Account auditing method and device for enterprise multi-source transaction data, and electronic equipment

Abstract

This invention discloses a method, apparatus, and electronic device for auditing accounts based on multi-source transaction data of enterprises, relating to the fields of financial technology or other related technical fields. The method includes: injecting accounting domain knowledge triples into a panoramic transaction graph and performing transaction association reasoning to obtain a knowledge-enhanced panoramic transaction graph; extracting structured semantic features of related transaction groups based on the panoramic transaction graph to obtain semantic feature codes; inputting the semantic feature codes into a large language model adjusted by financial domain corpus instructions; using the large language model to infer accounting subjects and generate standard accounting entries; verifying the standard accounting entries to obtain normal voucher data and abnormal voucher data; performing differentiated auditing and transaction posting processing to obtain final accounting vouchers and an abnormal account queue. This invention solves the technical problems of low efficiency and large differences in the quality of generated accounting vouchers in the processing of multi-source transaction data of enterprises in related technologies.

Description

Technical Field

[0001] This invention relates to the field of financial technology or other related fields, and more specifically, to a method, apparatus, and electronic device for auditing multi-source transaction data of enterprises. Background Technology

[0002] As enterprises deepen their digital transformation, the business systems of large enterprise groups (such as large multinational corporations) are becoming increasingly complex, and transaction data is characterized by multiple sources, heterogeneity, and high frequency. Data is usually scattered across multiple independent business modules such as direct bank-enterprise connections, internal settlement, bill management, credit management, and fund collection. In order to generate accurate financial vouchers and meet accounting standards, enterprises need to perform cross-module transaction data correlation, business substance understanding, and automatic inference of accounting subjects.

[0003] In related technologies, enterprise-level accounting automation primarily relies on manual experience for correlation and voucher preparation. Accountants log into different business systems to query transaction logs from modules such as bank-enterprise direct connection, settlement, and bills. Then, relying on their professional experience and memory, they manually compare and correlate transaction records belonging to the same economic transaction but scattered across different modules, finally manually preparing accounting entries. However, this approach has significant drawbacks: extremely low efficiency, limited daily processing capacity, and heavy reliance on personal experience, leading to inconsistent voucher quality. For large groups, this can easily result in missed, duplicate, or incorrect correlations, causing financial data distortion.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a method, apparatus, and electronic device for auditing accounts based on multi-source transaction data of an enterprise, in order to at least solve the technical problems of low efficiency and large differences in the quality of accounting vouchers generated during the processing of multi-source transaction data of an enterprise in related technologies.

[0006] To achieve the above objectives, according to one aspect of this application, a method for auditing accounts based on multi-source transaction data of an enterprise is provided, comprising: acquiring a pre-constructed panoramic transaction map, injecting accounting domain knowledge triples into the panoramic transaction map, and performing transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map, wherein each business entity node in the panoramic transaction map corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information; based on the panoramic transaction map, extracting structured semantic features of associated transaction groups in the multi-source transaction data of the enterprise to obtain semantic feature codes; inputting the semantic feature codes into a large language model adjusted by financial domain corpus instructions, using the large language model to infer accounting subjects and generate standard accounting entries; verifying the standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification; performing differentiated auditing and transaction posting processing on the normal voucher data and the abnormal voucher data to obtain the final posted accounting voucher and a queue of abnormal accounts awaiting manual auditing.

[0007] Optionally, before obtaining the pre-built panoramic transaction map, the process further includes: performing node abstraction on the business entities of each business module, and performing multi-layer progressive alignment on each business entity node to obtain a unified entity identifier node; modeling the transaction records in each business module as directed edges to obtain a subgraph associated with each business module, wherein the edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID; using the unified entity identifier node as the key, merging the subgraphs of each business module into a panoramic transaction map containing records of all business modules, wherein all transaction events of the same business entity in different business modules in the panoramic transaction map converge around the unified entity identifier node.

[0008] Optionally, the steps of abstracting the business entities of each business module into nodes and performing multi-layer progressive alignment processing on each business entity node to obtain a unified entity identifier node include: abstracting the business entities in each business module into multiple types of business entity nodes, wherein the multiple types of business entity nodes include: member units, financial accounts, bills, contracts, and credit lines; performing multi-layer progressive alignment processing on the same business entity node in different business modules, wherein the multi-layer progressive alignment processing includes: a first layer, matching using the unified social credit code, financial account number, and taxpayer identification number of each business entity node; a second layer, performing weighted fuzzy matching using multiple categories of enterprise attributes for entities without identifiers; a third layer, performing transaction relationship propagation alignment using the connection relationship of aligned business entity nodes in the panoramic transaction map; and a fourth layer, calling a large language model for auxiliary semantic judgment for abnormal business entity nodes that cannot be aligned in the first, second, and third layers.

[0009] Optionally, after performing multi-layer progressive alignment processing on the same business entity node in different business modules, the method further includes: performing alignment error propagation suppression processing on the business entity node after completing multi-layer progressive alignment processing; performing confidence decay processing during the alignment error propagation suppression processing, and performing hop-by-hop decay processing on the propagation path of the business entity node; performing propagation chain length limitation processing during the alignment error propagation suppression processing, and limiting the propagation depth of the business entity node by parameters; and performing sampling feedback closed-loop processing during the alignment error propagation suppression processing, sampling business entity nodes that have completed alignment processing according to a preset ratio, sending the sampling results to the reviewer, and updating the node feature vector and feature parameters of the business entity node based on the confirmation results given by the reviewer.

[0010] Optionally, the steps of injecting accounting domain knowledge triples into the panoramic transaction graph and performing transaction association reasoning to obtain a knowledge-enhanced panoramic transaction graph include: extracting static base knowledge triples composed of head entities, relationships, and tail entities from unstructured accounting standard text by matching trigger words through rule templates; statistically analyzing the co-occurrence frequency and conditional probability of various voucher items from a historical audited voucher database, and extracting voucher data with frequencies and conditional probabilities greater than preset thresholds to generate dynamically extended knowledge triples; forming the accounting domain knowledge triples based on the static base knowledge triples and the dynamically extended knowledge triples; injecting the accounting domain knowledge triples as prior edges into the panoramic transaction graph; performing transaction association reasoning tasks and association confidence scoring tasks on the panoramic transaction graph, wherein the transaction association reasoning tasks include: transaction event association, transaction process completion, and transaction relationship classification, and the association confidence scoring tasks include: structural similarity scoring, attribute consistency scoring, and semantic relevance scoring; and marking the association relationships between transaction events, confidence scores, and completed transaction process chains in the panoramic transaction graph.

[0011] Optionally, based on the panoramic transaction map, the step of extracting structured semantic features of related transaction groups in the enterprise's multi-source transaction data to obtain semantic feature encoding includes: selecting a set of transaction events with confidence scores higher than a preset confidence threshold from the panoramic transaction map to form related transaction groups; extracting multimodal feature encoding for each related transaction group, wherein the multimodal feature encoding includes: transaction type, relationship pattern between business entity nodes, direction and amount of money flow, transaction summary text, and time sequence pattern; and encoding the multimodal features into structured semantic feature vectors using an embedding model to obtain semantic feature encoding.

[0012] Optionally, the step of inputting the semantic feature encoding into a large language model adjusted by instructions from a financial domain corpus, and using the large language model to infer accounting subjects and generate standard accounting entries includes: converting the semantic feature encoding into query embeddings; performing an approximate nearest neighbor search in a vector database to obtain the top K historical typical cases with the highest similarity; injecting the matched accounting structure rules, the retrieved top K historical typical cases, and the currently pending related-party transaction group into the large language model; and using the large language model to infer accounting subjects and generate standard accounting entries that conform to the accounting structure rules.

[0013] Optionally, the step of verifying the standard accounting entries to obtain verified normal voucher data and verified abnormal voucher data includes: performing basic four-dimensional verification on the voucher data in the standard accounting entries, wherein the basic four-dimensional verification includes: checking whether the total debit and credit amounts are balanced, whether the relevant accounts are valid in the current accounting period, whether the amounts are within a reasonable range, and whether the business logic complies with accounting standards; performing case consistency verification on the voucher data in the standard accounting entries, comparing the output accounts, amount levels, and debit / credit directions with the top K historical typical cases retrieved in a structured manner, and calculating a consistency score; if the consistency score is lower than a preset score threshold or the basic four-dimensional verification fails, marking the voucher data as abnormal voucher data; if the consistency score is greater than or equal to the preset score threshold and the basic four-dimensional verification passes, marking the voucher data as normal voucher data.

[0014] Optionally, the steps of performing differentiated auditing and transaction posting processing on the normal voucher data and the abnormal voucher data to obtain the final posted accounting voucher and the abnormal account queue awaiting manual review include: for the normal voucher data, obtaining the voucher type; if the voucher type is a regular business transaction, performing automatic posting processing to obtain the final accounting voucher; if the voucher type is a specific business transaction, performing auditing review processing to obtain the final accounting voucher; for the abnormal voucher data that fails verification, generating an auditing task based on the voucher data of the abnormal voucher data, and entering the auditing task into the auditing queue, wherein the auditing task carries an inference trajectory including the transaction source, graph network association path, standard accounting entries output by the large language model, and verification results.

