with obvious position reversals, position mapping records are established. For example, the final state "confirmation" in ERP and the intermediate state "confirmation" in hospital consultation are saved as different role positions. Finally, a cross-industry approval responsibility chain map and a set of process role position mappings are generated. This map and mapping set explicitly record the problem of "same process word, different responsibility positions," preventing the semantic errors of the final state in industrial approvals from being inherited into the intermediate state semantics of medical consultations at the source.

[0035] Simultaneously, for each process term, based on its responsibility position and processing outcome in various industries, the degree of overlap in shared responsibility attributes is statistically analyzed and recorded as follows: , The value range is [0,1], which can be obtained by statistically analyzing the percentage of approval records with the same responsible role, the same preceding and following nodes, and the same processing result in various industries.

[0036] Through the above sub-steps, step S1 completes the construction of the cross-industry approval responsibility chain structure, providing systematic basic data and reference standards for subsequent semantic offset recognition and semantic isolation of responsibility chain attachment.

[0037] S2: Based on the cross-industry approval responsibility chain graph and process role position mapping set, the responsibility position parsing and stage matching of process words in the newly added training data are performed to identify the process role semantic offset set.

[0038] S2.1: Based on the process role position mapping set, perform responsibility position parsing on the process words in the newly added approval data and generate parsing results.

[0039] Figure 2 A flowchart of the process role semantic offset identification and risk scoring provided in an embodiment of this disclosure is shown, such as... Figure 2As shown, for each approval record in the newly added training data, they are merged according to the approval item number to form a new approval process unit. Process terms are extracted from each process unit, such as "submit," "reject," "review," "confirm," "archive," "correct," and "observe." Each extracted process term is compared with the process role position mapping set obtained in step S1 to determine the candidate responsibility position corresponding to that process term in the existing industry approval responsibility chain graph. Combining the time sequence of approval nodes, the role information of the approval entity, and subsequent processing results, the specific responsibility position to which the process term belongs in the current newly added data is determined, such as initiation position, review position, review position, confirmation position, archiving position, rollback position, or observation position.

[0040] To quantify this judgment process, a responsibility location analysis score is introduced. For each candidate responsibility position, its corresponding [position] is calculated according to the following formula. value: ; in, This indicates the newly added process term responsibility location analysis and scoring; This indicates the matching weight between process terms and existing responsibility positions, with a value ranging from 0.2 to 0.4. This indicates the number of times the process term appears in the historical chain of responsibility graph; This represents the credibility weight of the main role, with a value ranging from 0.2 to 0.4. This indicates the degree of matching between the approval body and the candidate's responsibility position, which is jointly determined by the position, department, and approval authority. Specifically, if the position and candidate position requirements are consistent, 0.5 points are awarded; if the department is consistent, 0.3 points are awarded; and if the authority is consistent, 0.2 points are awarded. The sum of the three items is normalized to [0,1]. This indicates the clarity of the subsequent processing results. If the subsequent operation is archiving or accounting, it is 1.0; if it is review or observation, it is 0.6; if it is correction, it is 0.4; and if there is no subsequent operation or it is rejection, it is 0.2.

[0041] In actual implementation, the calculation is performed separately for each candidate responsibility position. Value, and take the maximum. The candidate position corresponding to the value is used as the responsibility position parsing result for this process term, and the responsibility stage name of this candidate position is recorded as the position label for this process term. If the maximum Value below preset threshold For example, if the value is 0.5, the process term is marked as having an ambiguous responsibility location and is submitted for manual review or handled by a trigger rule. The above responsibility location parsing results and location tags are used as the core content of the new process semantic responsibility location parsing results. If there are ambiguous responsibility location markers, they are also included. This result will be passed to sub-step S2.2 for further processing.

[0042] S2.2: Based on the parsing results and the industry standard stage sequence in the cross-industry approval responsibility chain diagram, the responsibility position sequence of the newly added approval process is matched with the industry standard responsibility chain stage to obtain the semantic stage matching result of the newly added process.

[0043] Based on the obtained results of the semantic responsibility position analysis of the newly added process, responsibility chain stage matching is performed for each newly added approval process unit. According to the actual order of occurrence of process words in the approval items, a sequence of newly added responsibility positions is formed. This sequence is compared with the industry standard stage order in the cross-industry approval responsibility chain map established in step S1 to identify which type of stage structure the newly added process semantic belongs to in the enterprise ERP approval chain, hospital consultation approval chain, or other industry approval chain.

[0044] Furthermore, the actual stage position of the same process term in the newly added process is confirmed. For example, if a certain "confirmation" process term is located before financial accounting in the enterprise ERP, it is matched as the final confirmation stage; if it is located before supplementary examination in a hospital consultation, it is matched as the intermediate observation confirmation stage.

[0045] To quantitatively assess the degree of matching, a new semantic stage matching degree is introduced. The calculation formula is as follows: ; in, Indicates the semantic matching degree of the newly added process stage; This indicates the number of nodes in the newly added approval process; Indicates the first One approval node; Indicates the first The reliability of the responsibility position of each node is obtained by normalizing the responsibility position parsing score in sub-step S2.1; Indicates the first The degree of alignment between each node and the sequence of stages in the target industry's responsibility chain; Indicates the first The degree of deviation of each node from the target industry standard responsibility stage is determined by factors such as being ahead of schedule, behind schedule, duplicated, or missing.

