Financial document automatic auditing method and device and medium

Abstract

The invention discloses a financial document automatic auditing method and device and a medium, and relates to the technical field of financial reimbursement auditing. The method comprises the following steps: receiving a financial document to be audited and an associated attachment document, respectively extracting a first entity set, and extracting a second entity set from unstructured content; constructing a dynamic knowledge graph state space based on the entity set, wherein the dynamic knowledge graph state space comprises an entity vector generated by an entity embedding algorithm and a relation vector generated by a relation coding algorithm; defining a reinforcement learning action space, wherein the reinforcement learning action space comprises three types of atomic operations of newly adding and deleting a triple and adjusting confidence; in combination with the real-time document flow, the historical case library and the audit result data, dynamically evolving the knowledge graph through atomic operation, and calculating a value return value of each operation; pre-judging accumulated return values of different operation sequences by utilizing a Monte Carlo tree search algorithm, and pruning a low return sequence; executing the optimized operation sequence to update the knowledge graph; and finally, based on the updated atlas, triggering a logic verification rule to generate an auditing result.

Description

Technical Field

[0001] The present application relates to the technical field of financial reimbursement auditing, and in particular to a method, device and medium for automated auditing of financial documents. Background Art

[0002] Current mainstream financial sharing platforms generally adopt a rule-based model for document review, implementing basic compliance checks through pre-set field verification rules. However, this model has significant technical bottlenecks: First, the system can only process structured field data and lacks the ability to effectively parse unstructured attachment content (such as complex seals in invoice images and multi-page PDF related information), resulting in heavy reliance on manual spot checks; second, existing technologies are unable to capture deep logical contradictions between document fields and attachment content (such as semantic conflicts between reimbursement types and the actual content of bills), creating hidden audit loopholes; third, static rule bases require continuous manual maintenance of massive business rules, making it difficult to respond to high-frequency business changes, resulting in increased risk control lags.

[0003] At its core, the traditional architecture suffers from three technical flaws: first, the lack of heterogeneous data processing capabilities makes it impossible to collaboratively analyze multimodal data such as text, images, and tables; second, the lack of a cross-document semantic association mechanism makes it difficult to build a closed-loop verification chain for scattered content such as itineraries, accommodation vouchers, and transportation tickets; third, the rule base relies on manual updates from experts and is unable to autonomously learn and evolve based on real-time data streams, resulting in a long-term low coverage rate of abnormal patterns.

[0004] Therefore, how to break through the barriers of multimodal data analysis, establish a cross-document semantic-level verification mechanism and realize the autonomous dynamic evolution of audit rules, so as to comprehensively improve the recognition rate of hidden violations, has become a technical problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0005] The embodiments of the present application provide a method, device and medium for automated review of financial documents to solve the following technical problems: how to break through the barriers of multimodal data analysis, establish a cross-document semantic-level verification mechanism and realize the autonomous dynamic evolution of review rules, thereby comprehensively improving the recognition rate of hidden violations.

[0006] In a first aspect, an embodiment of the present application provides a method for automated review of financial documents, the method comprising: receiving financial documents to be reviewed and their associated attachment documents submitted by users, and extracting a first entity set from the structured fields of the financial documents to be reviewed, and extracting a second entity set from the unstructured content of the associated attachment documents; constructing a dynamic knowledge graph state space based on the first entity set and the second entity set; wherein the dynamic knowledge graph state space comprises: entity vectors generated by mapping each entity through an entity embedding algorithm, and relationship vectors generated by mapping relationships between entities through a relationship encoding algorithm; defining a reinforcement learning action space; wherein the reinforcement learning action space comprises three types of atomic operations for modifying the dynamic knowledge graph state space: adding a triple operation, deleting a triple operation, and adjusting a triple operation. Tuple confidence operation; based on real-time document flow data and preset historical case library data and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space, and after each atomic operation is executed, the operation value return value generated after the atomic operation is executed is calculated; using the Monte Carlo tree search algorithm, the expected cumulative return value of executing different atomic operation sequences is predicted in the reinforcement learning action space, and the atomic operation sequence with expected cumulative reward lower than the threshold is pruned according to the prediction result; execute the atomic operation sequence optimized and screened by the Monte Carlo tree search algorithm, and dynamically update the knowledge graph state space; based on the updated dynamic knowledge graph state space, trigger the logic verification rules to perform compliance verification on the financial documents to be audited, and generate the audit results.

[0007] In one embodiment of the present application, a second entity set is extracted from the unstructured content of the associated attachment document, specifically including: inputting the image or text content of the associated attachment document into a multimodal entity extraction model; processing the text information in the image or text content based on a pre-trained language model layer in the multimodal entity extraction model to generate context-aware text features; capturing the sequential dependencies of the text features based on a bidirectional long short-term memory network layer in the multimodal entity extraction model; and decoding the sequential dependencies based on a conditional random field layer in the multimodal entity extraction model to output a second entity set.

[0008] In one embodiment of the present application, a dynamic knowledge graph state space is constructed based on the first entity set and the second entity set, specifically including: mapping each entity in the first entity set and the second entity set into an entity vector of fixed dimension through the TransE algorithm; using the BERT-wwm-ext model to vector encode the relationship labels between entities to generate a relationship vector; constructing an initial triple set based on the entity vector and the relationship vector; analyzing the time series dependency of entities in historical documents through the LSTM module, and dynamically adjusting the confidence weight of the initial triple set.