[0015] According to another aspect of the present invention, an account auditing device for enterprise multi-source transaction data is also provided, comprising: a panoramic transaction map acquisition unit, configured to acquire a pre-constructed panoramic transaction map, inject accounting domain knowledge triples into the panoramic transaction map, and perform transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map, wherein each business entity node in the panoramic transaction map corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information; a semantic feature extraction unit, configured to extract structured semantic features of associated transaction groups in the enterprise multi-source transaction data based on the panoramic transaction map to obtain semantic feature codes; an accounting subject inference unit, configured to input the semantic feature codes into a large language model adjusted by financial domain corpus instructions, and use the large language model to infer accounting subjects and generate standard accounting entries; a verification unit, configured to verify the standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification; and an auditing unit, configured to perform differentiated auditing and transaction posting processing on the normal voucher data and the abnormal voucher data to obtain the final posted accounting voucher and a queue of abnormal accounts awaiting manual auditing.

[0016] Optionally, the accounting auditing device for enterprise multi-source transaction data further includes: a node abstraction unit, used to perform node abstraction processing on the business entities of each business module before obtaining the pre-built panoramic transaction map, and to perform multi-layer progressive alignment processing on each business entity node to obtain a unified entity identifier node; a directed edge modeling unit, used to model the transaction records in each business module as directed edges to obtain a subgraph associated with each business module, wherein the edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID; and a panoramic transaction map construction unit, used to merge the subgraphs of each business module into a panoramic transaction map containing records of all business modules, using the unified entity identifier node as the key, wherein all transaction events of the same business entity in different business modules in the panoramic transaction map converge around the unified entity identifier node.

[0017] Optionally, the node abstraction unit includes: a node abstraction module, used to abstract the business entities in each business module into multiple types of business entity nodes, wherein the multiple types of business entity nodes include: member units, financial accounts, bills, contracts, and credit lines; and a node alignment module, used to perform multi-layer progressive alignment processing on the same business entity node in different business modules, wherein the multi-layer progressive alignment processing includes: a first layer, matching using the unified social credit code, financial account number, and taxpayer identification number of each business entity node; a second layer, performing weighted fuzzy matching using multiple categories of enterprise attributes for unidentified entities; a third layer, performing transaction relationship propagation alignment using the connection relationship of aligned business entity nodes in the panoramic transaction map; and a fourth layer, calling a large language model for auxiliary semantic judgment for abnormal business entity nodes that cannot be aligned in the first, second, and third layers.

[0018] Optionally, the accounting auditing device for enterprise multi-source transaction data further includes: a propagation suppression module, used to perform alignment error propagation suppression processing on the business entity nodes after performing multi-level progressive alignment processing on the same business entity nodes in different business modules; a confidence decay module, used to perform confidence decay processing during the alignment error propagation suppression process, and perform hop-by-hop decay processing on the propagation path of the business entity nodes; a propagation chain length limiting module, used to perform propagation chain length limiting processing during the alignment error propagation suppression process, and perform parameter limiting on the propagation depth of the business entity nodes; and a sampling inspection feedback closed-loop processing module, used to perform sampling inspection feedback closed-loop processing during the alignment error propagation suppression process, sampling business entity nodes that have completed alignment processing according to a preset ratio, sending the sampling inspection results to the auditing station, and updating the node feature vector and feature parameters of the business entity nodes based on the confirmation results given by the auditing station.

[0019] Optionally, the panoramic transaction map acquisition unit includes: a static base knowledge triple extraction module, used to extract static base knowledge triples composed of head entities, relations, and tail entities from unstructured accounting standard text by matching trigger words through rule templates; a dynamic extended knowledge triple extraction module, used to statistically analyze the co-occurrence frequency and conditional probability of various voucher accounts from the historical audited voucher database, and extract voucher data with a frequency and conditional probability greater than a preset threshold to generate dynamic extended knowledge triples; and an accounting domain knowledge triple generation module, used to form... The accounting domain knowledge triples; the knowledge triple injection module, used to inject the accounting domain knowledge triples as prior edges into the panoramic transaction graph; performing transaction association reasoning tasks and association confidence scoring tasks on the panoramic transaction graph, wherein the transaction association reasoning tasks include: transaction event association, transaction process completion, and transaction relationship classification, and the association confidence scoring tasks include: structural similarity scoring, attribute consistency scoring, and semantic relevance scoring; the annotation module, used to annotate the association relationships between transaction events, confidence scores, and completed transaction process chains in the panoramic transaction graph.

[0020] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the accounting auditing method for enterprise multi-source transaction data described above.

[0021] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the accounting auditing method for enterprise multi-source transaction data described above.

[0022] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the accounting auditing method for multi-source transaction data of an enterprise as described above.

[0023] In this disclosure, a pre-constructed panoramic transaction graph is obtained, and accounting domain knowledge triples are injected into the panoramic transaction graph. Transaction association reasoning is then performed to obtain a knowledge-enhanced panoramic transaction graph. In this graph, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. Based on the panoramic transaction graph, structured semantic features of related transaction groups in multi-source enterprise transaction data are extracted to obtain semantic feature codes. These semantic feature codes are input into a large language model adjusted by financial domain corpus instructions. The large language model is used to infer accounting subjects and generate standard accounting entries. The standard accounting entries are verified to obtain verified normal voucher data and verified abnormal voucher data. Differential auditing and transaction posting processing are performed on the normal and abnormal voucher data to obtain the final posted accounting vouchers and a queue of abnormal accounts awaiting manual review.

[0024] Based on the aforementioned publicly available information, automatic association of transaction records across business modules can be achieved through panoramic transaction maps and progressive entity alignment processing. This reduces manual association costs, improves data integrity, and enhances transaction data processing efficiency. Furthermore, by associating the structured semantic features of transaction groups, a deeper understanding of the transaction's essence can be gained, reducing voucher quality issues caused by association errors in complex scenarios. Additionally, large language models can be used to infer accurate accounting subjects, reducing the number of misjudged subjects in complex scenarios and long-tail transactions. Finally, automatic voucher posting and abnormal voucher processing are performed through accounting entries, achieving a high proportion of automatic generation and accurate posting for routine business transactions. This significantly reduces manual intervention and improves accounting processing efficiency, thereby solving the technical problems of low efficiency and significant differences in the quality of generated accounting vouchers when processing multi-source transaction data in related technologies. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0026] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing an accounting auditing method for multi-source transaction data of an enterprise is shown.

[0027] Figure 2 This is a flowchart of an optional accounting auditing method for multi-source transaction data of an enterprise according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of an optional account auditing device for multi-source transaction data of an enterprise according to an embodiment of the present invention;

[0029] Figure 4This is a structural block diagram of an electronic device that performs an accounting auditing method for multi-source transaction data of an enterprise, according to an embodiment of this application. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0033] Holistic Transaction Graph (HTG) is a comprehensive transaction relationship graph built on a unified enterprise entity identifier and with transaction events as edges. It aggregates transaction data from multiple business modules, including direct bank-enterprise connections, internal settlement, bills, and credit.

[0034] The Knowledge-Enhanced Relational Graph Convolutional Network (KE-RGCN) is a relational graph convolutional network that injects knowledge triples from the accounting domain as prior edges to perform relational reasoning on a panoramic transaction graph.

[0035] The four-layer progressive entity alignment, or P-EA for short, is a method that maps heterogeneous entities from various modules to a unified entity ID through four progressive layers: precise matching with gold markers, multi-attribute weighted fuzzy matching, alignment based on transaction relationships, and alignment assisted by a large language model.

[0036] Semantic-Driven Account Inference (SDAI) encodes related transaction groups into structured semantic features, uses a large language model to understand the substance of the transaction business, dynamically infers the most appropriate accounting subject, and then performs four-dimensional verification to divert the data to differentiated audits.

[0037] The Knowledge Triple is a standardized representation of accounting knowledge, formatted as Head Entity-Relationship-Tail Entity, used for structured storage of logical relationships mined from accounting standards and historical vouchers.

[0038] Differentiated Audit Flow is a mechanism that automatically selects audit strategies based on voucher type, including automatic posting of routine transactions, marking special transactions for review, and sending abnormal transactions to a manual queue.

[0039] Transaction Relationship Propagation (TRP) is a third-layer method for entity alignment. It leverages the connections between aligned entities in the transaction graph to propagate identity information, thereby expanding the coverage of entity alignment.

[0040] Large Language Model (LLM) is used in entity alignment to handle semantic judgment of difficult entities, in account inference to understand the substance of transactions and output standard accounting entries, and in knowledge extraction to assist in extracting triples from unstructured text.

[0041] Retrieval-Augmented Generation (RAG) is a technique that injects case library search results into the LLM context. It guides the LLM to make subject inferences by searching for similar cases and supports the self-updating of the case library.

[0042] It should be noted that the accounting auditing method and apparatus for enterprise multi-source transaction data disclosed herein can be used in the fintech field for panoramic association and intelligent accounting auditing of enterprise multi-source transaction data based on knowledge-enhanced graph networks and semantic reasoning. It can also be used in any field other than fintech for panoramic association and intelligent accounting auditing of enterprise multi-source transaction data based on knowledge-enhanced graph networks and semantic reasoning. This disclosure does not limit the application field of the accounting auditing method and apparatus for enterprise multi-source transaction data.

