RPA government affair digital service platform based on large model driving

Abstract

The invention provides an RPA government affair digital service platform based on large model driving. The RPA government affair digital service platform comprises an acquisition module, a task atlas construction module, a semantic extraction module, a task-data matching module and a task completion module. According to the method, the logic architecture of the RPA in a government affair scene is reconstructed, so that the RPA has the capability of integrating semantic comprehension, task cognition, data perception and execution scheduling, and can automatically adapt to a complex application environment of multiple departments, multiple processes and multiple data sources.

Description

Technical Field

[0001] The present invention belongs to the field of big models, and in particular relates to an RPA government digital service platform driven by big models. Background Art

[0002] As the digital transformation of government continues to deepen, government agencies at all levels are committed to improving the intelligence and automation of government services. Against this backdrop, Robotic Process Automation (RPA) technology has been widely adopted in government scenarios to complete numerous repetitive, rule-based, and low-logic complexity tasks, such as document archiving, information entry, and cross-system data synchronization. However, traditional RPA systems still primarily rely on a "flowchart + rule template" approach, requiring developers to manually define operational paths based on business system interfaces and data interaction rules. This leads to extremely high system maintenance costs once the process changes. Furthermore, faced with inconsistent operational procedures across departments and regions, the portability and generalization capabilities of RPA systems are severely limited.

[0003] In recent years, the rise of artificial intelligence, particularly large language models (LLMs), has enabled significant advances in natural language understanding, unstructured data processing, and abstract modeling of complex tasks. This has enabled RPA systems to shift from the "execution layer" to the "decision-driving layer." However, in real-world government operations, a series of unresolved challenges remain. First, government processes often involve multiple roles and approval nodes (such as business initiation, departmental review, financial verification, and leadership signature). Traditional RPA systems are unable to automatically identify the full scope of the task flow and the required cross-role operation chain based on user natural language input. Second, government data formats are complex. In addition to standard structured forms, a large amount of information comes from unstructured content such as PDFs, Word documents, scanned documents, and email bodies. This content requires semantic understanding to be properly used in process execution. Third, during actual execution, task sequences, data access, and operational interfaces often vary depending on the specific business context, such as time, location, and personnel authority. Traditional RPA systems lack dynamic adjustment capabilities and are often unable to adapt to complex and changing execution environments. In addition, government affairs have extremely high requirements for system compliance and controllability. The system must not only be able to complete tasks, but also ensure the auditability of the execution process, the rationality of tasks and the compliance of authority.

[0004] There have been some exploratory attempts within the industry to introduce natural language processing technology or large models into RPA systems, such as guiding user operations through conversational interfaces or using pre-trained models to parse task descriptions. However, these methods are typically limited to semantic recognition of single-step operations or simple process matching, and are unable to support a complete government task chain involving multiple roles, multiple stages, and multiple systems. Furthermore, there is a lack of a systematic architectural design that can organically integrate the understanding capabilities of large models with process decomposition, data scheduling, and operational execution. Therefore, there is an urgent need for a new generation of "intelligent, highly generalizable, highly adaptable, and compliant" RPA government service platforms, with large models as the core driving engine and combined with mechanisms such as task cognitive modeling, data structured perception, and automated decision-making and scheduling. This platform can systematically address the shortcomings of current government RPA platforms in the three major areas of task understanding, process construction, and automated execution. Summary of the Invention

[0005] The purpose of this invention is to propose an RPA government digital service platform driven by a large model to address the problems existing in the government RPA automation process, such as difficulty in multi-role collaboration, difficulty in processing heterogeneous data, difficulty in dynamic process adaptation, and difficulty in ensuring operational compliance.

[0006] To achieve the above objectives, the present invention provides an RPA government digital service platform driven by a large model, the method comprising:

[0007] An acquisition module is used to obtain a user's natural language task description and parse it into a structured task intention object; wherein the structured task intention object includes: a task type, a set of participating roles, and an action chain sequence;

[0008] A task graph construction module is used to generate a task graph based on the structured task intent object by binding role-action semantics and reasoning on conditional edges; wherein the structure of the task graph includes: a plurality of task nodes and edges, wherein the edges are used to describe the dependency relationship of task execution, including sequential edges and conditional edges; and the task nodes are used to describe action descriptions, execution roles, and action numbers;

[0009] A semantic extraction module is used to extract data semantic blocks from heterogeneous government data sources based on several task nodes of the task graph, combined with action requirement vectors, role permissions and data timeliness, and form a data semantic block set;

[0010] A task-data matching module, configured to convert the task node into a corresponding action and action set; wherein the action includes an execution role, an action type, an operation target data block, and execution parameters;

[0011] The task completion module is used to execute a set of actions through a priority scheduling mechanism according to the dependency relationship of the task graph, monitor the process status in real time and perform dynamic rescheduling.