[0009] In one embodiment of the present application, a reinforcement learning action space is defined, specifically including: setting a new triple operation for inserting a new triple containing a head entity, a relationship, and a tail entity into a dynamic knowledge graph state space; setting a delete triple operation for removing redundant triples in the dynamic knowledge graph state space whose confidence is lower than a preset threshold; setting a triple confidence adjustment operation for dynamically modifying the confidence weight of the triple according to the real-time data stream, where the confidence weight is a continuous variable.

[0010] In one embodiment of the present application, based on real-time document flow data and preset historical case library data and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space, specifically including: receiving new financial document flow data in real time and extracting its associated entities and relationships; matching the triple structure of similar cases from the historical case library to generate candidate operation sequences; obtaining the triple features of violation cases from the audit result database and correcting the priority of the candidate operation sequence.

[0011] In one embodiment of the present application, the operation value return value generated after executing the atomic operation is calculated, specifically including: generating a historical value coefficient based on the improvement in the audit accuracy rate caused by the addition of new triples according to the historical case database; generating a risk control coefficient based on the change in the violation omission rate after deleting the triples verified in the audit result database; analyzing the coverage of the verification rules triggered by the confidence adjustment operation in the real-time document flow to generate a real-time response coefficient; and performing weighted summation of the historical value coefficient, the risk control coefficient and the real-time response coefficient to generate an operation value return value.

[0012] In one embodiment of the present application, a Monte Carlo tree search algorithm is used to predict the expected cumulative reward value of executing different atomic operation sequences in the reinforcement learning action space, specifically including: starting from the root node of the current dynamic knowledge graph state space, selecting an atomic operation expansion sub-node that has not been fully explored; simulating an operation sequence of a predefined number of steps through a random strategy to generate a leaf node state; and backtracking to calculate the cumulative reward value from the leaf node state to the root node.

[0013] In one embodiment of the present application, based on the updated dynamic knowledge graph state space, logical verification rules are triggered to perform compliance verification on the financial documents to be reviewed, specifically including: based on the entity association relationship in the dynamic knowledge graph state space, the travel voucher documents, accommodation voucher documents and transportation voucher documents in the attached documents are linked in series to generate a logical verification chain containing spatiotemporal entities; the geographic location entities of each voucher document in the logical verification chain are traversed to verify the topological connection relationship between the city name and the latitude and longitude coordinates on the itinerary route; the check-in / check-out time entity of the accommodation voucher and the start and end time entity of the itinerary voucher are extracted to calculate the coverage overlap between the number of accommodation days and the itinerary period; the triplets with broken geographic track nodes and time coverage deviations exceeding the preset tolerance range are marked as evidence of violations.

[0014] In a second aspect, an embodiment of the present application also provides an automated review device for financial documents, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute an automated review method for financial documents such as any one of the above items.

[0015] In a third aspect, an embodiment of the present application further provides a non-volatile computer storage medium storing computer-executable instructions. When the computer-executable instructions are executed, a method for automated review of financial documents as described above is implemented.

[0016] An automated audit method, device, and medium for financial documents provided in the embodiments of the present application have the following beneficial effects: by constructing a collaborative audit framework that integrates reinforcement learning and dynamic knowledge graphs, first, autonomous evolution of financial risk control rules at the minute level is achieved, and operation paths are predicted based on Monte Carlo tree search and triple confidence is dynamically updated, completely solving the fundamental defects of high maintenance costs and delayed responses of static rule bases; secondly, an innovative cross-document semantic verification chain mechanism is established, and the recognition rate of implicit logical vulnerabilities is increased to more than 5 times that of traditional methods through spatiotemporal entity topology verification and time coverage overlap analysis of itinerary vouchers, accommodation vouchers, and transportation vouchers; at the same time, the bottleneck of collaborative analysis of multimodal data is broken through, and the deep entity extraction and relationship prediction model are combined to increase the audit coverage rate of unstructured attachments from less than 30% in the manual spot check mode to fully automated processing, significantly improving the accuracy and timeliness of financial compliance reviews, and effectively ensuring the ability of the enterprise risk control system to dynamically adapt to complex business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0018] Figure 1 A flowchart of a method for automated review of financial documents provided in an embodiment of the present application;

[0019] Figure 2 A schematic diagram of the internal structure of an automated financial document review device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0021] The embodiments of the present application provide a method, device and medium for automated review of financial documents to solve the following technical problems: how to break through the barriers of multimodal data analysis, establish a cross-document semantic-level verification mechanism and realize the autonomous dynamic evolution of review rules, thereby comprehensively improving the recognition rate of hidden violations.

[0022] The technical solutions proposed in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0023] Figure 1 This is a flow chart of a method for automating the review of financial documents provided in this application embodiment. Figure 1 As shown, the embodiment of the present application provides a method for automated review of financial documents, which specifically includes the following steps:

[0024] Step 101: Receive a financial document to be reviewed and its associated attachment documents submitted by a user, extract a first entity set from the structured fields of the financial document to be reviewed, and extract a second entity set from the unstructured content of the associated attachment documents.