[0043] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data, etc.) collected in this public disclosure are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data are handled with corresponding access points, allowing users to choose to authorize or refuse. For example, this system has interfaces with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0044] It should be noted that in this disclosure, customer information is collected and analyzed, and users are provided with corresponding operation entry points to choose whether to agree to or reject the automated decision results; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0045] The following embodiments of the present invention can be applied to various systems / applications / devices for auditing multi-source transaction data of enterprises. The present invention is applicable to fintech scenarios and intelligent accounting scenarios, such as automated processing of multi-source heterogeneous transaction data of enterprise groups, real-time accounting voucher generation in direct bank-enterprise connection scenarios, semantic understanding and account inference of complex supply chain finance businesses, and compliant accounting systems that need to frequently adapt to new accounting standards.

[0046] This invention achieves high-precision automatic association of transaction records across business modules through a panoramic transaction map and a progressive entity alignment algorithm, reducing manual association costs and improving data integrity. Furthermore, it utilizes knowledge-enhanced graph networks and semantic reasoning technology to deeply understand the essence of transaction business, effectively solving the problem of misclassification of accounts caused by complex scenarios and long-tail transactions that traditional rule engines cannot handle. Through a self-updating mechanism driven by both accounting standards and typical cases, it achieves agile adaptation to accounting policies, quickly responding to standard changes without modifying the underlying code one by one. Ultimately, it achieves a high proportion of automatic generation and accurate posting of routine business transactions, significantly reducing the rate of manual intervention and improving the efficiency, accuracy, and compliance of accounting processing.

[0047] The present invention will now be described in detail with reference to various embodiments.

[0048] Example 1

[0049] According to an embodiment of the present invention, an embodiment of an accounting auditing method for multi-source transaction data of an enterprise is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0050] The accounting auditing method for multi-source transaction data of enterprises provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing an accounting auditing method for multi-source transaction data in an enterprise is shown. Figure 1 As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1 The processor 102 (which may include, but is not limited to, a microprocessor MCU (Microcontroller Unit) or a programmable gate array (FPGA)) is shown as 102a, 102b, ..., 102n. It also includes a memory 104 for storing data and a transmission device 106 for communication functions. In addition, it may include: a display, an input / output interface (I / O interface), a Universal Serial Bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0051] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0052] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the accounting audit method for enterprise multi-source transaction data in this embodiment of the application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-mentioned accounting audit method for enterprise multi-source transaction data. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0053] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0054] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0055] Under the aforementioned operating environment, this application provides the following: Figure 2 The method shown is for auditing accounts for multi-source transaction data of enterprises. Figure 2 This is a flowchart of an optional accounting auditing method for multi-source transaction data of an enterprise according to an embodiment of the present invention, such as... Figure 2 As shown, the method includes the following steps S201 to S205. The present invention will be described below in conjunction with each implementation step.

[0056] Optionally, before obtaining the pre-built panoramic transaction map, the process further includes: performing node abstraction on the business entities of each business module, and performing multi-layer progressive alignment on each business entity node to obtain a unified entity identifier node; modeling the transaction records in each business module as directed edges to obtain a subgraph relating each business module, wherein the edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID; merging the subgraphs of each business module into a panoramic transaction map containing records of all business modules, wherein all transaction events of the same business entity in different business modules in the panoramic transaction map converge around the unified entity identifier node.

[0057] This embodiment provides a preprocessing and fusion method for constructing a panoramic transaction graph (HTG). Specifically, the business entities of each business module are first abstracted into nodes, and the business entity nodes are then progressively aligned in multiple layers to obtain unified entity identifier nodes, which helps to eliminate the heterogeneity of entity identifiers between different business modules. Further, the transaction records in each business module are modeled as directed edges to obtain a subgraph connecting each business module. The edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID. By retaining these detailed edge attributes, it is helpful to accurately capture the time sequence of transactions, fund flow, and business source in the subsequent graph reasoning stage, thereby improving the granularity of association reasoning.

[0058] It should be noted that when constructing the panoramic transaction map, this embodiment uses the unified entity identifier node as the key to merge the sub-graphs of each business module into a panoramic transaction map containing records of all business modules. In the panoramic transaction map, all transaction events of the same business entity in different business modules converge around the unified entity identifier node. This entity-centric convergence method helps to form a complete transaction profile of the business entity and facilitates the identification of complex transaction links across modules.

[0059] Optionally, the steps of abstracting the business entities of each business module into nodes and performing multi-layer progressive alignment processing on each business entity node to obtain a unified entity identifier node include: abstracting the business entities in each business module into multiple types of business entity nodes, wherein the multiple types of business entity nodes include: member units, financial accounts, bills, contracts, and credit lines; performing multi-layer progressive alignment processing on the same business entity node in different business modules, wherein the multi-layer progressive alignment processing includes: first layer, matching using the unified social credit code, financial account number, and taxpayer identification number of each business entity node; second layer, for unidentified entities, performing weighted fuzzy matching using multiple types of enterprise attributes; third layer, performing transaction relationship propagation alignment using the connection relationship of aligned business entity nodes in the panoramic transaction map; fourth layer, for abnormal business entity nodes that cannot be aligned in the first, second, and third layers, calling a large language model for auxiliary semantic judgment.

[0060] The first layer of the multi-layered progressive alignment process uses the unified social credit code, financial account number, and taxpayer identification number of each business entity node for matching. This high-confidence unique identifier enables precise matching, quickly resolving alignment issues for most standard entities and improving processing efficiency. The second layer uses weighted fuzzy matching of unidentified entities based on multiple enterprise attributes. This layer calculates weighted similarity based on attributes such as enterprise name, amount, transaction time, summary, region, and contract number, helping to handle cases with missing or non-standard identifiers and expanding the scope of entity alignment. The third layer utilizes the connection relationships of aligned business entity nodes in the panoramic transaction graph to perform transaction relationship propagation alignment. By using the topological relationships in the graph structure to propagate identity information, it helps guide the alignment of unknown entities using known high-confidence entities, especially in scenarios with indirect relationships, further expanding the alignment results. Furthermore, for abnormal business entity nodes that cannot be aligned in the previous layers, the large language model is invoked for auxiliary semantic judgment. This layer serves as a fallback mechanism, utilizing the semantic understanding capabilities of the large language model to handle difficult entities. This helps to solve the alignment problems of low-frequency or complex entities that are difficult to cover by traditional rules, thereby improving the overall alignment coverage.

[0061] This embodiment, through the aforementioned multi-layer progressive entity alignment and panoramic view construction method, can effectively integrate multi-source heterogeneous transaction data, establish a unified enterprise transaction view, which helps the subsequent Knowledge Augmented Graph Network (KE-RGCN) to perform transaction association reasoning more accurately, and provides high-quality structured input for semantic-driven account inference, thereby improving the automation level and accuracy of accounting voucher generation.

[0062] To prevent the cascading amplification effect caused by initial alignment errors, this embodiment introduces a suppression mechanism to control the spread of errors during graph propagation, which helps improve the accuracy and reliability of entity identification in the panoramic transaction graph. Optionally, after performing multi-layer progressive alignment processing on the same business entity node in different business modules, the method further includes: performing alignment error propagation suppression processing on the business entity nodes after completing multi-layer progressive alignment processing; performing confidence decay processing during the alignment error propagation suppression processing to perform hop-by-hop decay processing on the propagation path of the business entity node; performing propagation chain length limitation processing during the alignment error propagation suppression processing to limit the propagation depth of the business entity node; and performing sampling feedback closed-loop processing during the alignment error propagation suppression processing, sampling business entity nodes that have completed alignment processing according to a preset ratio, sending the sampling results to the reviewer, and updating the node feature vector and feature parameters of the business entity node based on the confirmation results given by the reviewer.

[0063] During the alignment error propagation suppression process, confidence decay is performed, which decays the propagation path of the business entity nodes hop by hop. Specifically, this can be done according to the formula... Here, α is the decay factor, which can be selected by the user, for example, 0.7; d is the number of propagation hops. For example, when the calculated confidence of C(v) is lower than a preset threshold (such as 0.3), the entity node will not participate in subsequent propagation calculations. This hop-by-hop decay mechanism helps reduce the uncertainty caused by long-distance associations, avoids early errors being amplified indefinitely, and thus maintains the stability of graph reasoning. During the alignment error propagation suppression process, the propagation chain length is limited, and the propagation depth of business entity nodes is parameterized. Specifically, the maximum number of iteration rounds of the Transaction Relationship Propagation (TRP) algorithm (such as 3 rounds) and the maximum depth of a single chain (such as 5 hops) can be set. By limiting the depth and breadth of propagation, it helps to control computational complexity, prevent noise interference introduced by long-path propagation, and ensure that the alignment process converges within a controllable time window.

[0064] Furthermore, during the alignment error propagation suppression process, a sampling feedback loop is implemented. A predetermined percentage (e.g., 5%) of the aligned business entity nodes is sampled, and the results are sent to the reviewer. Based on the reviewer's confirmation, the node feature vectors and feature parameters of the business entity nodes can be updated. This human-machine collaborative feedback mechanism helps continuously optimize the weight parameters in multi-attribute weighted fuzzy matching (e.g., refitting through logistic regression), forming a data-driven model iteration loop, which helps continuously improve the accuracy and robustness of subsequent entity alignment. By suppressing error propagation and introducing human feedback, the risk of error accumulation in large-scale automated processing can be reduced, providing high-quality structured input to the Knowledge Augmented Graph Convolutional Network (KE-RGCN), thereby improving the overall accuracy of the accounting engine.

