A business and financial data integration processing platform fusing multi-modal models

Abstract

The application relates to the technical field of enterprise business and financial data processing, and discloses a business and financial data integrated processing platform fusing a multi-modal model, which comprises a workflow engine module for decoupling a directed acyclic graph topology and a flow state; a data analysis alignment module for extracting business and financial text semantic vectors by using a twin network; a feature classification fine-tuning module for outputting a difference classification result based on semantic and structural combined features and a low-rank self-adaptive adapter; a slice rule binding module for constructing a two-dimensional slice from the difference result and time sequence features, and calculating a slice amount according to priority weight; and a distributed scheduling module for executing cross-period patch mounting and preventing duplication control through a hash idempotent key and a prevention hanging mechanism. The application overcomes the matching obstacle of heterogeneous business and financial texts, reduces the iteration cost of multi-tenant difference models, eliminates cross-period rule calculation conflicts, prevents repeated accounting and account suspension caused by network disorder, and guarantees data final consistency.

Description

Technical Field

[0001] This invention relates to the field of enterprise business and financial data processing technology, specifically to an integrated business and financial data processing platform that integrates multimodal models. Background Technology

[0002] As enterprises continue to advance their digital transformation, the demand for collaborative processing between business and financial systems is growing. The integrated business and financial data processing platform, as the core hub connecting front-end transactions and back-end general ledger, undertakes the tasks of data aggregation, cleaning, transformation, and journal entry generation. In daily operations, the platform needs to receive heterogeneous data from multiple sources, including structured order details from the front end and unstructured invoices asynchronously obtained from the financial end.

[0003] Existing business and financial processing architectures rely heavily on traditional technical components for linear pipeline processing when handling the aforementioned tasks. In the data parsing phase, the system primarily uses text comparison methods involving regular expressions and keyword matching to establish a static mapping dictionary between business descriptions and financial accounts. Limited intelligent applications are confined to deploying a unified deep learning classification model globally for difference judgment. In the accounting calculation phase, the processing engine generally uses execution scripts with fixed logic, running various financial rules sequentially according to a pre-defined order. Regarding network communication management in a distributed environment, existing systems routinely extract the transaction serial number from the front-end business documents as the anti-duplicate primary key for database transactions and directly send write or update requests to the underlying general ledger database according to the physical order of instruction reception.

[0004] However, the severe heterogeneity between natural language on the business side and standardized accounts on the financial side in existing technologies makes traditional keyword exact matching prone to failure when encountering synonym replacements or texts of varying lengths. Introducing a single global classification model can lead to catastrophic forgetting when updating parameters in multi-tenant concurrent accounting scenarios, while deploying a full model independently for each tenant incurs unacceptable computational overhead. For complex cross-period businesses, a single document, after being split, often simultaneously hits multiple rules such as revenue allocation, tax deduction, and tail-difference tolerance. Because static execution scripts lack a mandatory conflict scheduling mechanism based on account compliance levels, the disordered parallel triggering of operator logic can easily cause algebraic calculation errors in the final amount. Therefore, this invention provides an integrated business and financial data processing platform that incorporates multimodal models to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an integrated business and financial data processing platform that incorporates a multimodal model. This platform solves the problems of existing business and financial data processing systems struggling to effectively align and parse unstructured financial documents and structured business orders when dealing with cross-period business and heterogeneous data. Furthermore, in distributed high-concurrency scenarios, cross-period data flow can easily lead to logical conflicts in accounting rules, and network retries and out-of-order instructions can cause duplicate accounting and suspended accounts.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an integrated business and financial data processing platform that incorporates multimodal models, comprising: The workflow engine module parses the received business instructions, instantiates a directed acyclic graph topology, and maintains the flow state data of task nodes in the state machine database. The data parsing and alignment module extracts time-series feature objects from business and financial documents associated with business instructions and generates a baseline semantic vector and a comparison semantic vector for the text. The feature classification fine-tuning module constructs combined features based on the baseline semantic vector and the contrast semantic vector, and outputs differential classification results by loading a classification model with a low-rank adaptive adapter. The slice rule binding module is driven by the difference classification results and time series feature objects to construct a two-dimensional slice data structure, and runs the financial rule engine to calculate the amount value of the slice unit; The distributed scheduling module generates journal entry requests or attaches transitional voucher patches based on the accounting period to which the slice unit belongs, and generates hash idempotent keys for the corresponding operations to perform anti-duplicate control.

[0007] Furthermore, the workflow engine module parses the business instructions to extract the tenant identifiers and currently active workflow templates, instantiates a directed acyclic graph (DAG) topology, and calls the topology sorting anti-loop checker to calculate the in-degree index of each task node to perform compliance checks. After the check passes, a mechanism for separating topology and state read / write is executed. A separate flow state record is created for each task node in an independent state machine database, and a mapping relationship between the DAG topology and flow state data is established based on the node identifier. The internally deployed node scheduler listens for state change events in the state machine database. When a preceding node is detected to have reached the completion state, the flow state update of subsequent task nodes that meet the conditions in the DAG topology is driven to proceed.

[0008] By decoupling the static topology from the dynamic running state, the graph model is prevented from carrying dynamic data during the flow, thereby reducing the memory overhead caused by lock conflicts in high-concurrency update scenarios.

[0009] Furthermore, the data parsing and alignment module calls the time-series parser to scan the explicit and relative time condition fields in the business instruction-related financial documents, extracts the start and end dates to generate time-series feature objects with probability confidence. When the probability confidence is lower than a preset threshold, a blocking signal is sent to the workflow engine module to drive the corresponding task node to change to manual review status. The business text description and financial text description in the corresponding financial documents are extracted as independent input data and transmitted to a Siamese text embedding network containing multiple layers of self-attention networks and sharing parameter weights. Through the global context feature extraction of the Siamese text embedding network, the heterogeneous text descriptions are mapped to the same metric space, and the baseline semantic vector and the comparative semantic vector are output respectively.

[0010] This mechanism utilizes a dual encoder structure with shared parameter weights to map business natural language descriptions and standardized financial account names to a high-dimensional feature space. It relies on spatial distance to measure deep semantics and overcomes matching interference caused by differences in word count and synonym substitution.

[0011] Furthermore, the feature classification fine-tuning module receives the baseline semantic vector and the comparison semantic vector output by the data parsing and alignment module, calculates the absolute residuals of each dimension to form a semantic residual vector, and concatenates the difference between the semantic residual vector and the structured order field on the feature dimension to generate a combined feature for downstream inference. In the computation sandbox environment allocated to the tenant with logical isolation, the module reads the corresponding tenant's private low-rank adaptive adapter and attaches it to the basic classification model side. It then uses the combined feature to perform forward propagation calculation in combination with the frozen basic model weights. The module projects the output of the hidden layer of the classification model to the probability distribution set of the preset business and financial difference categories, selects the category corresponding to the maximum probability value as the difference classification result output for downstream construction of the slice grid, and asynchronously updates the private low-rank adaptive adapter matrix parameters by generating a fine-tuning training set in combination with the manual review results when the confidence level is low.

[0012] This design fixes the basic model parameters during model inference and only updates the private dimensionality reduction and expansion matrices. While controlling local computing power overhead, it prevents catastrophic forgetting of the global model and achieves adaptive iteration of the differential model in a multi-tenant isolated environment.

[0013] Furthermore, the slicing rule binding module combines the difference classification results output by the feature classification fine-tuning module with the business categories extracted from the business instructions to form a row dimension, and sets the accounting period after discretization based on the time-series feature objects as the column dimension, cross-generating a two-dimensional slice data structure composed of slice units; extracts the set of financial rules that hit the current slice unit, calls the account set-level priority allocation function built into the financial rule engine, and calculates the priority weight score of each rule in combination with the sensitivity coefficient of the account set compliance strictness, generating a rigid execution sequence arranged in descending order; according to the order of the rigid execution sequence, extracts the operator logic contained in the rule in sequence to perform algebraic operations covering or superimposing on the total business amount, obtains and fills the amount value that should be allocated to the slice unit, and forcibly captures the rule metadata and writes it into the audit attribute set of the slice unit.

[0014] By leveraging a discretized time and business slice data structure and combining it with account set-level priority settings to determine the execution order of rules, conflicts in amount allocation calculations caused by the parallel triggering of multiple financial rules are eliminated.

