Administrative approval process automatic optimization system

Abstract

The invention discloses an automatic optimization system for an administrative approval process, which relates to the field of administrative approval and comprises a twin modeling module, a simulation prediction module, a scheme generation module, a feature modeling module, a dynamic routing module, a semantic analysis module, a change analysis module, a behavior intervention module, a fusion inference module and a decision support module. According to the method, a real process structure and an operation rule can be accurately copied, bottleneck nodes can be identified in different pressure scenes, intervention measures are ensured to consider efficiency, throughput and resource utilization rate, task allocation is efficient and fair by dynamically generating a routing instruction, the response speed and accuracy of approval to regulation updating are improved, and the method is suitable for popularization and application. Quantitative policy implementation effect prediction is provided for decision makers, the scheme is ensured to have transparency and auditing performance, and manual review and risk control are facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of administrative examination and approval, and in particular to an administrative examination and approval process automatic optimization system. BACKGROUND

[0002] The administrative examination and approval process is one of the most basic and key links in the government governance system, which not only carries the function of implementing public policies, but also reflects the efficiency and quality of government services. The efficiency and transparency of examination and approval are directly related to the access cost of market entities, the convenience of livelihood affairs, and the realization level of macro policy objectives. However, in the existing administrative practice, the traditional examination and approval mode still faces multiple challenges in structure and system: on the one hand, the process structure is complex and the cross-department coordination cost is high. A complete examination and approval matter often involves multiple departments and links, and the dependency relationship between nodes is complex, which is easy to cause the overall efficiency to decline due to single-point bottleneck. The examination and approval process has the situation of "stuck in one link, overall passive delay", and lacks real-time monitoring and dynamic optimization ability of the whole process. On the other hand, the frequent update of policies and regulations increases the difficulty of process adjustment. In reality, whenever new policies and regulations are introduced, or existing provisions are revised, the examination and approval process often needs to be adjusted manually, which consumes a lot of time and manpower, and is easy to miss or be inconsistent, causing potential compliance risks and implementation deviations. Lack of automatic analysis and computable modeling of the semantics of regulations leads to insufficient process adaptability, affecting the timeliness and accuracy of policy implementation. More importantly, there is a lack of scientific evaluation and feedback mechanism for policy effects. The existing process optimization or new policy implementation relies more on experience or pilot feedback, lacks causal inference and data-driven simulation verification, and it is difficult to quantify the real impact of the improvement of a certain examination and approval process on social and economic indicators (such as enterprise registration rate, employment level, investment activity, etc.), leading to lagging or even absence of policy effect evaluation.

[0003] The existing administrative examination and approval process automatic optimization system cannot accurately replicate the real process structure and operation rules, and cannot identify bottleneck nodes under different pressure scenarios, reduce intervention measures to balance efficiency, throughput and resource utilization, cannot guarantee efficient and fair task allocation, reduce the response speed and accuracy of examination and approval to regulations update, cannot provide quantitative policy implementation effect prediction for decision makers, reduce the transparency and auditability of the scheme, and is not convenient for manual review and risk control. Therefore, we propose an administrative examination and approval process automatic optimization system. SUMMARY

[0004] The purpose of the present application is to solve the defects in the prior art and provide an administrative examination and approval process automatic optimization system.

[0005] This invention proposes an automatic optimization system for administrative approval processes, the system comprising: a twin modeling module, a simulation prediction module, a scheme generation module, a feature modeling module, a dynamic routing module, a semantic parsing module, a change analysis module, a behavior intervention module, a fusion inference module, and a decision support module; The twin modeling module is used to construct a process simulation model that runs synchronously with the real approval system; The simulation prediction module is used to inject various extreme or sudden scenarios, dynamically deduce the chain reaction of the approval process, and predict the node where the bottleneck will first appear. The solution generation module automatically generates multiple feasible optimization solutions based on the bottleneck prediction results. The feature modeling module is used to generate deep features for each newly submitted application and to generate capability features for approvers and positions. The dynamic routing module is used to construct a matching scheme between applications and approvers, and convert it into dynamic routing instructions in real time; The semantic parsing module is used to perform semantic parsing on laws, regulations, departmental rules, and policy documents to construct a policy semantic network. The change analysis module updates the policy semantic network when a new policy is released or an old policy is revised, predicts the impact of policy changes on existing approval processes, and automatically generates draft process modifications. The behavioral intervention module is used to record the approver's approval behavior and establish a decision intervention model. When the approver handles a new case, it provides personalized suggestions and tracks the adoption status and results. The fusion inference module is used to integrate approval data with external data, construct a policy-effect correlation graph, and analyze the causal relationship between approval process optimization and changes in socio-economic indicators. The decision support module is used to perform virtual simulations of policies before the introduction of new policies or when optimizing existing policies, and to predict their potential effects in reality.

[0006] As a further aspect of the present invention, the specific steps of the simulation prediction module in predicting the node where the bottleneck first occurs are as follows: S1.1: Deploy a data acquisition pipeline to access raw event streams from various data sources, and perform field normalization processing on each raw record. Based on each approval process, set up a process node dictionary and build a corresponding mapping rule engine to map each raw event to one or more process nodes. Then, aggregate the accessed data into a case event sequence in chronological order and generate metadata for each event. Based on all case event sequences, extract the node set and directed edge set to establish an initial process directed graph, and count the historical number of outgoing and incoming edges for each node. S1.2: For paths with parallel or conditional branches in the directed graph of the process, calculate the conditional transition probability in layers according to the conditional features to generate multiple transition matrices indexed by the condition. Estimate the transition probability with time relationship by time window to obtain the changes of each process over time. Then extract the service time sample from the historical processing record of each process node, perform outlier detection and truncation on each service time sample, and then use the maximum likelihood or moment method to estimate the distribution parameters. S1.3: Calculate the concurrent processing capability and average single-person efficiency of the computing nodes. For nodes with a mixture of manual and automatic processing, establish manual service rate and automatic service rate respectively, and synthesize the mixed service rate. Set the resource set of each node, establish the available quantity time series for each type of resource, calculate the type and quantity of resources consumed by each case on the node, establish the corresponding resource consumption model, and embed resource constraints as hard constraints or soft constraints into the resource consumption model. S1.4: Use natural language processing to extract entities, actions and logical relationships from various documents such as policies, regulations, departmental rules and explanatory documents, then use rule expressions or logical predicates to represent each policy constraint, and generate application condition functions for each rule to make Boolean or scoring-based application judgments for specific cases. After that, all rules are constructed into a directed dependency graph to provide a priority resolution mechanism and a traceable audit chain for inconsistencies or conflicts between rules. S1.5: Establish a corresponding queue network based on the directed graph of the process, model each node as a service station with a finite number of concurrent connections, and calculate the queuing length and waiting time of each node under a given input load based on the transition probability matrix and service rate. Then, use the approximate queuing theory formula to evaluate the performance of each node under different injection traffic scenarios, so as to identify the nodes that first enter a high utilization or saturation state under specific pressure and mark them as bottleneck nodes.