[0046] The higher the value, the better the overall stage of the new process matches the target industry standard chain of responsibility stage. This result is directly output as the semantic stage matching result of the new process. When the matching result is obtained, the process is marked as a well-matched stage; when When it is marked as a partial match, the directional comparison needs to be strengthened in the subsequent step S2.3; when When this occurs, it is marked as a stage exception, triggering the chain of responsibility reconstruction process, which involves returning to step S1 to rebuild the cross-industry approval chain of responsibility structure. The above matching results and markings are the semantic stage matching results of the newly added process, and serve as the input for step S2.3.

[0047] S2.3: When the stage matching result is a partial match, compare the responsibility advancement direction of the same process words in different approval systems to obtain the result of the difference in responsibility advancement direction.

[0048] For the newly added process semantics marked as partially matched in step S2.2, the responsibility progression direction of the same process term in different approval systems is compared based on its stage matching results. Specifically, using the process term as the retrieval object, the preceding and subsequent responsibility stages in different industry approval chains and new approval chains are extracted to determine whether the responsibility progression direction corresponding to the process term is from the intermediate stage to the final stage, or from the confirmation stage to the review, observation, or correction stage. The responsibility progression directions of the same process term in different industries are compared side by side. For process terms with opposite directions, reversed stage relationships, or mixed use of final and intermediate stages, a record of direction differences is established.

[0049] To quantify this difference, a scoring system for the difference in the direction of responsibility advancement is introduced. The calculation formula is as follows: ; in, The scoring indicates the differences in the direction of responsibility and progress of process terms; This represents the stage reversal weight, with a value ranging from 0.3 to 0.5; Normalization strength indicates whether the same process term has undergone a stage reversal in different industries; This represents the normalized value of the frequency of reversals; This represents the divergence weight of subsequent results, with a value ranging from 0.2 to 0.4; Indicates the degree of difference in subsequent results for the same process term, such as subsequent archiving in one industry and subsequent observation in another; Indicates the degree of overlap in common subsequent results; This represents the weighting of the differences in principal responsibility, with a value ranging from 0.2 to 0.4. This indicates the degree of difference in the responsible parties.

[0050] This formula simultaneously evaluates whether the stage is reversed, whether the results are divergent, and whether the subjects are different, and amplifies the directional differences caused by significant differences in the responsible subjects through the squared term. Based on the calculation... The value is determined as follows: If If the direction is significantly different, the process word is marked as a candidate offset word, and the determination result is entered into the offset propagation chain construction process in sub-step S2.4; if Marked as having slight directional differences, these will not proceed to the offset propagation chain construction process for now; if They assumed the direction was consistent and did not make any special adjustments.

[0051] S2.4: Based on the results of the difference in the direction of responsibility advancement, establish a role offset propagation chain for the process semantics where there is a trend of responsibility position reversal.

[0052] For process words identified as having significant directional differences and marked as candidate offset words in step S2.3, a role offset propagation chain is established. Specifically, the preceding responsibility stage is traced forward along the approval process where the offset word is located, and the subsequent responsibility stage is traced backward, forming a local propagation segment consisting of a preceding node, an offset node, and a subsequent node. Multiple local propagation segments are then concatenated according to the approval item number, industry source, and responsibility stage order to form a complete role offset propagation chain. For recurring offset words, cross-industry propagation relationships are established in the propagation chain, such as the abnormal link where the final state "confirmation" in an enterprise ERP system is inherited by the intermediate state "confirmation" in a hospital consultation.

[0053] To measure the cumulative strength of the entire propagation chain, the role offset propagation chain strength is introduced. The calculation formula is as follows: ; in, This represents the strength of the role offset propagation chain, and is dimensionless. Indicates the number of offset segments in the propagation chain, in segments; Indicates the first One offset segment; Indicates the first The cross-industry propagation weight of each offset segment is dimensionless and is determined by the industry span, the degree of overlap of process terms, and the difference in responsibility stages. Indicates the first The responsibility advancement direction difference score for each offset segment, dimensionless, is obtained from sub-step S2.3; Indicates the first The duration of each offset segment that appears continuously within the training increment period, in hours; This represents the standard observation duration adjustment constant, ranging from 1 hour to 168 hours, used to normalize the time term; Indicates the first The degree to which an offset segment is absorbed by the correct industry boundary is dimensionless.

[0054] The formula uses the duration term to indicate whether the offset is long-lasting, and the degree of absorption in the denominator to suppress segments that have already been absorbed by the correct boundary. According to... The range of values ​​is handled as follows: If Mark the propagation chain as a strong offset chain; if Marked as the middle offset chain; if This is considered noise and is not output as part of the formal offset propagation chain. The final output character offset propagation chain only contains the strong offset chain and the medium offset chain.

[0055] S2.5: Based on the role offset propagation chain, perform risk statistics on the inheritance direction change status of process roles in different approval systems to obtain role inversion risk scores, and generate the process role semantic offset set based on the role inversion risk scores.

[0056] Based on the role offset propagation chain output in step S2.4, the changes in the inheritance direction of process roles in different approval systems are statistically analyzed. For each role offset propagation chain, the offset terms, source industry, target industry, preceding responsibility stage, subsequent responsibility stage, and offset persistence status are summarized. The propagation chain strength, responsibility advancement direction difference score, and matching degree of the newly added process semantic stage are comprehensively ranked, and low-risk, medium-risk, and high-risk offsets are respectively classified into different risk ranges. For high-risk offset chains, their inverted inheritance direction is explicitly marked, for example, "propagation from the final confirmation in the enterprise ERP system to the intermediate confirmation in the hospital consultation."