[0015] Furthermore, when implementing data splitting based on slice data, the distributed scheduling module extracts the column dimension identifier of the slice unit in the two-dimensional slice data structure and converts it into a timestamp. It reads the current open accounting period variable recorded by the current system through the interface and performs a time axis comparison. When it is determined that the column dimension identifier is equal to the current open accounting period, it extracts the amount value inside the slice unit and the bound audit attribute set, merges them to generate a formal accounting entry request, and submits it to the core general ledger database. When it is determined that the time corresponding to the column dimension identifier is later than the current open accounting period, it encapsulates the corresponding slice data as a transition voucher patch data package, locates the task node in the directed acyclic graph topology responsible for processing future periods to complete the patch mounting, and sends an instruction to change the flow status of the task node to a time-locked state.

[0016] By using patch mounting to temporarily store accounting instructions for future periods and performing node addressing, accounting errors can be prevented from being recorded prematurely for cross-period business data.

[0017] Furthermore, after generating a formal accounting entry request or patch mounting operation, the distributed scheduling module calls a hash idempotent key generator to extract the row dimension identifier, column dimension identifier, operation action code, and persistently stored environment key factor from the request context. The extracted factors are concatenated to form a basic string, and a cryptographic hash function with a fixed output length is used to perform operations on the basic string to generate a unique hash idempotent key for each independent accounting action. The generated hash idempotent key is encapsulated in the request message header, driving the downstream execution end to compare the key-value records in the persistent idempotent control table before any database action occurs, in order to identify and discard duplicate accounting write instructions caused by network retries.

[0018] By using multi-dimensional combination factors and fixed environment keys to generate anti-duplicate identifiers, the uniqueness of business and financial data requests can be controlled in distributed network retry scenarios.

[0019] Furthermore, the distributed scheduling module includes a calendar monitoring service. After the enterprise initiates a new billing cycle and triggers an open event, the calendar monitoring service captures the event and iterates through the task nodes in the time-locked state. It calculates the wake-up weight value of each task node by integrating the preset service level value and the waiting time indicator. Based on the calculated wake-up weight value, the locked nodes are distributed to multiple virtual queues. A weighted polling mechanism is used to periodically extract nodes to implement smooth flow limiting. Instructions are sent to the workflow engine module to synchronously update the flow status of the extracted nodes to the unlocked state. The driver executor reads the independent version number of the target task node from the state machine database, uses an optimistic locking mechanism to concurrently preempt the unlocked nodes, and executes the formal accounting operation after verifying that the state is open correctly during the secondary verification period.

[0020] By combining weighted smooth rate limiting with a non-blocking preemption mechanism, the risk of database lock waiting timeouts caused by massive concurrent wake-ups of nodes during the opening period can be avoided.

[0021] Furthermore, the distributed scheduling module establishes an anti-hanging consistency transaction interception mechanism. Before submitting accounting instructions to the downstream core ledger database, it extracts the generation timestamp and action sequence number of the currently arriving forward instruction and retrieves historical reverse reversal records under the same business root key in the independently maintained anti-hanging control table. The parameters of the forward instruction are compared with the parameters of the retrieved historical reverse reversal records. If it is found that the action sequence number of the current forward instruction is earlier than the reverse reversal record, or the generation timestamp is earlier than the processing timestamp of the reverse reversal record, it is determined that there is a hanging risk caused by network disorder. Based on the hanging risk determination result, the currently arriving forward instruction is directly discarded, and a processing result containing the anti-hanging interception identifier is returned to the requesting caller.

[0022] The above-mentioned interception and judgment prevents the reversal of the order of forward and reverse instruction processing caused by network delays and ensures the final consistency of global business and financial data.

[0023] Furthermore, the workflow engine module is equipped with a topology evolver driven by macro-level operational data. This evolver periodically collects the cumulative number of times each task node under a business category triggers anti-hangover interception and the cumulative number of times manual review is required. Based on the statistical values, algebraically weighted calculation is performed to obtain the comprehensive risk index of the corresponding node. The comprehensive risk index is compared with the system's preset topology variation threshold. When an over-limit is detected, the currently effective workflow template is extracted as the benchmark object. On the path preceding the high-risk node in the graph topology of the benchmark object, an enhanced audit control node with an independent identifier is dynamically inserted, thereby generating a new graph topology template with an incrementing version number.

[0024] Dynamic iteration of the execution topology version is based on historical operating indicators, so that the system business flow rules are automatically adjusted according to actual operating risks, forming a closed-loop risk control mechanism.

[0025] This invention provides an integrated business and financial data processing platform that incorporates multimodal models. It offers the following advantages: 1. This invention effectively overcomes the matching barrier in textual expression of heterogeneous business and financial documents by constructing a twin text embedding network containing shared parameter weights to extract benchmark and contrast semantic vectors, and concatenating them with structured field differences to form combined features. At the same time, a private low-rank adaptive adapter is mounted in a tenant-isolated computing sandbox to perform classification inference and parameter backflow update. Under the premise of keeping the global basic model weights frozen, independent iteration of multi-tenant difference classification logic is realized, reducing the computing power overhead of model fine-tuning and preventing data pollution between different sets of accounts.

[0026] 2. This invention utilizes differential classification results and time-series feature objects to construct a two-dimensional sliced ​​data structure, transforming complex inter-period business transactions into discrete time and business grids for layered processing. It also incorporates a built-in financial rule engine, calculating the priority weight score of rules based on the strictness of compliance at the account set level to generate a descending execution sequence. This eliminates algebraic operation conflicts caused by the parallel triggering of multiple financial rules during document processing, ensuring the orderliness and deterministic nature of inter-period account splitting and amount allocation calculations.

[0027] 3. In the distributed scheduling stage, this invention generates a hash idempotent key by extracting multiple correlation factors from the request dimension and a fixed environment key, thus avoiding duplicate accounting writes caused by network communication retries. At the same time, by comparing the action sequence number and timestamp in the anti-hanging control table, it intercepts the risk of reversed order of forward and reverse instructions caused by network latency. In the accounting period wake-up stage, it combines weighted polling smoothing rate limiting and optimistic locking preemption mechanism to effectively ensure the eventual consistency of business and financial transactions and the stability of system concurrent operation under high-concurrency flow environment. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the platform architecture of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of data parsing and semantic alignment processing according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating the topology instantiation and state maintenance process according to an embodiment of the present invention. Figure 5 This is a flowchart of the combined feature classification and fine-tuning update according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the principle of spatiotemporal slice construction and rule binding processing according to an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the principle of transition patch mounting and multidimensional idempotent key generation in one embodiment of the present invention. Figure 8 This is a comparison chart showing the effects of different business data processing schemes of the present invention.

[0029] Among them, 10 is the data parsing and alignment module; 20 is the workflow engine module; 30 is the feature classification fine-tuning module; 40 is the slicing rule binding module; and 50 is the distributed scheduling module. Detailed Implementation

[0030] The technical solutions in 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.

[0031] See attached document Figure 1 The present invention provides an integrated business and financial data processing platform that integrates multimodal models, including a data parsing and alignment module 10, a workflow engine module 20, a feature classification fine-tuning module 30, a slicing rule binding module 40, and a distributed scheduling module 50.

[0032] The data parsing and alignment module 10 receives and processes structured order data from the business system and asynchronously transmitted unstructured financial documents. The module incorporates a time-series parser and a Siamese text embedding network trained based on contrastive learning. The time-series parser extracts time-series feature objects containing start and end dates and confidence levels from the documents. The Siamese text embedding network maps business text descriptions and financial text descriptions to the same metric space with shared weights, outputting a baseline semantic vector and a contrastive semantic vector for calculating distance metrics.

[0033] Workflow engine module 20 is communicatively connected to data parsing and alignment module 10. Workflow engine module 20 instantiates a directed acyclic graph topology based on the tenant's currently active template. Workflow engine module 20 includes a topology sorting loop checker for performing topology loop detection. Workflow engine module 20 maintains the business flow state independently of the topology structure, storing it in a separate state machine database.