[0012] Furthermore, the acquisition module is used to obtain the user's natural language task description and parse it into a structured task intention object, specifically including:

[0013] The acquisition module uses the RoBERTa-large model to generate a context encoding matrix for each token corresponding to the user's natural language task description. At the same time, based on the context encoding matrix, a multi-channel structure parser is used to predict the task type, the set of participating roles and the preliminary action chain respectively, and the task type, the set of participating roles and the action chain sequence are combined into a structured task intention object.

[0014] Furthermore, the RoBERTa-large model is a 24-layer Transformer model with 1024 hidden units. It is fine-tuned using a joint optimization approach, with the AdamW optimizer using a linear warmup strategy, a batch size of 32, and a maximum input length of 512 tokens. An early stopping strategy is used during training.

[0015] The tasks of the RoBERTa-large model include:

[0016] Main task: task type classification;

[0017] Auxiliary task: action sequence extraction.

[0018] Furthermore, the task graph construction module is used to generate a task graph based on the structured task intention object by binding role-action semantics and reasoning on conditional edges, specifically including:

[0019] The task graph construction module uses the role binding function to determine the execution role corresponding to each operation action according to the action chain sequence, and performs role-action binding; constructs a set of sequential edges between actions according to the natural order of the action chain sequence; generates conditional edges based on the conditional edge reasoning; combines sequential edges and conditional edges, as well as task nodes, to generate a task graph.

[0020] Furthermore, constructing a set of sequential edges between actions according to the natural order of the action chain sequence specifically includes:

[0021] For adjacent action pairs in the action chain sequence, add a directed edge to indicate that the first action must be completed before the second action can be performed;

[0022] The generating of conditional edges according to the conditional edge reasoning specifically includes:

[0023] According to the business specification clauses corresponding to the task type, conditional edges are generated through rule matching reasoning; if the business specification clauses indicate that there is a conditional transfer from the current task node to the adjacent task node, it means that the two task nodes are connected by a conditional edge; otherwise, it means that the conditional edge does not exist.

[0024] Furthermore, the semantic extraction module is used to extract data semantic blocks from heterogeneous government data sources based on several task nodes of the task graph, combined with action requirement vectors, role permissions, and data timeliness, and form a data semantic block set. The steps specifically include:

[0025] Encode the action description of each task node, extract the task data requirement vector, construct a set of government heterogeneous data sources based on the task data requirement vector, and encode the data blocks of the government heterogeneous data sources using the BERT encoder to obtain a vector representation of each data block;

[0026] A multi-objective optimization matching function is constructed by combining the cosine similarity between the action requirement vector and the vector representation of each data block, the penalty term generated when the role authority does not match, and the regularization term of the difference between the current time and the data update time to generate a scoring function.

[0027] The optimal data block binding set for each task node is screened out through the scoring function as the data semantic block set.

[0028] Furthermore, the step of selecting the best data block binding set for each task node through the scoring function specifically includes:

[0029] For each task node, select the data block whose scoring function is greater than the set threshold as the associated data;

[0030] If there is no data block that meets the conditions, a supplementary request will be initiated or an abnormal data extraction prompt will be issued to ensure the integrity of the process.

[0031] Furthermore, the task-data matching module is used to convert the task node into a corresponding action and action set, and the specific steps include:

[0032] Obtain the task data requirement vector and the most suitable data block corresponding to each node in the task graph, and obtain the comprehensive matching score;

[0033] The data block of the task with the highest comprehensive matching score and exceeding the threshold is selected as the operation target, and a corresponding action is generated for the corresponding task node based on the task graph.

[0034] Furthermore, the task data requirement vector corresponding to each node in the task graph and the most suitable data block are obtained. If there is no qualified data block, the action completion mechanism is automatically triggered; wherein the action completion mechanism is expressed as:

[0035]

[0036] Among them, g 补 To complete the action, TemplateMatch represents the action template library In the example, according to node v i The action requirements and its upstream node v i-1 The context output matching generation; Context(v i-1 ) is the output data feature of the preceding node;

[0037] Insert the generated completion action as a new node into the task graph and update the corresponding edge to keep the process chain coherent.