[0057] To comprehensively quantify the role inversion risk of a specific process term, a role inversion risk score is introduced. The calculation formula is as follows: ; in, This indicates a risk score related to role reversal. This represents the propagation chain strength weight, with a value ranging from 0.3 to 0.5; The strength of the character offset propagation chain is indicated by sub-step S2.4; Indicates the degree of correct cross-domain sharing as confirmed by manual or rule-based methods, used to reduce misjudgment of terms in normal sharing processes; This indicates the stage inversion weight, with a value ranging from 0.3 to 0.5; The score indicating the difference in the direction of responsibility for process term progression is obtained from sub-step S2.3; This indicates the semantic matching degree of the newly added process stage, obtained from sub-step S2.2; The degree of overlap in the responsibility attributes of the same process term across different industries is obtained from step S1.4.

[0058] This formula integrates the continuity of the transmission chain, the inversion of responsibility stages, and the inhibitory effect of the degree of normal sharing and the degree of overlap in responsibility attributes, effectively avoiding the risk of misjudging reasonable cross-industry shared terms as role inversion. According to The range of values ​​is determined as follows: If If so, the process term is marked as a high-risk inversion term and added to the high-risk subset; if Marked as a medium-risk inverted word and added to the medium-risk subset; if Words marked as low-risk are temporarily excluded from the offset set. Meanwhile, for high-risk inverted words, their inversion inheritance direction is explicitly recorded. The union of the high-risk subset and the medium-risk subset is defined as the process role semantic offset set, where each process word contains its corresponding role inversion risk score. And the inverted inheritance direction (for high-risk inverted words). The set of semantic offsets for roles in this process will serve as the direct governance object for the subsequent S3 chain of responsibility dependency semantic isolation processing.

[0059] S3: Perform responsibility chain dependency marking, industry affiliation splitting, and stage isolation processing on the process words in the process role semantic offset set to generate a responsibility chain dependency semantic matrix.

[0060] S3.1: Extract process words from the process role semantic offset set, associate them with the role offset propagation chain, calculate the responsibility chain dependency strength of each process word with respect to the role offset propagation chain, and attach and bind the process words to the role offset propagation chain according to the responsibility chain dependency strength to generate a dependency mark set.

[0061] Obtain all process terms from the process role semantic offset set output in step S2. These process terms include, but are not limited to, "confirm," "review," "reject," "correct," and "archive." For each risk process term, associate it with the role offset propagation chain to determine the source industry, target industry, preceding responsibility stage, current responsibility stage, and subsequent responsibility stage of the process term.

[0062] To quantify the strength of the chain of responsibility attachment, we introduce the strength of the chain of responsibility attachment. The calculation formula is as follows: ; in, Indicates the strength of the chain of responsibility; This represents the risk weight for role reversal, with a value ranging from 0.3 to 0.5; This indicates the stability of a process term in the current responsibility phase, which is determined by whether the term consistently appears in the same phase in similar historical processes. This indicates the number of consecutive occurrences of the process term in the role offset propagation chain, which is converted to dimensionless after logarithmic processing. This represents the weighting for differences in responsibility stages, with a value ranging from 0.3 to 0.5. This indicates the degree of difference in the responsibility stage between the source and target industries for the process term; The degree of overlap in the shared responsibility attributes of the process term in the two industries is obtained from step S1.4.

[0063] This formula indicates that the higher the risk of role reversal, the greater the stage difference, and the more continuous the transmission, the stronger the attachment marker; while the higher the shared responsibility attribute, the lower the possibility of false isolation. Based on the calculations... The value is processed as follows: If Then, the process term is strongly attached to the role's offset propagation chain, i.e., forcibly bound; if If the condition is met, then weak dependency marking is performed, meaning only the dependency is recorded without mandatory binding. All dependency marking results, along with strong and weak markings, are stored in the chain of responsibility dependency marking set.

[0064] S3.2: Based on the set of dependency tags, the industry affiliation status of the same process term in different responsibility stages is split, the necessity of industry affiliation split is calculated, and the industry affiliation split result is generated.

[0065] Based on the responsibility chain dependency tag set, the industry affiliation status of the same process term in different responsibility stages is split. To determine whether industry affiliation splitting is necessary, an industry affiliation splitting necessity degree is introduced. The calculation formula is as follows: ; in, Indicates the necessity of industry classification; This represents the industry-specific weighting, with a value ranging from 0.2 to 0.4. The strength of the chain of responsibility is indicated by sub-step S3.1; This indicates the degree of divergence in the subsequent processing results of the same process term across different industries. The degree of overlap in the responsibility attributes of process terms with the same name in different industries is obtained from step S1.4; This represents the public attribute suppression weight, with a value ranging from 0.1 to 0.3.

[0066] This formula indicates that when the responsibility chain of process terms with the same name is strongly dependent and the subsequent results are highly divergent, the necessity of splitting increases; while when the degree of overlap in responsibility attributes is high, the denominator increases, thereby reducing the necessity of splitting and avoiding excessive splitting. According to The value is judged as follows: If Then, a splitting operation is performed, specifically as follows: using the process term as the main index, its attachment tags in the enterprise ERP approval chain, hospital consultation approval chain, and other approval chains are expanded separately; according to the responsibility stage, process terms with the same name are split into final confirmation, intermediate confirmation, rollback and review, observation and confirmation, and archiving completion categories; for each type of splitting result, clear industry and stage attribution labels are set so that "ERP final confirmation" and "hospital intermediate confirmation" no longer share the same semantic position; furthermore, minimal shared attributes are retained between process terms with the same name, only retaining their common attributes as "approval action terms," ​​and no longer inheriting their specific meanings of final state, archiving, observation, or correction; if If the item is not split, it will be retained as a shared semantic item. The splitting result (including whether it is split and the independent semantic items after splitting) is the industry affiliation splitting result, and is used as the input of sub-step S3.3.