[0025] In one embodiment of the present application, a second entity set is extracted from the unstructured content of the associated attachment document, specifically including: inputting the image or text content of the associated attachment document into a multimodal entity extraction model; processing the text information in the image or text content based on a pre-trained language model layer in the multimodal entity extraction model to generate context-aware text features; capturing the sequential dependencies of the text features based on a bidirectional long short-term memory network layer in the multimodal entity extraction model; and decoding the sequential dependencies based on a conditional random field layer in the multimodal entity extraction model to output a second entity set.

[0026] In this embodiment, a multimodal entity extraction technology solution is used to address the core flaw of traditional systems: their inability to parse unstructured attachments. It is understood that the financial documents to be reviewed are typically JSON / XML formatted data generated by ERP systems, with structured fields containing explicit semantic tags (such as "reimbursement ID," "expense type," and "amount"). Using regular expressions and predefined templates, the first entity set can be directly extracted. For example, "Transportation Fee" entities can be extracted from the "Expense Type" field of a travel expense report, while numerical entities can be extracted from the "Amount" field and associated with the unit of measurement.

[0027] It should be noted that the associated attachment documents include multimodal unstructured data such as scanned invoices, contract PDFs, and handwritten vouchers. In specific implementation, the image attachments are first converted into a text stream using the OCR engine, while the PDF document's embedded text layer is parsed. For example, for restaurant invoice images containing blurred seals, the image preprocessing module is used to enhance contrast and then identify the text area; for multi-page contract documents, the header and footer identifiers are automatically extracted to establish cross-page associations.

[0028] Specifically, the process of extracting the second entity set involves inputting the preprocessed text / image content into a multimodal entity extraction model. The language model layer utilizes a pretrained BERT architecture, which generates context-aware text features through a self-attention mechanism. For example, when recognizing "Shanghai Hongqiao International Airport," the co-occurrence of "Hongqiao" and "airport" enables the model to accurately identify this as a location entity rather than a common word. Sequential dependencies are then captured using a bidirectional long short-term memory (BiLSTM) network. For example, in the invoice code "044021800111," the first six digits (044021) represent the administrative region code, which is reinforced in the memory. Finally, a conditional random field (CRF) layer decodes and outputs the second entity set. This layer uses a transition probability matrix to avoid illegal label combinations (e.g., incorrectly inserting the "amount" label into the "invoice code" entity).

[0029] As you can understand, this step enables the coordinated processing of structured and unstructured data. For example, when extracting the "Expense Type: Accommodation Fee" entity from the structured fields of an expense report, the "Issuing Unit: A Catering Company" entity is simultaneously identified from a scanned hotel invoice. The resulting semantic contradiction provides the input basis for subsequent logical verification.

[0030] Step 102: Construct a dynamic knowledge graph state space based on the first entity set and the second entity set.

[0031] In an embodiment of the present application, the dynamic knowledge graph state space includes: entity vectors generated by mapping each entity through an entity embedding algorithm, and relationship vectors generated by mapping the relationships between entities through a relationship encoding algorithm.

[0032] In one embodiment of the present application, a dynamic knowledge graph state space is constructed based on the first entity set and the second entity set, specifically including: mapping each entity in the first entity set and the second entity set into an entity vector of fixed dimension through the TransE algorithm; using the BERT-wwm-ext model to vector encode the relationship labels between entities to generate a relationship vector; constructing an initial triple set based on the entity vector and the relationship vector; analyzing the time series dependency of entities in historical documents through the LSTM module, and dynamically adjusting the confidence weight of the initial triple set.

[0033] In this embodiment, the technical solution for constructing the state space of the dynamic knowledge graph is the core infrastructure for achieving dynamic evolution. It is understandable that the entity embedding algorithm adopts the TransE model, which converts the semantic relationship between entities into a translation operation in the vector space. Specifically, for the extracted entities such as "reimbursement person A" and "hotel B", after being mapped into a vector representation of a fixed dimension through TransE, if there is a relationship of "reimbursement person A-consumption place→hotel B", then the vector of "hotel B" should be approximately equal to the "reimbursement person A" vector plus the "consumption place" relationship vector. This mechanism effectively captures the topological correlation between entities. For example, in the travel scenario, "employee-department→sales department" and "sales department-business area→East China region" can deduce the implicit association of "employee-resident area→East China region".

[0034] It should be noted that the relationship encoding process implements the BERT-wwm-ext scheme. This model uses whole-word masking technology to enhance the semantic understanding of financial relationship labels. For example, it can distinguish the subtle differences between "attribution" relationships (such as "city-administrative affiliation → province") and "inclusion" relationships (such as "travel expenses - inclusion → transportation expenses"). In specific implementation, the relationship labels are input into the pre-trained model to generate high-dimensional relationship vectors. This ensures that semantically similar relationships (such as "ride" and "ride") are close in distance in the vector space, while semantically unrelated relationships (such as "payment" and "geographic location") are separated.

