Finance and tax task execution decision scheme generation method and device, equipment and storage medium

Abstract

The invention discloses a finance and taxation task execution decision scheme generation method and device, equipment and a storage medium, and relates to the technical field of finance and taxation, and the method comprises the steps: analyzing a natural language finance and taxation task to obtain a triad of a finance and taxation task semantic graph, an initial constraint set and a target weight, constructing a target rule twin graph containing time and version labels, and obtaining a target rule twin graph; converting a rule sub-graph in the graph into a rule constraint set with a space-time label, performing satisfiability judgment, performing sub-task decomposition and workflow arrangement based on the finance and tax task semantic graph to generate a target workflow skeleton after satisfiability judgment, determining a strategy set according to the finance and tax task semantic graph, and determining a target workflow skeleton according to the strategy set; a target rule twin map and a tense label are combined to form a tense caliber set, the two tense caliber sets form an evaluation space, strategy and tense anti-fact evaluation and Pareto optimization are performed on the evaluation space by using a target workflow skeleton to obtain a finance and taxation index evaluation vector, and a decision scheme set of finance and taxation tasks is generated through comparison and screening of a Pareto dominance relationship. The generation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of financial and tax technology, and in particular to a method, apparatus, equipment, and storage medium for generating execution decision schemes for financial and tax tasks. Background Technology

[0002] In recent years, large language models have made breakthroughs in understanding and generating natural language. Relying on large-scale corpora and autoregressive modeling, they possess the ability to learn context and follow instructions, and can be integrated with external tools for retrieval, computation, and data processing. However, relying solely on probabilistic generation cannot guarantee the determinism and consistency of multi-step tasks, and is susceptible to cue perturbations, knowledge timeliness, and context drift. When tasks involve multi-stage dependencies, cross-temporal scopes, and multi-source data coupling, the model output lacks verifiable evidence and stable error fallback paths, making it difficult to meet the requirements of engineering and auditability.

[0003] To improve usability, the industry has proposed an agent-based task planning framework centered on "planning-execution": a planner generates a sequence of sub-tasks, an executor schedules the computation, and a memory module or external knowledge base maintains intermediate states. While this paradigm enhances scalability, existing implementations generally suffer from three shortcomings: First, rules and definitions evolve over time, making it difficult for traditional static rule bases to express cross-version inheritance, conflicts, and dependencies, resulting in task orchestration being insensitive to temporal changes; second, multi-tool workflows lack unified formal constraints on data patterns, preconditions, and anomaly recovery, easily leading to fragile links and unreproducible experimental results; third, planning and reasoning lack satisfiability verification and traceability, making strategy comparison, risk quantification, and external verification difficult.

[0004] As can be seen from the above, how to improve the efficiency of generating decision-making plans for the implementation of fiscal and tax tasks is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for generating execution decision schemes for financial and tax tasks, which can improve the efficiency of generating execution decision schemes for financial and tax tasks. The specific solution is as follows:

[0006] Firstly, this application provides a method for generating execution decision-making schemes for fiscal and tax tasks, including:

[0007] The financial and tax tasks in natural language are parsed to obtain financial and tax task triples, which include a financial and tax task semantic graph, an initial constraint set, and a target weight. An initial rule twin graph, including time tags and version tags, is constructed based on the financial and tax task triples. Then, a consistency check is performed on the initial rule twin graph, and after the check passes, the initial rule twin graph is set as the target rule twin graph.

[0008] The rule subgraph in the target rule twin graph is determined, and the rule subgraph is transformed into a set of rule constraints including spatiotemporal labels. The satisfiability of the set of rule constraints is determined. After the determination is passed, the financial and tax task is decomposed into subtasks and the workflow is arranged based on the financial and tax task semantic graph using a preset hierarchical task network to obtain the target workflow skeleton.

[0009] Based on the semantic graph of the tax and finance task, a corresponding set of strategies is determined, and based on the twin graph of the target rules and the time label, a set of temporal definitions is determined, so as to form an evaluation space based on the set of strategies and the set of temporal definitions; the set of strategies includes different paths to execute the tax and finance task; the set of temporal definitions includes different interpretations of the same strategy at different time points;

[0010] The target workflow skeleton is used to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space to obtain target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space. The Pareto dominance relationship is then used to compare and filter the target fiscal and tax indicator evaluation vectors to obtain a set of decision schemes for executing the fiscal and tax task.

[0011] Optionally, the parsing of the natural language tax task yields a tax task triplet comprising a tax task semantic graph, an initial constraint set, and target weights, including:

[0012] The financial and tax task is parsed using dependency parsing to obtain the parsing result. Then, the event predicate, participants, time interval, and regional label are extracted from the parsing result to obtain a set of four-tuple elements including the event predicate, the participants, the time interval, and the regional label.

[0013] A semantic graph of financial and tax tasks is constructed based on the set of quadruple elements. Each node in the semantic graph represents the entity and subtasks corresponding to the financial and tax task. Each edge in the semantic graph represents the logical execution order and data transmission dependency between the nodes.

[0014] Based on the set of four elements, the financial and tax tasks are subject to time constraints, mutual exclusion constraints, numerical threshold constraints, and consistency constraints to obtain an initial constraint set. The time constraints are used to constrain the execution time of the financial and tax tasks. The mutual exclusion constraints are used to constrain the mutual exclusion between different operations. The numerical threshold constraints are used to set numerical thresholds for upper or lower limits of key values. The consistency constraints are used to ensure the consistency between data and logic.

[0015] Based on the optimization intent corresponding to the financial and tax task, the target weight corresponding to the financial and tax task is determined, and a financial and tax task triplet is constructed based on the financial and tax task semantic graph, the initial constraint set, and the target weight.

[0016] Optionally, the step of constructing an initial rule twin graph including time tags and version tags based on the tax task triplet, then performing a consistency check on the initial rule twin graph, and setting the initial rule twin graph as the target rule twin graph after the check passes, includes:

[0017] Based on the three-tuple of the financial and tax tasks, the effective time interval, applicable region and priority corresponding to the financial and tax tasks are determined, and the corresponding metadata is determined based on the effective time interval, the applicable region and the priority, and the corresponding rule nodes are constructed based on each of the metadata.

[0018] The revision relationships of rules over time, the dependency relationships of rule references, the conflict relationships of contradictory rule content, and the equivalence mapping relationships of semantically equivalent rules among the rule nodes are determined. Based on the revision relationships, dependency relationships, conflict relationships, and equivalence mapping relationships, edges between the rule nodes are determined. An initial rule twin graph is generated based on the rule nodes and their corresponding edges. Then, a consistency check is performed on the initial rule twin graph using a preset consistency check rule. After the check passes, the initial rule twin graph is set as the target rule twin graph.

[0019] Optionally, determining the rule subgraph in the target rule twin graph and converting the rule subgraph into a set of rule constraints including spatiotemporal labels, to determine the satisfiability of the set of rule constraints, includes:

[0020] Based on the aforementioned tax and finance task triplet, rule subgraphs that match in time and region are extracted from the target rule twin graph, and all applicable rule clauses in each rule subgraph are uniformly transformed into a set of rule constraints including spatiotemporal labels; the form corresponding to the rule clauses is composed of first-order logical predicates and linear arithmetic inequalities.

[0021] The set of rules and constraints is subjected to logical consistency and satisfiability determination using a satisfiability modular theory solver. If the determination result indicates that the set of rules and constraints does not meet the requirements of logical consistency and satisfiability, the rule subgraph that causes the logical conflict is located and extracted from the set of rules and constraints, and then the rule subgraph is processed.

[0022] Optionally, after the determination is passed, the step of using a preset hierarchical task network and based on the tax task semantic graph to perform sub-task decomposition and workflow orchestration on the tax task to obtain the target workflow skeleton includes:

[0023] The financial and tax tasks are decomposed from top to bottom using a pre-defined hierarchical task network and based on the financial and tax task semantic graph to obtain several task nodes and directed edges used to characterize the execution dependencies between the task nodes. Then, a target workflow skeleton is generated based on each task node and the corresponding directed edge.

[0024] Determine the accuracy of historical execution tasks corresponding to each candidate tool and the penalty for execution delay, and determine the historical average utility based on the accuracy and the penalty, and determine the historical call count corresponding to each candidate tool;

[0025] The selection score of each candidate tool in the current task node is determined by using the upper confidence bound algorithm and based on the historical utility mean and the historical call count. The candidate tool with the highest score among the selection scores is set as the target candidate tool. Then, the task node is bound to the target candidate tool to obtain the initial workflow skeleton.

[0026] Identify each upstream and downstream node pair with direct data dependencies in the initial workflow skeleton, and determine whether the output data pattern of the upstream node in the upstream and downstream node pair is semantically and structurally compatible with the input data pattern of the corresponding downstream node to obtain the data pattern compatibility verification result.

[0027] If the data pattern compatibility verification result indicates incompatibility, then a field conversion operator is inserted between the upstream and downstream node pairs using a preset field mapping rule to obtain the target workflow skeleton; the field conversion operator includes an operator for renaming data fields, an operator for filling missing data fields with default values, and an operator for standardizing and converting data units.