[0038] Furthermore, the task completion module is used to execute an action set through a priority scheduling mechanism according to the dependency relationship of the task graph, monitor the process status in real time and perform dynamic rescheduling, and the steps specifically include:

[0039] Build a dispatch ready queue and add all actions with an in-degree of 0 to the dispatch ready queue;

[0040] According to the dependency relationship of the task graph, each time an action is executed, the executable status of the subsequent actions is dynamically updated, and the action with the highest priority is selected from the scheduling ready queue each time for action execution; wherein the action execution uses the RPA control interface to perform task requests.

[0041] The beneficial technical effects of the present invention are at least as follows:

[0042] In response to the problems existing in the above-mentioned government RPA automation process, such as difficulty in multi-role collaboration, difficulty in handling data heterogeneity, difficulty in dynamic process adaptation, and difficulty in ensuring operational compliance, the present invention proposes a government digital service platform architecture that integrates task cognition and process scheduling driven by a large model. The core idea of ​​the present invention is to no longer use RPA only as a low-level executor, but to introduce a large model with natural language understanding and complex task modeling capabilities, and use it as the "cognitive engine" and "task generation core" of the system to achieve automatic construction of the entire process from user intention understanding to multi-role task chain generation, and then to data perception and action scheduling. Specifically, the present invention realizes the automatic identification of role responsibilities and task logical relationships from the user's natural language by constructing a task cognition structure oriented to the government context; solves the bottleneck of traditional RPA's inability to process unstructured data by abstractly modeling government data sources as a unified information unit; and dynamically generates the optimal execution path based on the system resource status and business compliance judgment by introducing a decision-making scheduling mechanism based on task chain generation.

[0043] Compared to existing methods, this invention does not focus on optimizing a single RPA module. Instead, it comprehensively reconstructs the logical architecture of RPA in government scenarios, enabling it to integrate the four capabilities of semantic understanding, task recognition, data perception, and execution scheduling, enabling it to automatically adapt to complex application environments involving multiple departments, multiple processes, and multiple data sources. This platform not only enhances automation capabilities but also significantly lowers the integration threshold between government systems, significantly enhancing the versatility, flexibility, and regulatory controllability of RPA systems. This provides a scalable, evolvable, and implementable intelligent service platform for governments in their digital transformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0045] Figure 1 This is a flowchart of the RPA government digital service platform driven by a large model in the present invention. DETAILED DESCRIPTION

[0046] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0047] like Figure 1 As shown, the embodiment of the present invention provides an RPA government digital service platform based on a large model drive, and the method includes the following modules 1 to 5:

[0048] Acquisition module 1 is used to obtain the user's natural language task description and parse it into a structured task intention object; wherein the structured task intention object includes: task type, participating role set and action chain sequence;

[0049] A task graph construction module 2 is configured to generate a task graph based on the structured task intent object by binding role-action semantics and reasoning on conditional edges; wherein the structure of the task graph includes: a plurality of task nodes and edges, wherein the edges are used to describe the dependency relationship of task execution, including sequential edges and conditional edges; and the task nodes are used to describe the action description, execution role, and action number.

[0050] Semantic extraction module 3 is used to extract data semantic blocks from heterogeneous government data sources based on several task nodes of the task graph, combined with action requirement vectors, role permissions and data timeliness, and form a data semantic block set;

[0051] A task-data matching module 4 is used to convert the task node into a corresponding action and action set; wherein the action includes an execution role, an action type, an operation target data block and an execution parameter;

[0052] The task completion module 5 is used to execute the action set through the priority scheduling mechanism according to the dependency relationship of the task map, monitor the process status in real time and perform dynamic rescheduling.

[0053] As one of the implementation methods, in the acquisition module 1, the purpose of this module is to parse the natural language task description q entered by the user into a structured task intention object T as the starting point of the automation process of the government RPA platform. The expression of government tasks is highly flexible and often includes multiple roles, multiple steps, and complex semantic sequences. Therefore, it is necessary to design a method that can parse the task type τ and the set of roles involved. and preliminary action chains methods, rather than just intent classification or slot extraction.

[0054] The input to this module is a natural language task description, q, such as "Please help me apply for a special subsidy for retirees in my unit. This requires approval from the Human Resources Department and the responsible leader." This text typically implicitly contains the task objective (applying for a special subsidy), the participating roles (the unit, the Human Resources Department, the responsible leader), and the action chain (application-submission-approval).