[0035] For example, an initial set of triples is constructed based on entity vectors and relationship vectors: when "Reimbursement recipient: Zhang" (entity vector V1), "Expense type: Accommodation fee" (entity vector V2), and "Relationship: Consumption type" (relation vector R1) are extracted, a triple (V1, R1, V2) is generated. This triple represents the business fact "Zhang - Consumption type → Accommodation fee" and becomes the basic reasoning unit of the knowledge graph.

[0036] Specifically, the LSTM module dynamically adjusts confidence levels. This module analyzes the time series patterns of entities in historical documents. For example, if a supplier repeatedly records "late-night consumption → hotel," the LSTM captures this temporal anomaly and reduces the confidence level of the associated triples. Conversely, if "high-speed rail ticket → travel expense" appears frequently in compliance cases, its confidence level is increased. This design enables the knowledge graph to perceive changes in the business environment, addressing the timeliness limitations of static rule bases.

[0037] It is understandable that the innovation of this step lies in building an evolvable state space foundation. For example, when a new business model generates the triple "online meeting platform - consumption venue → virtual space", its initial confidence is low; as compliance cases accumulate, LSTM gradually increases its weight until it becomes a stable rule, realizing the autonomous evolution of the knowledge graph.

[0038] Step 103: Define the reinforcement learning action space.

[0039] In an embodiment of the present application, the reinforcement learning action space includes three types of atomic operations for modifying the dynamic knowledge graph state space: adding a triplet operation, deleting a triplet operation, and adjusting a triplet confidence operation.

[0040] In one embodiment of the present application, a reinforcement learning action space is defined, specifically including: setting a new triple operation for inserting a new triple containing a head entity, a relationship, and a tail entity into a dynamic knowledge graph state space; setting a delete triple operation for removing redundant triples in the dynamic knowledge graph state space whose confidence is lower than a preset threshold; setting a triple confidence adjustment operation for dynamically modifying the confidence weight of the triple according to the real-time data stream, where the confidence weight is a continuous variable.

[0041] In this embodiment, the reinforcement learning action space definition scheme forms the operational basis for the dynamic evolution of the knowledge graph. It can be understood that the three types of atomic operations are like the "scissors-glue-correction tape" tool set of gene editing, enabling the system to autonomously optimize the knowledge structure. Specifically:

[0042] A new triple operation (ADD_TRIPLE) is added to capture new business relationships. For example, when "virtual currency platform - payment method → ​​digital currency" appears for the first time in an electronic payment scenario, a new triple containing this relationship is automatically inserted, with its initial confidence set to an intermediate value to balance exploration and reliability requirements.

[0043] The delete_triple operation (DELETE_TRIPLE) cleans up invalid rules. For example, if a hotel closes and "city X - accommodation → hotel Y" becomes redundant data, the system automatically removes the triplet to free up storage resources.

[0044] The confidence adjustment operation (MODIFY_SCORE) continuously optimizes the rule weights. For example, when a route is seasonally suspended, the system decays the confidence of the "city A-transportation mode→flight" triplet weekly, forming a progressive failure mechanism.

[0045] It's important to note that confidence weights are defined as continuous variables rather than discrete values. This design enables the system to accurately represent the gradient of rule reliability. For example, for the triple "travel expenses-association→visa fees," when cross-border business volume decreases, its confidence level slowly decays from a high level to a deletion threshold, rather than jumping to ineffectiveness.

[0046] For example, three types of operations work together in a business change scenario: when a supplier transitions from "hotel services" to "conference services," the system first adds a "supplier P - service type → conference venue" triple (ADD_TRIPLE), then decays the confidence level of the "supplier P - service type → accommodation" triple associated with the old business (MODIFY_SCORE), and finally deletes the old triple (DELETE_TRIPLE) when the confidence level falls below a threshold. This closed-loop operation chain enables smooth migration of business rules.

[0047] It's understandable that the innovation of this step lies in transforming the static rule base into a programmable state space. Specifically, the confidence adjustment operation supports fine-grained evolutionary control through a continuous variable mechanism (e.g., weight values ​​fluctuating in the interval [0, 1]), enabling the system to distinguish between "temporary changes" and "permanent failures"—the former simply requires lowering the confidence level and waiting for business recovery, while the latter triggers complete deletion.

[0048] Step 104: Based on the real-time document flow data and the preset historical case database data and audit result database data, the dynamic knowledge graph state space is dynamically evolved through the atomic operations in the reinforcement learning action space, and after each atomic operation is executed, the operation value return value generated after the execution of the atomic operation is calculated.

[0049] In one embodiment of the present application, based on real-time document flow data and preset historical case library data and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space, specifically including: receiving new financial document flow data in real time and extracting its associated entities and relationships; matching the triple structure of similar cases from the historical case library to generate candidate operation sequences; obtaining the triple features of violation cases from the audit result database and correcting the priority of the candidate operation sequence.

[0050] In one embodiment of the present application, the operation value return value generated after executing the atomic operation is calculated, specifically including: generating a historical value coefficient based on the improvement in the audit accuracy rate caused by the addition of new triples according to the historical case database; generating a risk control coefficient based on the change in the violation omission rate after deleting the triples verified in the audit result database; analyzing the coverage of the verification rules triggered by the confidence adjustment operation in the real-time document flow to generate a real-time response coefficient; and performing weighted summation of the historical value coefficient, the risk control coefficient and the real-time response coefficient to generate an operation value return value.