[0028] Optionally, the step of determining the corresponding strategy set based on the semantic graph of the tax and finance task, and determining the temporal caliber set based on the target rule twin graph and the time label, so as to construct an evaluation space based on the strategy set and the temporal caliber set, includes:

[0029] The semantic graph of the fiscal and tax task is parsed to obtain several corresponding variable decision points. A strategy set is generated based on each variable decision point. Based on the time label, the rule version corresponding to the historical rule that is related to or different from the task is selected from the twin graph of the target rule. A temporal caliber set is generated based on each rule version. Then, an evaluation space is constructed based on the strategy set and the temporal caliber set.

[0030] Accordingly, the counterfactual evaluation of strategy and temporality and Pareto optimization are performed on the evaluation space using the target workflow skeleton to obtain target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space. The Pareto dominance relationship is then used to compare and filter the target fiscal and tax indicator evaluation vectors to obtain a set of decision schemes for executing the fiscal and tax task, including:

[0031] Using the target workflow skeleton and based on preset indicators, the evaluation vectors of financial and tax indicators for each element in the evaluation space are determined to obtain the corresponding target financial and tax indicator evaluation vectors. The Pareto dominance relationship is used to filter the target financial and tax indicator evaluation vectors to obtain a set of decision schemes for executing the financial and tax tasks. The preset indicators include revenue indicators, cost indicators, risk indicators, time delay indicators, and stability indicators.

[0032] Optionally, after obtaining the set of decision schemes for performing the financial and tax tasks, the method further includes:

[0033] The upper confidence bound algorithm is used to perform a guided search on the set of decision schemes in the evaluation space to obtain candidate schemes and determine the Pareto front corresponding to each candidate scheme; the Pareto front includes several non-dominated solutions and their corresponding key influencing factors.

[0034] The set of rule constraints, the Pareto front, and the key influencing factors are merged to obtain a merged set of constraints. The satisfiability of the merged set of constraints is determined using a modular theory solver for satisfiability, and the satisfiability determination result is obtained.

[0035] When the satisfiability determination result is characterized as satisfiable, a corresponding evidence model is generated, and a structured proof summary is generated using the evidence model; the proof summary includes variable assignment, hash binding of input data, and a step-by-step derivation record from Boolean skeleton to numerical theory;

[0036] When the satisfiability determination result is characterized as unsatisfiable, the minimum unsatisfiable kernel is extracted from the merged constraint set, and each logical clause in the minimum unsatisfiable kernel is reverse-mapped to the specific rule entries in the target rule twin graph and the specific task nodes in the target workflow skeleton to obtain a source tracing report for locating and interpreting the root cause of the conflict.

[0037] Optionally, after comparing and filtering the evaluation vectors of the target fiscal and tax indicators using the Pareto dominance relationship to obtain a set of decision schemes for performing the fiscal and tax tasks, the method further includes:

[0038] The financial and tax tasks are executed based on the set of decision schemes to obtain the task execution results and the corresponding key event sequence. The key event sequence is then encoded using a preset encryption algorithm to obtain the corresponding ciphertext encoded sequence.

[0039] Each element in the ciphertext encoding sequence is concatenated with the hash value of the previous event and the ciphertext encoding of the current event in chronological order to obtain a secure hash value, and a linear hash chain is generated based on each of the secure hash values.

[0040] A Merkle tree is constructed based on each of the ciphertext encoding sequences, and the root value corresponding to the Merkle tree is determined. Then, the root value is set as the fingerprint identifier corresponding to the financial and tax task.

[0041] The key event sequence is proved using a preset proof circuit and based on the root value and the tax task triplet, and the proof result is obtained. The proof result is then sent to a preset external verification interface so that the external verifier can determine the verification result corresponding to the tax task based on the proof result.

[0042] The number of verification failures corresponding to the financial and tax task is determined. Then, a preset comprehensive loss function is used to determine the deviation value of the actual task execution delay corresponding to the financial and tax task exceeding the preset expected delay and the decrease value of the system stability index relative to the preset baseline based on the number of verification failures, so as to obtain the loss result.

[0043] The system strategy parameters corresponding to the decision scheme set are iteratively updated using a preset mirror descent algorithm and based on the loss results. Then, the updated system strategy parameters are verified using a preset historical regression dataset and each preset protection threshold to obtain the verification results. The preset protection thresholds include a decrease threshold for task verification success rate, an increase threshold for task execution latency, and a decrease threshold for system stability indicators.

[0044] If the verification result indicates that the updated system policy parameters meet each of the preset protection thresholds, then the updated system policy parameters are set as the target system policy parameters. If the verification result indicates that the updated system policy parameters do not meet each of the preset protection thresholds, then the update freeze mechanism is triggered.

[0045] Secondly, this application provides an apparatus for generating execution decision-making schemes for fiscal and tax tasks, comprising:

[0046] The rule twin graph generation module is used to parse the financial and tax tasks in natural language to obtain financial and tax task triples, which include a financial and tax task semantic graph, an initial constraint set, and target weights. Based on the financial and tax task triples, an initial rule twin graph including time tags and version tags is constructed. Then, a consistency check is performed on the initial rule twin graph, and after the check passes, the initial rule twin graph is set as the target rule twin graph.

[0047] The workflow skeleton generation module is used to determine the rule subgraph in the target rule twin graph, and convert the rule subgraph into a set of rule constraints including spatiotemporal labels, to determine the satisfiability of the set of rule constraints, and after the determination is passed, to use a preset hierarchical task network and based on the financial and tax task semantic graph to perform subtask decomposition and workflow orchestration of the financial and tax task to obtain the target workflow skeleton.

[0048] The space construction module is used to determine the corresponding strategy set based on the semantic graph of the financial and tax task, and to determine the temporal caliber set based on the twin graph of the target rule and the time label, so as to construct an evaluation space based on the strategy set and the temporal caliber set; the strategy set includes different paths to execute the financial and tax task; the temporal caliber set includes different interpretations of the same strategy at different time points.

[0049] The decision scheme set determination module is used to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space using the target workflow skeleton, to obtain the target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space, and to compare and filter the target fiscal and tax indicator evaluation vectors using the Pareto dominance relationship to obtain a set of decision schemes for executing the fiscal and tax task.

[0050] Thirdly, this application provides an electronic device, comprising:

[0051] Memory, used to store computer programs;

[0052] A processor is used to execute the computer program to implement the aforementioned method for generating execution decision schemes for financial and tax tasks.

[0053] Fourthly, this application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for generating execution decision schemes for financial and tax tasks.

[0054] As can be seen from the above, before generating the execution decision scheme for the financial and tax task, this application needs to parse the financial and tax task in natural language to obtain a financial and tax task triplet including a financial and tax task semantic graph, an initial constraint set, and target weights. Based on the financial and tax task triplet, an initial rule twin graph including time tags and version tags is constructed. Then, a consistency check is performed on the initial rule twin graph, and after passing the check, the initial rule twin graph is set as the target rule twin graph. The rule subgraphs in the target rule twin graph are determined, and the rule subgraphs are transformed into a set of rule constraints including spatiotemporal tags. The satisfiability of the rule constraint set is then determined, and after passing the determination, a preset hierarchical task network is used to process the financial and tax task based on the financial and tax task semantic graph. Subtask decomposition and workflow orchestration yield the target workflow skeleton. Based on the semantic graph of the financial and tax tasks, a corresponding strategy set is determined, and based on the target rule twin graph and time labels, a temporal caliber set is determined. An evaluation space is constructed based on the strategy set and the temporal caliber set. Then, the target workflow skeleton is used to perform counterfactual evaluation of the strategy and temporality within the evaluation space, as well as Pareto optimization, to obtain target financial and tax indicator evaluation vectors corresponding to each element in the evaluation space. The Pareto dominance relationship is then used to compare and filter the target financial and tax indicator evaluation vectors, resulting in a set of decision schemes for executing the financial and tax tasks. The strategy set includes different paths for executing the financial and tax tasks; the temporal caliber set includes different interpretations of the same strategy at different time points.

[0055] Therefore, this application first needs to parse the financial and tax tasks in natural language to obtain financial and tax task triples, including a financial and tax task semantic graph, an initial constraint set, and target weights. Based on these financial and tax task triples, an initial rule twin graph including time tags and version tags is constructed. Then, a consistency check is performed on the initial rule twin graph, and after passing the check, the initial rule twin graph is set as the target rule twin graph. Secondly, the rule subgraphs in the target rule twin graph are determined, and the rule subgraphs are transformed into a set of rule constraints including spatiotemporal tags. The satisfiability of the rule constraint set is then determined, and after passing the determination, a preset hierarchical task network is used, based on financial and tax task weights... The tax task semantic graph decomposes tax and finance tasks into subtasks and orchestrates workflows to obtain a target workflow skeleton. Then, based on the tax and finance task semantic graph, a corresponding set of strategies is determined, and based on the target rule twin graph and time labels, a set of temporal definitions is determined. An evaluation space is constructed based on the strategy set and the temporal definition set. Finally, the target workflow skeleton is used to perform counterfactual evaluation of strategies and temporal definitions in the evaluation space, as well as Pareto optimization, to obtain target tax and finance indicator evaluation vectors corresponding to each element in the evaluation space. Pareto dominance relationships are then used to compare and filter these target tax and finance indicator evaluation vectors, resulting in a set of decision schemes for executing the tax and finance tasks. This improves the efficiency of generating execution decision schemes for tax and finance tasks, thereby enhancing the user experience. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0057] Figure 1 This application discloses a flowchart for a method to generate execution decision schemes for fiscal and tax tasks.