[0065] Step S201: Obtain the pre-constructed panoramic transaction graph, inject accounting domain knowledge triples into the panoramic transaction graph, and perform transaction association reasoning to obtain the knowledge-enhanced panoramic transaction graph. In the panoramic transaction graph, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information.

[0066] Specifically, this embodiment can use the Knowledge Enhanced Relationship Graph Convolutional Network (KE-RGCN) to implement related transaction reasoning. Optionally, the steps of injecting accounting domain knowledge triples into the panoramic transaction graph and performing transaction association reasoning to obtain a knowledge-enhanced panoramic transaction graph include: extracting static base knowledge triples composed of head entities, relationships, and tail entities from unstructured accounting standard texts by matching trigger words through rule templates; statistically analyzing the co-occurrence frequency and conditional probability of various voucher items from a historical audited voucher database, and extracting voucher data with frequencies and conditional probabilities greater than preset thresholds to generate dynamically extended knowledge triples; forming accounting domain knowledge triples based on the static base knowledge triples and dynamically extended knowledge triples; injecting the accounting domain knowledge triples as prior edges into the panoramic transaction graph; performing transaction association reasoning tasks and association confidence scoring tasks on the panoramic transaction graph, wherein the transaction association reasoning tasks include: transaction event association, transaction process completion, and transaction relationship classification, and the association confidence scoring tasks include: structural similarity scoring, attribute consistency scoring, and semantic relevance scoring; and marking the association relationships between transaction events, confidence scores, and completed transaction process chains in the panoramic transaction graph.

[0067] From unstructured accounting standards texts, static base knowledge triples consisting of head entities, relations, and tail entities are extracted by matching trigger words using rule templates. Specifically, pre-set trigger word templates (such as accompanying, causing, corresponding, etc.) can be used for initial matching of standard provisions, combined with syntactic dependency analysis for disambiguation, to extract static base knowledge triples. This helps to structure authoritative accounting standards and provide standard logical constraints for reasoning. Furthermore, the co-occurrence frequency and conditional probability of various voucher items can be statistically analyzed from the historical audited voucher database, and voucher data with frequencies and conditional probabilities exceeding preset thresholds can be extracted to generate dynamically extended knowledge triples. Specifically, the co-occurrence of "decrease in financial institution deposits" accompanied by "increase in accounts receivable" can be statistically analyzed. If the frequency exceeds the threshold and the lift is greater than a set value (such as 1.5), a corresponding candidate triple is generated. This method based on historical data mining helps to capture implicit association rules in actual business operations and supplement specific business scenarios not covered by static standards.

[0068] Furthermore, accounting domain knowledge triples are injected as prior edges into the panoramic transaction graph. In the relational graph convolutional network (RGCN), these prior edges can serve as fixed or learnable weights, guiding the direction of information propagation between entities. This helps to utilize known accounting logic constraint graphs for reasoning, reducing invalid or erroneous association paths and improving the rationality of reasoning results. Further, transaction association reasoning tasks and association confidence scoring tasks are performed on the panoramic transaction graph. The transaction association reasoning tasks include: transaction event association (determining whether two transactions belong to the same economic activity), transaction process completion (detecting missing links in cross-module transaction chains), and transaction relationship classification (e.g., distinguishing between payment and receipt, loan and repayment). Performing these reasoning tasks helps to clarify the logical connections between multi-source transaction data, identify breakpoints or abnormal connections, and form a complete business chain view. It should be noted that the association confidence scoring task includes: structural similarity scoring, attribute consistency scoring, and semantic relevance scoring. Structural similarity scoring is based on graph topological distance calculation, attribute consistency scoring is based on the matching degree of fields such as time and amount, and semantic relevance scoring is based on the textual semantic similarity of the transaction summary. This multi-dimensional scoring mechanism helps to comprehensively assess the reliability of associations and provides a quantitative basis for subsequent decision-making.

[0069] Optionally, the relationships between transaction events, confidence scores, and the completed transaction process chain can be marked in the panoramic transaction map. The knowledge-enhanced panoramic transaction map not only retains the original transaction data, but also adds the inferred relationships and confidence information. This annotation result helps downstream modules (such as semantic-driven subject inference) to quickly obtain high-quality related transaction groups, reduce processing complexity, and improve the overall system processing efficiency.

[0070] Step S202: Based on the panoramic transaction map, extract the structured semantic features of related transaction groups in the enterprise's multi-source transaction data to obtain semantic feature encoding.

[0071] Optionally, based on the panoramic transaction map, the step of extracting structured semantic features of related transaction groups in multi-source transaction data of an enterprise to obtain semantic feature encoding includes: selecting a set of transaction events with confidence scores higher than a preset confidence threshold from the panoramic transaction map to form related transaction groups; extracting multimodal feature encoding for each related transaction group, wherein the multimodal feature encoding includes: transaction type, relationship pattern between business entity nodes, direction and amount of money flow, transaction summary text, and time sequence pattern; and using an embedding model to encode the multimodal features into structured semantic feature vectors to obtain semantic feature encoding.

[0072] Specifically, when selecting a set of transaction events with confidence scores higher than a pre-set confidence threshold from the panoramic transaction map to form related transaction groups, setting a confidence threshold can filter out low-quality or uncertain relationships, retaining highly reliable business chains. This screening mechanism helps reduce noise interference in downstream processing, allowing semantic feature encoding to focus more on transaction scenarios with clear economic substance. For each related transaction group, multimodal feature encoding is extracted, including: transaction type, relationship patterns between business entity nodes, direction and magnitude of money flow, transaction summary text, and temporal pattern. Transaction type and relationship patterns reflect the logical structure of the business, monetary information provides magnitude constraints, summary text carries natural language descriptions, and temporal pattern reflects the chronological order of business occurrences, helping to comprehensively capture the essential attributes of transaction events and providing rich information sources for semantic understanding.

[0073] Optionally, an embedding model can be used to encode multimodal features into structured semantic feature vectors, resulting in semantic feature encoding. The embedding model can map heterogeneous multimodal data into a unified vector space, preserving the semantic relationships between features. This helps large language models understand and process complex transaction contexts more efficiently, transforming unstructured business descriptions into machine-recognizable numerical representations, thereby supporting subsequent instruction fine-tuning for accurate account inference. By selecting high-confidence related transaction groups and extracting multimodal features, high-quality structured context can be provided to large language models, helping to improve the accuracy of accounting account inference and its adaptability to complex business scenarios.

[0074] Step S203: Input the semantic feature encoding into the large language model adjusted by the financial domain corpus instructions, use the large language model to infer the accounting subject, and generate standard accounting entries.

[0075] Optionally, the steps of inputting semantic feature encoding into a large language model adjusted by instructions from a financial domain corpus, and using the large language model to infer accounting subjects and generate standard accounting entries include: converting semantic feature encoding into query embeddings, performing approximate nearest neighbor retrieval in a vector database to obtain the top K historical typical cases with the highest similarity; injecting the matched accounting structure rules, the retrieved top K historical typical cases, and the currently pending related-party transaction group into the large language model; and using the large language model to infer accounting subjects and generate standard accounting entries that conform to the accounting structure rules.

[0076] The process involves encoding semantic features into query embeddings and performing near-nearest neighbor retrieval in a vector database to obtain the top K most similar historical cases. This can refer to using vector retrieval technology to quickly locate historical business records similar to the current transaction scenario. This retrieval enhancement generation mechanism helps provide the model with specific business reference examples, leveraging the effect of few-shot learning to guide the model to imitate historical standard processing patterns, thereby reducing the randomness of inference. Subsequently, the matched accounting structured rules, the retrieved top K historical cases, and the currently pending related transaction group can be injected into the large language model. The injected content includes system instructions, accounting standard rules, reference case summaries, and the currently input transaction features. This structured context injection method helps the large language model simultaneously consider legal constraints, historical experience, and current business details during the inference process, forming a multi-dimensional decision-making basis and improving the compliance and rationality of the output results.

[0077] Furthermore, based on injected contextual information, the large language model utilizes its pre-trained financial knowledge for logical reasoning, outputting standard accounting entries containing debit and credit accounts, amounts, and summaries. Employing a generation method based on instruction fine-tuning, it helps handle complex and non-standardized business scenarios, overcoming the limitations of traditional rule engines in covering long-tail business and achieving automated generation of accounting vouchers. By combining rule constraints, case references, and large model reasoning, the quality of accounting processing in complex transaction scenarios can be effectively improved, reducing the need for manual intervention and enhancing the automation level of accounting processing.

[0078] Step S204: Verify the standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification.