[0067] S3.3: Based on the industry affiliation splitting results, establish a final state inheritance restriction region for the process semantics at the final state responsibility position.

[0068] Based on the industry-specific classification results, candidate process semantics for establishing final state inheritance restriction regions are identified, namely, all process semantics that have been classified into final state confirmation, archiving completion, and settlement completion categories.

[0069] To quantify the strength of the inheritance constraint that should be applied, a final-state inheritance constraint strength is introduced. The calculation formula is as follows: ; in, Indicates the strength of the final state inheritance restriction; This represents the weight of the final state stage, with a value ranging from 0.3 to 0.5; The degree of necessity for industry classification is indicated by sub-step S3.2; This indicates the degree of irreversibility of the final state, which is determined by whether the process has entered the archiving, accounting, or closed-loop stages. This represents the propagation persistence weight, with a value ranging from 0.2 to 0.4; The frequency normalization value representing the propagation of final state semantics to non-final state stages; It represents the length of the continuous propagation of the final state semantics, in units of segments, and is converted to dimensionless after logarithmic processing; This indicates the degree of correct inheritance and absorption at the same stage, with a range of 0.7 to 1.

[0070] This formula shows that the more irreversible the final state semantics are and the more frequently they propagate to non-final states, the stronger the inheritance constraint; if its propagation has been correctly absorbed by the same stage, the constraint strength decreases. According to... The range of values ​​is divided into regions, and corresponding restriction operations are performed: If The region containing the final state semantics is designated as a strongly restricted region, and a final state inheritance boundary is set, ensuring that inheritance can only occur within the same industry, the same chain of responsibility stage, and the same subsequent result type, while prohibiting any cross-industry inheritance; if It is classified as a "medium-restricted area" and a final inheritance boundary is set, allowing inheritance only within the same industry and at the same stage; if The semantics of all final states are placed in a "weakly restricted area," and a final state inheritance boundary is set, allowing inheritance but requiring the inheritance path to be recorded. For all final state process semantics, their propagation relationships to intermediate review, observation retention, supplementary inspection, and other stages must be explicitly restricted. The set of all semantics placed in the restricted area constitutes the final state inheritance restricted area, which will serve as the input to sub-step S3.4.

[0071] S3.4: Based on the final state inheritance restriction region, establish a stage isolation region for the process semantics in the intermediate stage, and generate process stage isolation results.

[0072] Based on the final state inheritance restriction area, candidate process semantics for establishing stage isolation areas are determined: extract process words with the same name that are restricted from propagation from the final state inheritance restriction area, and reverse locate their intermediate stage position in the target industry; partition these process words into audit intermediate state, review intermediate state, observation and retention state, and correction and rollback state.

[0073] To quantify the reduction strength of cross-domain inheritance, a cross-domain inheritance reduction strength is introduced. The calculation formula is as follows: ; in, This indicates the cross-domain inheritance reduction strength, with a value range of [0,1]. The strength of the final state inheritance constraint is indicated by sub-step S3.3; This indicates the sensitivity of the target industry at intermediate stages, such as observation and confirmation or supplementary examination confirmation in medical consultations, and takes a higher value, usually set to 0.8; This indicates the degree of stable inheritance of the same process term within the target industry; This indicates a reasonable degree of weak sharing, with a value range of [0,1]. It can be initialized to 0.3 and then dynamically adjusted based on the proportion of correctly shared process terms across industries as manually annotated.

[0074] The formula indicates that the stronger the final state constraint and the more sensitive the intermediate stage of the target, the stronger the cross-domain inheritance reduction. Simultaneously, if the inheritance within the target industry is stable or confirmed to be reasonable sharing, the reduction magnitude decreases. According to... Isolate the range of values: if If so, then strong isolation is implemented for the process semantics of this intermediate stage, blocking the inheritance channel from the final state semantics of other industries; if Implement moderate isolation, retaining only weak sharing attributes; if No additional isolation is implemented. The specific scope of the isolation operation (which semantics and which industries are isolated) is the result of the process stage isolation.

[0075] S3.5: Summarize the dependency tag set, industry affiliation splitting results, final state inheritance restriction area, and process stage isolation results to generate the responsibility chain dependency semantic matrix.

[0076] Based on the process stage isolation results, the responsibility chain dependency markers, industry affiliation splitting status, final state inheritance restriction areas, and intermediate stage isolation areas are summarized and solidified. The industry source, responsibility stage, preceding responsibility role, subsequent responsibility role, final state restriction status, and inheritance reduction strength corresponding to each process term are structurally recorded. A responsibility chain dependency semantic matrix is ​​generated according to three dimensions: process term, industry, and responsibility stage. Each row of this matrix represents a combination of "process term plus industry source," such as "enterprise ERP confirmation" or "hospital consultation confirmation"; each column represents the responsibility stage, such as initiation stage, review stage, verification stage, confirmation stage, and archiving stage; each element in the matrix represents the dependency strength, restriction strength, and inheritance reduction status of a process term in a specific industry responsibility stage. By associating the final state inheritance restriction area with the stage isolation area, it is ensured that the final state semantics cannot directly penetrate into the intermediate stage semantics. For low-risk terms, a weak inheritance channel is retained; for high-risk inverted terms, a strong isolation channel is set.