[0058] Figure 2 This application discloses a flowchart of a specific method for generating decision-making schemes for fiscal and tax tasks.

[0059] Figure 3 This application discloses a flowchart for decomposing subtasks and orchestrating workflows.

[0060] Figure 4 This is a schematic diagram of the structure of a device for generating execution decision schemes for fiscal and tax tasks, as disclosed in this application.

[0061] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0062] 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.

[0063] Currently, in recent years, large language models have made breakthroughs in understanding and generating natural language. Relying on large-scale corpora and autoregressive modeling, they possess context learning and instruction following capabilities, and can be integrated with external tools through function calls to achieve retrieval, computation, and data processing. However, relying solely on probabilistic generation cannot guarantee the determinism and consistency of multi-step tasks, and is susceptible to cue perturbations, knowledge timeliness, and context drift. When tasks involve multi-stage dependencies, cross-temporal scopes, and multi-source data coupling, the model output lacks verifiable evidence and stable error fallback paths, making it difficult to meet the requirements of engineering and auditability. To address this, this application provides a method for generating execution decision schemes for financial and tax tasks, which can improve the efficiency of generating execution decision schemes for financial and tax tasks.

[0064] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for generating execution decision schemes for financial and tax tasks, including:

[0065] Step S11: Parse the financial and tax task in natural language to obtain a financial and tax task triple including a financial and tax task semantic graph, an initial constraint set, and a target weight. Based on the financial and tax task triple, construct an initial rule twin graph including time tags and version tags. Then, perform a consistency check on the initial rule twin graph and set it as the target rule twin graph after the check passes.

[0066] In this embodiment, during the generation of execution decision schemes for financial and tax tasks, the input tasks are first parsed into "task semantic graphs" and "initial constraints." Then, a "rule twin graph" with time and version labels is constructed and a consistency check is performed. Subsequently, a workflow is generated using a hierarchical task network and orchestrated using deterministic tool selection criteria. Then, counterfactual evaluation is performed on the "strategy set × temporal / caliber set" to form a multi-indicator candidate solution set and a Pareto front. Subsequently, the inference chain is translated into first-order logic and linear constraints, and satisfiability is checked in conjunction with rule constraints to generate a proof summary. Finally, a hash evidence chain is constructed for key inputs and intermediate evidence, and a zero-knowledge verification interface is provided. At the same time, steady-state evolution is achieved using an online update law based on runtime feedback.

[0067] It is worth mentioning that, Represents the task semantic graph. Indicates initial constraints. Indicates the target weight; Representing a rule-based twin graph, Indicates rule constraints. Indicates inference chain constraints; Indicates an executable workflow. and These represent the strategy set and the temporal / caliber set, respectively.

[0068] In this embodiment, the present application embodiment first selects to deterministically transcribe the natural language task into... This involves clearly defining entities, timeframes, geographic tags, and optimization objectives. Specifically, the embodiments of this application first annotate and extract elements from the text: performing deterministic dependency parsing and transformation on a natural language task to extract the "event-participant-time interval-region" quadruple, and unifying the time to "..." "Format. In one specific implementation, let the original text be..." The set of elements is obtained through analysis. ,in, Predicates indicating events Indicates the participating entity (subject, object, tool, etc.). Indicates the time of the closed interval. This represents a geographic label. The goal of this step is to constrain free text into structured facts, facilitating subsequent graph modeling and constraint generation.

[0069] Secondly, the task triples are constructed, that is, the task is transcribed into... .in, For the task semantic graph, Includes entities and subtasks. Indicates sequence and data dependency; The initial set of constraints (timing, mutual exclusion, threshold, consistency); Let the target weights be the values, and the expression for optimizing the target be as follows:

[0070] ;

[0071] in, For the joint decision of "step selection – data routing – operator parameters"; Let be the set of legal execution sequences allowed by the graph.

[0072] Specifically, the financial and tax tasks in natural language are parsed to obtain a financial and tax task triplet, which includes a financial and tax task semantic graph, an initial constraint set, and target weights. This can include: parsing the financial and tax tasks using dependency syntax to obtain the parsing results, and extracting event predicates, participants, time intervals, and location labels from the parsing results to obtain a quadruple element set including event predicates, participants, time intervals, and location labels; constructing a financial and tax task semantic graph based on the quadruple element set, where each node in the financial and tax task semantic graph represents the entity corresponding to the financial and tax task and each subtask; and each edge in the financial and tax task semantic graph represents the logical relationship between the nodes. The execution order and data transmission dependencies are considered. Based on the quadruple element set, temporal constraints, mutual exclusion constraints, numerical threshold constraints, and consistency constraints are applied to the financial and tax tasks to obtain an initial constraint set. Temporal constraints are used to constrain the execution sequence of the financial and tax tasks. Mutual exclusion constraints are used to constrain the mutual exclusion between different operations. Numerical threshold constraints are used to set numerical thresholds for upper or lower limits of key values. Consistency constraints are used to ensure consistency between data and logic. Based on the optimization intent corresponding to the financial and tax tasks, the target weights corresponding to the financial and tax tasks are determined, and financial and tax task triples are constructed based on the financial and tax task semantic graph, the initial constraint set, and the target weights.

[0073] In this embodiment, after obtaining the financial and tax task triplet, this application embodiment needs to perform multi-temporal rule twin construction and initial consistency check: establish a rule twin with temporal and version annotations. Extract relevant subgraphs and apply constraints. Perform theoretical satisfiability framework checks and locate the minimum unsatisfiability kernel.

[0074] Subsequently, this application embodiment requires the construction of multi-temporal rule twins and initial consistency checks: establishing a twin with temporal and version annotations. Extract relevant subgraphs and apply constraints. The theoretical satisfiability framework is checked, and the minimum unsatisfiability kernel is located. It's worth noting that the core objective mentioned above is to address the problem that static rule bases cannot characterize time and version evolution. By structuring the rule graph and performing satisfiability checks, it ensures that subsequent inference is based on conflict-free rule definitions.

[0075] Furthermore, rule graph modeling is performed: a rule twin graph with timestamps and version tags is constructed. It is worth mentioning that each of the above rules... The association contains metadata. ,in, Effective period For region, Priority; edge Divided into four categories: revision, dependency, conflict, and equivalence mapping, the above structure can explicitly express "when / where / which version of the rule applies, and the relationship between them".

[0076] Specifically, an initial rule twin graph, including time tags and version tags, is constructed based on the tax and finance task triples. Then, a consistency check is performed on the initial rule twin graph, and after passing the check, it is set as the target rule twin graph. This process may include: determining the effective time interval, applicable region, and priority corresponding to the tax and finance task based on the tax and finance task triples; determining the corresponding metadata based on the effective time interval, applicable region, and priority; constructing corresponding rule nodes based on each metadata; determining the revision relationships of rules over time, the dependency relationships of rule references, the conflict relationships of contradictory rule content, and the equivalence mapping relationships of semantically equivalent rules among the rule nodes; determining the edges between each rule node based on the revision relationships, dependency relationships, conflict relationships, and equivalence mapping relationships; generating an initial rule twin graph based on each rule node and its corresponding edges; and then performing a consistency check on the initial rule twin graph using preset consistency check rules. After passing the check, the initial rule twin graph is set as the target rule twin graph.

[0077] Step S12: Determine the rule subgraph in the target rule twin graph, and transform the rule subgraph into a set of rule constraints including spatiotemporal labels, so as to determine the satisfiability of the set of rule constraints. After the determination is passed, use a preset hierarchical task network and based on the financial and tax task semantic graph to perform subtask decomposition and workflow orchestration on the financial and tax task to obtain the target workflow skeleton.

[0078] In this embodiment, the present application requires subgraph extraction and constraint generation of the target rule twin graph: First, based on time and location labels from Extracting relevant subgraphs It also translates task-related entries into a set of first-order logic + linear arithmetic constraints. And the expression is as follows:

[0079] ;

[0080] in, For numerical variables, , The parameters are fixed after the rule text is parsed to ensure that the rule participates in subsequent reasoning in a "computable and verifiable" manner.

[0081] Then, a preliminary satisfiability check and conflict localization are performed: for The satisfactionability is determined using a satisfiability solver (linear integer arithmetic / linear real number arithmetic theory):

[0082] ;

[0083] If satisfiability If the constraint is unsatisfiable, a reduction-verification iteration is used to extract the minimum unsatisfiable kernel. Subsequently, It is mapped back to specific rule entries and time intervals, and is used to prioritize backtracking the "actually contradictory statements" in subsequent steps.

[0084] Specifically, the process involves identifying rule subgraphs in the target rule twin graph and transforming them into a set of rule constraints including spatiotemporal labels. This is followed by a satisfiability assessment of the rule constraint set. This process may include: extracting rule subgraphs that match in time and location from the target rule twin graph based on the tax task triplet; uniformly transforming all applicable rule clauses in each rule subgraph into a set of rule constraints including spatiotemporal labels; the rule clauses are in the form of first-order logical predicates and linear arithmetic inequalities; using a satisfiability module theory solver to assess the logical consistency and satisfiability of the rule constraint set, obtaining the assessment result; if the assessment result indicates that the rule constraint set does not meet logical consistency and satisfiability requirements, then locating and extracting the rule subgraphs causing logical conflicts from the rule constraint set, and then processing these rule subgraphs.