[0079] By constructing a multi-dimensional verification system, the preliminary results generated by the large language model can be automatically verified, which helps to filter out obvious logical errors or compliance deviations. Optionally, the steps of verifying standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification include: performing basic four-dimensional verification on the voucher data in the standard accounting entries, wherein the basic four-dimensional verification includes: checking whether the total debit and credit amounts are balanced, whether the relevant accounts are valid in the current accounting period, whether the amounts are within a reasonable range, and whether the business logic complies with accounting standards; performing case consistency verification on the voucher data in the standard accounting entries, comparing the output accounts, amount levels, and debit / credit directions with the top K historical typical cases retrieved in a structured manner, and calculating a consistency score; if the consistency score is lower than the preset score threshold or the basic four-dimensional verification fails, the voucher data is marked as abnormal voucher data; if the consistency score is greater than or equal to the preset score threshold and the basic four-dimensional verification passes, the voucher data is marked as normal voucher data.

[0080] The basic four-dimensional verification includes: checking whether the total amount of debits and credits is balanced, whether the relevant accounts are valid in the current accounting period, whether the amounts are within a reasonable range, and whether the business logic complies with accounting standards. Specifically, it can automatically calculate the sum of the debit and credit amounts to verify balance, query the account master data to verify timeliness, set threshold ranges to verify the reasonableness of the amounts, and verify the compliance of business logic based on a rule base. This basic verification helps to quickly intercept low-level errors that violate basic accounting equations or regulatory requirements.

[0081] It should be noted that, for the voucher data in standard accounting entries, a case consistency check is performed. The output account, amount level, and debit / credit direction are compared in a structured manner with the top K historical typical cases retrieved, and a consistency score is calculated. By comparing the model output with historical high-confidence successful cases in multiple dimensions, the degree of deviation between the current inference and historical experience can be assessed. This check mechanism helps to use the quality of historical data to constrain the generated results of the large model and improve the business relevance of the inference results.

[0082] Subsequently, a binary classification mechanism based on the dual verification results can be used to divide voucher data into high-confidence regular data and abnormal data requiring manual intervention. This helps optimize the allocation of audit resources, prioritizing high-pass-rate transactions while focusing on the manual review of abnormal accounts. This facilitates the construction of a closed-loop quality control mechanism. By combining basic rule verification with case consistency checks, low-quality or high-risk accounting entries can be effectively identified and isolated, improving the overall accuracy and compliance of the final output vouchers and supporting the implementation of differentiated audit strategies.

[0083] Step S205: Perform differentiated auditing and transaction posting processing on normal voucher data and abnormal voucher data to obtain the final accounting voucher and the queue of abnormal accounts awaiting manual review.

[0084] Dividing voucher data based on verification results helps to organically combine automated processing with manual intervention, optimizing resource allocation in the accounting process. Optionally, the steps of performing differentiated auditing and transaction posting processing on normal and abnormal voucher data to obtain the final posted accounting voucher and a queue of abnormal accounts awaiting manual review include: for normal voucher data, obtaining the voucher type; if the voucher type is a regular business transaction, performing automatic posting processing to obtain the final accounting voucher; if the voucher type is a specific business transaction, performing review processing by the auditing staff to obtain the final accounting voucher; for abnormal voucher data that failed verification, generating an auditing task based on the voucher data of the abnormal voucher data and entering the auditing task into the auditing queue, wherein the auditing task carries an inference trajectory including the transaction source, graph network association path, standard accounting entries output by the large language model, and verification results.

[0085] When the voucher type is a routine transaction, automatic posting is performed to obtain the final accounting voucher. Routine transactions typically refer to high-frequency, standardized transactions that have been verified and confirmed to be error-free. Direct posting helps reduce manual operations, improves the processing efficiency of routine transactions, and enables real-time entry of accounting data. However, when the voucher type is a specific transaction, a review process is performed to obtain the final accounting voucher. Specific transactions may include large-amount or special-nature transactions such as fixed asset purchases and external investments. By setting up a manual review process, it helps to maintain the efficiency of automation while increasing risk control points for key transactions, using human experience to conduct secondary confirmation of special transactions.

[0086] For abnormal voucher data that fails verification, an audit task is generated based on the voucher data of the abnormal voucher data, and the audit task is entered into the audit queue. Abnormal voucher data refers to data that fails the basic four-dimensional verification or case consistency verification. Transferring such data to the manual queue helps to centrally process complex, non-standard, or logically conflicting transactions and prevents erroneous vouchers from entering the formal accounting system.

[0087] Furthermore, the audit task includes a reasoning trajectory encompassing transaction sources, graph network connection paths, standard accounting entries output by the large language model, and verification results. This provides auditors with a complete decision-making context, including the basis for data sources, the logical connections within the graph structure, the model's inference results, and the specific reasons for failure. Providing a complete reasoning trajectory helps auditors quickly understand the causes of anomalies, improving the accuracy and efficiency of manual review, while also providing data feedback for subsequent rule optimization or model fine-tuning.

[0088] Through the above steps, a pre-constructed panoramic transaction graph can be obtained, and accounting domain knowledge triples can be injected into the panoramic transaction graph. Transaction association reasoning is then performed to obtain a knowledge-enhanced panoramic transaction graph. In this graph, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. Based on the panoramic transaction graph, structured semantic features of related transaction groups in the enterprise's multi-source transaction data are extracted to obtain semantic feature codes. These semantic feature codes are input into a large language model adjusted by financial domain corpus instructions. The large language model is used to infer accounting subjects and generate standard accounting entries. The standard accounting entries are verified to obtain verified normal voucher data and verified abnormal voucher data. Differential auditing and transaction posting processing are performed on the normal and abnormal voucher data to obtain the final posted accounting vouchers and a queue of abnormal accounts awaiting manual review. In this embodiment, automatic association of transaction records across business modules can be achieved through panoramic transaction maps and progressive entity alignment processing, reducing manual association costs and improving data integrity, thereby enhancing transaction data processing efficiency. Then, the structured semantic features of associated transaction groups are used to gain a deeper understanding of the transaction business substance, reducing voucher quality issues caused by association errors in complex scenarios. Furthermore, large language models are used to infer accurate accounting subjects, reducing the number of misjudgments of subjects caused by complex scenarios and long-tail businesses. Finally, voucher posting and abnormal voucher processing are performed through accounting entries, achieving a high proportion of automatic generation and accurate posting of routine business, significantly reducing the rate of manual intervention, and improving the efficiency of accounting processing. This solves the technical problems of low efficiency and large differences in the quality of generated accounting vouchers in the process of processing multi-source transaction data of enterprises in related technologies.

[0089] Optionally, in this embodiment, new vouchers that have been manually verified are anonymized, and their business scenario description, transaction group summary, and standard voucher template are stored as new cases in the typical case library. Finally, after the same type of scenario has been verified N times, the inference template for that scenario is solidified, and subsequent similar scenarios can be directly generated using the template without having to call the large language model again.

[0090] The following describes in detail another optional implementation method.

[0091] The present invention provides a method for panoramic association of enterprise multi-source transaction data and intelligent accounting engine based on knowledge-enhanced graph networks and semantic reasoning, which realizes automatic association of cross-module transaction records, semantic understanding of transaction substance, inference of account in complex scenarios, and adaptation of accounting standards.

[0092] This invention provides a two-stage architecture: the first stage, graph reasoning, utilizes a knowledge-enhanced relational graph convolutional network to infer transaction relationships on a panoramic transaction graph; the second stage, semantic reasoning, uses a large language model to infer accounting subjects based on the semantic features of related transaction groups. In the second stage, the large language model undertakes two types of tasks: subject inference and case retrieval-enhanced reasoning, both sharing a base model and distinguished by different prompt word templates. Specifically, the enterprise multi-source transaction data panoramic association and intelligent accounting engine method based on knowledge-enhanced graph networks and semantic reasoning includes the following five core modules.

[0093] Module 1: Construction of the M1 Panoramic Transaction Map

[0094] Establish a unified entity identification system (member units, bank accounts, bills, contracts, credit lines), align and integrate attributes across modules, and construct a complete transaction graph containing records of all business modules based on a unified enterprise graph and transaction events as edges. Nodes are enterprise entities / accounts / contracts / bills, and edges are transaction events (including time, amount, and type attributes).

[0095] The construction of the panoramic transaction map consists of three sub-steps:

[0096] Step 1, Entity Abstraction and Unified Identification: The business entities in each business module are abstracted into five types of entity nodes: member units, bank accounts, bills, contracts, and credit lines. Each type of entity obtains a unified entity ID through a four-layer progressive alignment algorithm, establishing a cross-module entity mapping index.

[0097] Step 2: Model transaction events as edges. Transaction records in each business module are modeled as directed edges, connecting the relevant entity nodes. Edge attributes include: timestamp, transaction amount, transaction type (payment / receipt / loan / aggregation / disbursement / endorsement / discounting, etc.), source module identifier, and original record ID.

[0098] Step 3: Cross-module graph fusion uses a unified entity ID as the key to merge the subgraphs of each module into a panoramic transaction graph. All transaction events of the same entity appearing in different modules converge around that entity, forming a complete transaction profile of that entity.

[0099] Part 1: Cross-module entity alignment algorithm, four-layer progressive alignment.

[0100] The identification format of the same entity varies in different business modules. A unified entity ID is established through a four-layer progressive alignment.

[0101] First layer: Gold identification alignment utilizes unique identifiers to directly match: unified credit code, financial account number, taxpayer identification number, achieving precise hash matching with a coverage rate of approximately 40%-60%.