[0007] As a further aspect of the present invention, the resource consumption model described in S1.3 is specifically manifested as follows: ; The resource constraints of this resource consumption model are specifically manifested as follows: ; In the formula, Represents time Time, node For resource types Utilization rate; Represents time Time, node Available resources Total amount; Represents time Constantly occupying , A set of active case indexes; Representative Cases At the node Resources Instantaneous demand; Represents time At any given moment, are all node resource constraints met simultaneously? The specific calculation formula for the application condition function described in S1.4 is as follows: ; In the formula, Representative Rules Regarding the case Applicable confidence scores; Representative Cases eigenvectors; as well as Representing rules respectively Weights and biases; This represents the activation function.

[0008] As a further aspect of the present invention, the data source mentioned in S1.1 specifically includes online application forms, office system logs, manual data entry, scanned document OCR results, transmission / reception timestamps, etc. The field standardization process described in S1.1 specifically includes date format standardization, field name mapping, and filling or marking missing fields; The resource set of each node described in S1.3 specifically includes personnel categories, number of approval terminals, number of review tool licenses, and necessary material inventory.

[0009] As a further aspect of the present invention, the specific steps of the scheme generation module in automatically generating multiple feasible optimization schemes are as follows: S2.1: Receive the output of each bottleneck node and its observations from the bottleneck prediction module, unify each observation to the same time scale and statistical caliber, and then perform exponential smoothing on the arrival rate and service rate of each bottleneck node to obtain a stable baseline estimate. After that, enumerate the feasible intervention types of each bottleneck node, establish a parameterized representation of each intervention type, and form a set of candidate vectors for each node. S2.2: Convert each candidate vector into a new equivalent service rate or parallel quantity that the corresponding node can provide after intervention. Based on the converted candidate vectors and corresponding nodes, establish multiple candidate schemes. Use the queuing approximation method to calculate the baseline latency and post-intervention latency at the node level in each candidate scheme. Then, use the traffic minimization principle to calculate the throughput change of the node before and after intervention in each candidate scheme, and propagate it to the connected upstream and downstream nodes to evaluate the impact of backlog and resource idleness. S2.3: Calculate the indicators of each candidate solution and establish a vectorized target set. Based on the Pareto dominance judgment, eliminate the solutions that are superior to other solutions in all objectives to establish a Pareto set. Then, for each solution in the Pareto set, recalculate the indicators in a set of representative scenarios set automatically or manually, and obtain the average performance and variance of each solution in the scenario set to evaluate the robustness and worst-case performance of each candidate solution. Then, use an interpretable weighted scaling method to calculate the corresponding comprehensive score based on the multidimensional objectives of each selected candidate solution. S2.4: If the overall score is higher than the preset threshold, the corresponding candidate solution is retained; otherwise, it is eliminated. The final decision support package for each candidate solution is generated, including key indicators and the mean and confidence interval in the scenario test, along with suggestions on the implementation order and the corresponding rolling implementation threshold.

[0010] As a further aspect of the present invention, the specific calculation formula for exponential smoothing described in S2.1 is as follows: ; ; In the formula, Representative node The original observation arrival rate during this observation window; This represents the smoothed estimated arrival rate; Represents the smoothed arrival rate of previous time points; Represents the arrival rate smoothing coefficient; This represents the smoothed baseline service rate; Represents the service rate smoothing coefficient; Representative node The observation service rate; A smoothed service rate estimate representing a previous time step; As a further aspect of the present invention, the feasible intervention types described in S2.1 specifically include temporary staffing, improving individual efficiency, increasing parallel processing units, adjusting priority / approval order, merging / splitting batches, temporarily transferring some work to other nodes, and introducing some automated steps, etc. The specific indicators for each candidate solution described in S2.3 include the expected total latency savings, throughput improvement, direct implementation costs, compliance / policy risks, implementation complexity, and expected labor input costs.