[0034] The feature classification fine-tuning module 30 is connected to the data parsing and alignment module 10. The feature classification fine-tuning module 30 concatenates the feature residuals of the aforementioned semantic vectors with the differences in structured fields to generate combined features. The feature classification fine-tuning module 30 internally deploys a classification model containing a low-rank adaptive architecture, which dynamically loads the corresponding private low-rank adaptive adapter in a multi-tenant sandbox environment. The feature classification fine-tuning module 30 outputs the classification results of the business differences and performs matrix parameter updates caused by manual review of sample feedback.

[0035] The slice rule binding module 40 is connected to the feature classification fine-tuning module 30 and the data parsing alignment module 10, respectively. The slice rule binding module 40 establishes a two-dimensional slice data structure based on business dimensions, difference classifications, and accounting periods. The slice rule binding module 40 is configured with a financial rule engine, which handles rule logic conflicts based on the account set-level priority allocation function, calculates the amount value of each slice unit, and binds the corresponding audit attributes.

[0036] The distributed scheduling module 50 serves as the system control center, connecting the aforementioned modules. The distributed scheduling module 50 includes a calendar service interface, a hash idempotent key generator, a weighted fair queue rate limiter, and a distributed transaction controller. The distributed scheduling module 50 generates globally unique hash idempotent keys, transforms and submits current period slice data as formal entry requests, generates transitional voucher patches for future period slice data, attaches them to future nodes, and executes time-series locking instructions.

[0037] See attached document Figure 2 This invention provides an integrated processing method for business and financial data that incorporates multimodal models, comprising the following steps: S10: Obtain multi-source business and financial document data, call the time series parser to extract time series feature objects from the documents, input the text description data into the twin text embedding network to generate a baseline semantic vector and a contrast semantic vector in the same metric space; S20: Based on the currently received business instructions, instantiate a directed acyclic graph (DAG) that separates the topology and state; perform a topology sorting algorithm on the instantiated DAG to check for the existence of loops; and create and maintain the flow state data of the corresponding task nodes in the state machine database. S30: Calculate the feature residual between the baseline semantic vector and the contrast semantic vector, combine the feature residual with the differences in structured fields to form a feature vector, input the feature vector into the classification model for inference to output the difference classification result, judge the confidence of the classification result, and trigger the manual review process and sample return mechanism based on the threshold condition. S40: Establish a sliced ​​data structure consisting of row and column dimensions, run the financial rule engine to handle logical conflicts between multiple rules according to the priority sequence, calculate the amount allocated to each rule and fill it into the sliced ​​cell, and record the audit attributes of the amount allocation. S50 determines the accounting period attribute of the slice unit based on the time-series feature object, generates journal entry requests for slice data belonging to the current accounting period, generates transition voucher patches for slice data belonging to future accounting periods and attaches the patches to the corresponding nodes of the directed acyclic graph, and calls the distributed scheduling module 50 to generate hash idempotent keys containing multiple association factors for each operation. S60 listens for period start events sent by the external calendar service, starts a weighted fair queue to perform a smooth wake-up process on nodes in a locked state, calls the executor to preempt the target node after wake-up using an optimistic locking mechanism, and completes the final accounting operation in combination with the distributed transaction controller and anti-dangling table. Based on the statistical indicators generated during node runtime, it regenerates the topology template version for subsequent execution.

[0038] The specific implementation methods of each part of the method and system of the present invention will be described in detail below with reference to specific embodiments.

[0039] See attached document Figure 3 The specific implementation of the aforementioned step S10 will be described in detail. This process is mainly divided into a multi-source data access and temporal feature extraction stage, and a twin network semantic alignment generation stage.

[0040] During the data access phase, the system employs differentiated acquisition strategies based on the characteristics of different data sources. The data parsing and alignment module 10 receives structured order data sent by the front-end business system through a pre-defined application programming interface (API). Simultaneously, for the financial side, the system listens to a message queue and asynchronously acquires unstructured financial text data converted by optical character recognition (OCR). The data parsing and alignment module 10 then calls an internally deployed time-series parser to extract basic features from the received business and financial documents.

[0041] The time-series parser comprises a rule-based configuration-based regular expression engine and a named entity recognition model. In specific business scenarios, unstructured documents often contain non-standardized or relative time expressions such as "this quarter" or "the 15th of the following month." For the specific regular expression matching strategies and the network layer construction methods of the named entity recognition model, those skilled in the art can perform conventional configuration and basic training based on the actual data format characteristics; these are well-known technologies in the field and will not be elaborated upon here.

[0042] The time series parser scans the explicit time field and relative time condition field in the input data, extracts the corresponding time start and end boundaries, and generates a standardized time series feature object. This time series feature object is defined as a vector containing multiple attribute dimensions: ; In the formula, This represents the generated time-series feature object; Indicates the start date of the accounting period obtained through parsing; Indicates the end date of this start and end range; A unit of measurement indicating the duration of a time period; This represents the probability confidence level of the analysis result, and its value range is set to [0,1].

[0043] Data parsing alignment module 10 compares probability confidence. The confidence level is set by the system administrator based on the fault tolerance and compliance requirements of the business account set, for example, a value of 0.85. If the probability confidence level is lower than the preset confidence threshold, the data parsing and alignment module 10 records the abnormal state parameters, the original text fragments, and the candidate period objects, and sends a blocking signal to the workflow engine module 20. Based on this, the workflow engine module 20 updates the business flow status of the corresponding task node to the manual review status to prevent low-quality time-series features from polluting the downstream accounting flow logic.

[0044] Regarding the execution process of semantic alignment generation in twin networks, the data parsing and alignment module 10 converts the text description content into a vector form that can be used for mathematical metric calculations. Business systems record natural language descriptions of goods or services, while financial documents record standardized accounting subjects or invoice item names; the two exhibit heterogeneous characteristics in their underlying text structure.

[0045] To overcome the failure of traditional keyword exact matching due to different business habits, the data parsing and alignment module 10 extracts business text descriptions from structured order data and financial text descriptions from unstructured financial text data. The data parsing and alignment module 10 transmits the extracted business text descriptions and financial text descriptions as independent input data to the twin text embedding network.

[0046] The Siamese text embedding network comprises two encoders with identical structures and shared parameter weights. Before the data input encoders execute, the module calls a pre-defined tokenizer to serialize the text and performs truncation or zero-padding to unify the input tensor dimension. The encoder employs a multi-layer self-attention network structure, which internally includes a multi-head self-attention layer, a feedforward neural network layer, and a pooling layer. After the input tensor extracts global contextual features through the self-attention layer, the variable-length features are compressed into a fixed-dimensional vector through the pooling layer (e.g., extracting features from classification marker bits or performing global average pooling). The basic architecture design of the self-attention network and the mechanism for calculating attention weights can be constructed using existing deep learning framework components, which are well-known technologies in the field and will not be elaborated upon here.

[0047] The data parsing and alignment module 10 uses a contrastive learning mechanism to pre-train the twin text embedding network. During the training process, the system extracts historically verified business and financial text groups as positive sample pairs and randomly samples unrelated text groups as negative sample pairs. The training process applies a contrastive loss function, introducing a margin parameter into the loss calculation formula and constraining the loss value through a gradient descent algorithm. This allows the system to use corresponding business and financial texts as positive sample pairs and uncorrelated business and financial texts as negative sample pairs, updating the network parameter weights by reducing the spatial distance between positive sample pairs and increasing the spatial distance between negative sample pairs.

[0048] During the actual inference phase, the data parsing and alignment module 10 calls the encoder with shared parameter weights to process the input text. Let the input business text be described as follows: The input financial text is described as follows: The encoder mapping function is denoted as... ,in This represents the set of parameters that have been solidified in the network after comparative learning training. The data parsing and alignment module 10 calculates the output vector as follows: ; ; In the formula, This represents the baseline semantic vector calculated for the business text description. This represents the comparative semantic vector calculated for the financial text description. The shared parameter weights structure allows the baseline and comparative semantic vectors to be projected into the same high-dimensional feature metric space.