[0051] In this embodiment, the dynamic graph evolution mechanism is the core driver for autonomous rule evolution. It can be understood that the real-time document flow acts as the system's "nerve endings" for sensing the environment. When a new travel expense claim is received, the system immediately extracts key entities (such as "Reimbursement Person: Mr. Li" and "Consumption City: Paris") and their relationships (such as "Accommodation Fee → Hotel A"). Specifically, it first matches the triple structure of similar cases from the historical case database. For example, if the abnormal pattern of "sudden increase in hotel expenses" in a historical Paris travel report is retrieved, a candidate operation (ADD_TRIPLE) for the "new consumption amount verification rule" is automatically generated.

[0052] It should be noted that the audit results database plays a corrective role in this process. The database is used to obtain the priority of corrective operations based on the characteristics of violation cases: when the system detects a high frequency of "accommodation invoices without itinerary matching" in recent fictitious travel cases, it increases the weight of the "hotel invoice → itinerary" verification operation, so that the newly added amount verification rules and existing verification rules form a joint defense system. For example, a certain operation sequence may include: adding a hotel consumption limit rule (ADD_TRIPLE) → adjusting the confidence level of the associated itinerary (MODIFY_SCORE) → deleting expired traffic rules (DELETE_TRIPLE), and the execution order is dynamically adjusted by audit data feedback.

[0053] Specifically, the return value of the operation value is calculated immediately after each atomic operation. The multi-source feedback mechanism of this process: first, the operation benefits are statistically analyzed based on the historical case database, such as the improvement in the violation recognition rate of similar historical cases after the new consumption limit rules are added, and the historical value coefficient is generated; at the same time, the operation risks are verified through the audit result database, such as the change in the relevant missed detection rate after the old traffic rules are deleted, and the risk control coefficient is calculated; then, the rule coverage change (such as the triggering ratio of the new rule) is analyzed in combination with the real-time document flow to generate a real-time response coefficient. Finally, the three coefficients are weighted and fused to form a comprehensive return value - for example, when the new rules significantly increase the capture rate of high-consumption anomalies in the current document flow, but historical data suggests that normal business accommodation may be accidentally affected, the system will balance the positive and negative benefits and give an accurate valuation.

[0054] Understandably, this step achieves a closed loop of "perception-decision-evaluation" driven by data. For example, if a city hosts an international conference and hotel prices temporarily increase, the system won't directly impose new restrictions based on the sudden increase in spending. Instead, it will prioritize a flexible strategy of increasing the confidence level (MODIFY_SCORE) based on historical price fluctuations and real-time audit feedback to avoid frequent fluctuations in the rule base.

[0055] Step 105: Use the Monte Carlo tree search algorithm to predict the expected cumulative reward values ​​of executing different atomic operation sequences in the reinforcement learning action space, and prune the atomic operation sequences whose expected cumulative rewards are lower than the threshold based on the prediction results.

[0056] In one embodiment of the present application, a Monte Carlo tree search algorithm is used to predict the expected cumulative reward value of executing different atomic operation sequences in the reinforcement learning action space, specifically including: starting from the root node of the current dynamic knowledge graph state space, selecting an atomic operation expansion sub-node that has not been fully explored; simulating an operation sequence of a predefined number of steps through a random strategy to generate a leaf node state; and backtracking to calculate the cumulative reward value from the leaf node state to the root node.

[0057] In this embodiment, the intelligent decision-making mechanism solves the path optimization problem of rule evolution through forward-looking simulation. It can be understood that Monte Carlo Tree Search (MCTS) equips the system with a "time telescope," enabling it to foresee the business impact of multiple operations. Specifically, starting from the root node of the current knowledge graph state space, it prioritizes expanding subnodes of atomic operations that have not been fully explored. For example, when the system faces the branch choice of "adding a new electronic invoice rule" (ADD_TRIPLE) and "adjusting the confidence level of paper invoices" (MODIFY_SCORE), it prioritizes exploring operation paths with low recent execution frequency to balance exploration and utilization.

[0058] It should be noted that the expanded simulation phase uses a randomized strategy to execute a predefined sequence of operations. For example, starting from the newly added e-invoice rule node, the simulation's subsequent five steps might be: Add → Associated Enterprise Tax Number Verification Rule (ADD_TRIPLE) → Adjust the Confidence Level of Old-Format Invoices (MODIFY_SCORE) → Delete the Voided Invoice Code Library (DELETE_TRIPLE) → Add Digital Currency Payment Association (ADD_TRIPLE) → Adjust the Cross-Border Tax Rate Calculation Weight (MODIFY_SCORE). Each step is based on probabilistic sampling of the real-time document flow. For example, when simulating the deletion of a voided invoice code, the frequency of occurrence of the code in historical documents is randomly sampled to assess the impact.

[0059] Specifically, the backtracking calculation phase evaluates the cumulative return of leaf node states. Starting from the leaf node where the simulation terminates and working backwards to the current root node, the expected value of all operations along the path is accumulated. For example, if a leaf node indicates that "new cross-border tax rate rules have led to an increase in the misjudgment rate of downstream documents," this negative return will be used to reduce the value of each node along the path. This process pays special attention to the coupling effects of the operation chain—for example, if deleting an operation causes the newly added operation to fail, the value of the entire path will be significantly reduced.

