Financial data auditing method, device and equipment

Abstract

The invention provides a financial data auditing method, device and equipment. The financial data auditing method comprises the steps of obtaining expense financial data and bill financial data of a target user; performing data auditing on the cost financial data through a preset first big language model to obtain a cost auditing result; performing data auditing on the bill financial data through a preset multi-modal large model to obtain a bill auditing result; and obtaining an input data query statement, and generating a data query reply corresponding to the data query statement according to the data query statement, the cost auditing result and the bill auditing result. The efficiency and accuracy of financial data auditing can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence (AI) technology, and in particular to a financial data auditing method, apparatus and equipment. Background Technology

[0002] As businesses expand, various reimbursement-related financial expenses, such as travel, office expenses, hospitality, and training expenses, continue to grow. Expense auditing has become a crucial aspect of cost control and risk management for enterprises. To ensure expense compliance, online auditing processes are typically used to audit reimbursement-related financial expenses.

[0003] However, online audit processes require multiple approval nodes, each of which requires manual review. As a result, there are issues such as low processing efficiency, missed detection of real anomalies, or misjudgment of normal behavior. Summary of the Invention

[0004] This application provides a financial data auditing method, apparatus, and equipment, which can solve the problems of low efficiency in financial data auditing and inaccurate processing results such as missed detection of real anomalies or misjudgment of normal behavior, thereby improving the efficiency and accuracy of financial data auditing.

[0005] Firstly, this application provides a financial data auditing method, which includes: acquiring the expense financial data and invoice financial data of the target user; performing data auditing on the expense financial data using a preset Large Language Model (LLM) to obtain expense auditing results; performing data auditing on the invoice financial data using a preset Multimodal Large Language Model (MLLM) to obtain invoice auditing results; acquiring the input data query statement, and generating a data query response corresponding to the data query statement based on the data query statement, the expense auditing results, and the invoice auditing results.

[0006] Secondly, this application provides a financial data auditing device, which includes: a data acquisition module for acquiring expense financial data and invoice financial data of a target user; a data auditing module for performing data auditing on the expense financial data using a preset first large language model to obtain expense auditing results; and performing data auditing on the invoice financial data using a preset multimodal large model to obtain invoice auditing results; and a data querying module for acquiring an input data query statement and generating a data query response corresponding to the data query statement based on the data query statement, the expense auditing results, and the invoice auditing results.

[0007] Thirdly, this application provides an electronic device, including: a processor and a memory, the memory for storing a computer program, and the processor for calling and running the computer program stored in the memory to perform the methods as described in the first aspect or its various implementations.

[0008] Fourthly, this application provides a computer-readable storage medium for storing a computer program that causes a computer to perform the methods described in the first aspect or its various implementations.

[0009] Fifthly, this application provides a computer program product including computer program instructions that cause a computer to perform the methods as described in the first aspect or its various implementations.

[0010] Sixthly, this application provides a computer program that causes a computer to perform the methods described in the first aspect or its various implementations.

[0011] Other technical features and effects involved in this application will be described in subsequent embodiments, and will not be repeated here to avoid repetition. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0013] Figure 1 A flowchart of a financial data auditing method provided in this application embodiment; Figure 2 A schematic diagram illustrating a financial data auditing method provided in an embodiment of this application; Figure 3 A schematic diagram illustrating another financial data auditing method provided in this application embodiment; Figure 4 A schematic diagram of a financial data auditing device 400 provided for an embodiment of this application; Figure 5 This is a schematic diagram of an electronic device 500 provided in an embodiment of this application. Detailed Implementation

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application 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 this application 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 non-exclusive inclusion; for example, a process, method, system, product, or server 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 devices.

[0015] It should be understood that the technical solution of this application can be applied to the following scenarios, but is not limited to: In one embodiment, the technical solution of this application can be applied to the auditing of financial data. For example, it can be applied to the auditing of employee travel expenses and the querying of audit results, but is not limited thereto.

[0016] In one embodiment, the solution provided in this application can be executed by any electronic device with data processing capabilities. For example, the electronic device can be a server, specifically a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Alternatively, the electronic device can be a terminal device, specifically a tablet computer, laptop computer, or desktop computer. Furthermore, the electronic device can be a combination of a server and a terminal device, wherein the server and terminal device in the combination can communicate wirelessly or via wired means. This application does not impose specific limitations on the electronic device.