[0049] The data parsing alignment module 10 calculates the spatial distance between the baseline semantic vector and the comparison semantic vector in the metric space to measure their semantic alignment. From a theoretical perspective, the vector direction in high-dimensional space can represent the deep semantics of the text, and calculating the spatial angle can effectively shield against interference caused by differences in word count or synonym substitution. The data parsing alignment module 10 uses cosine distance calculation logic to calculate the semantic distance; the specific calculation formula is as follows: ; In the formula, The numerical value representing the calculated semantic distance is theoretically between [0, 2]. A value closer to 0 indicates a higher semantic correlation between the two text segments. The numerator represents the inner product of the baseline semantic vector and the comparison semantic vector; the denominator represents the product of the L2 norm of the baseline semantic vector and the L2 norm of the comparison semantic vector. The data parsing and alignment module 10 temporarily stores the calculated baseline and comparison semantic vectors as the basic input feature parameters for subsequent feature classification. See attached document Figure 4 The specific implementation of the aforementioned step S20 will be described in detail. This process covers the generation of directed acyclic graphs and anti-loop verification mechanisms, as well as the decoupling and independent storage maintenance mechanism for business flow status.

[0050] In this embodiment, the workflow engine module 20 receives business instructions triggered by the front end, parses the tenant identifier carried in the instructions to extract the currently active workflow template for that tenant, and instantiates the corresponding directed acyclic graph structure accordingly. The workflow template pre-defines standard task nodes and temporal dependencies between nodes for different business scenarios. This directed acyclic graph is mathematically defined as a graph model. ,in This represents a set of task nodes that encompass various business processing logics. This represents the set of directed edges that characterize the dependencies between nodes in terms of execution order.

[0051] In system operation scenarios, circular dependencies in the graph topology can cause task flow to stall. To address this, the workflow engine module 20 incorporates a topology sorting anti-loop checker to check the instantiated graph model. Perform compliance checks. The specific check logic relies on calculating the in-degree metric of each node. The workflow engine module iterates through the node set 20 times. Count the number of directed edges that point to each node from other nodes, i.e., calculate the in-degree of each node. ,in For set Any node in the array.

[0052] Workflow engine module 20 establishes an empty queue and pushes all nodes with an initial in-degree of zero into the queue, while simultaneously initializing the visited node counter. When the queue is not empty, the workflow engine module 20 pops the head node of the queue in sequence and performs an increment operation on the counter, that is, according to the formula. Update the process. For all adjacent nodes pointed to by the newly popped node, the workflow engine module 20 decrements their in-degree values ​​by one. If the in-degree value of an adjacent node drops to zero, then that adjacent node is pushed into the queue.

[0053] After the above loop operation is completed, the workflow engine module 20 compares the counter value. With sets Total number of nodes If the two are equal, it indicates that the graphical model... There are no loops in the topology, so topology generation is effective. If the counter value... Less than the total number of nodes This indicates the existence of a circular dependency node that could not be traversed. The workflow engine module 20 determines that there is an internal loop in the graph, blocks the execution flow of the business instruction, and returns an alarm message indicating a topology abnormality to the calling end.

[0054] After confirming the validity of the graph model topology, workflow engine module 20 executes the node state stripping and independent maintenance mechanism. In the graph model... Under the premise of only being used to solidify the business flow topology path and not carrying dynamic data, the workflow engine module 20 targets the set in an independent state machine database. Each task node within the graph has its own unique flow state record. A mapping relationship between directed acyclic graph nodes and state records is established using globally unique node identifiers. Through these node identifiers, the system can independently read or update the current execution state of each node while querying static graph dependencies.

[0055] Workflow engine module 20 assigns state variable functions to each node. Its value range is limited to a preset set of standard financial turnover states. Inside. This set of states is defined as: ; In the formula, This indicates a pending state, waiting for the preceding task to complete. This indicates a manual review status triggered due to data anomalies. This indicates a time-locked state that is not enabled during the accounting period; This indicates the unlocked state after the calendar constraint has been removed; This indicates the accounting status where an underlying database write operation is being performed; This indicates that the downstream core general ledger system has returned a voucher number or a successfully confirmed posting status. When using Outbox or Saga mode, the workflow engine module 20 will only update the corresponding node status after receiving a successful confirmation from the downstream general ledger system. ; This indicates a canceled state triggered when a business return occurs; This indicates an abnormal state where a system call fails or rolls back. The transitions between states are governed by the transition matrix rules built into the state machine engine.

[0056] , and This is not a completed state that can trigger the advancement of subsequent nodes; for accounting nodes, the completed state is... For offsetting nodes, the completed state is... For nodes requiring manual review, after the manual review is completed, the state machine updates the node state based on the review results. , , or The review results will be written to the audit log and the corresponding tenant's private sample library.

[0057] For the decoupled state-driven logic, workflow engine module 20 includes a node scheduler. The node scheduler monitors state change events in the state machine database in real time. When a node... The status has changed to At that time, the node scheduler uses the graph model. The set of directed edges in The node scheduler queries all successor nodes of the given node and checks if all predecessor nodes of each successor node have reached the corresponding completion state. If the condition is met, the node scheduler changes the state of the successor node that meets the condition from [previous state]. Update to the next stage state, thereby driving the business to continue flowing according to the diagram structure.

[0058] When the graph topology instantiation is complete, the workflow engine module 20 will set the version number of the template used. Core metadata is written as immutable attributes to the persistent configuration table of the directed acyclic graph instance. The design logic of separating topology structure from state read and write helps reduce the memory overhead caused by lock conflicts in high-concurrency update scenarios, and ensures that even if the global template is iterated and upgraded later, existing business instances can still execute independently along the original dependency path based on snapshot-level metadata.

[0059] See attached document Figure 5 The specific implementation of step S30 will be described in detail below. This process includes the splicing and construction of multimodal combined features and the sandboxed low-rank adaptive inference and model parameter backflow update mechanism.

[0060] During the feature processing stage, the system receives the baseline semantic vector and the comparison semantic vector transmitted by the data parsing and alignment module 10 through the feature classification fine-tuning module 30. The feature classification fine-tuning module 30 calculates the absolute residuals of the two vectors in each dimension to form a semantic residual vector. Simultaneously, the feature classification fine-tuning module 30 extracts corresponding numerical and categorical difference fields from the structured order data and the unstructured financial text data. These fields include, but are not limited to, the difference in amount between the business and financial ends, the tax rate comparison deviation, and the time span difference. The feature classification fine-tuning module 30 normalizes the extracted difference fields and combines them into a structured feature vector. For the specific normalization algorithm, those skilled in the art can use existing max-min normalization or standardization components, which are well-known technologies in the field and will not be elaborated upon here.

[0061] The feature classification fine-tuning module 30 concatenates the semantic residual vector and the structured feature vector along the feature dimension to generate multimodal combined features for downstream model inference. By fusing text residuals containing deep semantics with intuitive numerical differences, the system can simultaneously capture inconsistencies in natural language descriptions and rigid deviations in business data. The formula for the concatenation operation is: ; In the formula, This represents the combined feature vector generated by concatenation; Represents the baseline semantic vector; Represents a contrast semantic vector; This represents the semantic residual vector composed of the element-wise absolute differences between the two. Represents structured feature vectors; This represents the vector concatenation operator.

[0062] After generating the combined feature vector, the feature classification fine-tuning module 30 performs differential classification inference. In an enterprise-level multi-tenant architecture, the differences in business and financial accounting among tenants are specific, and data contamination between different tenants needs to be prevented. The feature classification fine-tuning module 30 allocates logically isolated computing sandbox environments for each tenant within the system.

[0063] Within this computational sandbox, the specific classification model deployed by the system includes an input layer, multiple fully connected hidden layers, and a normalized exponential function output layer. This classification model consists of two parts: a globally shared base model and tenant-private low-rank adapters. The low-rank adapters are connected in parallel via bypass and are mounted on the weight matrix side of the fully connected hidden layers.

[0064] The feature classification fine-tuning module 30 dynamically loads the low-rank adapter parameters of the corresponding tenant. During the forward propagation calculation of the classification model, the parameter matrix of the base model is frozen and does not participate in gradient calculation. The output of the hidden layer of the model is obtained by summing the calculation results of the base model branch and the calculation results of the low-rank adapter branch. The forward inference calculation formula is: ; In the formula, This represents the hidden layer output vector; The dimension is The frozen base model weight matrix; The dimension is Tenant-private low-rank reduced-dimensional matrix; The dimension is Tenant-private low-rank up-dimensional matrix; The rank parameter of the matrix is ​​much smaller than the feature dimension. With output dimension In practical applications, it is usually set to a smaller integer such as 8 or 16 based on computing power limitations and feature complexity experience.