[0055] First, in order to transform the input q into a contextual semantic encoding that can be structured and parsed, we use a RoBERTa-large model that has been fine-tuned in the government affairs field, denoted as Model M .

[0056] Among them, Model M The model was constructed as follows: The model was initialized with weights based on the open-source RoBERTa-large (24-layer Transformer, 1024-dimensional hidden units), and then fine-tuned on a government data set. The fine-tuning data sources included government policy documents (such as the National Policy Release Platform); standard operating manuals from government service centers; and government office logs and conversational data.

[0057] Furthermore, the fine-tuning task is designed as follows:

[0058] Main task: task type classification;

[0059] Auxiliary tasks: action sequence extraction;

[0060] Fine-tuning was performed using a joint optimization approach, with AdamW as the optimizer, a linear warmup strategy, a batch size of 32, and a maximum input length of 512 tokens. Training was performed using an early stopping strategy (stopping if the validation set loss did not decrease continuously).

[0061] It is understandable that through this fine-tuning, the Model M It can better understand proprietary concepts such as task objectives, role names, and operational actions in government expressions, and has strong domain adaptability.

[0062] Input q passes through Model M Process and output the context encoding matrix H of each token:

[0063] H=Model M (q);

[0064] Where H=[h1,h2,...,h n ], n is the number of tokens, each h i It is a 1024-dimensional vector containing rich contextual dependencies and domain knowledge.

[0065] Then, based on H, a multi-channel structure parser is designed to predict the task type τ, the set of participating roles, and and action chain sequences In order to ensure the logical consistency of the analysis results, the following joint optimization objectives are proposed:

[0066]

[0067] in, Cross entropy loss for task type classification; Sigmoid binary classification loss for multi-label role recognition; CRF sequence annotation loss for action phrase extraction; Action sequence reasoning loss ensures the reasonable arrangement of action chains; Ω: character action consistency regularization term to prevent character binding actions with logical conflicts.

[0068] In actual reasoning, By constructing an action phrase dependency graph, we can infer the order of actions. For example, the action "returning materials" must precede the action "resubmitting." Ω uses role-action mutual information penalty to prevent the "approver" from mistakenly executing the action "supplementing materials."

[0069] Example deduction:

[0070] For the input: "Help me apply for special subsidies for retirees, which requires preliminary review by the personnel department and approval by the responsible leader",M Output H;

[0071] Multi-channel structure analysis yields τ = "special subsidy application";

[0072] Multi-label recognition

[0073] Action chain analysis

[0074] The action sequence module ensures a reasonable flow from "fill in" -> "preliminary review" -> "approval";

[0075] The role consistency rule ensures that the HR department only performs the "preliminary review" action and is not bound to the "approval" action.

[0076] The final output is a structured task intent object:

[0077]

[0078] Where: τ is the task type (such as "special subsidy application"); For the collection of participating roles; An action chain sequence.

[0079] As one of the implementation methods, in the task graph construction module 2, the purpose of this module is to construct the structured task intention object based on the output of the previous module. Automatically construct an executable multi-role task graph G = (V, E). The task graph G is the structural skeleton of the entire government RPA process, where the node set V represents each detailed specific operation unit (such as "filling out the application form" and "uploading information"), and the edge set E represents the sequential dependency relationship between the operation nodes (such as "the form must be submitted before approval"). To adapt to the characteristics of multi-role collaboration and conditional jumps or branches in task processes in government scenarios, this module proposes an innovative graph construction method based on role-action association graph reasoning. This ensures that the task graph structure can not only accurately express task details but also flexibly adapt to slightly different process execution modes between different departments.

[0080] In order to construct a multi-role task map, we first need to We use the action list extracted from the function to determine the execution role corresponding to each action. We introduce a simple and effective role binding function BindRole to determine the binding relationship between actions and roles:

[0081]

[0082] Among them, g i Represents a preliminary action list The i-th action in rj Is a collection of characters The jth character in Sim(g i ,r j ) is the semantic similarity score between the action and the role. It is calculated by encoding the action phrase and role description into vectors respectively (such as using the Word2Vec model in the government affairs field) and calculating the cosine similarity; argmax means selecting the role with the highest similarity as the executor of the action.

[0083] For example, the action "Submit Approval Form" is usually performed by "Applicant", while "Approval Process" is usually performed by "Approval Leader". This method can automatically assign the appropriate execution role to each action without manual intervention.