[0085] In this embodiment, the subtasks are decomposed and the workflow is orchestrated, and the corresponding flowchart is shown below. Figure 3 As shown: A hierarchical task network is used to... Decompose and generate workflow The algorithm uses the confidence bounds criterion to make a deterministic selection from the candidate toolset, while performing compatibility checks and necessary field mappings. The core objective of these steps is to address the vulnerability of the workflow caused by a lack of formal constraints in tool orchestration. Through hierarchical decomposition and deterministic tool selection, the workflow is ensured to be executable and reproducible. The process of decomposing the hierarchical task network is as follows: The Hierarchical Task Network (HTN) is applied to decompose abstract tasks into executable units from top to bottom, resulting in a Work Breakdown Structure (WBS), and a workflow skeleton is formed accordingly. This is to ensure that "complex tasks are broken down into implementable nodes and dependencies".

[0086] Subsequently, the deterministic criteria for tool selection were determined: for each node in the candidate toolset The upper confidence bound algorithm (UCB1) is used for scoring:

[0087] ;

[0088] in, The historical average utility (calculated as "correctness - delay penalty"). For the number of times it is used, This represents the total number of rounds. Then, the round with the highest score is selected, ensuring a balance between stability and exploration within limited data, and the selection logic is fixed and reproducible.

[0089] In this embodiment, the data pattern compatibility and mapping of this application embodiment need to be performed: each edge needs to be verified. When incompatibility occurs, three atomic operators are inserted according to the static mapping table: field renaming, default filling, and unit conversion, until compatibility is achieved, thereby realizing "node execution" as "data interoperability" and eliminating common link vulnerabilities.

[0090] Specifically, after the initial assessment, a pre-defined hierarchical task network is used to decompose and orchestrate the financial and tax tasks based on the financial and tax task semantic graph, resulting in the target workflow skeleton. This process includes: decomposing the financial and tax tasks from top to bottom using the pre-defined hierarchical task network and the financial and tax task semantic graph, obtaining several task nodes and directed edges representing the execution dependencies between task nodes; generating the target workflow skeleton based on each task node and its corresponding directed edges; determining the accuracy of historical execution tasks corresponding to each candidate tool and the penalties incurred due to execution delays; determining the historical average utility based on the accuracy and penalties; and determining the historical call count corresponding to each candidate tool; and using the upper confidence bound algorithm and based on the historical average utility and historical call count to determine the selection of each candidate tool in the current task node. Select a score and set the candidate tool corresponding to the highest score among the selected scores as the target candidate tool. Then, bind the task node to the target candidate tool to obtain the initial workflow skeleton. Determine each upstream and downstream node pair with direct data dependency in the initial workflow skeleton, and determine whether the output data pattern of the upstream node in the upstream and downstream node pair is semantically and structurally compatible with the input data pattern of the corresponding downstream node to obtain the data pattern compatibility verification result. If the data pattern compatibility verification result indicates incompatibility, insert field conversion operators between the upstream and downstream node pairs using preset field mapping rules to obtain the target workflow skeleton. Field conversion operators include operators for renaming data fields, operators for filling missing data fields with default values, and operators for standardizing data units.

[0091] Step S13: Determine the corresponding strategy set based on the semantic graph of the financial and tax task, and determine the temporal caliber set based on the twin graph of the target rule and the time label, so as to form an evaluation space based on the strategy set and the temporal caliber set; the strategy set includes different paths to execute the financial and tax task; the temporal caliber set includes different interpretations of the same strategy at different time points.

[0092] In this embodiment, strategy-temporal counterfactual evaluation and Pareto optimization are performed: for each Calculate multi-index vectors And an approximate Pareto front is obtained by using the upper confidence bound algorithm (UCT) applied to tree search. The core objective of this step is to address the difficulty of balancing benefits, costs, and risks in single-indicator evaluations by providing interpretable multiple alternatives through counterfactual evaluation and Pareto optimization.

[0093] The evaluation space and indicator definitions are then defined, and a strategy set is constructed. With tense / caliber set In the workflow Next, for each Calculate the multi-index vector:

[0094] ;

[0095] Among them, revenue / cost / latency / stability are calculated according to the system's configured metric function, and risk is calculated using... Quantify "tail loss".

[0096] Specifically, based on the semantic graph of the tax and finance task, a corresponding set of strategies is determined, and based on the target rule twin graph and time labels, a set of temporal definitions is determined. The evaluation space is then constructed based on the strategy set and the temporal definition set. This process may include: parsing the semantic graph of the tax and finance task to obtain several corresponding variable decision points; generating a strategy set based on each variable decision point; selecting rule versions corresponding to historical rules related to and different from the task from the target rule twin graph based on time labels; generating a set of temporal definitions based on each rule version; and then constructing the evaluation space based on the strategy set and the temporal definition set. Correspondingly, the target workflow skeleton is used to perform counterfactual evaluation of the evaluation space on the strategies and temporal aspects. The evaluation process involves estimation and Pareto optimization to obtain target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space. Pareto dominance is then used to compare and filter these target fiscal and tax indicator evaluation vectors, resulting in a set of decision-making schemes for executing fiscal and tax tasks. This includes: determining the fiscal and tax indicator evaluation vectors for each element in the evaluation space based on the target workflow skeleton and preset indicators, obtaining corresponding target fiscal and tax indicator evaluation vectors, and then using Pareto dominance to filter these target fiscal and tax indicator evaluation vectors to obtain a set of decision-making schemes for executing fiscal and tax tasks. The preset indicators include revenue indicators, cost indicators, risk indicators, time delay indicators, and stability indicators.

[0097] Step S14: Use the target workflow skeleton to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space to obtain the target financial and tax indicator evaluation vectors corresponding to each element in the evaluation space. Then, use the Pareto dominance relationship to compare and filter the target financial and tax indicator evaluation vectors to obtain a set of decision schemes for executing the financial and tax task.

[0098] In this embodiment, guided search and frontier determination are required: the Upper Confidence Bound (UCT) algorithm (constant) applied to tree search is used. )exist The space is used to search for representative candidates, and the Pareto front is given using dominance relationships:

[0099] ;

[0100] It is worth mentioning that the output of the above steps is not just a "unique optimal" but a "comparable and interpretable" set of frontiers, along with key influencing factors (rules, data sources, or nodes that contribute the most to the frontier position).

[0101] Then, formal verification and proof generation are performed: and The merging process uses SMT-LIA / LRA in DPLL(T) to solve the decision. Generate evidence models and proof summaries The core objective of this step is to address the problem of unprovable conclusions due to a lack of satisfiability verification. By transforming logical constraints and solving using SMT (Simplified Methods and Techniques), a formal proof for the conclusions is provided.

[0102] Subsequently, inference chain constraints are performed: the obtained execution and intermediate conclusions are translated into inference chain constraints. .in, The same predicate + linear constraint expression is used to ensure that the "learning conclusion" enters the "logically verifiable" channel.

[0103] In this embodiment, a decision and evidence model is used to establish merging rules and inference chain constraints, and the expressions are as follows:

[0104] ;

[0105] in, Assign values ​​to specific models; Hash binding for input / evidence; This document records the step-by-step derivation from the Boolean skeleton to numerical theory. If... Extract the minimum kernel It also maps back to rule entries and task steps, providing "precise and reproducible" backtracking clues.

[0106] Specifically, after obtaining the set of decision-making schemes for performing fiscal and tax tasks, the process may further include: using the upper confidence bound algorithm to perform a guided search of the decision-making scheme set in the evaluation space to obtain candidate schemes and determine the Pareto front corresponding to each candidate scheme; the Pareto front includes several non-dominated solutions and their corresponding key influencing factors; merging the rule constraint set, Pareto front, and key influencing factors to obtain a merged constraint set, and using a modular theory solver for satisfiability to determine the satisfiability of the merged constraint set, obtaining a satisfiability determination result; when the satisfiability determination result indicates satisfiability, generating a corresponding evidence model to generate a structured proof summary; the proof summary includes variable assignment, hash binding of input data, and a step-by-step derivation record from the Boolean skeleton to numerical theory; when the satisfiability determination result indicates unsatisfiability, extracting the minimum unsatisfiable kernel from the merged constraint set, and mapping each logical clause in the minimum unsatisfiable kernel in reverse to the specific rule entries in the target rule twin graph and the specific task nodes in the target workflow skeleton, obtaining a source tracing report for locating and explaining the root causes of conflicts.

[0107] In this embodiment, auditable output and evidence chain retention are required: an evidence chain fingerprint is formed using a SHA-256 (Secure Hash Algorithm 256-bit) chain and a Merkle root, and an external verification interface is provided using zk-SNARK (Zero-Knowledge Succinct Non-Interactive Argument of Knowledge). The core objective of this step is to solve the problem of "insufficient process traceability making external verification difficult" by balancing privacy protection and auditability through hash chains and zero-knowledge proofs.