[0065] The feature classification fine-tuning module 30 inputs the hidden layer output vector into the normalized exponential function layer, and maps the output to a set of probability distributions for preset business and financial difference categories. These preset business and financial difference categories correspond to specific physical states of financial accounting anomalies, such as price discrepancies, tax rate anomalies, inter-period recognition, and account mismatches. The feature classification fine-tuning module 30 selects the category corresponding to the maximum probability as the classification result and uses this maximum probability value as the confidence level of the classification result.

[0066] The feature classification fine-tuning module 30 compares the confidence level with a preset risk control interception threshold. This threshold is configured by the system based on the audit rigor level of different account sets; for example, it can be set to 0.90 under strict audit requirements. When the confidence level is lower than the risk control interception threshold, it indicates that the classification model lacks sufficient certainty in determining the current difference category. The feature classification fine-tuning module 30 sends a command to the workflow engine module 20 to change the business flow status of the current processing node to manual review status. After business personnel complete manual review and enter the correct difference classification labels, the feature classification fine-tuning module 30 triggers an asynchronous model parameter backflow update mechanism. Before the manually reviewed samples enter the training set, the system performs deduplication, desensitization, label consistency verification, and abnormal sample filtering. Each model inference and backflow update records the base model version number, tenant low-rank adapter version number, training sample version number, and feature engineering version number, and writes the above version information into the audit record of the corresponding difference classification result.

[0067] The feature classification fine-tuning module 30 extracts the combined feature vector of the current sample and the correct classification label to construct a fine-tuning training set. The module uses the cross-entropy loss function to calculate the loss between the predicted output and the true label, and combines this with the Adam optimizer to set the learning rate parameter. During backpropagation, since the weights of the base model are frozen, the system only calculates the gradients of the tenant-private low-rank reduced-dimensional matrix and low-rank increased-dimensional matrix, and updates their parameters. This local parameter update mechanism helps mitigate the catastrophic forgetting phenomenon that easily occurs when modifying the global base model, thereby supporting adaptive iteration for different tenant business differences.

[0068] See attached document Figure 6 The specific implementation of the aforementioned step S40 will be described in detail. This process involves the establishment of a multi-dimensional slice data structure and the conflict handling and amount allocation mechanism of rigid priority rules.

[0069] For documents spanning multiple periods and involving multiple business transactions, the total amount often needs to be broken down and deconstructed based on different business attributes and time spans. The slicing rule binding module 40 receives the difference classification results processed by the feature classification fine-tuning module 30, and combines them with the business category in the business instruction and the accounting period in the time-series feature object to construct a two-dimensional slice data structure. Specifically, the slicing rule binding module 40 uses the combination of business category and difference classification as the row dimension, and the accounting period discretized according to the time span (e.g., by accounting calendar month) as the column dimension, generating a matrix structure composed of multiple slice units.

[0070] Each slice unit in the system is defined as a data tuple containing multiple attributes, and its data structure expression formula is as follows: ; In the formula, Indicates an independent slice unit; This indicates the row dimension identifier to which the slice belongs. Its value is calculated by combining the business category and the classification result using a well-known hash algorithm. This indicates the column dimension identifier, which corresponds to a specific accounting calendar month or financial cycle. This represents the amount of money allocated to this slice unit; the initial value is set to zero. This represents the set of audit attributes bound to the slice. Through the above data organization logic, the system can map business and financial documents that span multiple periods and contain differences to discrete time and business grids.

[0071] After establishing the multi-dimensional slice data structure, the system needs to perform amount calculations and rule binding for each slice unit. The enterprise's accounting system has multiple pre-defined financial rules, such as revenue allocation rules, tax deduction rules, and tolerance rules for minor discrepancies. In actual calculations, a slice unit may simultaneously trigger multiple financial rules, leading to conflicts in the calculation logic. The slice rule binding module 40 incorporates a financial rule engine, using an accounting system-level priority allocation function to sort and force-schedule the triggered rules.

[0072] The slice rule binding module 40 extracts the set of rules that match the current slice unit. For any rule in the rule set... The financial rules engine calls the account set-level priority allocation function to calculate the corresponding weight score. This weight score calculation depends not only on the basic rule settings but also on the compliance requirements of the current account set, and incorporates the rule's own sensitivity coefficient to the stringency of compliance. The calculation formula is as follows: ; In the formula, This represents the priority weight score calculated from the output. This represents the preset basic priority constant in the rule dictionary, with a value range set to the positive integer interval [1, 100]. This indicates the compliance strictness rating of the current account set. It is statically read from the system configuration table and its value is set to an integer within the range [1, 5]. The larger the value, the stricter the compliance audit requirements. Representation rules The sensitivity coefficient to the level of compliance strictness is used to adjust the priority of different rules under the same set of account compliance levels. and These represent the adjustment coefficients corresponding to the basic priority and the compliance stringency, respectively, and both satisfy the constraint conditions. as well as , Its value is manually configured by the system administrator based on the control strategies of different account sets.

[0073] After calculation, the financial rule engine sorts the rule set in descending order based on priority weight scores. To handle the special case of identical scores, the financial rule engine incorporates a secondary comparator in its sorting logic. When two rules are found to have equal weight scores, the secondary comparator extracts the system's preset global rule codes corresponding to the two rules and compares them again according to the lexicographical order of the code strings to establish a unique order. Through this dual sorting operation, the financial rule engine generates a rigid execution sequence.

[0074] Based on this rigid execution sequence, the slice rule binding module 40 sequentially extracts the operator logic contained in the rule, performs algebraic operations on the total amount, and calculates the amount that should be allocated to the current slice unit. The rule execution results include at least four processing types: overriding, overlaying, mutual exclusion, and blocking. If the processing type of a high-priority rule is blocking, then the low-priority rule will not be executed. If the processing type is overlaying, then the low-priority rule continues to calculate based on the output of the previous rule. If the processing type is mutual exclusion, then only the rule with the highest priority is executed. If the processing type is overriding, then the output of the subsequently executed rule overwrites the same type of output fields of the previous rule.

[0075] During each amount allocation calculation, the slice rule binding module 40 forcibly extracts the rule code, execution timestamp, and operator identifier of the currently effective rule, and writes them as immutable metadata into the audit attribute set. This priority-weight-based conflict handling mechanism helps control the orderliness of accounting logic operations within the same grid at the same time, reducing the probability of amount allocation conflicts caused by parallel rule triggering.

[0076] See attached document Figure 7 The specific implementation of the aforementioned step S50 will be described in detail. This process involves data diversion and cross-period patch mounting operations based on accounting period characteristics, as well as a multidimensional hash idempotent key anti-reconstruction mechanism.

[0077] After the multi-dimensional slice data structure is established and the amount allocation is completed, the system needs to determine the exact accounting timing for each slice unit. The distributed scheduling module 50 reads the current open accounting period variable recorded in the system for the current accounting set through an interface. The distributed scheduling module 50 extracts the column dimension identifiers from the slice units, converts them into standard timestamp format, and compares them with the timeline of the current open accounting period.

[0078] If the accounting period corresponding to the column dimension identifier of a slice unit is equal to the current open accounting period, and the status of that period in the target ledger is Open, the distributed scheduling module 50 determines that the slice belongs to the current valid processing scope. Under this determination, the distributed scheduling module 50 extracts the monetary values ​​and audit attribute set contained within the slice, and combines them with the accounting subjects corresponding to the business to assemble them into a formal accounting entry request, which is then submitted to the downstream core general ledger database for accounting processing. If the accounting period corresponding to the slice unit is earlier than the current open accounting period, the distributed scheduling module 50 queries the ledger status of that historical period; if the historical period has been closed, direct modification of the historical ledger is prohibited, and adjusting or reversing entries are generated and attached to the compensation node corresponding to the current open period.

[0079] If the column dimension identifier of a slice unit corresponds to a time later than the current open accounting period, the system prohibits early posting operations based on financial compliance constraints. The distributed scheduling module 50 determines that the slice belongs to a future period and encapsulates the slice data as a transitional voucher patch data package. The distributed scheduling module 50 determines the target future accounting period based on the column dimension identifier and addresses and locates the specific task node in the directed acyclic graph instance responsible for processing the accounting for that future period.