[0084] After binding the roles, you need to determine the task graph node set V. Each node v i Includes: action description (such as "fill out the form"); execution role (such as "applicant"); action number (for sorting and indexing).

[0085] Then, according to the natural order of actions (given by Given), we preliminarily construct a set of sequential edges E between actions. Specifically, for The adjacent action pairs in (g i ,g i+1 ), add a directed edge (v i ,v i+1 ), indicating that the action v must be completed first i , then v i+1 ".

[0086] However, actual government processes often involve conditional branches (e.g., applications that fail to pass the initial review and require resubmission). To address this, this module innovatively introduces a conditional jump rule inference mechanism based on task type τ. By analyzing the standard process specifications for task type τ (e.g., in the case of special subsidy approvals, applications must be returned for supplementary materials if the initial review fails), conditional branch edges are automatically added to the graph. Conditional branch edges control execution path switching by attaching judgment conditions (e.g., "failed the initial review")

[0087] The specific conditional edge reasoning method is as follows:

[0088]

[0089] Among them, CondEdge(v i ,v j ) indicates whether the task node v i to v jA conditional jump edge is established between them; the judgment basis is inferred based on the business specification clauses corresponding to τ through rule matching (such as the government process library or the preset business rule table).

[0090] For example, in the "Special Subsidy Application" process, if the "Initial Review Rejected" action is received, the "Supplement Materials" action must be re-executed. When constructing the task graph, this process specification is retrieved based on τ and a conditional edge is automatically added between the "Initial Review Rejected" node and the "Supplement Materials" node.

[0091] Finally, by combining sequential edges and conditional edges, the task graph G = (V, E) is constructed.

[0092] This module outputs the task graph G = (V, E), where:

[0093] V is a set of task nodes, each node contains action description, role binding, and action number;

[0094] E is an edge set, which includes action sequence dependency edges and conditional jump edges inferred based on task specifications.

[0095] This diagram fully expresses the logical dependencies between all operation modules, participating roles and actions required from the initiation to the completion of a government task, and is the direct basis for subsequent action generation and process execution scheduling.

[0096] As one implementation method, in the semantic extraction module 3, this module is based on the task graph G = (V, E) output by the previous module, where each node v i Describes an action performed by a specific role. The goal is to extract a set of structured, context-related, and permission-controlled data semantic blocks from heterogeneous government data sources based on the specific needs of each action. This module does not rely on general data retrieval, but rather relies closely on task-driven data perception based on the node actions and role requirements of the task graph, ensuring accurate data extraction and directly supporting the generation of subsequent RPA actions. To address the challenges of heterogeneous data sources, multiple levels of permissions, and rapidly changing data in government scenarios, this module introduces an innovative multi-objective matching regularized extraction mechanism to achieve high-quality, low-redundancy extraction of data blocks.

[0097] Furthermore, in order to achieve efficient data extraction for action requirements, firstly for each task node v i Encode the action description and extract the task data requirement vector d i The encoder adopts a two-layer bidirectional GRU (256-dimensional hidden units per layer) structure, denoted as Encoder D , implemented as follows:

[0098] d i =EncoderD (Action(v i ));

[0099] Among them: Action(v i ): node v i Action text, such as "Upload approval materials"; Encoder D : Two-layer Bi-GRU encoder, outputting the low-dimensional demand feature d of the action i .

[0100] Next, from the pre-standardized heterogeneous government data source set Retrieve potential data blocks from each d j It consists of information such as field name, field value, data source, ownership system, version number, etc., and is fine-tuned by the BERT encoder in the field. B Encode and get the vector representation e of each data block j .

[0101] To achieve efficient matching between action requirements and data blocks, we designed the following innovative multi-objective optimization matching function. This function introduces data permission consistency regularization and temporal freshness regularization, comprehensively considering three factors: semantic similarity, access rights, and data timeliness:

[0102]

[0103] Among them, cos(d i ,e j ): cosine similarity between action requirement vector and data block vector; AccessPenalty(v i ,d j ): Penalty item generated when role permissions do not match, if v i Execution role no access j If the data source has permission, the penalty is 1, otherwise it is 0; StalePenalty(d j ): data block d j The time freshness penalty is defined as the regularization term of the difference between the current time and the data update time. The larger the time difference, the higher the penalty. λ1 and λ2 are the regularization term weight parameters, which are used to regulate the balance between semantic matching and security compliance. The empirical settings are 0.5 and 0.3.