[0102] The second layer: When there is no unique identifier for multi-attribute weighted fuzzy alignment, weighted similarity is calculated for multiple attributes. The weights are determined by logistic regression fitting of historical labeled data or AHP (Analogous Hierarchical Analysis). The specific matching method is illustrated in Table 1 below.

[0103] Table 1 Multi-attribute weighted fuzzy alignment table

[0104]

[0105] Overall score: Score = Σwi·Simi, if Score > 0.85, automatic alignment; 0.60-0.85, manual candidate selection.

[0106] The values ​​are marked as expected values, based on simulation tests and theoretical derivations.

[0107] The third layer: Transaction Relationship Propagation Alignment (TRP) uses the connections between aligned entities in the transaction graph to propagate identity information.

[0108] The fourth layer: LLM-assisted alignment (Offline-Assisted) is used to perform semantic judgment on the difficult entities that cannot be aligned by the first three layers (about 10%-15%). LLM is called offline in batches only for candidate pairs with confidence in the range of [0.60, 0.85]. The output is manually confirmed and then stored in the database.

[0109] Part Two: Suppression of Alignment Error Propagation.

[0110] Errors during entity alignment can be amplified through TRP iterations, requiring a three-layer suppression mechanism:

[0111] The first layer of inhibition mechanism, confidence decay: Aligned confidence is decayed hop-by-hop along the TRP propagation path, C(v) = C(seed)·α^d, where α = 0.7 (attenuation factor, empirical value, which can be optimized by searching the labeled data grid) d = propagation hop count. When C(v) < 0.3, the entity does not participate in subsequent propagation.

[0112] The second layer of suppression mechanism is the propagation chain length limit: TRP can only execute a maximum of 3 rounds of iteration, and the maximum depth of a single chain cannot exceed 5 hops, to prevent the initial seed error from being amplified indefinitely.

[0113] The third layer of suppression mechanism is the sampling feedback closed loop: the automatic alignment results are randomly sampled at a rate of 5% for manual confirmation, and the confirmation results are injected back to update the feature vector and weight parameters, forming a continuous optimization closed loop.

[0114] Module Two: Transaction-Related Reasoning

[0115] Module 2 can be used to encode accounting knowledge as prior edges of a graph convolutional network in the form of knowledge triples.

[0116] The triplet adopts a static base + dynamic expansion two-layer structure:

[0117] Static base (approximately 70%): derived from core provisions of accounting standards, extracted through rule template matching, and managed in a versioned manner.

[0118] Dynamic expansion (approximately 30%): derived from historical voucher mining (frequency statistics + association rule mining), can be promoted to static base after N≥5 confirmations.

[0119] When accounting policies change, the corresponding standard triples in the static base are updated, and the dynamic extension part is automatically re-statistically adapted without the need for overall reconstruction. The triples come from three sources, as illustrated in Table 2 below.

[0120] Table 2. Sources of the Triplet

[0121]

[0122] Extracting triples from unstructured accounting standards texts:

[0123] Accounting standard provision: If converted into a ternary form through a two-stage assembly line:

[0124] Phase 1 - Rule Template Matching: Pre-defined trigger word templates, such as: "accompanying" → relation_type=ACCOMPANY; "causing" → relation_type=CAUSE; "corresponding to" → relation_type=CORRESPOND; "reducing" → relation_type=DECREASE; "increasing" → relation_type=INCREASE. The trigger word mapping is only for initial matching and may have ambiguity in the accounting context. The specific relation type is actually determined through syntactic dependency analysis for disambiguation.

[0125] For example, “Enterprises should recognize revenue when they have fulfilled their performance obligations under the contract” → template matching “Fulfill performance obligations → Recognize revenue” → triple: [Performance obligations fulfilled] - [CAUSE] -> [Revenue recognition].

[0126] Phase 2 - Large Language Model-Assisted Extraction (only supplementing what was missed in Phase 1).

[0127] For complex clauses that are not matched in Phase 1, the large language model is called offline in batch mode. The input is: "Recognize long-term lease liabilities as the present value of lease payments". The output is: [Lease Liability Recognition]-[CORRESPOND_TO]->[Present Value of Lease Payments]. The results are manually reviewed and then stored in the database.

[0128] Mining triples from historical voucher database: For historically audited vouchers, count the co-occurrence frequency and conditional probability of account pairs, P(Account B|Account A) = count(A→B) / count(A). If P>θ_min(0.3) and lift>1.5, then a candidate triple relationship is identified. For example, in historical vouchers, the frequency of "decrease in bank deposits" accompanied by "increase in accounts receivable" is >500, which is identified as a candidate triple: [decrease in bank deposits] - [ACCOMPANY] -> [increase in accounts receivable], which is confirmed as a valid triple after auditing.

[0129] Three reasoning tasks: transaction event association (determining whether two transactions are related to the same economic activity), transaction process completion (detecting missing links in cross-module transaction chains), and transaction relationship classification (payment vs. receipt, lending vs. repayment, aggregation vs. disbursement). Association confidence is assessed using three dimensions: structural similarity (graph topological distance), attribute consistency (degree of time-amount matching), and semantic relevance (summary semantic similarity).

[0130] Module 3, Semantic Reasoning Subject Inference.

[0131] Module 3 allows the semantic features of related transaction groups (transaction type + entity relationships involved + amount direction and size + transaction summary text + time sequence pattern) to be encoded, input into a large language model fine-tuned by instructions from a financial domain corpus, and output standard accounting entries.

[0132] Four-dimensional verification: debit and credit balance (total debit and credit amounts are equal), account applicability (accounts are valid in the current accounting period), amount reasonableness (within a reasonable range), and business logic (complies with accounting standards and company accounting policies).

[0133] Module 4, Voucher Generation and Verification.

[0134] Differentiated review: Routine transactions are automatically approved and posted; special transactions (fixed asset purchases, external investments) are marked for manual review; abnormal transactions are entered into the manual review queue with an explanation of the reasoning process.

[0135] Audit trail: Each voucher records a complete generated trail - transaction source module and ID list, graph network association path, intermediate output of large language model, and verification results.

[0136] Module 5: Dual-driven self-updating of criteria and case studies.

[0137] Accounting standards are coded as structured reasoning rules, and a typical case library stores standard cases ranging from business scenarios to multi-module transaction records to accounting vouchers. New scenarios are guided by a large language model to infer by searching for similar cases. When accounting policies change, only the standards knowledge base needs to be updated to adapt.

[0138] New vouchers, after manual verification and confirmation, are automatically added to the database as new cases after being anonymized. When the same type of scenario is confirmed more than N times (N is a configurable parameter, default N=3), the reasoning for that scenario no longer requires a large language model and can be directly generated using a template (cold start → bootstrapping → rule solidification). Scenarios of the same type are determined by a scenario classification model, with determination dimensions including: transaction type, entity relationship patterns involved, amount range, and summary keyword clustering.

[0139] The following is a detailed description with reference to another embodiment.

[0140] Example 2

[0141] The accounting auditing device for multi-source transaction data of an enterprise provided in this embodiment includes multiple implementation units, each of which corresponds to the implementation steps in the above embodiment one. The specific implementation method and beneficial effects can be referred to the foregoing method embodiment, and will not be repeated here.

[0142] Figure 3 This is a schematic diagram of an optional account auditing device for multi-source transaction data of an enterprise according to an embodiment of the present invention, such as... Figure 3 As shown, the accounting auditing device for enterprise multi-source transaction data may include: a panoramic transaction map acquisition unit 31, a semantic feature extraction unit 32, an accounting subject inference unit 33, a verification unit 34, and an auditing unit 35.

[0143] Among them, the panoramic transaction map acquisition unit 31 is used to acquire a pre-constructed panoramic transaction map, inject accounting domain knowledge triples into the panoramic transaction map, and perform transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map. In the panoramic transaction map, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information.

[0144] The semantic feature extraction unit 32 is used to extract the structured semantic features of related transaction groups in multi-source transaction data of enterprises based on the panoramic transaction map, and obtain semantic feature encoding.

[0145] The accounting subject inference unit 33 is used to input semantic feature encoding into a large language model that has been adjusted by instructions from a financial domain corpus, and use the large language model to infer accounting subjects and generate standard accounting entries.

[0146] Verification unit 34 is used to verify standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification.

[0147] Audit unit 35 is used to perform differentiated audits and transaction postings on normal and abnormal voucher data, and to obtain the final accounting vouchers and a queue of abnormal accounts awaiting manual review.