[0080] During the addressing and location process, if the distributed scheduling module 50 detects that the target task node has not yet been instantiated in the current graph topology, the distributed scheduling module 50 will call the workflow engine module 20 to initiate a request to add a future node within the instance. Before adding the future node and its corresponding directed dependency edge, the workflow engine module 20 first performs a topology sorting anti-loop check; if the check passes, it dynamically generates the future node and establishes a directed dependency edge from the current node to the future node, while writing the added node ID, trigger slice ID, addition time, and the instance topology version number after the addition to the audit log; if the check fails, it rejects the addition and sets the current node status to manual review. or failure state This is to ensure the connectivity and auditability of the topology graph.

[0081] The distributed scheduling module 50 attaches the transition credential patch data package to the target task node. After attachment, the distributed scheduling module 50 writes a timing lock tag to the target task node and sends a state change instruction to the workflow engine module 20, synchronously changing the flow state of the corresponding node in the state machine database to the timing lock state. Nodes in this state pause their progress and wait for event-driven wake-up from the subsequent calendar service interface.

[0082] In a distributed service architecture, network interruptions and retry / retransmission mechanisms can lead to duplicate execution of accounting write commands, resulting in duplicate accounting data. To address this issue, a hash idempotent key generator is deployed within the distributed scheduling module 50. For the generated formal accounting entry requests and patch mounting operations, the hash idempotent key generator is responsible for generating a globally unique anti-duplicate identifier for each independent operation.

[0083] Considering that a single business order number cannot distinguish the split execution actions of the same order in different accounting periods, the hash idempotent key generator adopts a multi-dimensional combination extraction strategy when constructing factors. The hash idempotent key generator extracts the tenant ID, legal entity ID, account set ID, source system ID, transaction ID, source document line ID, slice ID, accounting period, operation action code, patch version number, and environment key from the context. The hash idempotent key generator then concatenates the extracted factors in a preset order to form a basic string.

[0084] To enhance tamper resistance and avoid identifier conflicts between different environments, the hash idempotency key generator uses a fixed environment key pre-generated and persistently stored at the deployment environment or tenant level. This environment key does not change with each request. For repeated requests with the same business semantics, the same operational behavior, and the same patch version, the hash idempotency key generator outputs the same idempotency key. Subsequently, the system performs calculations using the HMAC-SHA256 algorithm based on this environment key. The specific idempotency key calculation formula is expressed as follows: ; In the formula, This represents the hash idempotent key that satisfies the uniqueness verification requirement within the system's business space, as calculated in the output. This represents a keyed cryptographic hash function, specifically the HMAC-SHA256 algorithm with a fixed output length of 256 bits. Indicates the original serial number of the extracted business document; This indicates the row dimension identifier associated with the operation; The column dimension identifier indicating the association; This indicates the action code, such as "ACT_BOK" for an accounting request and "ACT_MNT" for a patch mount. This refers to a fixed environment key that is pre-generated and persistently stored at the deployment environment or tenant level. This represents the string concatenation operator.

[0085] In specific implementations, the splicing factor also includes tenant ID, legal entity ID, account set ID, source system ID, source document line ID, slice ID, accounting period, and patch version number, to distinguish different accounting processing actions in multi-tenant, multi-account set, multi-period, and multi-version scenarios. The specific underlying bitwise AND XOR operation steps of the SHA-256 algorithm and the random number generation logic for the environmental salt value can be constructed using existing cryptographic standard library components by those skilled in the art; these are well-known technologies in the field and will not be elaborated upon here.

[0086] After calculating the hash idempotency key, the distributed scheduling module 50 encapsulates it in the request header parameters of the corresponding request message and transmits it to the downstream execution end. Before performing substantive database operations, the downstream system verifies the persistent idempotency control table based on the extracted hash idempotency key and accelerates queries by combining caching. The idempotency control table is indexed uniquely by the hash idempotency key and records the processing status, which includes at least Processing, Success, Failed, and Cancelled.

[0087] If the same idempotent key is matched in the idempotency control table and the status is "Success", the downstream system determines that the current request is a duplicate request caused by a network retry. In this case, the system directly intercepts and discards the database write operation within the request, and returns the previous successful processing result recorded in the idempotency control table to the calling end. If a record with a status of "Processing" is matched, the system returns a processing status or refuses to repeat the execution. If a record with a status of "Failed" is matched, the system determines whether to allow re-execution based on the failure reason and the preset retry strategy to meet the idempotency requirements of distributed transactions. The above-mentioned multi-dimensional hash idempotent key construction mechanism helps support the uniqueness control of cross-period splitting processing requests for business and financial data in complex network environments.

[0088] The specific implementation of the aforementioned step S60 will be described in detail. This process involves weighted smooth wake-up based on calendar events, anti-hanging interception of consistent transactions, and an adaptive evolution mechanism for the system's macro topology.

[0089] The distributed scheduling module 50 has a built-in calendar monitoring service. When a company starts a new financial period, the calendar monitoring service captures the opening events during that period. To handle the sudden surge in patch wake-up requests that may be triggered during the opening of a period, the distributed scheduling module 50 incorporates a weighted fair queue algorithm for smooth rate limiting and batch wake-up. The distributed scheduling module 50 iterates through all task nodes in a time-locked state and calculates the wake-up weight value for each task node. The specific weight calculation formula is as follows: ; In the formula, This represents the calculated wake-up weight value. This represents the preset service level value of the tenant to which the node belongs, and the value range is set to an integer in the range of [1, 5]. This represents the total cost of the patches mounted on the node. This indicates the waiting time for the node to be in a time-locked state; , , The weight coefficients for each factor are all set to real numbers greater than 0 and satisfy the following conditions: By introducing a logarithmic function to handle the amount factor, we can avoid the overflow of weight values ​​caused by extremely large transactions.

[0090] The distributed scheduling module 50 allocates locked nodes to multiple virtual queues based on the calculated wake-up weight values. According to the system's preset concurrent throughput threshold, it periodically extracts nodes from multiple virtual queues using a weighted round-robin or deficit round-robin mechanism. This ensures that high-weight queues receive a higher wake-up frequency while guaranteeing that low-weight queues receive the minimum wake-up quota within a preset period, preventing low-weight nodes from being starved for a long time. The module also updates the flow status of the extracted nodes from the time-locked state to the unlocked state.

[0091] In a microservice architecture where multiple executors concurrently consume tasks from nodes, the distributed scheduling module 50 utilizes a version-based optimistic locking mechanism for resource preemption. The distributed scheduling module 50 maintains an independent version number field for each node in the state machine database. When an executor attempts to retrieve an unlocked node, it reads the current version number and uses it as a filter condition when executing a database update statement. If the database returns zero affected rows, it indicates that the node has been preempted by another concurrent executor, and the current executor abandons its operation on that node. This non-blocking locking mechanism helps control the risk of database lock wait timeouts in high-concurrency scenarios.

[0092] Before the executor successfully preempts and prepares to submit the voucher, the distributed scheduling module 50 queries the status of the accounting period corresponding to the target account set again; if the period is still Open, the accounting continues; if the period has been changed to Close, the direct posting is terminated, and adjusting entries or reversing entries are generated by adding compensation nodes.

[0093] In actual business execution, to address occasional network out-of-order issues or instruction errors caused by retry mechanisms in distributed network communication, the system deploys a mechanism to prevent suspended consistency transactions. In asynchronous business and finance processing scenarios, if a forward accounting transaction triggers a reverse compensation or cancellation instruction due to a timeout, and due to network routing delays, the reverse instruction arrives at the downstream database before the forward instruction, then if a subsequent forward instruction arrives again and is blindly executed, it will cause a suspended accounting error.

[0094] The distributed scheduling module 50 maintains an independent anti-hanging control table on the downstream core general ledger database side. This anti-hanging control table records the business root key, action type, action sequence number, hash idempotent key, and processing status. The business root key is used to associate forward, reverse, compensating, or canceling actions within the same business transaction chain; the hash idempotent key is used to prevent the repeated execution of a single action; and the action type is used to identify whether the current action is a forward posting, reverse cancellation, or compensating process. Before executing any substantive accounting instruction, the distributed scheduling module 50 extracts the hash idempotent key generated in the aforementioned steps, and extracts the business root key and action sequence number corresponding to the same transaction chain, performing an existence check in the anti-hanging control table.