[0104] Among them, through Scoring can filter out each node v i The best set of data block bindings Right now:

[0105] For each v i , select score Data blocks d that are larger than the set threshold θ j as linked data;

[0106] If there is no data block that meets the conditions, a supplementary request will be initiated or an abnormal data extraction prompt will be issued to ensure the integrity of the process.

[0107] For example, node v i If the action is "HR Officer Upload Retired Employee List", the system will automatically retrieve data blocks from the "Personnel File System" that contain the field "Retired Employee List" and whose update time is within the last 30 days. At the same time, ensure that the "HR Officer" role has read permission in the system access control table; otherwise, the data block will be penalized and excluded.

[0108] Through this multi-objective optimization mechanism, this module not only ensures the high correlation between extracted data and task actions, but also strictly controls the compliance and real-time nature of data usage, greatly improving the security and accuracy of RPA process execution.

[0109] The final output is a set of task-driven data semantic blocks Each b i A specific data unit, including field name, field value, data source system, data timestamp, and access role information, for direct call by subsequent action generation modules

[0110] As one of the implementation methods, in the task-data matching module 4, the core task of this module is to convert the abstract task node v i Converted into specific actions that can be directly executed by the government RPA system i . Actions must not only be bound to the correct data blocks, but also contain compliant execution parameters and operation paths, and ensure the logical integrity of the action chain. Taking into account the complex situations that often occur in government processes, such as data inconsistency, cross-system authority fragmentation, and task node changes, this module proposes an innovative mechanism based on data-driven action instantiation + anomaly detection and dynamic completion, and further introduces the action chain closed-loop regularization term to ensure that the generated action set A can be fully executed without interruption.

[0111] Input comes from the previous module: Task graph G = (V, E): Each node v i Contains action description, role, action number; data semantic block collection Each b i Contains field name, field value, belonging system, belonging role, and update timestamp.

[0112] First, for each task graph node v i , extract its action demand vector d i (Defined in step ③, through Encoder D Encoding), and then To ensure that the generated action not only matches the data but also conforms to the execution context, this step introduces a comprehensive matching score function:

[0113] Score(v i ,b j )=cos(d i ,e j )-λ1·RoleMismatch(v i ,b j )-λ2·Staleness(b j )+λ3·

[0114] ContextFit(v i ,b j );

[0115] Among them, cos(d i ,e j ) is the semantic similarity between the action requirement vector and the data block vector; RoleMismatch(v i ,b j ) Penalty execution role is inconsistent with data permission role; Staleness (b j ) Penalize data blocks that are outdated (based on the difference between timestamp and current time); ContextFit(v i ,b j ) is an innovative context adaptation item that measures the relationship between the data block field content and the local context of the task graph (such as the preceding node v i-1 data output) to avoid isolated actions.

[0116] Understandably, by incorporating ContextFit, we not only focus on matching individual actions, but also consider the rationality of the data flow between action chains, thereby improving overall process coherence. This feature is particularly designed for the data-intensive nature of approval chains and document submission processes in government RPA.

[0117] Then, for each v i , select the data block b with the highest score exceeding the threshold θ j As the operation target. If there is no qualified data block, the action completion mechanism is automatically triggered. The completion mechanism depends on the context node of the task graph G and is based on the predefined action template library. (Such as "Supplementary materials", "Initiate approval", "Resubmit") Generate supplementary actions 补 , and correct G and E.

[0118] The completion action generation process is formally described as follows:

[0119]

[0120] Among them, TemplateMatch represents the action template library In the example, according to node v i The action requirements and its upstream node v i-1 The context output matching generation; Context(v i-1 ) is the output data feature of the preceding node; the generated completion action g 补 Insert it into G as a new node and update edge E to keep the process chain coherent.

[0121] For example, node v i The action is "Submit Retirement Certificate", if If no data is available, then by analyzing v i-1 Output (for example, when the "Upload ID Card" action is completed), the system automatically infers the need for a new "Retirement Certificate" action and inserts it into the flow chart. This anomaly detection and completion mechanism ensures that even if data sources are missing or changed, the action chain can be dynamically corrected, improving the integrity and intelligent adaptability of the RPA system.

[0122] Finally, for each task node v i Generate action a i , including: execution role (bound from v i ); Action type (such as fill, upload, approve); Operation target data block (select or complete generation); execution parameters (such as form field ID, API interface path, input format requirements, etc.).