[0060] Understandably, the pruning mechanism is the core innovation of this step. The system automatically removes paths whose cumulative returns fall below the average level of the same layer. For example, if the expected return of the "add electronic invoice → adjust the confidence level of paper invoices" path is only one-third of the "directly delete paper invoice rules" path, the inefficient branch is immediately pruned. This mechanism effectively avoids the common "rule iteration trap" in business scenarios. For example, when an airline changes its invoice format, it avoids the ineffective cycle of "retaining the old format → fine-tuning the confidence level → repeated revisions."

[0061] Step 106: Execute the atomic operation sequence optimized and screened by the Monte Carlo tree search algorithm to dynamically update the knowledge graph state space.

[0062] In this embodiment, the core execution phase of the main process in this step transforms the predicted and optimized operation sequence into actual knowledge graph evolution. It can be understood that this operation sequence is like a precisely calculated "construction blueprint." For example, the "add e-invoice rule → verify associated enterprise tax number → delete old invoice code" operation chain, which is retained after pruning in step 105, will be executed strictly in sequence. This ensures transactional integrity during execution—that is, all steps within a single operation sequence must either succeed or roll back to the pre-operation state, preventing graph state disruption caused by local failures.

[0063] It's important to note that each atomic operation modifies the knowledge graph topology in real time. For example, when performing an add_triple operation (ADD_TRIPLE), a node "electronic invoice - verification method → ​​blockchain verification" is created in the vector space, with the initial confidence set to an intermediate value. When performing a delete_triple operation (DELETE_TRIPLE), invalid nodes such as "paper invoice - version → 2018 standard" and their associated edges are completely removed. When performing a confidence adjustment operation (MODIFY_SCORE), transitional rules such as "cross-border payment - tax rate association → ordinary trade" are weighted attenuated.

[0064] Specifically, the dynamic update process adheres to the principle of non-transactional operations. For example, in a digital currency payment scenario, when the system executes the sequence "add virtual currency platform payment rules → adjust cash payment confidence → delete check payment rules," even if the second adjustment step fails due to document flow conflicts, the successful addition will be retained. A compensation mechanism will automatically delay the deletion until the confidence level decays to a certain level. This design ensures business continuity while avoiding the "all-or-nothing" risk of rule iteration.

[0065] Understandably, the updated knowledge graph immediately enters the verification-ready state. For example, once the invoice rule iteration is complete, newly received electronic invoices in the document flow will directly trigger the newly added blockchain verification rules, while paper invoices that have not yet been cleared will continue to participate in the calculation based on the attenuated confidence level, achieving seamless migration of business rules.

[0066] Step 107: Based on the updated dynamic knowledge graph state space, trigger the logic verification rules to perform compliance verification on the financial documents to be audited and generate audit results.

[0067] In one embodiment of the present application, based on the updated dynamic knowledge graph state space, logical verification rules are triggered to perform compliance verification on the financial documents to be reviewed, specifically including: based on the entity association relationship in the dynamic knowledge graph state space, the travel voucher documents, accommodation voucher documents and transportation voucher documents in the attached documents are linked in series to generate a logical verification chain containing spatiotemporal entities; the geographic location entities of each voucher document in the logical verification chain are traversed to verify the topological connection relationship between the city name and the latitude and longitude coordinates on the itinerary route; the check-in / check-out time entity of the accommodation voucher and the start and end time entity of the itinerary voucher are extracted to calculate the coverage overlap between the number of accommodation days and the itinerary period; the triplets with broken geographic track nodes and time coverage deviations exceeding the preset tolerance range are marked as evidence of violations.

[0068] In this embodiment, the cross-document semantic chain verification scheme is the core technology implementation of implicit violation identification. It can be understood that the logic verification rules are not static preset field matching, but dynamic reasoning chains generated in real time based on the knowledge graph. Specifically, a cross-voucher semantic verification chain is first constructed - for example, from the travel reimbursement scenario, the "departure A → destination B" in the itinerary, the "location B" in the hotel invoice, and the "destination B → return A" in the boarding pass are connected in series to form a geographical closed loop of "departure → arrival → accommodation → return".

[0069] It's important to note that geographic trajectory continuity verification is achieved through topological relationships. For example, if the itinerary shows "Beijing → Shanghai" and the hotel invoice shows "Shanghai Pudong," but the boarding pass lacks the "Shanghai → Beijing" entry and instead shows "Guangzhou → Beijing," the system automatically marks the "Shanghai → Beijing" segment as a broken node. The system also verifies for logical geolocation conflicts—for example, if the hotel location shows "Hangzhou Xihu District" but Hangzhou isn't included in the itinerary, a "fictitious accommodation" violation alert will be triggered. This topological verification overcomes the limitations of traditional rule engines, which can only verify explicit fields.