[0017] It should be noted that all technical solutions in this application can be combined in any way to form optional embodiments of this application, which will not be elaborated upon here. Furthermore, all information, data, instructions, and signals, including expense and financial data, invoice financial data, and query statements, involved in this application are authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0018] The technical solution of this application will be described below: Figure 1 A flowchart of a financial data auditing method provided in this application embodiment is shown below. Figure 1 As shown, the method may include the following steps: S110: Obtain the target user's expense and invoice financial data; S120: Perform data audit on expense and financial data using the preset first language model to obtain the expense audit results; S130: Perform data audit on the financial data of bills through a preset multimodal large model to obtain the bill audit results; S140: Obtain the input data query statement, and generate the corresponding data query response based on the data query statement, the expense audit result, and the invoice audit result.

[0019] The following sections will introduce the auditing of user expense data (corresponding to expense financial data) and user invoice data (corresponding to invoice financial data).

[0020] In one embodiment, an electronic device can audit user fee data through the following steps: S120-1: Obtain user expense data of the target user within a preset time period. The user expense data includes at least one of the following: expense order data and attendance data. S120-2: Aggregate user fee data to obtain aggregated fee data; S120-3: Format the aggregated cost data to obtain cost financial data in natural language format; S120-4: Add the expense and financial data to the preset prompt word template to obtain expense and financial prompt words. The prompt word template is used to instruct the first language model to output the expense audit results. S120-5: Using the first major language model, calculate the expense financial prompts based on at least one expense audit dimension to obtain the expense audit results.

[0021] The expense audit includes at least one of the following dimensions: regional dimension (e.g., the geographical location where the expense is generated, the geographical location in the expense order, etc.), time dimension (e.g., the time point, time period, or time series in which the expense is generated, etc.), expense type dimension (e.g., the dimension under the nature of the expense business such as transportation, accommodation, catering, entertainment, etc.), expense item dimension (e.g., the dimension of the business project to which the expense is generated), budget control dimension (e.g., the dimension under the budget amount corresponding to the business project, the department mentioned in the business, or the personnel), payment method dimension (e.g., the dimension under corporate / personal, virtual card, prepayment / reimbursement), personnel role dimension (e.g., the dimension under the job level travel / entertainment standard), and expense correlation dimension (referring to the correlation dimension between expense records from different expense sources; for example, accommodation orders and transportation orders are corresponding and correlated, and the regions and times involved in accommodation order expenses and transportation order expenses should be correlated). The expense audit result includes the audit result under at least one expense audit dimension, the anomaly type, and the data on the reason for generating the corresponding audit result.

[0022] For example, the above-mentioned data aggregation of user fee data to obtain aggregated fee data may include: data cleaning of user fee data, format adjustment using a preset format, and aggregation based on a preset key combination. Specifically, it may include at least one of the following, but is not limited to: field standardization and format adjustment (e.g., time field, etc.), deduplication and missing value imputation, currency conversion, and aggregation according to the key combination of "user identifier + date / item number".

[0023] As can be understood, user expense data refers to the expense data generated by the spending behavior of target users within the financial data. For example, procurement expenses incurred from purchasing activities, and travel expenses incurred from business trips. Correspondingly, expense order data can include: reimbursement / financial system data (reimbursement records, payment application forms, payment voucher data), travel / trip type order data (airline / train / bus ticket orders and itineraries, hotel accommodation orders, taxi / ride-hailing services), and consumption payment data (business trip application and approval records), etc.

[0024] For example, the above-mentioned calculation of expense financial prompts based on at least one expense audit dimension using the first major language model to obtain expense audit results can include: calculating expense financial prompts based on at least one expense audit dimension and company internal policy data using the first major language model to obtain expense audit results. The company internal policy data can include at least one of the following, but is not limited to: travel standard data, reimbursement limit data, and subsidy policy data.

[0025] One approach is to first fine-tune the primary language model using historical expense and financial data, and then use the fine-tuned primary language model to perform data analysis on expense and financial prompts to obtain expense audit results.

[0026] Furthermore, the template database can contain multiple prompt word templates. Before executing S120-4, a matching prompt word template matching the expense and financial data can be found in the template database based on the specific content contained in the expense and financial data and at least one expense audit dimension mentioned above. S120-4 is then executed based on this matching prompt word template. The matching prompt word template contains all fields of the specific content contained in the expense and financial data, as well as at least one expense audit dimension mentioned above. Alternatively, the matching prompt word template can be obtained by combining at least one sub-matching prompt word template, each sub-matching prompt word template containing at least one field of the specific content contained in the expense and financial data, and at least one expense audit dimension.