[0148] The aforementioned account auditing device for enterprise multi-source transaction data can acquire a pre-constructed panoramic transaction map through the panoramic transaction map acquisition unit 31, inject accounting domain knowledge triples into the panoramic transaction map, and perform transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map. In this panoramic transaction map, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. Based on the panoramic transaction map, the semantic feature extraction unit 32 extracts the structured semantic features of the associated transaction groups in the enterprise multi-source transaction data to obtain semantic feature codes. The accounting subject inference unit 33 inputs the semantic feature codes into a large language model adjusted by financial domain corpus instructions, uses the large language model to infer accounting subjects, and generates standard accounting entries. The verification unit 34 verifies the standard accounting entries to obtain normal voucher data that passes verification and abnormal voucher data that fails verification. The auditing unit 35 performs differentiated auditing and transaction posting processing on the normal voucher data and abnormal voucher data to obtain the final posted accounting voucher and a queue of abnormal accounts awaiting manual auditing. In this embodiment, automatic association of transaction records across business modules can be achieved through panoramic transaction maps and progressive entity alignment processing, reducing manual association costs and improving data integrity, thereby enhancing transaction data processing efficiency. Then, the structured semantic features of associated transaction groups are used to gain a deeper understanding of the transaction business substance, reducing voucher quality issues caused by association errors in complex scenarios. Furthermore, large language models are used to infer accurate accounting subjects, reducing the number of misjudgments of subjects caused by complex scenarios and long-tail businesses. Finally, voucher posting and abnormal voucher processing are performed through accounting entries, achieving a high proportion of automatic generation and accurate posting of routine business, significantly reducing the rate of manual intervention, and improving the efficiency of accounting processing. This solves the technical problems of low efficiency and large differences in the quality of generated accounting vouchers in the process of processing multi-source transaction data of enterprises in related technologies.

[0149] Optionally, the accounting auditing device for enterprise multi-source transaction data further includes: a node abstraction unit, used to perform node abstraction processing on the business entities of each business module before obtaining the pre-built panoramic transaction map, and to perform multi-layer progressive alignment processing on each business entity node to obtain a unified entity identifier node; a directed edge modeling unit, used to model the transaction records in each business module as directed edges to obtain a subgraph related to each business module, wherein the edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID; and a panoramic transaction map construction unit, used to merge the subgraphs of each business module into a panoramic transaction map containing records of all business modules, using the unified entity identifier node as the key, wherein all transaction events of the same business entity in different business modules in the panoramic transaction map converge around the unified entity identifier node.

[0150] Optionally, the node abstraction unit includes: a node abstraction module, used to abstract the business entities in each business module into multiple types of business entity nodes, wherein the multiple types of business entity nodes include: member units, financial accounts, bills, contracts, and credit lines; and a node alignment module, used to perform multi-layer progressive alignment processing on the same business entity node in different business modules, wherein the multi-layer progressive alignment processing includes: a first layer, matching using the unified social credit code, financial account number, and taxpayer identification number of each business entity node; a second layer, for unidentified entities, using weighted fuzzy matching using multiple types of enterprise attributes; a third layer, using the connection relationship of aligned business entity nodes in the panoramic transaction map to perform transaction relationship propagation alignment; and a fourth layer, for abnormal business entity nodes that cannot be aligned in the first, second, and third layers, calling a large language model for auxiliary semantic judgment.

[0151] Optionally, the accounting auditing device for enterprise multi-source transaction data further includes: a propagation suppression module, used to perform alignment error propagation suppression processing on the business entity nodes after performing multi-level progressive alignment processing on the same business entity nodes in different business modules; a confidence decay module, used to perform confidence decay processing during the alignment error propagation suppression process, and perform hop-by-hop decay processing on the propagation path of the business entity nodes; a propagation chain length limiting module, used to perform propagation chain length limiting processing during the alignment error propagation suppression process, and perform parameter limiting on the propagation depth of the business entity nodes; and a sampling inspection feedback closed-loop processing module, used to perform sampling inspection feedback closed-loop processing during the alignment error propagation suppression process, sampling business entity nodes that have completed alignment processing according to a preset ratio, sending the sampling inspection results to the auditing station, and updating the node feature vector and feature parameters of the business entity nodes based on the confirmation results given by the auditing station.

[0152] Optionally, the panoramic transaction map acquisition unit includes: a static base knowledge triple extraction module, used to extract static base knowledge triples composed of head entities, relations, and tail entities from unstructured accounting standard text by matching trigger words through rule templates; a dynamic extended knowledge triple extraction module, used to statistically analyze the co-occurrence frequency and conditional probability of various voucher accounts from the historical audited voucher database, and extract voucher data with a frequency and conditional probability greater than a preset threshold to generate dynamic extended knowledge triples; and an accounting domain knowledge triple generation module, used to generate dynamic extended knowledge triples based on the static base knowledge triples and dynamic extended knowledge triples. The system consists of tuples, forming accounting domain knowledge triples; a knowledge triple injection module, used to inject accounting domain knowledge triples as prior edges into the panoramic transaction graph; and a transaction association reasoning task and an association confidence scoring task, performed on the panoramic transaction graph. The transaction association reasoning task includes: transaction event association, transaction process completion, and transaction relationship classification. The association confidence scoring task includes: structural similarity scoring, attribute consistency scoring, and semantic relevance scoring. The annotation module is used to annotate the association relationships between transaction events, confidence scores, and the completed transaction process chain in the panoramic transaction graph.

[0153] Optionally, the semantic feature extraction unit includes: a related transaction group generation module, used to filter out a set of transaction events with confidence scores higher than a preset confidence threshold from the panoramic transaction map to form related transaction groups; a feature encoding extraction module, used to extract multimodal feature codes for each related transaction group, wherein the multimodal feature codes include: transaction type, relationship patterns between business entity nodes, direction and amount of money flow, transaction summary text, and time sequence pattern; and a feature encoding module, used to encode the multimodal features into structured semantic feature vectors using an embedding model to obtain semantic feature codes.

[0154] Optionally, the accounting subject inference unit includes: a feature encoding conversion module, used to convert semantic feature encoding into query embedding, perform approximate nearest neighbor retrieval in the vector database, and obtain the top K historical typical cases with the highest similarity; a model injection module, used to inject the matched accounting structured rules, the retrieved top K historical typical cases, and the currently pending related transaction group into the large language model; and an accounting subject inference module, used to infer accounting subjects using the large language model and generate standard accounting entries that conform to the accounting structured rules.

[0155] Optionally, the verification unit includes: a basic four-dimensional verification module, used to perform basic four-dimensional verification on the voucher data in standard accounting entries, wherein the basic four-dimensional verification includes: checking whether the total debit and credit amounts are balanced, whether the relevant accounts are valid in the current accounting period, whether the amount is within a reasonable range, and whether the business logic complies with accounting standards; a consistency verification module, used to perform case consistency verification on the voucher data in standard accounting entries, comparing the output accounts, amount levels, and debit / credit directions with the top K historical typical cases retrieved in a structured manner, and calculating a consistency score; an abnormal voucher data marking module, used to mark voucher data as abnormal voucher data when the consistency score is lower than a preset scoring threshold or the basic four-dimensional verification fails; and a normal voucher data marking module, used to mark voucher data as normal voucher data when the consistency score is greater than or equal to a preset scoring threshold and the basic four-dimensional verification passes.

[0156] Optionally, the audit unit includes: a voucher type acquisition module for acquiring the voucher type for normal voucher data; an automatic posting module for performing automatic posting processing to obtain the final accounting voucher when the voucher type is a regular business transaction; an audit post review processing module for performing audit post review processing to obtain the final accounting voucher when the voucher type is a specific business transaction; and an audit task generation module for generating audit tasks based on the voucher data of abnormal voucher data that failed verification, and entering the audit tasks into the audit queue. The audit tasks carry an inference trajectory containing the transaction source, graph network association path, standard accounting entries output by the large language model, and verification results.

[0157] The aforementioned accounting auditing device for enterprise multi-source transaction data may also include a processor and a memory. The aforementioned panoramic transaction map acquisition unit 31, semantic feature extraction unit 32, accounting subject inference unit 33, verification unit 34, auditing unit 35, etc. are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0158] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, a comprehensive association of multi-source transaction data and intelligent accounting auditing can be achieved based on knowledge-enhanced graph networks and semantic reasoning.

[0159] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0160] Example 3

[0161] Embodiments of this application may provide an electronic device. Figure 4 This is a structural block diagram of an electronic device for performing an accounting auditing method for multi-source transaction data of an enterprise, according to an embodiment of this application. Figure 4 As shown, the electronic device may include: one or more ( Figure 4 (Only one is shown) processor 402, memory 404, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0162] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the accounting audit method and apparatus for enterprise multi-source transaction data in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned accounting audit method for enterprise multi-source transaction data. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, enterprise intranets, local area networks, mobile communication networks, and combinations thereof.

[0163] The processor can access information and applications stored in memory via a transmission device to execute the following steps: acquire a pre-constructed panoramic transaction map, inject accounting domain knowledge triples into the panoramic transaction map, and perform transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map. In this map, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. Based on the panoramic transaction map, extract structured semantic features of related transaction groups from multi-source enterprise transaction data to obtain semantic feature codes. Input the semantic feature codes into a large language model adjusted by financial domain corpus instructions, and use the large language model to infer accounting subjects and generate standard accounting entries. Verify the standard accounting entries to obtain verified normal voucher data and verified abnormal voucher data. Perform differentiated auditing and transaction posting processing on the normal and abnormal voucher data to obtain the final posted accounting vouchers and a queue of abnormal accounts awaiting manual review.

[0164] Those skilled in the art will understand that Figure 4 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones, tablets, handheld computers, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 4This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 4 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 4 The different configurations shown.

[0165] Those skilled in the art will understand that all or part of the steps in the various accounting methods for multi-source transaction data of enterprises in the above embodiments can be implemented by a program instructing the hardware of the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0166] Example 4

[0167] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the accounting auditing method for multi-source transaction data of an enterprise provided in Embodiment 1.