[0011] As a further aspect of the present invention, the specific steps of the dynamic routing module in constructing a matching scheme between applications and approvers and converting it into dynamic routing instructions in real time are as follows: S3.1: Read the descriptive and structured features of each application from the access system, as well as the capabilities, experience, authority, and current pending tasks of each approver. Then, perform Min-Max normalization on the two types of features, and then perform L2 standardization on the processed features. Finally, fill in the missing values ​​in the standardized features by interpolation. Based on the preprocessed features, calculate the fit of each application-approver pair by cosine similarity. S3.2: Based on the premise that each application is assigned at least once, the total workload processed by each approver within a unit of time cannot exceed its capacity, and if the approval path allows multiple hops, the assignment is regarded as an allocation to a position or queue, rather than individual individual needs. Decision variables and hard constraints are set, the delay and risk of each application-approver pair are calculated, and a multi-objective function of the process is constructed. Then, based on the adaptation benefits, work costs and overload penalties, a corresponding local utility function is constructed for each approver. S3.3: The final application-approver matching scheme is transformed into a combinatorial optimization problem with capacity constraints. One objective is selected as the main constraint and a series of upper bounds are set on it. The remaining objectives are used as optimization objects, and different solutions are obtained to obtain the set of applications that each approver is willing to accept. Then, based on the set of applications that each approver is willing to accept, each decision variable is updated, and the next round of iteration is carried out until the changes in the multi-objective function values ​​of multiple rounds converge to the preset interval. S3.4: Collect all candidate solutions generated after the iteration, calculate the index vector of each candidate solution on each objective, and remove candidate solutions that are "better or not inferior and at least one better" than other candidate solutions on all objectives. Use linear programming relaxation to obtain the baseline solution of each candidate solution. Then, discretize each relaxed candidate solution through greedy correction. If any approver is overloaded after discretization, backtrack according to the candidate solution with the best multi-objective function value, and at the same time, reassign the unaccepted tasks to the next candidate approver. S3.5: Convert each candidate solution into an actual routing instruction. Each instruction must include at least the target approver, task ID, routing priority, expected completion time or deadline, and processing suggestions. Calculate the priority score for each routing instruction and then schedule them in descending order of the calculated priority scores.

[0012] As a further aspect of the present invention, the specific steps for the semantic parsing module to construct the policy semantic network are as follows: S4.1: Collect text data of various laws, regulations, and notices in different formats. Then, use OCR technology to convert the paper-based originals into electronic formats, unify the character set, extract the structured units of each text data at the document level (document, clause / section, sentence, and clause), and retain metadata such as file ID, issuing authority, and publication time. Perform word segmentation, word form restoration, and stop word removal on each text data. At the same time, protect proper nouns and terms with a dictionary. Finally, use a word vector model to generate an embedding vector for each sentence. S4.2: Identify the boundaries of legal clauses and intra-sentence condition separators in the legal text, perform syntactic dependency analysis on parallel or subordinating conjunctions, locate the main conditional phrases and subject-predicate structures, and split each sentence into multiple candidate clauses based on the analysis results. Calculate the sentence break boundary confidence score of each candidate clause and retain candidate clauses with sentence break boundary confidence scores higher than the preset confidence threshold. S4.3: For each split clause, legal domain entity categories are extracted using named entity recognition technology. For each retained clause, named entity recognition is performed to extract legal domain entity categories. Then, attributes are added to the extracted entities, and the confidence score and source location index of the entities are given. Then, synonymous entities are merged into a unified entity identifier using a terminology database or thesaurus. After that, the conditional components and action components are located in each clause, and the association strength of each pair (conditional fragment, action fragment) is calculated to establish candidate condition-action pairs. S4.4: Identify the quantifiers, time limits, and permission levels in each candidate condition-action pair, and map each candidate condition-action pair into a formal logical expression. For clauses containing complex logic, convert them into a unified standard form. Then, assign corresponding rule confidence and scope of application to each logical rule. Then, represent the extracted entities, conditions, actions, and their relationships in the form of tuples. Finally, construct each tuple into a directed labeled graph and establish the corresponding adjacency structure.

[0013] As a further aspect of the present invention, the specific steps of the change analysis module in predicting the impact of policy changes on existing approval processes and automatically generating draft process modifications are as follows: S5.1: Receive newly released or revised policy documents. If the document is a scanned copy, perform OCR processing. If it is HTML, perform DOM cleaning. Then, extract the document metadata from each policy document after processing and record the source link. Then, divide the document into articles, clauses, and clause-level units according to the legal hierarchy. Perform language preprocessing on each unit, calculate the weighted text vector of each clause, and write the segments, vectors, and metadata into the candidate node pool of the semantic network. At the same time, retain the original text index of the policy for source tracing. S5.2: Extract the entity, condition, and action of each clause using NER technology, and establish corresponding candidate rule fragments. Then, retain the original text fragment and source index of each candidate rule fragment, calculate the confidence value of each candidate rule fragment, and mark the candidate rule fragments with confidence values ​​lower than the preset threshold as requiring manual review. Then, establish a candidate rule item list according to the confidence value from largest to smallest. The candidate rule list includes information such as condition, action, entity, confidence value, source, and effective time interval. S5.3: Alignment objects for each candidate rule fragment in the semantic network are searched in parallel through text similarity matching, entity / term homonym or normalization matching, and logical matching. Corresponding evidence scores and mapping suggestions are generated for each type of matching. Evidence fusion is performed on multiple mapping suggestions, and the final mapping action decision is generated. The source information of all mapping actions is recorded. The attributes, edges and confidence of each node after mapping are written into the semantic network. The edge weights and node confidences of existing edges are updated according to the evidence accumulation rule. An incremental index is triggered after the update is completed. S5.4: Set the injection value for each newly added rule node, simulate disturbance, and use the graph propagation algorithm to propagate the disturbance along the weighted directed edges to other rule nodes, entity nodes and nodes that are ultimately mapped to the process. Iterate for multiple rounds until the preset propagation depth limit is reached. Record the intermediate impact value of each round of propagation, generate the cumulative impact score and propagation path source of each semantic network node, and pass the cumulative impact value of the semantic node to the corresponding approval process node according to the mapping relationship to obtain the semantic impact score of each process node. Generate an impact summary for each affected process node and sort it from high to low according to business priority. S5.5: Process nodes whose impact score or expected quantitative changes exceed a predetermined threshold are marked as high-impact. A process modification draft is automatically generated according to a set of rule-based editing templates. The draft items include suggested modification points, reasons for modification, expected quantitative effects, implementation risks and confidence levels, and rollback and pilot recommendations. Then, the implementation confidence score of each draft is calculated. If the implementation confidence score is higher than the automatic on-chain threshold and the expected cost is lower than the cost limit, it is automatically issued to the pilot. Otherwise, it is submitted to manual review, and an interpretable traceability is generated that can be viewed by the reviewer. The draft is then written to the change management queue and the version and approval flow route are recorded. Monitoring indicators are set for the automatically issued pilot changes. At the same time, the effects are automatically summarized after the pilot period for final decision-making.