[0123] Complete action set A={a1,a2,...,a k}, you can proceed to the subsequent action scheduling and execution steps.

[0124] As one implementation method, in the task completion module 5, this step is based on the action set A={a1, a2,…, a k}, combined with the task graph G = (V, E), the actions are scheduled, sent, executed, and monitored according to the action dependency sequence, ultimately completing the closed loop of the entire government process. To address the large action granularity, high execution risk, and strong system heterogeneity in government RPA processes, this step innovatively proposes a priority scheduling mechanism under action dependency constraints, combined with a dynamic rescheduling strategy based on real-time feedback, to ensure the executable and fault-tolerance of the action chain.

[0125] The input includes: action set A = {a1, a2, ..., a k}, each a iIt includes roles, action types, operation data blocks, and execution parameters. The task graph G = (V, E), where E describes the dependencies (sequence / conditions) of task execution. Each action must be scheduled for execution based on its dependencies.

[0126] First, create a scheduling ready queue Q and queue all actions a with in-degree 0. i After adding an action to the queue, actions without pre-dependencies can be executed immediately. Afterwards, action scheduling strictly follows the edge set E defined in G. After each action is executed, the executable status of subsequent actions is dynamically updated.

[0127] To optimize the scheduling order, we design the following comprehensive priority scoring function, integrating dimensions such as action depth, number of subsequent dependencies, data freshness, and action execution risk:

[0128]

[0129] Where: Depth(a i ): the depth of the action in the task graph; Fanout (a i ): the number of action successor dependent nodes; Staleness (a i ): Time staleness of the bound data block; Risk(a i ) is the risk value of action failure (based on statistical estimation of action type); α, β, γ, δ are weight parameters, and empirical values ​​are (1.0, 0.5, 0.8, 1.2).

[0130] And each time select the action a with the highest priority from Q i , perform the action.

[0131] Action execution uses the RPA control interface (such as HTTP API, Selenium, UiPath robot call, etc.). The execution request structure is as follows:

[0132] Action type (e.g. "fill out a form");

[0133] Target system (such as "social security system portal");

[0134] Operation element ID (such as field ID, button ID);

[0135] Enter data (such as the retiree's name "Zhang San").

[0136] Example execution call process: Construct an HTTP POST request with the URL pointing to the RPA controller interface, such as http: / / rpa-controller / execute;

[0137] The request body includes action parameters, as shown below:

[0138] {

[0139] "actionType":"fillForm",

[0140] "targetSystem":"SocialSecurityPortal",

[0141] "fieldId":"input_name",

[0142] "value":"Zhang San",

[0143] "role": "clerk"

[0144] }

[0145] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the embodiment scheme, and no limitation is made here.

[0146] In addition, for technical details not fully described in this embodiment, please refer to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0147] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0148] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0149] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0150] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. The RPA government digital service platform driven by a large model is characterized by: The platform includes: An acquisition module is used to obtain a user's natural language task description and parse it into a structured task intention object; wherein the structured task intention object includes: a task type, a set of participating roles, and an action chain sequence; A task graph construction module is used to generate a task graph based on the structured task intent object by binding role-action semantics and reasoning on conditional edges; wherein the structure of the task graph includes: a plurality of task nodes and edges, wherein the edges are used to describe the dependency relationship of task execution, including sequential edges and conditional edges; and the task nodes are used to describe action descriptions, execution roles, and action numbers; A semantic extraction module is used to extract data semantic blocks from heterogeneous government data sources based on several task nodes of the task graph, combined with action requirement vectors, role permissions and data timeliness, and form a data semantic block set; A task-data matching module, configured to convert the task node into a corresponding action and action set; wherein the action includes an execution role, an action type, an operation target data block, and execution parameters; The task completion module is used to execute a set of actions through a priority scheduling mechanism according to the dependency relationship of the task graph, monitor the process status in real time and perform dynamic rescheduling.

2. The RPA government digital service platform based on large model drive according to claim 1 is characterized in that: The acquisition module is used to obtain the user's natural language task description and parse it into a structured task intention object, specifically including: The acquisition module uses the RoBERTa-large model to generate a context encoding matrix for each token corresponding to the user's natural language task description. At the same time, based on the context encoding matrix, a multi-channel structure parser is used to predict the task type, the set of participating roles and the preliminary action chain respectively, and the task type, the set of participating roles and the action chain sequence are combined into a structured task intention object.