[0095] If the anti-hanging control table already contains a record with a corresponding hash idempotent key and marked as canceled or compensated, the distributed scheduling module 50 extracts the generation timestamp of the currently arriving instruction and compares it with the processing timestamp of the existing record in the table. If the action sequence number corresponding to the current forward instruction is earlier than or equal to the action sequence number of the existing reverse record, the distributed scheduling module 50 determines that a typical hanging risk caused by network out-of-order delivery has occurred. If the action sequence number is missing or cannot be read, the system uses the generation timestamp of the current instruction and the processing timestamp of the existing reverse record as auxiliary judgment criteria; if the generation timestamp of the current forward instruction is earlier than or equal to the processing timestamp of the existing reverse record, it is also determined that a hanging risk exists. Based on this judgment result, the distributed scheduling module 50 directly discards the current forward instruction and returns an interception flag to prevent invalid forward data from being written, thereby maintaining the eventual consistency of global data.

[0096] During the long-term operation of the system, the topology evolver built into the workflow engine module 20 performs adaptive version iterations of the macro topology graph based on historical operating data. The topology evolver statistically analyzes operating metrics, calculates risk indices, and generates template versions within the tenant sandbox. Operating metrics, template versions, risk thresholds, and evolution results are isolated between different tenants. The topology evolver periodically collects operating statistics for each task node, specifically including the number of anti-hangover interceptions and the number of manual reviews triggered by low confidence. The topology evolver calculates a comprehensive risk index for specific categories of nodes in the graph structure, using the following formula: ; In the formula, This represents the calculated overall risk index of the nodes. This indicates the cumulative number of times the anti-suspension interception was triggered at this node within the statistical period; This indicates the cumulative number of times the node was downgraded to a manually reviewed state within the statistical period; This indicates the total number of business documents that flow through this node within the statistical period; and These represent the adjustment parameters corresponding to the interception rate and the manual intervention rate, respectively, and their constraints satisfy... and, .

[0097] The topology evolver compares the overall risk index of nodes with the preset topology mutation threshold in the system configuration table. If the overall risk index of a node for a certain business category exceeds the topology mutation threshold, the topology evolver triggers template cloning and mutation operations. The topology evolver extracts the currently effective workflow template as a baseline file and locates the high-risk node in the corresponding dependency relationship of the graph model.

[0098] The topology evolver dynamically inserts enhanced audit nodes with independent node identifiers along the execution path preceding the high-risk node, such as adding mandatory dual-signature tasks, thereby generating a new graph topology template with an incrementing version number. When receiving subsequent business instructions of the same type, the workflow engine module 20 will call the updated template to instantiate the directed acyclic graph, thus supporting closed-loop control where business rules are dynamically adjusted based on operational risk data. Historical DAGs that have been instantiated and are in operation continue to execute according to the template version number fixed in their instance metadata. The topology evolver does not directly modify the nodes and edges of running historical DAGs, thereby ensuring the version stability and auditability of the running DAG topology.

[0099] In one specific application scenario, this is applied to supply chain cost settlement. The business system receives a supplier service order, which includes the supplier code, service items, service amount, tax rate, service description, and related document number. The finance department then uploads the contract text and invoice scans related to the order. The contract text specifies the service coverage and cost settlement terms, while the invoice scans specify the invoice item name, invoice amount, tax amount, and supplier information.

[0100] After receiving the aforementioned business order data and unstructured financial documents, the data parsing and alignment module 10 first calls the layout parsing and entity extraction engine to extract supplier codes, expense types, amounts, tax rates, and text descriptions from the contract text and scanned invoices. The time-series parser parses the period conditions in the contract text and generates a time-series feature object by combining them with the associated basic fields in the business documents. If the parsing confidence level is lower than a preset threshold, the corresponding task node is set to manual review; if the parsing confidence level meets the threshold requirement, the time-series feature object is sent to the subsequent slicing processing stage.

[0101] In the semantic alignment stage, the data parsing and alignment module 10 uses the business description in the business order as the business text description and the invoice item name and contract fee description as the financial text description. These are input into a twin text embedding network with shared parameter weights to obtain a baseline semantic vector and a contrast semantic vector. The system calculates the semantic distance based on the two semantic vectors and uses the element-wise absolute residual as the semantic residual vector. Simultaneously, the system constructs structured difference features from fields such as the difference between the order amount and the invoice amount, the difference between the business tax rate and the invoice tax rate, whether the supplier code is consistent, and whether there is an offset in the period objects.

[0102] The feature classification fine-tuning module 30 concatenates the semantic residual vector with the structured difference features into a combined feature vector and loads the corresponding private low-rank adapter within the current tenant's sandbox. The classification model outputs the difference category, classification confidence, and evidence field. When the classification result points to a difference category such as "freight surcharge" or "abnormal tax rate," and the confidence meets the risk control threshold, the system continues to the slicing rule binding stage; when the confidence is lower than the threshold or hits the high-risk supplier condition, the system sets the task node to manual review and feeds the review result back to the current tenant's private sample library.

[0103] The slicing rule binding module 40 establishes a two-dimensional slice data structure based on business categories, difference classification results, and accounting period units. For the slice unit corresponding to the current open period, the system generates a formal accounting entry request based on the account mapping rules, tax separation rules, allocation rules, and tail difference processing rules configured in the accounting set. For the slice unit corresponding to subsequent periods, the system generates a transitional voucher patch with deferred or accrued attributes and attaches the patch to the corresponding future period node in the DAG instance. If the target future node has not yet been generated, the workflow engine module 20 adds the node and its corresponding directed edge after the topology sorting anti-loop verification passes, and writes the added node information to the audit log.

[0104] The distributed scheduling module generates hash idempotent keys for each of the 50 pairs of formal accounting entry requests and transitional voucher patches. These idempotent keys are generated from the tenant ID, legal entity ID, accounting set ID, source system ID, transaction ID, source document line ID, slice ID, accounting period, operation code, patch version number, and fixed environment key. Before performing any substantive write operations, the downstream general ledger system queries the persistent idempotency control table to prevent duplicate postings caused by network retries or repeated submissions.

[0105] When the enterprise accounting calendar service releases a new period opening event, the distributed scheduling module 50 smoothly wakes up nodes in a time-locked state through a weighted fair queue. The executor preempts unlocked nodes using an optimistic locking mechanism and re-verifies the target accounting period status before submitting vouchers. If the period status meets the posting conditions, the posting continues; if the period status does not meet the posting conditions, direct posting is terminated, and adjusting or reversing entries are generated through compensation nodes. The system maintains business root keys, action sequence numbers, and anti-hanging control tables at the transaction chain level to intercept the risk of accounting hang caused by reverse instructions arriving before forward instructions.

[0106] Within the tenant sandbox, the topology evolver continuously tracks the number of anti-hangover interceptions and manual reviews for nodes of specific business categories, and calculates a comprehensive risk index. When the comprehensive risk index for a certain business category or supplier dimension exceeds a preset threshold, the system generates a new topology template version based on the currently effective template and strengthens audit nodes in front of high-risk nodes. Subsequent business instructions of the same type instantiate the DAG based on the new template; historical DAGs that have been instantiated and are in operation continue to execute according to the template version fixed in their instance metadata.

[0107] Comparison plan: The first approach is a solution based on manual review and fixed rule templates; The second approach is a processing solution based on general text vector matching and a rule engine; The third approach is the processing scheme proposed in this embodiment, which integrates twin semantic alignment, sandboxed differential classification, slice patch mounting, and distributed anti-duplicate mechanism.

[0108] The test data is shown in the table below: Comparison table of the effects of different treatment schemes See attached document Figure 8 The above results demonstrate that the proposed solution does not rely solely on text keywords or general semantic similarity. Instead, it combines twin semantic residuals, structured field differences, and tenant private adapters to form a combined feature classification mechanism for business and financial difference scenarios, thereby improving the stability and interpretability of difference recognition.