[0014] The beneficial effects of this invention are: This invention accesses multi-source raw event streams through data acquisition pipelines, standardizes fields, and maps them to process nodes, thereby constructing a directed process graph and transition matrix. Simultaneously, it extracts node service time, concurrent processing capacity, and resource constraints to establish a resource consumption model. Subsequently, it combines legal semantic parsing to extract entities, conditions, and actions from policy and regulatory texts, transforming them into logical rules and constructing a semantic dependency graph to provide priority and tracing mechanisms for conflicts. Then, it uses queuing theory and transition matrix analysis to identify bottleneck nodes under load, and generates candidate solutions through candidate intervention vectors and multi-objective optimization. After Pareto screening, robustness evaluation, and weighted comprehensive scoring, it outputs an optimization suggestion package. In the task allocation stage, it normalizes and calculates the similarity of the characteristics of applicants and approvers, combining capacity and constraints... This approach constructs a combinatorial optimization problem, iteratively solves for matching solutions, and transforms them into dynamic routing instructions. Then, it performs OCR, word segmentation, NER, and semantic mapping on newly released or revised regulations, integrating them into a semantic network. Finally, it uses a graph propagation algorithm to calculate the ripple effect of policy changes and generates a traceable draft of process modifications for automated pilot testing or manual review. This approach accurately replicates real-world process structures and operational patterns, identifies bottlenecks under different pressure scenarios, and ensures that intervention measures balance efficiency, throughput, and resource utilization. By dynamically generating routing instructions, it ensures efficient and fair task allocation, improves the responsiveness and accuracy of approval processes to regulatory updates, provides decision-makers with quantitative predictions of policy implementation effects, guarantees transparency and auditability of the solutions, and facilitates manual review and risk control. Attached Figure Description

[0015] The present invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a framework diagram of an automated administrative approval process optimization system. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0019] This invention provides an automated optimization system for administrative approval processes. See also... Figure 1 , Figure 1This is a framework diagram of an automatic optimization system for administrative approval processes provided in an embodiment of the present invention. The system includes: a twin modeling module, a simulation prediction module, a scheme generation module, a feature modeling module, a dynamic routing module, a semantic parsing module, a change analysis module, a behavior intervention module, a fusion inference module, and a decision support module.

[0020] The twin modeling module is used to build a process simulation model that runs synchronously with the real approval system.

[0021] Specifically, a data acquisition pipeline is deployed to access raw event streams from various data sources. Field normalization is performed on each raw record. Based on each approval process, a process node dictionary is set up, and a corresponding mapping rule engine is built to map each raw event to one or more process nodes. Then, the accessed data is aggregated into a case event sequence in chronological order, and metadata is generated for each event. Based on all case event sequences, a node set and a directed edge set are extracted to establish an initial process directed graph. Simultaneously, the historical frequency of outgoing and incoming edges for each node is counted. For paths in the process directed graph with parallelism or conditional branches, conditional transition probabilities are calculated hierarchically according to conditional features to generate multiple transition matrices indexed by conditions. Transition probabilities with temporal relationships are estimated by time window segmentation to obtain the changes in each process over time. Then, service time samples are extracted from the historical processing records of each process node, and outlier detection and truncation are performed on each service time sample. Finally, the maximum likelihood or moment method is used to estimate distribution parameters, calculate the concurrent processing capacity and average single-person efficiency of the nodes, and establish manual service rate and automatic service rate for nodes with a mix of manual and automatic processing. The system calculates the service rate and synthesizes a hybrid service rate. It sets the resource set for each node, establishes a time-series sequence of available resources for each type, calculates the type and amount of resources consumed by each case on the node, and builds a corresponding resource consumption model. Resource constraints are embedded as hard or soft constraints into the resource consumption model. Natural language processing is used to extract entities, actions, and logical relationships from various documents, including policies, regulations, departmental rules, and interpretative documents. Each policy constraint is then represented using rule expressions or logical predicates, and an application condition function is generated for each rule to perform Boolean or scoring-based application judgments for specific cases. All rules are then constructed into a directed dependency graph, providing a priority resolution mechanism and a traceable audit chain for inconsistencies or conflicts between rules. A corresponding queue network is established based on the directed graph, modeling each node as a service station with a finite number of concurrent connections. Based on the transition probability matrix and service rate, the queuing length and waiting time of each node under a given input load are calculated. An approximate queuing theory formula is then used to evaluate the performance of each node under different injection traffic scenarios to identify nodes that first enter a high utilization or saturation state under specific pressure and mark them as bottleneck nodes.

[0022] The specific form of the resource consumption model is as follows: ; The resource constraints of this resource consumption model are specifically manifested as follows: ; In the formula, Represents time Time, node For resource types Utilization rate; Represents time Time, node Available resources Total amount; Represents time Constantly occupying , A set of active case indexes; Representative Cases At the node Resources Instantaneous demand; Represents time At any given moment, are all node resource constraints met simultaneously? The specific calculation formula for the applied conditional function is as follows: ; In the formula, Representative Rules Regarding the case Applicable confidence scores; Representative Cases eigenvectors; as well as Representing rules respectively Weights and biases; Represents the activation function; In addition, it should be noted that the data sources specifically include online application forms, office system logs, manual data entry, scanned document OCR results, transmission / reception timestamps, etc.; field standardization processing specifically includes unified date format, field name mapping, and filling or labeling missing fields; the resource set of each node specifically includes personnel categories, number of approval terminals, number of review tool licenses, and necessary material inventory, etc.

[0023] The simulation prediction module is used to inject various extreme or sudden scenarios, dynamically simulate the chain reaction of the approval process, and predict the node where the bottleneck will first appear; the solution generation module automatically generates multiple feasible optimization solutions based on the bottleneck prediction results.