[0027] Furthermore, the financial risk characteristics corresponding to at least one of the aforementioned expense audit dimensions can be determined first. These financial risk characteristics represent specific financial risk behaviors occurring under the corresponding expense audit dimension. For example, the financial risk characteristics of the time dimension can characterize the financial risk behavior of consumption outside of working hours, while the financial risk characteristics of the expense type dimension can characterize the financial risk behavior of expenditures with a pre-defined specific expense generator. This specific expense generator could be a fake supplier, a fake hotel, etc., and could be obtained in advance through searches on various online platforms. Specifically, the specific financial risk behaviors corresponding to each expense audit dimension can be pre-defined. Then, the behavioral text entities of these specific financial risk behaviors can be determined (e.g., the name of the expense generator or the time range corresponding to non-working hours). Entity extraction is performed on the expense financial data using these behavioral text entities (extracting entities with a text similarity greater than a similarity threshold to the behavioral text entities) to obtain expense risk entities. Financial risk characteristics are then generated based on these expense risk entities. If no expense risk entity can be extracted, it is determined that the financial risk characteristics corresponding to the corresponding expense audit dimension are not involved. Next, it can be determined whether the aforementioned expense audit results and / or expense financial data involve financial risk characteristics corresponding to at least one expense audit dimension. If any target financial risk characteristic is involved among the financial risk characteristics corresponding to at least one expense audit dimension, then other expense financial data associated with the target financial risk characteristic (the expense financial data of users other than the target user; associated with the target financial risk characteristic means that the expense financial data involves data related to the target financial risk characteristic, such as data involving the target financial risk characteristic or the financial risk behavior or expense risk entity corresponding to the target financial risk characteristic) can be queried and obtained, and data auditing can be performed on the other expense financial data. In this way, when any target user is found to be involved in financial risk behavior, other users can be detected simultaneously in a timely manner to see if they are also involved in financial risk behavior, thereby detecting all users involved in financial risk behavior and improving financial risk control capabilities.

[0028] In addition, the cost audit results can be written to the audit result database for subsequent visualization and Q&A modules to use.

[0029] For example, taking employees as the target users and travel expenses as the user expense data, we can first obtain all travel order information for employees within a certain natural day, such as air tickets, train tickets, hotel check-ins, and car rental orders. We can also obtain employee attendance data, such as check-in data for that natural day (including check-in location and time). Next, using "Employee ID (employee identifier) ​​+ Travel Date" as the primary key, we can aggregate all travel orders within that natural day to obtain aggregated expense data. This aggregated expense data includes fields for each travel order, such as departure location, destination, departure time, arrival time, mode of transportation, booking person, and user. Then, we can use a prompt template to format the above order information into natural language, resulting in expense financial prompts as shown below: An employee had the following itinerary on [date]: departing from Region 1 to Region 2 by high-speed rail at 9:00 AM, departing from Region 2 to Region 3 by car at 2:00 PM, and checking into a hotel in Region 3 at 6:00 PM. Please determine if this itinerary is reasonable and whether it involves any unusual travel. The analysis results should be returned in plain JSON format (a lightweight data exchange format), including the following fields: "is_reasonable (whether it is abnormal): true / false (normal / abnormal); "detailed_reason(generated reason)": "Detailed analysis of the reasons".

[0030] The "Is it abnormal?" field is a boolean type, used to indicate whether abnormal travel behavior has been detected. `true` indicates an anomaly exists, and `false` indicates the trip is reasonable. The "Generation Reason" field is a string type, a natural language explanation generated by the large model, describing the semantic reasons, logical conflicts, or suspicious behaviors for the anomaly judgment, such as spatiotemporal contradictions, route jumps, or mismatched modes of transportation.