[0070] Specifically, time series dependency matching performs coverage overlap analysis. The check-in / check-out time entities on the accommodation voucher are compared with the start and end time entities on the itinerary to calculate the effective coverage interval. For example, if the itinerary period is "January 1st - January 5th," but the hotel invoice shows "January 4th check-in - January 8th check-out," the system automatically identifies the abnormal coverage deviation, indicating that "January 6th-8th is outside the itinerary." It is worth noting that this mechanism can also catch complex violations caused by artificially splitting itineraries. For example, if a consecutive stay is split into two invoices, "January 1st-3rd Hotel A" and "January 3rd-5th Hotel B," the system can detect the "single-day duplicate reimbursement" vulnerability by verifying the overlap between the "January 3rd check-in" on the Hotel B invoice and the "January 3rd check-out" on Hotel A.

[0071] As you can understand, violation evidence generation utilizes a triple tagging mechanism. Specifically, conflicting triples (e.g., <itinerary, missing, Shanghai→Beijing flight>) are linked to specific voucher location coordinates, creating a visual audit trail. For example, for a detected "time coverage discrepancy" violation, the departure timestamp is highlighted on the boarding pass scan, and the actual geofence is annotated on the hotel invoice at the location anomaly, ensuring traceability of audit results.

[0072] The above is an embodiment of the method proposed in this application. Based on the same inventive concept, this application embodiment also provides a financial document automatic review device, the structure of which is as follows: Figure 2 shown.

[0073] Figure 2 This is a schematic diagram of the internal structure of a financial document automation review device provided in the embodiment of this application. Figure 2 As shown, the equipment includes:

[0074] at least one processor 201;

[0075] and, a memory 202 communicatively coupled to the at least one processor;

[0076] The memory 202 stores instructions that can be executed by at least one processor, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to:

[0077] Receive a financial document to be reviewed and its associated attachment document submitted by a user, and extract a first entity set from a structured field of the financial document to be reviewed, and extract a second entity set from unstructured content of the associated attachment document;

[0078] Based on the first entity set and the second entity set, a dynamic knowledge graph state space is constructed; wherein the dynamic knowledge graph state space includes: entity vectors generated by mapping each entity through an entity embedding algorithm, and relationship vectors generated by mapping relationships between entities through a relationship encoding algorithm;

[0079] Define a reinforcement learning action space; the reinforcement learning action space includes three types of atomic operations for modifying the state space of the dynamic knowledge graph: adding a triplet, deleting a triplet, and adjusting the triplet confidence.

[0080] Based on real-time document flow data, preset historical case database data, and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space. After each atomic operation is executed, the operational value return generated by the execution of the atomic operation is calculated.

[0081] Using the Monte Carlo tree search algorithm, we predict the expected cumulative reward of executing different atomic operation sequences in the reinforcement learning action space, and prune atomic operation sequences whose expected cumulative reward is below a threshold based on the prediction results.

[0082] Execute the atomic operation sequence optimized and filtered by the Monte Carlo tree search algorithm, and dynamically update the knowledge graph state space;

[0083] Based on the updated dynamic knowledge graph state space, the logic verification rules are triggered to perform compliance verification on the financial documents to be audited and generate audit results.

[0084] Some embodiments of the present application provide corresponding Figure 1 A non-volatile computer storage medium stores computer executable instructions, wherein the computer executable instructions are configured as follows:

[0085] Receive a financial document to be reviewed and its associated attachment document submitted by a user, and extract a first entity set from a structured field of the financial document to be reviewed, and extract a second entity set from unstructured content of the associated attachment document;

[0086] Based on the first entity set and the second entity set, a dynamic knowledge graph state space is constructed; wherein the dynamic knowledge graph state space includes: entity vectors generated by mapping each entity through an entity embedding algorithm, and relationship vectors generated by mapping relationships between entities through a relationship encoding algorithm;

[0087] Define a reinforcement learning action space; the reinforcement learning action space includes three types of atomic operations for modifying the state space of the dynamic knowledge graph: adding a triplet, deleting a triplet, and adjusting the triplet confidence.

[0088] Based on real-time document flow data, preset historical case database data, and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space. After each atomic operation is executed, the operational value return generated by the execution of the atomic operation is calculated.

[0089] Using the Monte Carlo tree search algorithm, we predict the expected cumulative reward of executing different atomic operation sequences in the reinforcement learning action space, and prune atomic operation sequences whose expected cumulative reward is below a threshold based on the prediction results.

[0090] Execute the atomic operation sequence optimized and filtered by the Monte Carlo tree search algorithm, and dynamically update the knowledge graph state space;

[0091] Based on the updated dynamic knowledge graph state space, the logic verification rules are triggered to perform compliance verification on the financial documents to be audited and generate audit results.

[0092] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences from the other embodiments. In particular, the IoT device and media embodiments are generally similar to the method embodiments, so their description is relatively simple. For relevant portions, refer to the description of the method embodiments.

[0093] The system and medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and medium also have similar beneficial technical effects to their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be repeated here.