[0108] In one specific implementation, the key event sequence Encode to obtain And construct a linear hash chain with a Merkle root:

[0109] ;

[0110] ;

[0111] Among them, chained hashing ensures that the order cannot be tampered with, while Merkle root facilitates the coexistence of "local evidence collection and overall verification".

[0112] Subsequently, an external zero-knowledge verification is performed: the public inputs are Root, caliber label, and time interval, and the statement to be proved is "there exists an assignment that satisfies..." Furthermore, the evidence binding is consistent with the Root. The verifier only learns "true / false" and does not access plaintext data, thus striking a balance between privacy and verifiability. Then, operational feedback and parameter updates are performed: the policy parameters are adjusted using an online mirror. Post-approval updates set three safeguard thresholds for correctness, latency, and stability to ensure that evolution does not regress. The core objective is to address the problem of a lack of feedback loops hindering system evolution, and to ensure monotonic performance improvements through online learning and safeguard thresholds.

[0113] Furthermore, this application embodiment requires loss definition and online updating: let the strategy parameter be... The weighted loss is calculated based on "verification failure penalty + latency exceeding limit + instability". Perform online mirror dropdown:

[0114] ;

[0115] in, Step size, The Bregman distance is induced by a strongly convex function. The above update guarantees a move "towards higher accuracy, lower latency, and stronger stability".

[0116] Furthermore, implement a protection threshold and rollback mechanism: verify the three protection constraints on the regression set after each update.

[0117] ;

[0118] ;

[0119] ;

[0120] It is worth mentioning that if any of the above constraints are not met, the process will roll back to... It also freezes the update window to avoid the phenomenon of "getting worse with each modification" and to ensure the monotonicity and controllability of the system's long-term evolution.

[0121] Specifically, after comparing and filtering the evaluation vectors of target fiscal and tax indicators using Pareto dominance relations to obtain a set of decision schemes for executing fiscal and tax tasks, the process may further include: executing fiscal and tax tasks based on the set of decision schemes, obtaining the task execution results and corresponding key event sequences, and encoding the key event sequences using a preset encryption algorithm to obtain corresponding ciphertext encoded sequences; concatenating the hash value of the previous event with the ciphertext encoding of the current event in chronological order for each element in the ciphertext encoded sequence to obtain a secure hash value, and generating a linear hash chain based on each secure hash value; constructing a Merkle tree based on each ciphertext encoded sequence, determining the root value corresponding to the Merkle tree, and then setting the root value as a fingerprint identifier corresponding to the fiscal and tax task; using a preset proof circuit and based on the root value and the fiscal and tax task triplet to prove the key event sequence, obtaining the proof result, and sending the proof result to a preset external verification interface so that the external verification party can determine the fiscal and tax task based on the proof result. The verification results for the task are determined. The number of verification failures corresponding to the financial and tax task is determined, and then, using a preset comprehensive loss function and based on the number of verification failures, the deviation value of the actual task execution delay exceeding the preset expected delay and the decrease value of the system stability index relative to the preset baseline are determined to obtain the loss result. A preset mirror descent algorithm is used to iteratively update the system strategy parameters corresponding to the decision scheme set based on the loss result. Then, the updated system strategy parameters are verified using a preset historical regression dataset and each preset protection threshold to obtain the verification result. The preset protection thresholds include a decrease threshold for task verification success rate, an increase threshold for task execution delay, and a decrease threshold for the system stability index. If the verification result indicates that the updated system strategy parameters meet each preset protection threshold, the updated system strategy parameters are set as the target system strategy parameters. If the verification result indicates that the updated system strategy parameters do not meet each preset protection threshold, an update freeze mechanism is triggered.

[0122] In one specific implementation, this application embodiment requires task parsing and context modeling: First, the text is labeled and elements are extracted: Based on the proposed natural language requirement, "Complete the 2024 corporate income tax settlement for Company A, comply with the latest 2024 policy, submit before May 31, 2025, maximize tax benefits, and ensure manual review time is ≤2 hours," the system performs deterministic dependency parsing to extract quadruples. ,in, For the "2024 Corporate Income Tax Settlement and Payment", Includes Company A (taxpayer), Beijing Tax Bureau (competent authority), and the 2024 profit statement (data object). These are the tax period [2024-01-01, 2024-12-31] and the filing deadline [2025-01-01, 2025-05-31]. This refers to Announcement No. 3 of 2024 issued by the State Taxation Administration and Announcement No. 7 of 2024 issued by the Ministry of Finance and the State Taxation Administration.

[0123] Then, task triples are constructed: further constructing task triples. , middle It includes nodes such as "data collection, qualification assessment of micro and small enterprises, and calculation of R&D expense deduction". This indicates a dependency relationship such as "data collection → qualification assessment". Including the constraints of "manual review time ≤ 2 hours, taxable income ≤ 3 million yuan (standard for micro and small enterprises)," the expression for the objective function is:

[0124] ;

[0125] Then, multi-temporal rule twin construction and initial consistency check are performed: First, rule graph modeling is carried out: a rule twin graph containing the corporate income tax policy for 2023-2024 is constructed. Each rule node Related metadata For example, the node "2024 small and micro enterprise income tax rate 2.5%" , =“Nationwide” =1 (high priority), the node "2023 Small and Micro Enterprise Income Tax Rate 5%" =[2023-01-01, 2023-12-31]、 =2 (low priority), edge Includes the revision border for "2023 tax rate → 2024 tax rate".

[0126] Subsequently, subgraph extraction and constraint generation are performed: subgraphs are extracted based on Company A's "2024" time label. Translated into a set of rule constraints ,in, The tax rate is calculated as follows: "If the number of employees is less than or equal to 300, the total assets are less than or equal to 50 million yuan, and the taxable income is less than or equal to 3 million yuan, then the tax rate is 2.5%". The additional deduction amount for R&D expenses y is 2y. The expression is "z=280-2y";

[0127] Furthermore, a preliminary satisfiability check and conflict localization are performed: the SMT-LRA solver is used to determine SAT(Φ). Substituting the data from Company A (85 employees, total assets of 28 million yuan, y = 800,000 yuan), we get z = 1.2 million yuan ≤ 3 million yuan, satisfying all constraints. No rule conflicts.

[0128] Then, subtask decomposition and workflow orchestration are performed: First, a hierarchical task network decomposition is carried out: In The application uses a hierarchical task network (HTN) to decompose the abstract task of "corporate income tax final settlement" from top to bottom into a workflow framework of "data collection (using a financial and tax data import tool to read the Excel profit and loss statement) → qualification verification (using a small and micro enterprise qualification verification tool) → R&D expense compliance verification (using a special verification tool) → additional deduction calculation (using an additional deduction calculator) → tax calculation (using a corporate income tax calculator) → tax return generation (using an electronic tax return tool)". Then, determine the deterministic criteria for tool selection: for the selection of "micro and small enterprise qualification verification tools", from the candidate set The algorithm used in this calculation is UCB1 (Upper Confidence Bound1), where the total number of rounds N=150, and tool X... =0.95 (historical mean utility) =100 (number of uses), score 1.21, Tool Y =0.92、 =50, score 1.29, finally choose tool Y.

[0129] Subsequently, data model compatibility and mapping were performed: the compatibility between the output of the "R&D Expense Compliance Verification Tool" ("R&D expenses - personnel labor of RMB 450,000, R&D expenses - direct investment of RMB 350,000") and the input of the "Additional Deduction Calculator" ("Total deductible R&D expenses") was verified, the "Field Summary Operator" was automatically inserted, and the two types of detailed fields were summed to "Total deductible R&D expenses = RMB 800,000" to ensure seamless data model integration.

[0130] In this embodiment, strategy-temporal counterfactual evaluation and Pareto optimization are performed: First, spatial evaluation and indicator definition are carried out: an evaluation space is constructed. Among them, the strategy set include (Full R&D tax deduction applies) (Enjoy a 50% R&D tax deduction) (Not eligible for R&D tax deduction), and the expression for the tense / caliber set is as follows:

[0131] ;

[0132] Then, guided search and frontier determination are performed: for each Calculate the multi-index vector:

[0133] ;

[0134] In one specific implementation, The indicators of discount amount, cost, risk, latency, and stability are defined as follows:

[0135] =40,000 yuan;

[0136] =1.5 hours (manual review time);

[0137] =05,000 yuan ( );

[0138] =0.5 hours (total time);

[0139] =98%;

[0140] The definitions of the amount, cost, and risk indicators are as follows:

[0141] =10,000 yuan;

[0142] =1.2 hours;

[0143] =02,000 yuan;

[0144] In this embodiment, the UCT algorithm (constant) is used. )exist Searching in space based on dominance relationships and Screening Pareto frontiers ,in, Because the 100% additional tax rate policy in 2024 offers the highest preferential treatment. Because reducing the number of ledger verification items lowers the risk, it provides Company A with an interpretable set of decision-making options.

[0145] Subsequently, formal verification and proof generation are performed: First, the inference chain is constrained: The execution process is translated into inference chain constraints. , The tool Y determines that Company A meets the criteria for a micro-enterprise: 85 ≤ 300 employees, total assets 2800 ≤ 5000 million yuan, z = 120 ≤ 3000 million yuan. The R&D expense compliance verification tool determined that y=800,000 yuan is fully deductible → additional deduction amount = 1,600,000 yuan. The formula is: "z = 280 - 160 = 1.2 million yuan → tax = 1.2 million × 2.5% = 30,000 yuan".