[0031] Then, the electronic device can input the aforementioned expense and financial prompts into the first language model (e.g., Qwen-32B or an equivalent model). The first language model then performs semantic extraction and cross-order logical reasoning to determine if the employee has exhibited abnormal travel behavior. Specifically, the first language model can be based on a Transformer architecture and utilize a self-attention mechanism to jointly analyze the following factors (corresponding to at least one expense audit dimension): spatial context (e.g., the distance between departure and destination, which can be inferred based on geographical knowledge in the model training data and / or external geographic databases), temporal context (e.g., consistency of departure and arrival times, travel duration, and time intervals between consecutive trips), and mode of transportation (e.g., the suitability of the selected mode of transportation to time and space). Finally, the expense audit results output by the first language model can be obtained, as shown below: Is it abnormal: true; Reason for generation: This itinerary contains obvious spatial and temporal logical errors. An employee traveling over 1200 kilometers from Region 1 to Region 2 in 3.5 hours is physically impossible; even the fastest flight plus ground travel time would take at least 5-6 hours. Furthermore, the route planning from Region 1 to Region 2 to Region 3 is unreasonable; normal business itineraries should be arranged according to geographical location. In summary, this employee is likely to have fabricated an entire itinerary for reimbursement purposes. It is recommended that a thorough audit be initiated immediately, requiring complete travel documentation and identity verification records.

[0032] In actual testing, by simulating a real business trip dataset from a company, the technical solution of this application was compared with rule-based auditing methods in related technologies. The results showed that the accuracy of travel anomaly identification improved from 67.4% to 91.2%, and the recall rate improved from 62.3% to 88.5%. Particularly in issues such as "multi-city jumps across days" and "route conflicts," 81% of the abnormal behaviors that rule-based auditing methods could not identify were successfully captured by the technical solution of this application. At the same time, the proportion of reliance on manual review was reduced (by approximately 60%).

[0033] It is understandable that auditing expense and financial data based on the first major language model can be understood as: comparing the consistency of numerical values ​​and text in expense and financial data under at least one expense audit dimension based on the first major language model. Specifically, it can be a comparison of the numerical values ​​and text contained in the expense and financial data with the actual numerical values ​​and actual text (the actual numerical value can be a specific numerical value or a range of numerical values, and the actual text can be a specific text or a set of texts) to determine whether the numerical values ​​and text contained in the expense and financial data are the actual numerical values ​​and actual text and / or whether there is a compatibility between the numerical values ​​and text contained in the expense and financial data (for example, if transportation mode 1 is used for travel, the travel time is 5 hours). For example, this could involve comparing the travel region data (departure and destination), travel time data (departure and arrival times), and modes of transportation used in the travel financial data within the dimensions of region, time, and mode of transportation. For instance, under normal circumstances, the actual travel region data between departure and destination indicates the travel order as: departure, transit point 1, destination; the actual travel time interval is normally 5-7 hours; and the actual mode of transportation is 5 hours using mode 1 or 7 hours using mode 2. However, the user-submitted travel financial data shows departure, transit point 2, destination; the travel time interval is 2 hours; and the mode of transportation is 2 hours using mode 3. Then, a primary language model is used to perform semantic analysis on the travel region data, time intervals, and modes of transportation in the travel financial data to determine inconsistencies with the actual travel region data, actual time intervals, and actual modes of transportation. Therefore, the audit result can be determined to be that the financial data contains anomalies or that the financial data audit fails.

[0034] In the aforementioned process, automated identification of expense behaviors such as travel costs can be achieved through cross-order data aggregation and the construction of semantic reasoning chains using large language models. For example, auditing of issues such as consecutive city jumps, abnormal route sequences, unreachable modes of transportation, and conflicting time intervals can be achieved through a "semantic-space-time" ternary reasoning link, enabling expense behavior analysis across documents, modalities, and spatiotemporal dimensions. Moreover, the above process is automatically executed by the large language model without the need for explicit definition of path rules or enumeration conditions, overcoming the technical limitations of insufficient semantic understanding and limited rule expressiveness in related technologies. Therefore, it possesses stronger adaptability and generalization capabilities, as well as high readability and traceability, improving the depth and breadth of the system's identification of hidden risks.

[0035] In one embodiment, an electronic device can audit user ticket data through the following steps: S130-1: Obtain user ticket data of at least one modality in at least one format for the target user; S130-2: Perform at least one of the following processes on user bill data: clarity detection, noise reduction, tilt correction, and resolution enhancement, to obtain bill financial data; S130-3: Extract financial data from bills using a multimodal large language model to obtain structured bill field information; S130-4: Compare the structured invoice field information with the recorded invoice fee data to obtain the invoice audit results.

[0036] The results of the invoice audit can include: whether it is abnormal, the type of abnormality, the comparison fields, and the reason for generation.

[0037] It is understandable that user receipt data refers to receipt data generated from the spending behavior of target users within financial data. Furthermore, receipt audit results can be written to an audit result database for subsequent use by visualization and Q&A modules. User receipt data can be in at least one modality, such as image, text, or voice; that is, user receipt data can be user receipt images, user receipt text, or user receipt audio.