[0109] This embodiment utilizes accounting period slicing, transitional voucher patch mounting, hash idempotent keys, anti-dangling control tables, and optimistic locking preemption mechanisms to create a traceable, deduplication-preventable, and compensable execution chain for cross-period data processing and distributed ledger operations. This mechanism reduces the impact of duplicate entries, out-of-order execution, and topology anomalies on accounting results and allows manual review results to be fed back to the tenant's private sample library, supporting subsequent local adaptive updates to template versions and classification models.

[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal data processing platform integrating business and financial data, characterized in that, include: The workflow engine module parses the received business instructions, instantiates a directed acyclic graph topology, and maintains the flow state data of task nodes in the state machine database. The data parsing and alignment module, the specific execution process of which includes: The time sequence parser is invoked to scan the explicit and relative time condition fields in the business instruction-related financial documents, extract the start and end dates, and generate a time sequence feature object with probability confidence. When the probability confidence is lower than a preset threshold, a blocking signal is sent to the workflow engine module to drive the corresponding task node to change to manual review status. The business text description and financial text description extracted from the corresponding business and financial documents are used as independent input data and transmitted to a twin text embedding network containing a multi-layer self-attention network and sharing parameter weights. By extracting global contextual features through a Siamese text embedding network, heterogeneous text descriptions are mapped to the same metric space, and baseline semantic vectors and comparative semantic vectors are output respectively. The feature classification fine-tuning module constructs combined features based on the baseline semantic vector and the contrast semantic vector, and outputs differential classification results by loading a classification model with a low-rank adaptive adapter. The slice rule binding module, the specific execution process of which includes: The difference classification results output by the feature classification fine-tuning module are combined with the business categories extracted from the business instructions to form the row dimension, and the accounting period after discretization based on the time-series feature objects is set as the column dimension, and a two-dimensional slice data structure composed of slice units is generated by cross-generating the two-dimensional slice data structure. Extract the set of financial rules that hit the current slice unit, call the account set-level priority allocation function built into the financial rule engine, calculate the priority weight score of each rule by combining the sensitivity coefficient of the account set compliance strictness, and generate a rigid execution sequence in descending order; Based on the order of the rigid execution sequence, the operator logic contained in the rule is extracted in sequence, and algebraic operations of covering or superimposing are performed on the total amount of business to obtain and fill the amount value that should be allocated to the slice unit. At the same time, the rule metadata is forcibly captured and written into the audit attribute set of the slice unit. The distributed scheduling module generates journal entry requests or attaches transition voucher patches based on the accounting period to which the slice unit belongs, and generates hash idempotent keys for the corresponding operations to perform anti-duplicate control. The specific execution process of the distributed scheduling module when implementing anti-duplication control includes: After generating a formal accounting entry request or patch mounting operation, call the hash idempotent key generator to extract the row dimension identifier, column dimension identifier, operation action code, and persistently stored environment key factor from the request context. The extracted multiple factors are concatenated to form a basic string. A cryptographic hash function with a fixed output length is used to perform operations on the basic string to generate a unique hash idempotent key for each independent accounting action. The generated hash idempotent key is encapsulated in the request message header, driving the downstream execution end to compare the key-value record in the persistent idempotent control table before the database action occurs, so as to identify and discard duplicate accounting write instructions caused by network retries.

2. The integrated business and financial data processing platform merging multimodal models according to claim 1, characterized in that, The specific execution process of the workflow engine module includes: The business instructions are parsed to extract the tenant identifiers and currently active workflow templates, a directed acyclic graph topology is instantiated, and the topology sorting anti-loop checker is called to calculate the in-degree index of each task node to perform compliance checks. After the verification is passed, the mechanism of separating topology structure and state read and write is executed. A separate flow state record is created for each task node in an independent state machine database, and a mapping relationship between the directed acyclic graph topology structure and the flow state data is established based on the node identifier. The system uses an internally deployed node scheduler to monitor state change events in the state machine database. When a preceding node is detected to have reached a completed state, it drives the subsequent task nodes in the directed acyclic graph topology that meet the conditions to advance their state updates.

3. The integrated business and financial data processing platform merging multimodal models according to claim 1, characterized in that, The specific execution process of the feature classification fine-tuning module includes: Receive the baseline semantic vector and the comparison semantic vector output by the data parsing and alignment module, calculate the absolute residuals of each dimension to form a semantic residual vector, and concatenate the semantic residual vector with the difference of the structured order field on the feature dimension to generate a combined feature for downstream inference. Within the computation sandbox environment that is logically isolated for the tenant, the low-rank adaptive adapter private to the corresponding tenant is read and attached to the basic classification model. The combined features are then used to perform forward propagation computation in combination with the frozen basic model weights. The output of the hidden layer of the classification model is projected onto a set of probability distributions of predefined business and financial difference categories. The category corresponding to the highest probability value is selected as the difference classification result output for downstream construction of slice grids. When the confidence level is low, the private low-rank adaptive adapter matrix parameters are updated asynchronously by generating a fine-tuning training set in combination with the results of manual review.

4. The integrated business and financial data processing platform merging multimodal models according to claim 1, characterized in that, The specific execution process of the distributed scheduling module when implementing data splitting based on sliced ​​data includes: Extract the column dimension identifiers of the slice units in the two-dimensional slice data structure and convert them into timestamps. Then, read the current open accounting period variables recorded by the current system through the interface and perform time axis comparison. When the column dimension identifier is equal to the current open accounting period, extract the amount value inside the slice cell and the bound audit attribute set, merge them to generate a formal accounting entry request and submit it to the core general ledger database. When the time corresponding to the column dimension identifier is later than the current open accounting period, the corresponding slice data is encapsulated as a transition voucher patch data package, located in the directed acyclic graph topology structure, and the task node responsible for processing future periods is mounted to complete the patch. Then, an instruction is sent to change the flow state of the task node to the time-locked state.

5. The integrated business and financial data processing platform merging multimodal models according to claim 4, characterized in that, The distributed scheduling module also includes a calendar monitoring service, the execution process of which includes: After the enterprise opens the event during the new billing period, the calendar listening service captures the event and traverses the task nodes in the time-locked state, and calculates the wake-up weight value of each task node by combining the preset service level value and the waiting time index. Based on the calculated wake-up weight value, the locked nodes are distributed to multiple virtual queues. Nodes are periodically extracted through a weighted polling mechanism to implement smooth flow limiting. Instructions are sent to the workflow engine module to synchronously update the flow status of the extracted nodes to the unlocked state. The actuator reads the independent version number of the target task node from the state machine database, uses an optimistic locking mechanism to concurrently preempt the unlocked node, and performs the formal accounting operation after the state is confirmed to be open correctly during the secondary verification period.

6. The integrated business and financial data processing platform merging multimodal models according to claim 1, characterized in that, The distributed scheduling module also includes a mechanism to prevent suspended consistency transactions, the execution process of which includes: Before submitting accounting instructions to the downstream core general ledger database, extract the generation timestamp and action sequence number of the current incoming positive instruction, and retrieve the historical reverse reversal records under the same business root key in the independently maintained anti-hanging control table. The parameters of the forward instruction are compared with the parameters of the retrieved historical reverse reversal records. If it is found that the action sequence number of the current forward instruction is earlier than the reverse reversal record, or the generation timestamp is earlier than the processing timestamp of the reverse reversal record, it is determined that there is a hanging risk caused by network disorder. Based on the suspension risk assessment result, the currently arriving positive instruction is directly discarded, and a processing result containing the anti-suspension interception flag is returned to the requesting caller.

7. The integrated business and financial data processing platform merging multimodal models according to claim 2, characterized in that, The workflow engine module also deploys a topology evolver driven by macroscopic operational data, and the execution process includes: The cumulative number of times anti-hangover interception is triggered and the cumulative number of times manual review is required for each task node under the business category are collected periodically. The comprehensive risk index of the corresponding node is calculated by algebraically weighting the statistical values. By comparing the comprehensive risk index with the system's preset topology variation threshold, when an over-limit is detected, the currently active workflow template is extracted as the benchmark object. On the path preceding high-risk nodes in the graph topology of the baseline object, enhanced audit control nodes with independent identifiers are dynamically inserted to generate new graph topology templates with incrementing version numbers.

Images