[0146] Furthermore, SMT decision and evidence modeling is performed: merging rule constraints and inference chain constraints:

[0147] ;

[0148] Subsequently, SMT-LRA was used to solve and determine the solution. Assignment exists All constraints are satisfied:

[0149] ;

[0150] Then, proof summary generation is performed:

[0151] ;

[0152] Among them, bindings is "Profit Statement Hash = SHA256(2024 Profit Statement.xlsx) = XXX, R&D Ledger Hash = YYY", and justifications records "by..." and The applicable tax rate of 2.5% is derived from... and The complete logical chain of "deriving the additional deduction amount and finally the taxable income of 30,000 yuan" is explained.

[0153] In this embodiment, auditable output and evidence chain retention are performed: First, hash chaining and root fingerprint generation are performed: the key event sequence during task execution ( (Basic information about Company A) (Results of the qualification assessment for micro and small enterprises) (Research and Development Expense Verification Results) (Tax calculation result) (Electronic application form), and then use the AES-256 algorithm to encode the key event sequence to obtain Then, a linear hash chain is constructed, where the initial hash... The expression for the subsequent hash is as follows:

[0154] ;

[0155] In one specific implementation, the corresponding expression is as follows:

[0156] ;

[0157] Subsequently, based on Construct the Merkle tree and calculate the Merkle Root, as shown in the following formula:

[0158] ;

[0159] Then, external zero-knowledge verification is performed: a zk-SNARK external verification interface is provided, requiring only the input of the publicly available information "Root" and the tag "2024 Small and Micro Enterprises + R&D Tax Credit Policy" to verify that "a valid event sequence exists that satisfies..." Furthermore, the evidence binding is consistent with "Root", and the system only returns "established / invalid", without disclosing any plaintext data, thus balancing privacy protection and auditability.

[0160] Furthermore, this embodiment requires operational feedback and parameter updates: First, the loss domain is updated online: defining policy parameters. The loss function for (utility weights of tool selection) is:

[0161] ;

[0162] Among them, in the process During the execution, the number of verification failures was 0, the latency was 0.5 hours, and the stability was 98%. =0.6, and then the online mirror descent algorithm is used to update the parameters, as shown in the following formula:

[0163] ;

[0164] Among them, step size (k represents the update round) Using the Bregman distance, the utility weight of tool Y increases from 0.92 to 0.94 after the update.

[0165] Finally, a guardian threshold rollback mechanism is implemented: after the update, the three guardian constraints are validated on a historical regression dataset. =0 (Verification success rate remains 100%) =-0.1 hours (increase ≤5%) =0 (decrease ≤ 2%), all constraints are satisfied, retain. This is to ensure that the system evolves without regression.

[0166] As can be seen from the above, the embodiments of this application first need to parse the financial and tax tasks in natural language to obtain financial and tax task triples, including a financial and tax task semantic graph, an initial constraint set, and target weights. Based on these financial and tax task triples, an initial rule twin graph including time tags and version tags is constructed. Then, a consistency check is performed on the initial rule twin graph, and after passing the check, the initial rule twin graph is set as the target rule twin graph. Next, the rule subgraphs in the target rule twin graph are determined, and the rule subgraphs are transformed into a set of rule constraints including spatiotemporal tags. The satisfiability of the rule constraint set is then determined, and after passing the determination, a preset hierarchical task network is used and... The financial and tax task semantic graph is used to decompose the task into subtasks and orchestrate the workflow, resulting in a target workflow skeleton. Then, based on the semantic graph, a corresponding set of strategies is determined, and a set of temporal definitions is determined based on the target rule twin graph and time labels. An evaluation space is constructed based on the strategy set and the temporal definition set. Finally, counterfactual evaluation and Pareto optimization are performed on the evaluation space using the target workflow skeleton to obtain target financial and tax indicator evaluation vectors corresponding to each element in the evaluation space. Pareto dominance is then used to compare and filter these evaluation vectors, resulting in a set of decision schemes for executing the financial and tax task. This improves the efficiency of generating execution decision schemes for the financial and tax task, thereby enhancing the user experience.

[0167] Accordingly, see Figure 4 As shown, this application also provides an apparatus for generating execution decision schemes for financial and tax tasks, including:

[0168] The rule twin graph generation module 11 is used to parse the financial and tax tasks in natural language, obtain financial and tax task triples including a financial and tax task semantic graph, an initial constraint set and a target weight, and construct an initial rule twin graph including time tags and version tags based on the financial and tax task triples. Then, a consistency check is performed on the initial rule twin graph, and after the check passes, the initial rule twin graph is set as the target rule twin graph.

[0169] Workflow skeleton generation module 12 is used to determine the rule subgraph in the target rule twin graph and convert the rule subgraph into a set of rule constraints including spatiotemporal labels, to determine the satisfiability of the set of rule constraints, and after the determination is passed, to use a preset hierarchical task network and based on the financial and tax task semantic graph to perform subtask decomposition and workflow arrangement of the financial and tax task to obtain the target workflow skeleton.

[0170] The space construction module 13 is used to determine the corresponding strategy set based on the semantic graph of the financial and tax task, and to determine the temporal caliber set based on the twin graph of the target rule and the time label, so as to construct an evaluation space based on the strategy set and the temporal caliber set; the strategy set includes different paths to execute the financial and tax task; the temporal caliber set includes different interpretations of the same strategy at different time points.

[0171] The decision scheme set determination module 14 is used to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space using the target workflow skeleton, to obtain the target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space, and to compare and filter the target fiscal and tax indicator evaluation vectors using the Pareto dominance relationship to obtain a set of decision schemes for executing the fiscal and tax task.

[0172] In some specific embodiments, the rule twin graph generation module 11 may specifically include:

[0173] The element set generation unit is used to parse the financial and tax task using dependency syntax, obtain the parsing result, and extract the event predicate, participants, time interval and regional label from the parsing result to obtain a four-tuple element set including the event predicate, the participants, the time interval and the regional label.

[0174] The task semantic graph construction unit is used to construct a financial and tax task semantic graph based on the set of quadruple elements. Each node in the financial and tax task semantic graph is used to represent the entity and subtasks corresponding to the financial and tax task. Each edge in the financial and tax task semantic graph is used to represent the logical execution order and data transmission dependency between the nodes.

[0175] The constraint set generation unit is used to apply time-series constraints, mutual exclusion constraints, numerical threshold constraints, and consistency constraints to the financial and tax tasks based on the quadruple element set, to obtain an initial constraint set; the time-series constraints are used to constrain the execution sequence of the financial and tax tasks; the mutual exclusion constraints are used to constrain the mutual exclusion between different operations; the numerical threshold constraints are used to set numerical thresholds for upper or lower limits of key values; and the consistency constraints are used to ensure the consistency between data and logic.

[0176] The target weight determination unit is used to determine the target weight corresponding to the financial and tax task based on the optimization intention corresponding to the financial and tax task, so as to construct a financial and tax task triple based on the financial and tax task semantic graph, the initial constraint set and the target weight.

[0177] In some specific embodiments, the rule twin graph generation module 11 may specifically include:

[0178] The rule node construction unit is used to determine the effective time interval, applicable region and priority corresponding to the financial and tax task based on the financial and tax task triple, to determine the corresponding metadata based on the effective time interval, the applicable region and the priority, and to construct the corresponding rule nodes based on each of the metadata.

[0179] An equivalence mapping relationship determination unit is used to determine the revision relationship of rules over time, the dependency relationship of reference relationships between rules, the conflict relationship of contradictory rule content, and the equivalence mapping relationship of semantic equivalence between the rule nodes. Based on the revision relationship, the dependency relationship, the conflict relationship, and the equivalence mapping relationship, it determines the edges between each rule node. Based on each rule node and its corresponding edge, it generates an initial rule twin graph. Then, it performs a consistency check on the initial rule twin graph using a preset consistency check rule. After the check passes, it sets the initial rule twin graph as the target rule twin graph.

[0180] In some specific embodiments, the workflow skeleton generation module 12 may specifically include:

[0181] The rule constraint set generation unit is used to extract each rule subgraph that matches in time and region from the target rule twin graph based on the fiscal and tax task triple, and to uniformly transform the rule clauses of all applicable rules in each rule subgraph into a rule constraint set including spatiotemporal labels; the form corresponding to the rule clauses is a form composed of first-order logical predicates and linear arithmetic inequalities;

[0182] The judgment result generation unit is used to perform logical consistency and satisfiability judgment on the rule constraint set using a satisfiability modular theory solver to obtain a judgment result. If the judgment result indicates that the rule constraint set does not meet the requirements of logical consistency and satisfiability, the unit locates and extracts the rule subgraph to be processed that causes logical conflict from the rule constraint set, and then processes the rule subgraph to be processed.

[0183] In some specific embodiments, the workflow skeleton generation module 12 may specifically include:

[0184] The workflow skeleton generation subunit is used to decompose the financial and tax tasks from top to bottom using a preset hierarchical task network and based on the financial and tax task semantic graph to obtain several task nodes and directed edges used to characterize the execution dependencies between the task nodes. Then, the target workflow skeleton is generated based on each task node and the corresponding directed edge.