[0038] For example, structured invoice field information may include at least one of the following, but not limited to: invoice header, reimbursement recipient / user, invoice date, amount, currency, tax rate, description of goods or services, invoice number, tax ID, and travel itinerary information.

[0039] For example, comparing the structured invoice field information with the recorded invoice expense data may include at least one of the following, but not limited to: consistency between the header's legality and the affiliated unit, consistency between the user and the expense applicant, consistency between the invoice date and the travel / attendance data, comparison of amounts and anomaly judgment based on a preset deviation rate threshold, semantic matching between the product / service description and the expense type, uniqueness of the invoice number / tax number and identification of duplicate reimbursements, and consistency between the passenger's name and the order registrant in transportation invoices.

[0040] One approach is to first fine-tune the multimodal large language model using historical bill financial data, and then use the fine-tuned multimodal large language model to extract data from the bill financial data.

[0041] Specifically, we can first obtain a manually labeled first training sample (containing first historical invoice financial data and first data extraction results of the first invoice financial data; the data extraction results refer to structured invoice field information, and the first data extraction results are obtained by extracting data from the first historical invoice financial data using the target data extraction rule) and an unlabeled second training sample (containing only second historical invoice financial data), wherein the amount of data in the first historical invoice financial data is less than the amount of data in the second historical invoice financial data. Then, we can determine the target data extraction rule based on the first training sample. The target data extraction rule indicates the field identification method, field type, and generation method of the extraction results for the fields to be extracted from the invoice financial data. It can be a regular expression, but is not limited to it (for example, we can determine the original position of the first data extraction result in the first historical invoice financial data, obtain the context features of the adjacent positions in the first historical invoice financial data; obtain the preset historical regular expression used for data extraction and the historical context features corresponding to the data at the specific original position extracted in the historical regular expression; adjust the historical regular expression according to the context features and historical context features to obtain the regular expression corresponding to the first historical invoice financial data). Next, the multimodal large language model can be trained using the first training sample, the second training sample, and the actual second data extraction result obtained by extracting data from the second training sample using the target data extraction rule, to obtain the fine-tuned multimodal large language model.

[0042] Furthermore, a user ticket fingerprint can be generated for each ticket in the user ticket data. The ticket fingerprint is used to represent the ticket format information (e.g., visual style features, layout features, font texture features, character shape features, seal edge features, color features) and ticket content information (e.g., amount, date, merchant, unit, header, applicant). Specifically, it can be a high-dimensional vector obtained by deep feature extraction of the user ticket data using an encoder. Then, the similarity between the user ticket fingerprint and the fingerprint of a specific ticket can be calculated to obtain the fingerprint similarity. If the fingerprint similarity is greater than the similarity threshold, it is determined that the corresponding ticket has a risk of forgery (the risk of forgery means that the information represented by the ticket fingerprint involves or contains sensitive / unhealthy / discordant information), that is, the corresponding ticket fails the detection and needs to be reviewed again (e.g., detecting the numerical values ​​or text in the ticket, specifically indicating that the numerical values ​​and text in the ticket need to be reviewed again by manual review). Specifically, a specific invoice fingerprint refers to the fingerprint of any of the following types of invoices: fake / illegal / non-compliant / reimbursed / sensitive / risky / inconsistent invoices found on online platforms (e.g., fake invoices obtained from an illegally sold fake invoice database found on online platforms) or historical invoices with forgery risks. For example, when calculating the similarity between a user's invoice fingerprint and a specific invoice fingerprint, the two invoices may have different merchant names, but their invoice format information is similar, and they may have been printed by the same printer, by the same person, forged using the same forgery method, or from the same fake invoice database.

[0043] For example, electronic devices can first collect receipt data in various formats, such as train tickets, plane tickets, taxi receipts, taxi vouchers, accommodation invoices, office supplies purchase invoices, restaurant receipts, and VAT electronic general invoices in formats such as PDF, PNG, JPG, photographs, scanned images, and compressed files, through single upload or batch scanning. Then, the collected receipt images can undergo the following preprocessing: image sharpness detection, noise reduction, tilt correction, and resolution enhancement, thereby improving the subsequent multimodal large-scale model's accuracy in resolving key field regions. Next, the preprocessed receipt images can be input into the multimodal large-scale model (such as Qwen-VL or other multimodal large-scale models). The multimodal large-scale model can combine visual features and language pre-training capabilities to directly output structured field information (i.e., structured receipt field information), including: the name of the person claiming reimbursement / user; the invoice header and company name; the invoice date; the amount (including / excluding tax); product details / purpose description; service items (such as train number, itinerary, catering, accommodation, etc.); and receipt-specific fields such as tax number / invoice code / receipt number.