[0077] To provide an executable integrated isolation control parameter for each element in the matrix, an integrated isolation value is introduced. The calculation formula is as follows: ; in, This represents the overall isolation value of a single element in the chain of responsibility dependency semantic matrix; Indicates the strength of the chain of responsibility; Indicates the necessity of industry classification; Indicates the strength of the final state inheritance restriction; Indicates the intensity of cross-domain inheritance reduction; This represents the dependency strength weight, with a value ranging from 0.2 to 0.3; This represents the weight of the necessity of splitting, with a value ranging from 0.2 to 0.3; This represents the final state constraint weight, with a value ranging from 0.2 to 0.4; This indicates that the inherited weights are reduced, with a value ranging from 0.2 to 0.3; and the sum of the four weights is 1.

[0078] The value, serving as the comprehensive isolation value of the matrix elements, is directly written into the corresponding position in the responsibility chain dependency semantic matrix. Regarding parameter settings, when the risk of role inversion is high, the final state constraint weight is appropriately increased; in scenarios with strong consequences such as medical consultations and financial settlements, the stage sensitivity is increased; for low-risk, common process terms such as "submit" and "view," the degree of reasonable sharing is increased to avoid excessive isolation. The final generated responsibility chain dependency semantic matrix will serve as the direct basis for the subsequent S4 step of cross-industry approval semantic dynamic governance processing.

[0079] Through the above sub-steps, step S3 completes the semantic isolation processing of the responsibility chain dependency of the role inversion risk process words, providing an executable and callable isolation control data structure for subsequent dynamic governance.

[0080] S4: Based on the aforementioned responsibility chain dependency semantic matrix, perform real-time responsibility chain positioning, attribution verification, and order correction on newly added approval data to generate dynamic responsibility governance results.

[0081] S4.1: For each batch of newly added approval data entering the training system, the extracted process words are compared with the responsibility chain dependency semantic matrix, real-time responsibility chain positioning is performed, and real-time responsibility chain positioning results are generated.

[0082] For each batch of newly added approval data entering the training system, they are merged according to the approval item number to form new approval process units. Information such as process terms, approval subject, approval time, preceding nodes, following nodes, and approval results are extracted from each process unit. The extracted process terms are compared with the responsibility chain dependency semantic matrix generated in step S3 to determine the dependency position of the process term in the corresponding industry and corresponding responsibility stage. Furthermore, based on the process stage isolation results obtained in step S3, it is determined whether the process term falls into the final state inheritance restriction region, the intermediate stage isolation region, or the weakly shared region.

[0083] Through the above operations, each process term in the newly added approval data will be promptly placed into the responsibility chain structure, generating real-time responsibility chain positioning results.

[0084] S4.2: Based on the real-time responsibility chain positioning results, verify the responsibility stage attribution of newly added process words, mark process words with questionable attribution, and generate responsibility stage attribution verification results.

[0085] Based on the real-time responsibility chain location results, the responsibility stage attribution status of newly added process terms is verified. The real-time responsibility chain location information corresponding to each process term is read, including its responsibility stage, industry origin, and the type of isolation area it is located in.

[0086] To quantify the credibility of responsibility stage attribution, a responsibility stage attribution credibility index is introduced. The calculation formula is as follows: ; in, Indicates the credibility of attribution of responsibility at each stage; This represents the overall isolation value read from the chain of responsibility dependency semantic matrix; This indicates the stability of the stage within the target industry, which can be obtained by calculating the frequency of the process term being in the same responsibility stage in the N most recent approvals within the target industry, for example, taking N=100; This indicates the intensity of stage conflicts for process terms with the same name across industries, and is determined by the proportion of different responsibility stage labels for that process term in different industries. This indicates the degree of divergence in the subsequent processing results of the same process term. The value is the percentage of inconsistent subsequent processing results (archiving, reviewing, observing, etc.) of the process term in different industries.

[0087] This formula indicates that the more stable the internal stages within a target industry, the higher the reliability of attribution; conversely, the greater the cross-industry stage conflict and subsequent outcome discrepancies, the lower the reliability of attribution. According to The value is judged as follows: If If the process term's responsibility stage attribution is verified, no correction is needed; if If the verification fails, it is marked as having questionable attribution, and the inversion trend identification in sub-step S4.3 is triggered. The above verification results and markings are the attribution verification results for the responsibility stage.

[0088] S4.3: Based on the responsibility stage attribution verification results, dynamically adjust the process semantics that show a role inversion trend, adjust the industry inheritance scope, and generate industry inheritance scope adjustment results.

[0089] Based on the results of the responsibility stage attribution verification, candidate process words that need to be subjected to inversion trend analysis are identified, namely those process words marked as having questionable attribution.

[0090] To quantify the strength of the role reversal trend, we introduce the role reversal trend strength. The calculation formula is as follows: ; in, Indicates the strength of the role reversal trend; Indicates the credibility of attribution of responsibility at each stage; Indicates the intensity of cross-industry process term conflicts at different stages; The cross-domain inheritance reduction strength is indicated by the output of step S3; The strength of the character offset propagation chain is represented by sub-step S2.4, and its calculation already includes the time duration factor. Indicates a reasonable degree of weak sharing; and is a weighting constant, with values ​​ranging from 0.3 to 0.7, and the sum of the two is 1.

[0091] The formula indicates that the more unreliable the stage attribution, the stronger the cross-industry conflict, and the more obvious the persistent deviation, the stronger the inversion trend; while the higher the reasonable degree of weak sharing, the lower the risk of misadjustment. According to The value will undergo the corresponding inheritance scope adjustment: if This significantly narrows the industry inheritance scope of the process term, allowing it to be inherited only within the same industry, the same responsibility stage, and the same subsequent processing result; if Appropriately narrow the scope of inheritance, for example, excluding the industry with the greatest differences in responsibility stages; if The original scope of inheritance will be maintained. The adjusted scope of inheritance, together with the specific restrictions, constitutes the result of the industry's scope of inheritance adjustment.