[0185] The historical call count generation unit is used to determine the accuracy of the historical execution tasks corresponding to each candidate tool and the penalty caused by execution delay, and to determine the historical utility average based on the accuracy and the penalty, and to determine the historical call count corresponding to each candidate tool.

[0186] The selection score generation unit is used to determine the selection score of each candidate tool in the current task node using the upper confidence bound algorithm and based on the historical utility mean and the historical call count, and to set the candidate tool with the highest score among the selection scores as the target candidate tool. Then, the task node is bound to the target candidate tool to obtain the initial workflow skeleton.

[0187] The compatibility verification unit is used to determine each upstream and downstream node pair with direct data dependency in the initial workflow skeleton, and to determine whether the output data pattern of the upstream node in the upstream and downstream node pair is semantically and structurally compatible with the input data pattern of the corresponding downstream node, so as to obtain the data pattern compatibility verification result.

[0188] The conversion operator insertion unit is used to insert a field conversion operator between the upstream and downstream node pairs according to a preset field mapping rule if the data pattern compatibility verification result indicates incompatibility, thereby obtaining the target workflow skeleton. The field conversion operator includes an operator for renaming data fields, an operator for filling missing data fields with default values, and an operator for standardizing and converting data units.

[0189] In some specific embodiments, the space construction module 13 may specifically include:

[0190] The strategy set generation unit is used to parse the semantic graph of the financial and tax task to obtain several corresponding variable decision points, generate a strategy set based on each variable decision point, and select the rule version corresponding to the historical rules that are related to and different from the task from the target rule twin graph based on the time label, generate a temporal caliber set based on each rule version, and then construct an evaluation space based on the strategy set and the temporal caliber set.

[0191] The evaluation vector filtering unit is used to determine the financial and tax indicator evaluation vectors for each element in the evaluation space based on the target workflow skeleton and preset indicators, thereby obtaining the corresponding target financial and tax indicator evaluation vectors. The Pareto dominance relationship is used to filter each target financial and tax indicator evaluation vector to obtain a set of decision schemes for executing the financial and tax task. The preset indicators include revenue indicators, cost indicators, risk indicators, time delay indicators, and stability indicators.

[0192] In some specific embodiments, the device for generating execution decision schemes for fiscal and tax tasks may further include:

[0193] The candidate solution generation unit is used to perform a guided search on the set of decision solutions in the evaluation space using the upper confidence bound algorithm to obtain candidate solutions and determine the Pareto front corresponding to each candidate solution; the Pareto front includes several non-dominated solutions and their corresponding key influencing factors.

[0194] The satisfiability determination unit is used to merge the set of rule constraints, the Pareto front and the key influence factors to obtain a merged set of constraints, and use a modular theory solver of satisfiability to determine the satisfiability of the merged set of constraints to obtain a satisfiability determination result.

[0195] The proof summary generation unit is used to generate a corresponding evidence model when the satisfiability determination result is characterized as satisfiable, so as to generate a structured proof summary using the evidence model; the proof summary includes variable assignment, hash binding of input data, and a step-by-step derivation record from Boolean skeleton to numerical theory;

[0196] The source tracing report generation unit is used to extract the minimum unsatisfiable kernel from the merged constraint set when the satisfiability determination result is characterized as unsatisfiable, and to reverse map each logical clause in the minimum unsatisfiable kernel to the specific rule entries in the target rule twin graph and the specific task nodes in the target workflow skeleton to obtain a source tracing report for locating and interpreting the root cause of the conflict.

[0197] In some specific embodiments, the device for generating execution decision schemes for fiscal and tax tasks may further include:

[0198] The ciphertext encoding sequence generation unit is used to execute the financial and tax task based on the set of decision schemes, obtain the task execution result and the corresponding key event sequence, and encode the key event sequence using a preset encryption algorithm to obtain the corresponding ciphertext encoding sequence.

[0199] The hash chain generation unit is used to concatenate the hash value of the previous event with the ciphertext code of the current event in chronological order for each element in the ciphertext encoding sequence to obtain a secure hash value, so as to generate a linear hash chain based on each of the secure hash values;

[0200] The root value generation unit is used to construct a Merkle tree based on each of the ciphertext encoding sequences, determine the root value corresponding to the Merkle tree, and then set the root value as a fingerprint identifier corresponding to the financial and tax task.

[0201] The verification result generation unit is used to use a preset proof circuit and based on the root value and the tax task triplet to prove the key event sequence, obtain the proof result, and send the proof result to a preset external verification interface so that the external verification party can determine the verification result corresponding to the tax task based on the proof result.

[0202] The loss result generation unit is used to determine the number of verification failures corresponding to the financial and tax task, and then use a preset comprehensive loss function and based on the number of verification failures to determine the deviation value of the actual task execution delay corresponding to the financial and tax task exceeding the preset expected delay and the decrease value of the system stability index relative to the preset baseline, so as to obtain the loss result.

[0203] The reduction threshold determination unit is used to iteratively update the system strategy parameters corresponding to the decision scheme set based on the loss result using a preset mirror descent algorithm, and then verify the updated system strategy parameters using a preset historical regression dataset and each preset protection threshold to obtain the verification result; the preset protection thresholds include the reduction threshold of task verification success rate, the increase threshold of task execution latency, and the reduction threshold of system stability index;

[0204] The strategy parameter determination unit is used to set the updated system strategy parameters as the target system strategy parameters if the verification result indicates that the updated system strategy parameters meet each of the preset protection thresholds, and to trigger the update freeze mechanism if the verification result indicates that the updated system strategy parameters do not meet each of the preset protection thresholds.

[0205] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the method for generating execution decision schemes for financial and tax tasks disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0206] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0207] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0208] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the execution decision-making scheme generation method for the financial and tax tasks performed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0209] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for generating execution decision schemes for financial and tax tasks. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0210] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0211] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0212] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0213] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0214] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for generating execution decision-making schemes for fiscal and tax tasks, characterized in that, include: The financial and tax tasks in natural language are parsed to obtain financial and tax task triples, which include a financial and tax task semantic graph, an initial constraint set, and a target weight. An initial rule twin graph, including time tags and version tags, is constructed based on the financial and tax task triples. Then, a consistency check is performed on the initial rule twin graph, and after the check passes, the initial rule twin graph is set as the target rule twin graph. The rule subgraph in the target rule twin graph is determined, and the rule subgraph is transformed into a set of rule constraints including spatiotemporal labels. The satisfiability of the set of rule constraints is determined. After the determination is passed, the financial and tax task is decomposed into subtasks and the workflow is arranged based on the financial and tax task semantic graph using a preset hierarchical task network to obtain the target workflow skeleton. Based on the semantic graph of the tax and finance task, a corresponding set of strategies is determined, and based on the twin graph of the target rules and the time label, a set of temporal definitions is determined, so as to form an evaluation space based on the set of strategies and the set of temporal definitions; the set of strategies includes different paths to execute the tax and finance task; the set of temporal definitions includes different interpretations of the same strategy at different time points; The target workflow skeleton is used to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space to obtain target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space. The Pareto dominance relationship is then used to compare and filter the target fiscal and tax indicator evaluation vectors to obtain a set of decision schemes for executing the fiscal and tax task.

2. The method for generating execution decision schemes for fiscal and tax tasks according to claim 1, characterized in that, The process of parsing the financial and tax task in natural language yields a financial and tax task triplet comprising a semantic graph of the financial and tax task, an initial constraint set, and target weights, including: The financial and tax task is parsed using dependency parsing to obtain the parsing result. Then, the event predicate, participants, time interval, and regional label are extracted from the parsing result to obtain a set of four-tuple elements including the event predicate, the participants, the time interval, and the regional label. A semantic graph of financial and tax tasks is constructed based on the set of quadruple elements. Each node in the semantic graph represents the entity and subtasks corresponding to the financial and tax task. Each edge in the semantic graph represents the logical execution order and data transmission dependency between the nodes. Based on the set of four elements, the financial and tax tasks are subject to time constraints, mutual exclusion constraints, numerical threshold constraints, and consistency constraints to obtain an initial constraint set. The time constraints are used to constrain the execution time of the financial and tax tasks. The mutual exclusion constraints are used to constrain the mutual exclusion between different operations. The numerical threshold constraints are used to set numerical thresholds for upper or lower limits of key values. The consistency constraints are used to ensure the consistency between data and logic. Based on the optimization intent corresponding to the financial and tax task, the target weight corresponding to the financial and tax task is determined, and a financial and tax task triplet is constructed based on the financial and tax task semantic graph, the initial constraint set, and the target weight.

3. The method for generating execution decision schemes for fiscal and tax tasks according to claim 1, characterized in that, The process of constructing an initial rule twin graph based on the tax task triplet, including time tags and version tags, then performing a consistency check on the initial rule twin graph, and setting the initial rule twin graph as the target rule twin graph after the check passes, includes: Based on the three-tuple of the financial and tax tasks, the effective time interval, applicable region and priority corresponding to the financial and tax tasks are determined, and the corresponding metadata is determined based on the effective time interval, the applicable region and the priority, and the corresponding rule nodes are constructed based on each of the metadata. The revision relationships of rules over time, the dependency relationships of rule references, the conflict relationships of contradictory rule content, and the equivalence mapping relationships of semantically equivalent rules among the rule nodes are determined. Based on the revision relationships, dependency relationships, conflict relationships, and equivalence mapping relationships, edges between the rule nodes are determined. An initial rule twin graph is generated based on the rule nodes and their corresponding edges. Then, a consistency check is performed on the initial rule twin graph using a preset consistency check rule. After the check passes, the initial rule twin graph is set as the target rule twin graph.