[0024] Specifically, the system receives the output of the bottleneck prediction module, including each bottleneck node and its observations. These observations are then standardized to the same time scale and statistical caliber. Exponential smoothing is applied to the arrival rate and service rate of each bottleneck node to obtain a stable baseline estimate. Next, feasible intervention types for each bottleneck node are enumerated, and a parameterized representation for each intervention type is established, forming a set of candidate vectors for each node. Each candidate vector is converted into a new equivalent service rate or number of parallel connections that the corresponding node can provide after intervention. Based on the converted candidate vectors and corresponding nodes, multiple candidate schemes are established. A queuing approximation method is used to calculate the baseline latency and post-intervention latency at the node level for each candidate scheme. Finally, using the flow minimization principle, the throughput change before and after node intervention in each candidate scheme is calculated and propagated to connected upstream and downstream nodes to assess congestion and resource idleness. The process involves calculating various indicators for each candidate solution and establishing a vectorized target set. Based on the Pareto dominance criterion, solutions that are superior to other solutions in all objectives are eliminated to establish a Pareto set. Then, for each solution in the Pareto set, the indicators are recalculated under a set of representative scenarios set automatically or manually, and the average performance and variance of each solution on the scenario set are obtained to evaluate the robustness and worst-case performance of each candidate solution. Subsequently, an interpretable weighted scaling method is used to calculate the corresponding comprehensive score based on the multidimensional objectives of each selected candidate solution. If the comprehensive score is higher than a preset threshold, the corresponding candidate solution is retained; otherwise, it is eliminated. The final decision support package for each candidate solution is generated, including key indicators and the mean and confidence interval on the scenario test, along with suggestions on the implementation order and the corresponding rolling implementation threshold.

[0025] It should be noted that the specific formula for calculating exponential smoothing is as follows: ; ; In the formula, Representative node The original observation arrival rate during this observation window; This represents the smoothed estimated arrival rate; Represents the smoothed arrival rate of previous time points; Represents the arrival rate smoothing coefficient; This represents the smoothed baseline service rate; Represents the service rate smoothing coefficient; Representative node The observation service rate; A smoothed service rate estimate representing a previous time step; As a further solution in this embodiment, feasible intervention types specifically include temporary staffing, improving individual efficiency, increasing parallel processing units, adjusting priority / approval order, merging / splitting batches, temporarily transferring some work to other nodes, and introducing some automated steps; the specific indicators for each candidate solution include expected total latency savings, throughput improvement, direct implementation costs, compliance / policy risks, implementation complexity, and expected labor input costs.

[0026] The feature modeling module generates deep features for each newly submitted application and capability features for approvers and positions; the dynamic routing module constructs a matching scheme between applications and approvers and converts it into dynamic routing instructions in real time.

[0027] Specifically, the system reads descriptive and structured features of each application from the access system, as well as the capabilities, experience, authority, and current backlog of each approver. Then, it performs Min-Max normalization on both types of features, followed by L2 standardization. Missing values ​​in the standardized features are filled using interpolation. Based on the preprocessed features, the suitability of each application-approvers pair is calculated using cosine similarity. This is based on the following constraints: each application must be assigned at least once; the total workload processed by each approver per unit time cannot exceed their capacity; and if the approval path allows multiple hops, the assignment is considered to be to a position or queue, rather than individual requirements. Decision variables and hard constraints are set, and the latency and risk of each application-approvers pair are calculated. A multi-objective function for the process is constructed. Based on suitability benefits, work costs, and overload penalties, a corresponding local utility function is constructed for each approver. The final application-approvers matching scheme is transformed into a combinatorial optimization problem with capacity constraints. One objective is selected as the primary constraint, and a series of upper bounds are set on it. The remaining objectives are used as optimizations. The system iterates through the objects to obtain different solutions, thus obtaining the set of applications that each approver is willing to accept. Then, based on the set of applications each approver is willing to accept, it updates the decision variables and proceeds to the next iteration until the changes in the multi-objective function values ​​converge to a preset interval. After the iterations, it collects the candidate solutions generated, calculates the index vector of each candidate solution on each objective, and eliminates candidate solutions that are "better or not inferior, and at least one better" than other candidate solutions on all objectives. Linear programming relaxation is used to obtain the baseline solution for each candidate solution. Then, a greedy correction is used to discretize each relaxed candidate solution. If, after discretization, any approver is overloaded, the candidate solution with the optimal multi-objective function value is used for backtracking. Simultaneously, unaccepted tasks are reassigned to the next candidate approver. The solutions of each candidate solution are transformed into actual routing instructions, where each instruction includes at least the target approver, task ID, routing priority, expected completion time or deadline, and processing suggestions. The priority score of each routing instruction is calculated, and then the tasks are scheduled sequentially from largest to smallest priority score. Example 2

[0028] This invention provides an automated optimization system for administrative approval processes. See also... Figure 1 , Figure 1 This is a framework diagram of an automatic optimization system for administrative approval processes provided in an embodiment of the present invention. The system includes: a twin modeling module, a simulation prediction module, a scheme generation module, a feature modeling module, a dynamic routing module, a semantic parsing module, a change analysis module, a behavior intervention module, a fusion inference module, and a decision support module.

[0029] The semantic parsing module is used to perform semantic parsing on laws, regulations, departmental rules, and policy documents to construct a policy semantic network.