[0092] S4.4: Based on the industry inheritance scope adjustment results, perform consistency verification on the responsibility advancement order in the cross-industry approval reasoning results, and generate responsibility advancement order verification results.

[0093] Based on the adjusted industry inheritance scope, the consistency of the responsibility progression order in the cross-industry approval reasoning results is verified. The adjusted industry inheritance scope of the process terms is read. Then, the responsibility progression path is reconstructed according to the typical order of initiation, review, verification, confirmation, and archiving in the approval process.

[0094] To quantitatively assess the consistency of the responsibility progression sequence throughout the entire approval chain, a responsibility progression sequence consistency score is introduced. The calculation formula is as follows: ; in, Indicates the consistency score of the order of responsibility advancement; This indicates the number of nodes in the approval process, expressed in units of individual nodes. Indicates the sequence number of adjacent approval nodes; Indicates the first The degree of consistency of the phase sequence of adjacent nodes; Indicates the first The degree of responsibility of the responsible parties of each adjacent node; Indicates the first The anomalous inversion intensity of each adjacent node is obtained from cases such as stage forward shift, stage backward shift, and final state penetration.

[0095] The higher the value, the more consistent the order of responsibility progression throughout the entire chain of accountability. According to The value is determined as follows: If The verification passed, with no abnormalities; if If the verification passes, the sequence of responsibility for anomalies is detected and recorded. The anomalies are then categorized into stage forward shift anomalies, stage backward shift anomalies, final state penetration anomalies, and responsible entity misconnection anomalies, and the specific location of the anomaly node is recorded. If... If the verification fails, the entire chain of responsibility is marked as having an abnormal sequence, and is subsequently categorized and recorded. The above responsibility progression sequence verification results include a consistency score. And a list of abnormal nodes.

[0096] S4.5: Based on the result of the responsibility advancement sequence verification, the approval semantics that have abnormal responsibility sequence are re-executed for stage isolation correction, and the dynamic responsibility governance result is generated.

[0097] Based on the results of the responsibility advancement sequence verification, the candidate approval semantics that need to be isolated and corrected are identified, namely, the process words and their associated responsibility chain fragments that are partially approved or not approved by the responsibility advancement sequence verification results.

[0098] To quantify the strength of the need to modify the succession of responsibilities, a strength of responsibility succession modification is introduced. The calculation formula is as follows: ; in, Indicates the strength of the inheritance of responsibility modification; Indicates the strength of the role reversal trend; Indicates the consistency score of the order of responsibility advancement; Indicates the intensity of cross-domain inheritance reduction; Indicates a reasonable degree of weak sharing; , , These are weighting constants, each ranging from 0.2 to 0.4, and the sum of the three is 1.

[0099] according to The range of values ​​will trigger a correction operation of appropriate intensity: if Perform a strong correction, that is, rewrite the isolation value at the corresponding position in the chain of responsibility dependency semantic matrix and forcibly block the exception inheritance path; if Perform a moderate correction, adjusting only the boundaries of the isolated region without rewriting the entire matrix; if No correction will be made. For anomalies that require correction, first identify the type of anomaly source (final state inheritance penetration, intermediate stage miscompression, excessively broad industry inheritance scope, or misamplification of weak shared channels), and then perform the corresponding correction operation according to the correction intensity mentioned above.

[0100] The corrected matrix and isolated regions are merged and output as dynamic responsibility governance results. At the same time, the responsibility advancement order verification results are also output for the subsequent cross-industry approval semantic stability output processing in step S5.

[0101] Through the above sub-steps, step S4 completes the real-time chain of responsibility location, attribution verification, inversion trend identification, sequence consistency verification, and anomaly isolation correction for the newly added approval training data, providing dynamically governed training data and governance results for subsequent stable output.

[0102] S5: Based on the dynamic responsibility governance results, perform responsibility chain consistency verification, process role final state verification, and reasoning stability detection on the governed approval data, and output cross-industry approval semantic governance results.

[0103] S5.1: Based on the dynamic responsibility governance results, perform a responsibility chain consistency review on the governed approval data and calculate the responsibility chain consistency score.

[0104] Based on the dynamic responsibility governance results, responsibility order verification results, and responsibility inheritance correction results output in step S4, a new responsibility chain consistency analysis is performed on the governed approval training data. Specifically, the responsibility progression relationship between each node in each approval process—initiation, review, verification, confirmation, and archiving—is re-examined.

[0105] To comprehensively quantify the stability of the approval chain after governance, a responsibility chain consistency score is introduced. The calculation formula is as follows: ; in, Indicates the chain of responsibility consistency score; The consistency score for the order of responsibility progression is obtained from sub-step S4.4; The confidence level of responsibility attribution at each stage is indicated by sub-step S4.2; The strength of the responsibility inheritance modification is indicated by sub-step S4.5; The strength of the role reversal trend is indicated by sub-step S4.3; , , These are weighting constants, each ranging from 0.2 to 0.4, and the sum of the three is 1.

[0106] This formula integrates the consistency of responsibility progression order, the credibility of responsibility stage attribution, and the strength of inheritance correction, while appropriately suppressing the scoring through the strength of the inverted trend in the denominator. The calculated... The value represents the result of the chain of responsibility consistency verification.

[0107] S5.2: Based on the chain of responsibility consistency score, perform final state verification on the process role attribution status in different approval systems, and calculate the credibility of the process role final state.