4. The method for generating execution decision schemes for fiscal and tax tasks according to claim 1, characterized in that, The step of determining the rule subgraph in the target rule twin graph and converting the rule subgraph into a set of rule constraints including spatiotemporal labels, and then performing a satisfiability determination on the set of rule constraints, includes: Based on the aforementioned tax and finance task triplet, rule subgraphs that match in time and region are extracted from the target rule twin graph, and all applicable rule clauses in each rule subgraph are uniformly transformed into a set of rule constraints including spatiotemporal labels; the form corresponding to the rule clauses is composed of first-order logical predicates and linear arithmetic inequalities. The set of rules and constraints is subjected to logical consistency and satisfiability determination using a satisfiability modular theory solver. If the determination result indicates that the set of rules and constraints does not meet the requirements of logical consistency and satisfiability, the rule subgraph that causes the logical conflict is located and extracted from the set of rules and constraints, and then the rule subgraph is processed.

5. The method for generating execution decision schemes for fiscal and tax tasks according to claim 1, characterized in that, After the judgment is passed, the financial and tax tasks are decomposed into subtasks and the workflow is arranged based on the preset hierarchical task network and the financial and tax task semantic graph to obtain the target workflow skeleton, including: The financial and tax tasks are decomposed from top to bottom using a pre-defined hierarchical task network and based on the financial and tax task semantic graph to obtain several task nodes and directed edges used to characterize the execution dependencies between the task nodes. Then, a target workflow skeleton is generated based on each task node and the corresponding directed edge. Determine the accuracy of historical execution tasks corresponding to each candidate tool and the penalty for execution delay, and determine the historical average utility based on the accuracy and the penalty, and determine the historical call count corresponding to each candidate tool; The selection score of each candidate tool in the current task node is determined by using the upper confidence bound algorithm and based on the historical utility mean and the historical call count. The candidate tool with the highest score among the selection scores is set as the target candidate tool. Then, the task node is bound to the target candidate tool to obtain the initial workflow skeleton. Identify each upstream and downstream node pair with direct data dependencies in the initial workflow skeleton, and determine whether the output data pattern of the upstream node in the upstream and downstream node pair is semantically and structurally compatible with the input data pattern of the corresponding downstream node to obtain the data pattern compatibility verification result. If the data pattern compatibility verification result indicates incompatibility, then a field conversion operator is inserted between the upstream and downstream node pairs using a preset field mapping rule to obtain the target workflow skeleton; the field conversion operator includes an operator for renaming data fields, an operator for filling missing data fields with default values, and an operator for standardizing and converting data units.

6. The method for generating execution decision schemes for fiscal and tax tasks according to claim 1, characterized in that, The process of determining the corresponding strategy set based on the semantic graph of the tax and finance task, and determining the temporal caliber set based on the target rule twin graph and the time label, to construct an evaluation space based on the strategy set and the temporal caliber set, includes: The semantic graph of the fiscal and tax task is parsed to obtain several corresponding variable decision points. A strategy set is generated based on each variable decision point. Based on the time label, the rule version corresponding to the historical rule that is related to or different from the task is selected from the twin graph of the target rule. A temporal caliber set is generated based on each rule version. Then, an evaluation space is constructed based on the strategy set and the temporal caliber set. Accordingly, the counterfactual evaluation of strategy and temporality and Pareto optimization are performed on the evaluation space using the target workflow skeleton to obtain target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space. The Pareto dominance relationship is then used to compare and filter the target fiscal and tax indicator evaluation vectors to obtain a set of decision schemes for executing the fiscal and tax task, including: Using the target workflow skeleton and based on preset indicators, the evaluation vectors of financial and tax indicators for each element in the evaluation space are determined to obtain the corresponding target financial and tax indicator evaluation vectors. The Pareto dominance relationship is used to filter the target financial and tax indicator evaluation vectors to obtain a set of decision schemes for executing the financial and tax tasks. The preset indicators include revenue indicators, cost indicators, risk indicators, time delay indicators, and stability indicators.

7. The method for generating execution decision schemes for fiscal and tax tasks according to claim 6, characterized in that, After obtaining the set of decision schemes for performing the financial and tax tasks, the method further includes: The upper confidence bound algorithm is used to perform a guided search on the set of decision schemes in the evaluation space to obtain candidate schemes and determine the Pareto front corresponding to each candidate scheme; the Pareto front includes several non-dominated solutions and their corresponding key influencing factors. The set of rule constraints, the Pareto front, and the key influencing factors are merged to obtain a merged set of constraints. The satisfiability of the merged set of constraints is determined using a modular theory solver for satisfiability, and the satisfiability determination result is obtained. When the satisfiability determination result is characterized as satisfiable, a corresponding evidence model is generated, and a structured proof summary is generated using the evidence model; the proof summary includes variable assignment, hash binding of input data, and a step-by-step derivation record from Boolean skeleton to numerical theory; When the satisfiability determination result is characterized as unsatisfiable, the minimum unsatisfiable kernel is extracted from the merged constraint set, and each logical clause in the minimum unsatisfiable kernel is reverse-mapped to the specific rule entries in the target rule twin graph and the specific task nodes in the target workflow skeleton to obtain a source tracing report for locating and interpreting the root cause of the conflict.

8. The method for generating execution decision schemes for fiscal and tax tasks according to claim 7, characterized in that, After comparing and filtering the evaluation vectors of the target fiscal and tax indicators using Pareto dominance to obtain a set of decision schemes for performing the fiscal and tax tasks, the method further includes: The financial and tax tasks are executed based on the set of decision schemes to obtain the task execution results and the corresponding key event sequence. The key event sequence is then encoded using a preset encryption algorithm to obtain the corresponding ciphertext encoded sequence. Each element in the ciphertext encoding sequence is concatenated with the hash value of the previous event and the ciphertext encoding of the current event in chronological order to obtain a secure hash value, and a linear hash chain is generated based on each of the secure hash values. A Merkle tree is constructed based on each of the ciphertext encoding sequences, and the root value corresponding to the Merkle tree is determined. Then, the root value is set as the fingerprint identifier corresponding to the financial and tax task. The key event sequence is proved using a preset proof circuit and based on the root value and the tax task triplet, and the proof result is obtained. The proof result is then sent to a preset external verification interface so that the external verifier can determine the verification result corresponding to the tax task based on the proof result. The number of verification failures corresponding to the financial and tax task is determined. Then, a preset comprehensive loss function is used to determine the deviation value of the actual task execution delay corresponding to the financial and tax task exceeding the preset expected delay and the decrease value of the system stability index relative to the preset baseline based on the number of verification failures, so as to obtain the loss result. The system strategy parameters corresponding to the decision scheme set are iteratively updated using a preset mirror descent algorithm and based on the loss results. Then, the updated system strategy parameters are verified using a preset historical regression dataset and each preset protection threshold to obtain the verification results. The preset protection thresholds include a decrease threshold for task verification success rate, an increase threshold for task execution latency, and a decrease threshold for system stability indicators. If the verification result indicates that the updated system policy parameters meet each of the preset protection thresholds, then the updated system policy parameters are set as the target system policy parameters. If the verification result indicates that the updated system policy parameters do not meet each of the preset protection thresholds, then the update freeze mechanism is triggered.

9. A device for generating execution decision schemes for fiscal and tax tasks, characterized in that, include: The rule twin graph generation module is used to parse the financial and tax tasks in natural language to obtain financial and tax task triples, which include a financial and tax task semantic graph, an initial constraint set, and target weights. Based on the financial and tax task triples, an initial rule twin graph including time tags and version tags is constructed. Then, a consistency check is performed on the initial rule twin graph, and after the check passes, the initial rule twin graph is set as the target rule twin graph. The workflow skeleton generation module is used to determine the rule subgraph in the target rule twin graph, and convert the rule subgraph into a set of rule constraints including spatiotemporal labels, to determine the satisfiability of the set of rule constraints, and after the determination is passed, to use a preset hierarchical task network and based on the financial and tax task semantic graph to perform subtask decomposition and workflow orchestration of the financial and tax task to obtain the target workflow skeleton. The space construction module is used to determine the corresponding strategy set based on the semantic graph of the financial and tax task, and to determine the temporal caliber set based on the twin graph of the target rule and the time label, so as to construct an evaluation space based on the strategy set and the temporal caliber set; the strategy set includes different paths to execute the financial and tax task; the temporal caliber set includes different interpretations of the same strategy at different time points. The decision scheme set determination module is used to perform counterfactual evaluation of strategy and temporality and Pareto optimization on the evaluation space using the target workflow skeleton, to obtain the target fiscal and tax indicator evaluation vectors corresponding to each element in the evaluation space, and to compare and filter the target fiscal and tax indicator evaluation vectors using the Pareto dominance relationship to obtain a set of decision schemes for executing the fiscal and tax task.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the method for generating execution decision schemes for financial and tax tasks as described in any one of claims 1 to 8.

Images