[0168] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the account auditing method for multi-source transaction data of an enterprise as described in any of the above embodiments.

[0169] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0170] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the account auditing method for multi-source transaction data of an enterprise as described in various embodiments of this application.

[0171] This application also provides a computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the accounting auditing method for multi-source transaction data of an enterprise as described in various embodiments of this application.

[0172] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0173] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0174] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0175] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0177] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0178] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for auditing accounts based on multi-source transaction data of an enterprise, characterized in that, include: A pre-constructed panoramic transaction graph is obtained, and accounting domain knowledge triples are injected into the panoramic transaction graph. Transaction association reasoning is then performed to obtain a knowledge-enhanced panoramic transaction graph. In the panoramic transaction graph, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. Based on the panoramic transaction map, the structured semantic features of related transaction groups in the enterprise's multi-source transaction data are extracted to obtain semantic feature encoding; The semantic features are encoded and input into a large language model that has been adjusted by instructions from a financial domain corpus. The large language model is then used to infer accounting subjects and generate standard accounting entries. The standard accounting entries are verified to obtain normal voucher data that passes verification and abnormal voucher data that fails verification. Differential auditing and transaction posting are performed on the normal voucher data and the abnormal voucher data to obtain the final accounting voucher and the abnormal account queue awaiting manual review.

2. The method for auditing accounts for multi-source transaction data of enterprises according to claim 1, characterized in that, Before obtaining the pre-built panoramic transaction map, the following is also included: The business entities of each business module are abstracted into nodes, and the business entity nodes are progressively aligned in multiple layers to obtain a unified entity identifier node. The transaction records in each business module are modeled as directed edges to obtain a subgraph that connects each business module. The edge attributes of the directed edges include: timestamp, transaction amount, transaction type, source module identifier, and original record ID. Using the unified entity identifier node as the key, the subgraphs of each business module are merged into a panoramic transaction graph containing records of all business modules. In the panoramic transaction graph, all transaction events of the same business entity in different business modules converge with the unified entity identifier node as the center.

3. The method for auditing accounts for multi-source transaction data of enterprises according to claim 2, characterized in that, The steps involved in abstracting the business entities of each business module into nodes and performing multi-layered progressive alignment on each business entity node to obtain a unified entity identifier node include: The business entities in each business module are abstracted into multiple types of business entity nodes, including: member units, financial accounts, bills, contracts, and credit lines. A multi-layer progressive alignment process is performed on the same business entity node in different business modules. The multi-layer progressive alignment process includes: a first layer, matching based on the unified social credit code, financial account number, and taxpayer identification number of each business entity node; a second layer, weighted fuzzy matching based on multiple enterprise attributes for unidentified entities; a third layer, alignment based on the connection relationships of aligned business entity nodes in the panoramic transaction map; and a fourth layer, auxiliary semantic judgment by invoking a large language model for abnormal business entity nodes that cannot be aligned in the first, second, and third layers.

4. The method for auditing accounts for multi-source transaction data of enterprises according to claim 3, characterized in that, After performing multi-level progressive alignment processing on the same business entity node in different business modules, it also includes: Perform alignment error propagation suppression processing on business entity nodes after completing multi-layer progressive alignment processing; During the alignment error propagation suppression process, confidence decay processing is performed to decay the propagation path of the business entity nodes hop by hop. During the alignment error propagation suppression process, propagation chain length limitation is performed to limit the propagation depth of business entity nodes using parameters. During the alignment error propagation suppression process, a sampling feedback closed-loop process is performed. Business entity nodes that have completed alignment processing are sampled according to a preset ratio, and the sampling results are sent to the reviewer. Based on the confirmation results given by the reviewer, the node feature vector and feature parameters of the business entity nodes are updated.

5. The method for auditing accounts for multi-source transaction data of an enterprise according to claim 1, characterized in that, The steps of injecting accounting domain knowledge triples into the panoramic transaction graph and performing transaction association reasoning to obtain the knowledge-enhanced panoramic transaction graph include: From unstructured accounting standards texts, static base knowledge triples consisting of head entities, relations, and tail entities are extracted by matching trigger words using rule templates. The co-occurrence frequency and conditional probability of various voucher items are statistically analyzed from the historical audited voucher database, and voucher data with a frequency and conditional probability greater than the preset threshold are extracted to generate dynamic extended knowledge triples. Based on the static base knowledge triplet and the dynamic extended knowledge triplet, the accounting domain knowledge triplet is formed. The accounting domain knowledge triples are injected as prior edges into the panoramic transaction graph. The transaction association reasoning task and the association confidence scoring task are performed on the panoramic transaction map. The transaction association reasoning task includes: transaction event association, transaction process completion and transaction relationship classification. The association confidence scoring task includes: structural similarity scoring, attribute consistency scoring and semantic relevance scoring. The panoramic transaction map marks the relationships between transaction events, confidence scores, and the completed transaction process chain.

6. The method for auditing accounts for multi-source transaction data of an enterprise according to claim 1, characterized in that, Based on the aforementioned panoramic transaction map, the steps of extracting structured semantic features of related transaction groups from multi-source enterprise transaction data to obtain semantic feature encoding include: The set of transaction events with confidence scores higher than a preset confidence threshold are selected from the panoramic transaction map to form an associated transaction group; For each of the associated transaction groups, multimodal feature codes are extracted, wherein the multimodal feature codes include: transaction type, relationship pattern between business entity nodes, direction and amount of money flow, transaction summary text, and time sequence pattern; The multimodal features are encoded into structured semantic feature vectors using an embedding model, resulting in semantic feature encoding.

7. The method for auditing accounts for multi-source transaction data of enterprises according to claim 1, characterized in that, The steps of inputting the semantic feature encoding into a large language model adjusted by instructions from a financial domain corpus, and using the large language model to infer accounting subjects and generate standard accounting entries include: The semantic features are encoded into query embeddings, and approximate nearest neighbor retrieval is performed in the vector database to obtain the top K most similar historical typical cases. The matching accounting structure rules, the retrieved top K historical typical cases, and the currently pending related transaction groups are injected into the large language model; The large language model is used to infer accounting subjects and generate standard accounting entries that conform to accounting structure rules.

8. The method for auditing accounts for multi-source transaction data of an enterprise according to claim 7, characterized in that, The steps of verifying the standard accounting entries to obtain verified normal voucher data and verified abnormal voucher data include: Perform basic four-dimensional verification on the voucher data in the standard accounting entries, wherein the basic four-dimensional verification includes: checking whether the total debit and credit amounts are balanced, whether the relevant accounts are valid in the current accounting period, whether the amounts are within a reasonable range, and whether the business logic complies with accounting standards; Perform case consistency verification on the voucher data in the standard accounting entries, and compare the output account, amount level, debit / credit direction with the top K historical typical cases retrieved in a structured manner to calculate the consistency score; If the consistency score is lower than the preset score threshold or the basic four-dimensional verification fails, the voucher data is marked as abnormal voucher data; If the consistency score is greater than or equal to the preset score threshold and the basic four-dimensional verification is passed, the voucher data is marked as normal voucher data.

9. The method for auditing accounts for multi-source transaction data of an enterprise according to claim 1, characterized in that, The steps of performing differentiated auditing and transaction posting processing on the normal voucher data and the abnormal voucher data to obtain the final posted accounting voucher and the queue of abnormal accounts awaiting manual review include: For the normal voucher data, obtain the voucher type; if the voucher type is a regular business transaction, perform automatic posting to obtain the final accounting voucher; if the voucher type is a specific business transaction, perform review by the auditing staff to obtain the final accounting voucher. For the abnormal voucher data that fails to be verified, an audit task is generated based on the voucher data of the abnormal voucher data, and the audit task is entered into the audit queue. The audit task carries an inference trajectory that includes the transaction source, graph network association path, standard accounting entries output by the large language model, and verification results.

10. An account auditing device for multi-source transaction data of an enterprise, characterized in that, include: The panoramic transaction map acquisition unit is used to acquire a pre-constructed panoramic transaction map, inject accounting domain knowledge triples into the panoramic transaction map, and perform transaction association reasoning to obtain a knowledge-enhanced panoramic transaction map. In the panoramic transaction map, each business entity node corresponds to a unified entity identifier, and each connecting edge contains complete transaction attribute information. The semantic feature extraction unit is used to extract the structured semantic features of related transaction groups in the enterprise's multi-source transaction data based on the panoramic transaction map, and obtain semantic feature encoding. The accounting subject inference unit is used to input the semantic feature encoding into a large language model that has been adjusted by instructions from a financial domain corpus, and use the large language model to infer accounting subjects and generate standard accounting entries. The verification unit is used to verify the standard accounting entries to obtain normal voucher data that has passed verification and abnormal voucher data that has failed verification. The auditing unit is used to perform differentiated auditing and transaction posting processing on the normal voucher data and the abnormal voucher data, so as to obtain the final accounting voucher posted and the abnormal account queue awaiting manual auditing.

11. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the accounting auditing method for multi-source transaction data of an enterprise as described in any one of claims 1 to 9.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the accounting auditing method for multi-source transaction data of an enterprise as described in any one of claims 1 to 9.

Images