[0108] Based on the consistency verification results of the chain of responsibility, the final state verification of the process role attribution status in different approval systems is performed. Specifically, process terms such as "confirm," "review," "reject," and "archive" are bound to their respective responsibility stages to determine whether the process term has truly entered an irreversible final state. For example, in an enterprise ERP system, if "confirm" is followed directly by subsequent financial accounting or archiving operations, its final state attribute is retained; however, in a hospital consultation system, if "confirm" is followed by observation, supplementary examination, or re-review operations, it is not judged as a final state, but rather retains its intermediate state attribute.

[0109] To quantify whether a process term truly embodies the meaning of final state responsibility, we introduce the process role final state credibility. The calculation formula is as follows: ; in, Indicates the credibility of the final state of a process role; The chain of responsibility consistency score is obtained from sub-step S5.1; This indicates the final state stability within the target industry, which can be obtained by statistically analyzing the frequency percentage of the process term marked as a final state within the target industry, typically ranging from 0.8 to 1. This indicates the intensity of the final state conflict of process terms with the same name across industries. The value is the difference in the proportion of the process term being marked as the final state in different industries. This indicates the degree of disagreement in the subsequent processing results, which is the same as the definition in sub-step S4.2.

[0110] according to The final state attribute of the process term is adjusted accordingly: if Preserve its final state properties; if The final state attribute will not be adjusted for the time being; if Remove its final state attribute and adjust it to an intermediate state or observation state. The calculated... The value serves as the credibility of the final state of the process role, and the adjusted final state attribute is used separately as the result of the adjustment of the final state attribute of the process role.

[0111] S5.3: Based on the credibility of the final state of the process roles, the stability of the cross-industry approval reasoning results is tested, and an approval reasoning stability score is calculated.

[0112] Based on the credibility of the final state of the aforementioned process roles, a stability test is performed on the responsibility progression logic in the cross-industry approval reasoning results. The test examines whether the reasoning chain conforms to the inherent stage sequence of the target industry, such as whether it follows a progressive relationship from initiation to review, then to verification, confirmation, and archiving.

[0113] To quantify the stability of the reasoning results, an approval reasoning stability score is introduced. The calculation formula is as follows: ; in, This indicates the stability score of the approval reasoning; The final state credibility of the process role is indicated by sub-step S5.2; The consistency score for the order of responsibility progression is obtained from sub-step S4.4; Indicates a reasonable degree of weak sharing; The score representing the difference in the direction of responsibility advancement is obtained from sub-step S2.3; , , These are weighting constants, each ranging from 0.2 to 0.4, and the sum of the three is 1.

[0114] This formula checks whether the inference result maintains both industry boundaries and retains reasonable weak-sharing attributes. The calculated result... The value represents the result of the approval reasoning stability test.

[0115] S5.4: Based on the stability score of the approval reasoning, backtrack the abnormal responsibility propagation path of the reasoning chain with role inversion inheritance trend, calculate the backtracking strength, and output the cross-industry approval semantic governance results.

[0116] Based on the stability test results of approval reasoning, backtracking is performed on the abnormal responsibility propagation path of reasoning chains exhibiting a role inversion inheritance trend. Here, a reasoning chain exhibiting a role inversion inheritance trend refers to an approval reasoning stability score... The reasoning chain. Based on the backtracking strength, perform the appropriate level of backtracking operation to correct the abnormal chain of responsibility dependencies.

[0117] To quantify the strength of the need for backtracking and re-correction of the chain of responsibility dependencies, anomaly propagation backtracking strength is introduced. The calculation formula is as follows: ; in, Indicates the intensity of the backtracking of abnormal responsibility propagation; The approval reasoning stability score is obtained from sub-step S5.3; The strength of the role reversal trend is indicated by sub-step S4.3; The score representing the difference in the direction of responsibility advancement is obtained from sub-step S2.3; Indicates a reasonable degree of weak sharing; , , These are weighting constants, each ranging from 0.2 to 0.4, and the sum of the three is 1.

[0118] according to The value executes the corresponding backtracking operation: if If so, a strong backtracking is executed, that is, the source industry and source node of the abnormal propagation are located, the isolation processing in step S3 is re-executed, and the responsibility chain dependency semantic matrix is ​​updated; if Performing a weak backtracking only adjusts the inheritance permissions on the current propagation path, without updating the global matrix; if Without performing backtracking, the current governance result is directly output. After backtracking, the final matrix, isolated regions, and inheritance paths output are merged into cross-industry approval semantic governance results, chain of responsibility stability results, and correct inheritance results of process roles.

[0119] Through the above sub-steps, step S5 completes the final consistency review, final state verification, stability detection, and backtracking correction of abnormal propagation paths for the post-governance approval data. This achieves stable isolation and correct inheritance of process role semantics in multi-stage approval flow joint training scenarios, providing a complete solution for reliable governance of cross-industry approval semantics.

[0120] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions stored in the memory to execute the methods provided in the above embodiments.

[0121] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0122] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0123] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0124] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0125] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for governing large model data based on cross-domain semantic drift, characterized in that, The method includes: S1. Based on the process actions, responsible roles and transmission relationships in cross-industry approval data, construct a cross-industry approval responsibility chain structure, and generate a cross-industry approval responsibility chain map and a set of process role location mappings. S2. Based on the cross-industry approval responsibility chain graph and process role position mapping set, the responsibility position is parsed and the stage is matched for the process words in the newly added training data to identify the process role semantic offset set. S3. Perform responsibility chain dependency marking, industry affiliation splitting and stage isolation processing on the process words in the process role semantic offset set to generate a responsibility chain dependency semantic matrix. S4. Based on the responsibility chain dependency semantic matrix, perform real-time responsibility chain positioning, attribution verification and order correction on the newly added approval data to generate dynamic responsibility governance results. S5. Based on the dynamic responsibility governance results, perform responsibility chain consistency verification, process role final state verification, and reasoning stability detection on the governed approval data, and output cross-industry approval semantic governance results.