[0030] Specifically, the process involves collecting text data of various laws, regulations, and notices in different formats. Paper-based originals are then converted to electronic formats using OCR technology. Character sets are unified, and structured units are extracted from each text data at the document level (document, clause / section, sentence, and clause). Document ID, issuing authority, and publication time metadata are retained. Each text data item undergoes word segmentation, lemmatization, and stop word removal. Proper nouns and terms are protected using a dictionary. Word vector models are then used to generate embedding vectors for each sentence. Legal clause boundaries and intra-sentence condition separators are identified within the legal text. Syntactic dependency analysis is performed on parallel or subordinating conjunctions to locate subject-conditional phrases and subject-predicate structures. Based on the analysis results, each sentence is broken down into multiple candidate clauses. A confidence score for each candidate clause's segmentation boundary is calculated, and candidate clauses with segmentation boundary confidence scores higher than a pre-set confidence threshold are retained. Named entity recognition technology is then used to further refine the analysis. For each split clause, legal domain entity categories are extracted. Named entity recognition is performed on each retained clause to extract legal domain entity categories. Attributes are then added to the extracted entities, and confidence scores and source location indexes are provided. Synonymous entities are merged into a unified entity identifier using a terminology database or thesaurus. Conditional and action components are located within each clause, and the association strength of each pair (condition fragment, action fragment) is calculated to establish candidate condition-action pairs. Quantifiers, time limits, and permission levels in each candidate condition-action pair are identified, and each candidate condition-action pair is mapped to a formal logical expression. Clauses containing complex logic are converted into a unified normative form. Each logical rule is then assigned a corresponding rule confidence and scope of application. The extracted entities, conditions, actions, and their relationships are represented in the form of tuples. Each tuple is then constructed into a directed labeled graph, and a corresponding adjacency structure is established.

[0031] When a new policy is released or an old policy is revised, the change analysis module updates the policy semantic network, predicts the impact of policy changes on existing approval processes, and automatically generates draft process modifications.

[0032] Specifically, the system receives newly released or revised policy documents. If the document is a scanned copy, it undergoes OCR processing; if it is HTML, it performs DOM cleaning. Then, it extracts the document metadata from each policy document and records the source link. Next, it segments the document into clause, paragraph, and clause-level units according to the regulatory hierarchy. Language preprocessing is then performed on each unit, calculating the weighted text vector for each clause. The segments, vectors, and metadata are written into the candidate node pool of the semantic network, while retaining the original text index for source tracing. Entities, conditions, and actions of each clause are extracted using NER technology, and corresponding candidate rule fragments are created. Finally, the original text fragment and source index of each candidate rule fragment are retained, and the calculation of each candidate rule fragment... The confidence value of each segment is calculated, and candidate rule segments with confidence values ​​below a preset threshold are marked as requiring manual review. A candidate rule list is then created, ranked from highest to lowest confidence value. This list includes information on conditions, actions, entities, confidence values, sources, and effective time intervals. Alignment objects for each candidate rule segment in the semantic network are searched in parallel using text similarity matching, entity / term name homonym or normalized matching, and logical matching. Corresponding evidence scores and mapping suggestions are generated for each type of matching. Evidence fusion is performed on multiple mapping suggestions to generate the final mapping action decision. The source information of all mapping actions is recorded. Finally, the attributes, edges, and confidence values ​​of each mapped node are written into the semantic network. For existing edges, update edge weights and node confidence according to the evidence accumulation rule. Simultaneously, trigger an incremental index after the update, set the injection value for each newly added rule node, simulate disturbances, and use a graph propagation algorithm to propagate the disturbances along weighted directed edges to other rule nodes, entity nodes, and nodes ultimately mapped to the process. Iterate multiple times until the preset propagation depth limit is reached. Record the intermediate impact value of each propagation round, generate the cumulative impact score and propagation path source for each semantic network node, and pass the cumulative impact value of the semantic nodes to the corresponding approval process nodes according to the mapping relationship to obtain the semantic impact score of each process node. Generate an impact summary for each affected process node, sort them from highest to lowest business priority, and classify the impact... Process nodes whose quantified changes exceed a predetermined threshold are marked as high-impact. A draft process modification is automatically generated using a set of standardized editing templates. Each draft includes suggested modifications, reasons for modifications, quantified expected effects, implementation risks and confidence levels, and rollback and pilot recommendations. An implementation confidence score is then calculated for each draft. If the implementation confidence score is higher than the automatic on-chain threshold and the expected cost is lower than the cost ceiling, it is automatically sent to the pilot program. Otherwise, it is submitted for manual review, generating an interpretable traceable document that can be viewed by the reviewer. The draft is then written to the change management queue, and its version and approval flow route are recorded. Monitoring indicators are set for automatically sent pilot changes, and the results are automatically summarized after the pilot period for final decision-making.

[0033] The behavioral intervention module records the approval behavior of approvers and establishes a decision intervention model. When approvers handle new cases, it provides personalized suggestions and tracks the adoption status and results. The fusion inference module integrates approval data with external data to construct a policy-effect correlation graph and analyze the causal relationship between approval process optimization and changes in socio-economic indicators. The decision support module performs virtual simulations of policies before the introduction of new policies or when optimizing existing policies to predict their potential effects in reality.

[0034] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An automated optimization system for administrative approval processes, characterized in that, include: The module includes twin modeling, simulation prediction, scheme generation, feature modeling, dynamic routing, semantic parsing, change analysis, behavioral intervention, fusion inference, and decision support. The twin modeling module is used to construct a process simulation model that runs synchronously with the real approval system; The simulation prediction module is used to inject various extreme or sudden scenarios, dynamically deduce the chain reaction of the approval process, and predict the node where the bottleneck will first appear. The solution generation module automatically generates multiple feasible optimization solutions based on the bottleneck prediction results. The feature modeling module is used to generate deep features for each newly submitted application and to generate capability features for approvers and positions. The dynamic routing module is used to construct a matching scheme between applications and approvers, and convert it into dynamic routing instructions in real time; The semantic parsing module is used to perform semantic parsing on laws, regulations, departmental rules, and policy documents to construct a policy semantic network. The change analysis module updates the policy semantic network when a new policy is released or an old policy is revised, predicts the impact of policy changes on existing approval processes, and automatically generates draft process modifications. The behavioral intervention module is used to record the approver's approval behavior and establish a decision intervention model. When the approver handles a new case, it provides personalized suggestions and tracks the adoption status and results. The fusion inference module is used to integrate approval data with external data, construct a policy-effect correlation graph, and analyze the causal relationship between approval process optimization and changes in socio-economic indicators. The decision support module is used to perform virtual simulations of policies before the introduction of new policies or when optimizing existing policies, and to predict their potential effects in reality.