3. The RPA government digital service platform based on large model drive according to claim 2 is characterized in that: The RoBERTa-large model is a 24-layer Transformer model with 1024 hidden units. It is fine-tuned using a joint optimization approach, with AdamW as the optimizer, a linear warmup strategy, a batch size of 32, and a maximum input length of 512 tokens. An early stopping strategy is used during training. The tasks of the RoBERTa-large model include: Main task: task type classification; Auxiliary task: action sequence extraction.

4. The RPA government digital service platform based on large model drive according to claim 1 is characterized in that: The task graph construction module is used to generate a task graph based on the structured task intention object by binding role-action semantics and reasoning on conditional edges, specifically including: The task graph construction module uses the role binding function to determine the execution role corresponding to each operation action according to the action chain sequence, and performs role-action binding; constructs a set of sequential edges between actions according to the natural order of the action chain sequence; generates conditional edges based on the conditional edge reasoning; combines sequential edges and conditional edges, as well as task nodes, to generate a task graph.

5. The RPA government digital service platform based on large model drive according to claim 4 is characterized in that: The step of constructing a set of sequential edges between actions according to the natural order of the action chain sequence specifically includes: For adjacent action pairs in the action chain sequence, add a directed edge to indicate that the first action must be completed before the second action can be performed; The generating of conditional edges according to the conditional edge reasoning specifically includes: According to the business specification clauses corresponding to the task type, conditional edges are generated through rule matching reasoning; if the business specification clauses indicate that there is a conditional transfer from the current task node to the adjacent task node, it means that the two task nodes are connected by a conditional edge; otherwise, it means that the conditional edge does not exist.

6. The RPA government digital service platform based on large model drive according to claim 1 is characterized in that: The semantic extraction module is used to extract data semantic blocks from heterogeneous government data sources based on several task nodes in the task graph, combined with action requirement vectors, role permissions, and data timeliness, and form a data semantic block set. The steps specifically include: Encode the action description of each task node, extract the task data requirement vector, construct a set of government heterogeneous data sources based on the task data requirement vector, and encode the data blocks of the government heterogeneous data sources using the BERT encoder to obtain a vector representation of each data block; A multi-objective optimization matching function is constructed by combining the cosine similarity between the action requirement vector and the vector representation of each data block, the penalty term generated when the role authority does not match, and the regularization term of the difference between the current time and the data update time to generate a scoring function. The optimal data block binding set for each task node is screened out through the scoring function as the data semantic block set.

7. The RPA government digital service platform based on large model drive according to claim 6 is characterized in that: The step of screening out the best data block binding set for each task node through the scoring function specifically includes: For each task node, select the data block whose scoring function is greater than the set threshold as the associated data; If there is no data block that meets the conditions, a supplementary request will be initiated or an abnormal data extraction prompt will be issued to ensure the integrity of the process.

8. The RPA government digital service platform based on large model drive according to claim 6 is characterized in that: The task-data matching module is used to convert the task node into a corresponding action and action set. The specific steps include: Obtain the task data requirement vector and the most suitable data block corresponding to each node in the task graph, and obtain the comprehensive matching score; The data block of the task with the highest comprehensive matching score and exceeding the threshold is selected as the operation target, and a corresponding action is generated for the corresponding task node based on the task graph.

9. The RPA government digital service platform based on large model drive according to claim 8 is characterized in that: The task data requirement vector corresponding to each node in the task graph and the most suitable data block are obtained. If there is no qualified data block, the action completion mechanism is automatically triggered; wherein the action completion mechanism is expressed as: Among them, g 补 To complete the action, TemplateMatch represents the action template library In the example, according to node v i The action requirements and its upstream node v i-1 The context output matching generation; Context(v i-1 ) is the output data feature of the preceding node; Insert the generated completion action as a new node into the task graph and update the corresponding edge to keep the process chain coherent.

10. The RPA government digital service platform based on large model drive according to claim 1 is characterized in that: The task completion module is used to execute the action set through the priority scheduling mechanism according to the dependency relationship of the task graph, monitor the process status in real time and perform dynamic rescheduling. The steps specifically include: Build a dispatch ready queue and add all actions with an in-degree of 0 to the dispatch ready queue; According to the dependency relationship of the task graph, each time an action is executed, the executable status of the subsequent actions is dynamically updated, and the action with the highest priority is selected from the scheduling ready queue each time for action execution; wherein the action execution uses the RPA control interface to perform task requests.

Images