2. The method according to claim 1, characterized in that, S1 includes: Acquire cross-industry approval data and divide the cross-industry approval data into a set of semantically coherent industry approval process segments according to the approval items; Each node in the industry approval process segment set is marked with a responsibility role, forming a responsibility role marked process set; Based on the set of responsibility role marking processes, the responsibility transfer relationship between adjacent nodes under the same approval item is associated to generate a set of responsibility transfer chains; Based on the set of responsibility transfer chains, the responsibility positions of the same process words in different approval chains are mapped and compared, and the degree of overlap of common responsibility attributes is counted to generate the cross-industry approval responsibility chain map and the process role position mapping set.

3. The method according to claim 1, characterized in that, S2 includes: Based on the aforementioned process role location mapping set, the responsibility location of process terms in the newly added approval data is parsed to generate parsing results; Based on the analysis results and the industry standard stage sequence in the cross-industry approval responsibility chain diagram, the responsibility position sequence of the newly added approval process is matched with the industry standard responsibility chain stage to obtain the semantic stage matching result of the newly added process. When the stage matching result is a partial match, the responsibility advancement direction of the same process words in different approval systems is compared to obtain the result of the difference in responsibility advancement direction. Based on the results of the differences in the direction of responsibility advancement, a role offset propagation chain is established for the process semantics where there is a trend of responsibility position reversal; Based on the role offset propagation chain, risk statistics are performed on the inheritance direction change status of process roles in different approval systems to obtain role inversion risk scores, and the process role semantic offset set is generated based on the role inversion risk scores.

4. The method according to claim 3, characterized in that, The step of matching the responsibility position sequence of the newly added approval process with the stages of the industry standard responsibility chain to obtain the semantic stage matching result of the newly added process includes: The sequence of responsibility positions for process terms in the newly added approval process is compared with the industry standard stage sequence in the cross-industry approval responsibility chain diagram; The matching status is determined based on a preset threshold, and the matching status includes good stage matching, partial matching, and abnormal stage matching.

5. The method according to claim 1, characterized in that, S3 includes: Extract process words from the process role semantic offset set, associate them with the role offset propagation chain, calculate the responsibility chain dependency strength of each process word with respect to the role offset propagation chain, and attach and bind the process words to the role offset propagation chain according to the responsibility chain dependency strength to generate a dependency tag set. Based on the aforementioned dependency tag set, the industry affiliation status of the same process term in different responsibility stages is split, the necessity degree of industry affiliation splitting is calculated, and the industry affiliation splitting result is generated. Based on the industry affiliation decomposition results, a final state inheritance restriction region is established for the process semantics at the final state responsibility position; Based on the final state inheritance restriction region, a stage isolation region is established for the process semantics in the intermediate stage, and a process stage isolation result is generated. The dependency mark set, industry affiliation splitting results, final state inheritance restriction region, and process stage isolation results are summarized to generate the responsibility chain dependency semantic matrix.

6. The method according to claim 5, characterized in that, Based on the dependency tag set, the industry affiliation status of the same process term in different responsibility stages is split, the necessity degree of industry affiliation splitting is calculated, and the industry affiliation splitting result is generated, including: Calculate the industry affiliation necessity for each process term; Based on whether the necessity of industry classification exceeds a preset threshold, determine whether to perform industry classification for the process term; When performing the split, the process term is used as the main index, and its responsibility chain attachment mark in different industries is split into at least one of the following categories according to the responsibility stage: final confirmation, intermediate confirmation, backtracking and review, observation confirmation, or archiving completion. A clear industry affiliation label and stage affiliation label are set for each type of split result, and the process term is retained as the minimum shared attribute of approval action term.

7. The method according to claim 1, characterized in that, S4 includes: For each batch of newly added approval data entering the training system, the extracted process words are compared with the responsibility chain dependency semantic matrix, real-time responsibility chain positioning is performed, and real-time responsibility chain positioning results are generated. Based on the real-time responsibility chain location results, the responsibility stage attribution of newly added process words is verified, process words with questionable attribution are marked, and responsibility stage attribution verification results are generated. Based on the responsibility stage attribution verification results, the semantics of processes with role inversion trends are dynamically adjusted, the industry inheritance scope is adjusted, and the industry inheritance scope adjustment results are generated. Based on the industry inheritance scope adjustment results, the consistency verification of the responsibility advancement order in the cross-industry approval reasoning results is performed, and the responsibility advancement order verification results are generated. Based on the results of the responsibility advancement sequence verification, the approval semantics that show abnormal responsibility sequence are re-executed in the isolation correction stage to generate the dynamic responsibility governance results.

8. The method according to claim 1, characterized in that, S5 includes: Based on the dynamic responsibility governance results, the responsibility chain consistency of the governed approval data is reviewed, and a responsibility chain consistency score is calculated. Based on the chain of responsibility consistency score, the final state verification of the process role attribution status in different approval systems is performed, and the credibility of the final state of the process role is calculated. Based on the credibility of the final state of the process roles, the stability of the cross-industry approval reasoning results is detected, and an approval reasoning stability score is calculated. Based on the stability score of the approval reasoning, the abnormal responsibility propagation path of the reasoning chain with the role inversion inheritance trend is backtracked, the backtracking strength is calculated, and the cross-industry approval semantic governance results are output.

9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the large model data governance method based on cross-domain semantic drift as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, It stores a computer program, which, when executed, implements the large model data governance method based on cross-domain semantic drift according to any one of claims 1-8.

Images