2. The administrative approval process automatic optimization system according to claim 1, characterized in that, The specific steps of the simulation prediction module in predicting the node where the bottleneck will first occur are as follows: S1.1: Deploy a data acquisition pipeline to access raw event streams from various data sources, and perform field normalization processing on each raw record. Based on each approval process, set up a process node dictionary and build a corresponding mapping rule engine to map each raw event to one or more process nodes. Then, aggregate the accessed data into a case event sequence in chronological order and generate metadata for each event. Based on all case event sequences, extract the node set and directed edge set to establish an initial process directed graph, and count the historical number of outgoing and incoming edges for each node. S1.2: For paths with parallel or conditional branches in the directed graph of the process, calculate the conditional transition probability in layers according to the conditional features to generate multiple transition matrices indexed by the condition. Estimate the transition probability with time relationship by time window to obtain the changes of each process over time. Then extract the service time sample from the historical processing record of each process node, perform outlier detection and truncation on each service time sample, and then use the maximum likelihood or moment method to estimate the distribution parameters. S1.3: Calculate the concurrent processing capability and average single-person efficiency of the computing nodes. For nodes with a mixture of manual and automatic processing, establish manual service rate and automatic service rate respectively, and synthesize the mixed service rate. Set the resource set of each node, establish the available quantity time series for each type of resource, calculate the type and quantity of resources consumed by each case on the node, establish the corresponding resource consumption model, and embed resource constraints as hard constraints or soft constraints into the resource consumption model. S1.4: Use natural language processing to extract entities, actions and logical relationships from various documents such as policies, regulations, departmental rules and explanatory documents, then use rule expressions or logical predicates to represent each policy constraint, and generate application condition functions for each rule to make Boolean or scoring-based application judgments for specific cases. After that, all rules are constructed into a directed dependency graph to provide a priority resolution mechanism and a traceable audit chain for inconsistencies or conflicts between rules. S1.5: Establish a corresponding queue network based on the directed graph of the process, model each node as a service station with a finite number of concurrent connections, and calculate the queuing length and waiting time of each node under a given input load based on the transition probability matrix and service rate. Then, use the approximate queuing theory formula to evaluate the performance of each node under different injection traffic scenarios, so as to identify the nodes that first enter a high utilization or saturation state under specific pressure and mark them as bottleneck nodes.

3. The administrative approval process automatic optimization system according to claim 2, characterized in that, The specific form of the resource consumption model described in S1.3 is as follows: ； The resource constraints of this resource consumption model are specifically manifested as follows: ； In the formula, Represents time Time, node For resource types Utilization rate; Represents time Time, node Available resources Total amount; Represents time Constantly occupying , A set of active case indexes; Representative Cases At the node Resources Instantaneous demand; Represents time At any given moment, are all node resource constraints met simultaneously? The specific calculation formula for the application condition function described in S1.4 is as follows: ； In the formula, Representative Rules Regarding the case Applicable confidence scores; Representative Cases eigenvectors; as well as Representing rules respectively Weights and biases; This represents the activation function.

4. The administrative approval process automatic optimization system according to claim 2, characterized in that, The specific steps by which the scheme generation module automatically generates multiple feasible optimization schemes are as follows: S2.1: Receive the output of each bottleneck node and its observations from the bottleneck prediction module, unify each observation to the same time scale and statistical caliber, and then perform exponential smoothing on the arrival rate and service rate of each bottleneck node to obtain a stable baseline estimate. After that, enumerate the feasible intervention types of each bottleneck node, establish a parameterized representation of each intervention type, and form a set of candidate vectors for each node. S2.2: Convert each candidate vector into a new equivalent service rate or parallel quantity that the corresponding node can provide after intervention. Based on the converted candidate vectors and corresponding nodes, establish multiple candidate schemes. Use the queuing approximation method to calculate the baseline latency and post-intervention latency at the node level in each candidate scheme. Then, use the traffic minimization principle to calculate the throughput change of the node before and after intervention in each candidate scheme, and propagate it to the connected upstream and downstream nodes to evaluate the impact of backlog and resource idleness. S2.3: Calculate the indicators of each candidate solution and establish a vectorized target set. Based on the Pareto dominance judgment, eliminate the solutions that are superior to other solutions in all objectives to establish a Pareto set. Then, for each solution in the Pareto set, recalculate the indicators in a set of representative scenarios set automatically or manually, and obtain the average performance and variance of each solution in the scenario set to evaluate the robustness and worst-case performance of each candidate solution. Then, use an interpretable weighted scaling method to calculate the corresponding comprehensive score based on the multidimensional objectives of each selected candidate solution. S2.4: If the overall score is higher than the preset threshold, the corresponding candidate solution is retained; otherwise, it is eliminated. The final decision support package for each candidate solution is generated, including key indicators and the mean and confidence interval in the scenario test, along with suggestions on the implementation order and the corresponding rolling implementation threshold.

5. The administrative approval process automatic optimization system according to claim 4, characterized in that, The specific steps of the dynamic routing module in constructing a matching scheme between applications and approvers and converting it into dynamic routing instructions in real time are as follows: S3.1: Read the descriptive and structured features of each application from the access system, as well as the capabilities, experience, authority, and current pending tasks of each approver. Then, perform Min-Max normalization on the two types of features, and then perform L2 standardization on the processed features. Finally, fill in the missing values ​​in the standardized features by interpolation. Based on the preprocessed features, calculate the fit of each application-approver pair by cosine similarity. S3.2: Based on the premise that each application is assigned at least once, the total workload processed by each approver within a unit of time cannot exceed its capacity, and if the approval path allows multiple hops, the assignment is regarded as an allocation to a position or queue, rather than individual individual needs. Decision variables and hard constraints are set, the delay and risk of each application-approver pair are calculated, and a multi-objective function of the process is constructed. Then, based on the adaptation benefits, work costs and overload penalties, a corresponding local utility function is constructed for each approver. S3.3: The final application-approver matching scheme is transformed into a combinatorial optimization problem with capacity constraints. One objective is selected as the main constraint and a series of upper bounds are set on it. The remaining objectives are used as optimization objects, and different solutions are obtained to obtain the set of applications that each approver is willing to accept. Then, based on the set of applications that each approver is willing to accept, each decision variable is updated, and the next round of iteration is carried out until the changes in the multi-objective function values ​​of multiple rounds converge to the preset interval. S3.4: Collect all candidate solutions generated after the iteration, calculate the index vector of each candidate solution on each objective, and remove candidate solutions that are "better or not inferior and at least one better" than other candidate solutions on all objectives. Use linear programming relaxation to obtain the baseline solution of each candidate solution. Then, discretize each relaxed candidate solution through greedy correction. If any approver is overloaded after discretization, backtrack according to the candidate solution with the best multi-objective function value, and at the same time, reassign the unaccepted tasks to the next candidate approver. S3.5: Convert each candidate solution into an actual routing instruction. Each instruction must include at least the target approver, task ID, routing priority, expected completion time or deadline, and processing suggestions. Calculate the priority score for each routing instruction and then schedule them in descending order of the calculated priority scores.