[0094] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0095] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0098] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0099] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0100] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0101] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0102] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for automated auditing of financial documents, characterized in that: The method comprises: Receive a financial document to be reviewed and its associated attachment document submitted by a user, and extract a first entity set from a structured field of the financial document to be reviewed, and extract a second entity set from unstructured content of the associated attachment document; Based on the first entity set and the second entity set, a dynamic knowledge graph state space is constructed; wherein the dynamic knowledge graph state space includes: entity vectors generated by mapping each entity through an entity embedding algorithm, and relationship vectors generated by mapping relationships between entities through a relationship encoding algorithm; Defining a reinforcement learning action space; wherein the reinforcement learning action space includes three types of atomic operations for modifying the dynamic knowledge graph state space: adding a triplet operation, deleting a triplet operation, and adjusting a triplet confidence operation; Based on real-time document flow data and preset historical case database data and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space, and after each atomic operation is executed, the operation value return value generated by executing the atomic operation is calculated; Using a Monte Carlo tree search algorithm, the expected cumulative reward values ​​of executing different atomic operation sequences are predicted in the reinforcement learning action space, and based on the prediction results, the atomic operation sequences whose expected cumulative rewards are lower than a threshold are pruned; Execute the atomic operation sequence optimized and screened by the Monte Carlo tree search algorithm to dynamically update the knowledge graph state space; Based on the updated dynamic knowledge graph state space, the logic verification rules are triggered to perform compliance verification on the financial documents to be audited and generate audit results.

2. A method for automated auditing of financial documents according to claim 1, characterized in that: Extracting a second entity set from the unstructured content of the associated attachment document specifically includes: Inputting the image or text content of the associated attachment document into a multimodal entity extraction model; Processing text information in the image or text content based on a pre-trained language model layer in the multimodal entity extraction model to generate context-aware text features; Capturing the sequential dependencies of the text features based on the bidirectional long short-term memory network layer in the multimodal entity extraction model; The sequence dependency is decoded based on the conditional random field layer in the multimodal entity extraction model to output the second entity set.

3. The method for automated auditing of financial documents according to claim 1, characterized in that: Based on the first entity set and the second entity set, a dynamic knowledge graph state space is constructed, specifically including: Mapping each entity in the first entity set and the second entity set into an entity vector of fixed dimension using the TransE algorithm; The BERT-wwm-ext model is used to vector-encode the relationship labels between entities and generate relationship vectors; Constructing an initial triple set based on the entity vector and the relationship vector; The time series dependencies of entities in historical documents are analyzed through the LSTM module, and the confidence weights of the initial triple set are dynamically adjusted.

4. The method for automated auditing of financial documents according to claim 1, characterized in that: Define the reinforcement learning action space, including: Setting a new triple operation for inserting a new triple containing a head entity, a relationship, and a tail entity into the dynamic knowledge graph state space; Setting a triple deletion operation to remove redundant triples whose confidence level is lower than a preset threshold in the dynamic knowledge graph state space; The triplet confidence adjustment operation is set to dynamically modify the confidence weight of the triplet according to the real-time data stream, where the confidence weight is a continuous variable.

5. The method for automated auditing of financial documents according to claim 1, characterized in that: Based on real-time document flow data, preset historical case database data, and audit result database data, the dynamic knowledge graph state space is dynamically evolved through atomic operations in the reinforcement learning action space, specifically including: Receive new financial document flow data in real time and extract its related entities and relationships; Match the triple structures of similar cases from the historical case database to generate candidate operation sequences; The triplet features of the violation cases are obtained from the audit result database, and the priorities of the candidate operation sequences are modified.

6. The method for automated auditing of financial documents according to claim 1, characterized in that: Calculate the return value of the operation value generated after executing the atomic operation, specifically including: Generate a historical value coefficient based on the improvement in audit accuracy caused by the addition of new triples to the historical case database; Based on the audit result database, the change in the violation omission rate after deleting the triples is verified and the risk control coefficient is generated; Analyze the coverage of validation rules triggered by confidence adjustment operations in real-time document flows and generate real-time response coefficients; The historical value coefficient, risk control coefficient and real-time response coefficient are weighted and summed to generate the operation value return value.

7. The method for automated auditing of financial documents according to claim 1, characterized in that: Using the Monte Carlo tree search algorithm, the expected cumulative reward value of executing different atomic operation sequences in the reinforcement learning action space is predicted, specifically including: Starting from the root node of the current dynamic knowledge graph state space, select the under-explored atomic operation to expand the child node; Generate leaf node states by simulating a sequence of operations with a predefined number of steps through random strategies; Backtrack and calculate the cumulative return value from the leaf node state to the root node.

8. The method for automated auditing of financial documents according to claim 1, characterized in that: Based on the updated dynamic knowledge graph state space, the logic verification rules are triggered to perform compliance verification on the financial documents to be reviewed, specifically including: Based on the entity association relationship in the state space of the dynamic knowledge graph, the itinerary voucher documents, accommodation voucher documents, and transportation voucher documents in the attached documents are serially associated to generate a logical verification chain containing spatiotemporal entities; Traversing the geographic location entities of each credential document in the logical verification chain to verify the topological connection relationship between the city name and the latitude and longitude coordinates on the itinerary; Extract the check-in / check-out time entity of the accommodation voucher and the start and end time entity of the itinerary voucher, and calculate the coverage overlap between the number of accommodation days and the itinerary period; Triplet with broken nodes in geographic trajectories and time coverage deviation exceeding the preset tolerance range are marked as violation evidence.

9. An automated financial document auditing device, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method for automated review of financial documents as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: When the computer-executable instructions are executed, a method for automated review of financial documents as described in any one of claims 1 to 8 is implemented.

Images