6. The administrative approval process automatic optimization system according to claim 1, characterized in that, The specific steps for constructing the policy semantic network using the semantic parsing module are as follows: S4.1: Collect text data of various laws, regulations, and notices in different formats. Then, use OCR technology to convert the paper-based originals into electronic formats, unify the character set, extract the structured units of each text data at the document level (document, clause / section, sentence, and clause), and retain metadata such as file ID, issuing authority, and publication time. Perform word segmentation, word form restoration, and stop word removal on each text data. At the same time, protect proper nouns and terms with a dictionary. Finally, use a word vector model to generate an embedding vector for each sentence. S4.2: Identify the boundaries of legal clauses and intra-sentence condition separators in the legal text, perform syntactic dependency analysis on parallel or subordinating conjunctions, locate the main conditional phrases and subject-predicate structures, and split each sentence into multiple candidate clauses based on the analysis results. Calculate the sentence break boundary confidence score of each candidate clause and retain candidate clauses with sentence break boundary confidence scores higher than the preset confidence threshold. S4.3: For each split clause, legal domain entity categories are extracted using named entity recognition technology. For each retained clause, named entity recognition is performed to extract legal domain entity categories. Then, attributes are added to the extracted entities, and the confidence score and source location index of the entities are given. Then, synonymous entities are merged into a unified entity identifier using a terminology database or thesaurus. After that, the conditional components and action components are located in each clause, and the association strength of each pair (conditional fragment, action fragment) is calculated to establish candidate condition-action pairs. S4.4: Identify the quantifiers, time limits, and permission levels in each candidate condition-action pair, and map each candidate condition-action pair into a formal logical expression. For clauses containing complex logic, convert them into a unified standard form. Then, assign corresponding rule confidence and scope of application to each logical rule. Then, represent the extracted entities, conditions, actions, and their relationships in the form of tuples. Finally, construct each tuple into a directed labeled graph and establish the corresponding adjacency structure.

7. The administrative approval process automatic optimization system according to claim 1, characterized in that, The specific steps of the change analysis module in predicting the impact of policy changes on existing approval processes and automatically generating draft process modifications are as follows: S5.1: Receive newly released or revised policy documents. If the document is a scanned copy, perform OCR processing. If it is HTML, perform DOM cleaning. Then, extract the document metadata from each policy document after processing and record the source link. Then, divide the document into articles, clauses, and clause-level units according to the legal hierarchy. Perform language preprocessing on each unit, calculate the weighted text vector of each clause, and write the segments, vectors, and metadata into the candidate node pool of the semantic network. At the same time, retain the original text index of the policy for source tracing. S5.2: Extract the entity, condition, and action of each clause using NER technology, and establish corresponding candidate rule fragments. Then, retain the original text fragment and source index of each candidate rule fragment, calculate the confidence value of each candidate rule fragment, and mark the candidate rule fragments with confidence values ​​lower than the preset threshold as requiring manual review. Then, establish a candidate rule item list according to the confidence value from largest to smallest. The candidate rule list includes information such as condition, action, entity, confidence value, source, and effective time interval. S5.3: Alignment objects for each candidate rule fragment in the semantic network are searched in parallel through text similarity matching, entity / term homonym or normalization matching, and logical matching. Corresponding evidence scores and mapping suggestions are generated for each type of matching. Evidence fusion is performed on multiple mapping suggestions, and the final mapping action decision is generated. The source information of all mapping actions is recorded. The attributes, edges and confidence of each node after mapping are written into the semantic network. The edge weights and node confidences of existing edges are updated according to the evidence accumulation rule. An incremental index is triggered after the update is completed. S5.4: Set the injection value for each newly added rule node, simulate disturbance, and use the graph propagation algorithm to propagate the disturbance along the weighted directed edges to other rule nodes, entity nodes and nodes that are ultimately mapped to the process. Iterate for multiple rounds until the preset propagation depth limit is reached. Record the intermediate impact value of each round of propagation, generate the cumulative impact score and propagation path source of each semantic network node, and pass the cumulative impact value of the semantic node to the corresponding approval process node according to the mapping relationship to obtain the semantic impact score of each process node. Generate an impact summary for each affected process node and sort it from high to low according to business priority. S5.5: Process nodes whose impact score or expected quantitative changes exceed a predetermined threshold are marked as high-impact. A process modification draft is automatically generated according to a set of rule-based editing templates. The draft items include suggested modification points, reasons for modification, expected quantitative effects, implementation risks and confidence levels, and rollback and pilot recommendations. Then, the implementation confidence score of each draft is calculated. If the implementation confidence score is higher than the automatic on-chain threshold and the expected cost is lower than the cost limit, it is automatically issued to the pilot. Otherwise, it is submitted to manual review, and an interpretable traceability is generated that can be viewed by the reviewer. The draft is then written to the change management queue and the version and approval flow route are recorded. Monitoring indicators are set for the automatically issued pilot changes. At the same time, the effects are automatically summarized after the pilot period for final decision-making.

Images