[0044] Next, anomaly detection and semantic comparison can be performed on the structured ticket field information. Specifically, the structured invoice field information can be semantically compared with the corresponding expense application data, budget information, employee files, etc. in the financial system, and the following audit logic can be executed: User consistency verification (identifying whether the reimbursement person in the invoice is consistent with the registered user), header legality verification (comparing whether the header matches the name of the employee's company / subsidiary), time consistency comparison (verifying whether the invoice time conflicts with the expense occurrence time and business trip time), amount deviation judgment (whether the actual reimbursement amount is consistent with the invoice amount, and whether there are cases of splitting invoices, merging, underreporting, overreporting, etc.), project / purpose semantic matching (e.g., if the invoice content for "office supplies" is used for "travel reimbursement", determine whether the purposes are consistent), duplicate reimbursement identification (comprehensively judging whether there is duplicate submission through invoice number, invoice code, amount, etc.), passenger consistency verification (verifying whether the "passenger" field in the invoice is consistent with the registered person in the order, and identifying whether there are issues of proxy booking, proxy reimbursement, or non-personal use), and amount and train number consistency verification (whether the extracted amount, departure time, train number, etc. are consistent with the reimbursement order, and identifying risks such as replaced invoices and duplicate reimbursements). Finally, the invoice audit results can be output as follows (similar to the expense audit results mentioned above): Is it abnormal?: true; "Reason for generation": The invoice header does not match the company's standard name; the passenger is not the employee registered in the order, posing a risk of invoice discrepancy.

[0045] In actual testing, using enterprise invoice samples, the accuracy rate of the proposed technical solution reached 97.7%, far exceeding the 83.1% of the traditional Optical Character Recognition (OCR) method; the anomaly recognition accuracy rate reached 93.2%, effectively reducing the amount of manual review by nearly 70%.

[0046] In the above process, a multimodal large model can be used to audit various types of invoices, such as travel, office, and reception invoices. It can achieve the ability to build integrated text and image data (i.e., automatically identify invoices submitted by employees and perform semantic comparison and consistency verification with the expense data already registered in the system). Compared with the method of using OCR for invoice auditing, the multimodal large model can complete visual understanding, semantic normalization, and structural output in one go. Moreover, it has a certain semantic fault tolerance and field position adaptation capability, and can deal with problems such as non-standard formats, layout drift, missing fields, or rearrangement, which can improve accuracy and robustness.

[0047] In one embodiment, large language models and multimodal model inference services can be deployed using AI acceleration chips (e.g., Ascend 910B) to provide high throughput and parallel computing capabilities. For example, large language models such as Qwen3-32B and multimodal models such as Qwen-VL can be directly deployed on servers or clusters composed of Ascend 910B chips, which will perform all calculations to obtain cost audit results and invoice audit results.

[0048] Furthermore, audits can be conducted in conjunction with expense financial data and invoice financial data, and / or, expense audit results and invoice audit results can be conducted in conjunction with each other. For example, an audit of invoice financial data and / or invoice audit results can be triggered when an anomaly is detected in expense audit results, and an audit of expense financial data and / or expense audit results can be triggered when an anomaly is detected in invoice audit results. Specifically, an audit of invoice financial data and / or invoice audit results for a specific time period can be triggered when an anomaly is detected in expense audit results for a certain period of time; similarly, an audit of expense financial data and / or expense audit results related to a particular merchant can be triggered when an anomaly is detected in invoice audit results for that merchant.

[0049] Regarding S140, in one embodiment, the electronic device can further acquire the input data query statement; determine the query category of the data query statement; and generate a data query response corresponding to the data query statement based on the query type and the data query statement. The data query statement can be a query statement in natural language form input through a terminal. The query type includes at least one of the following: financial audit query category, financial knowledge Q&A category, and non-financial related type.

[0050] For example, in response to a query of type "financial audit query," a data query response can be generated based on the data query statement, expense audit results, and invoice audit results. Specifically, information is extracted from the data query statement using a pre-defined second language model to obtain key query information; this key information is then converted into query language; and the query language is used to perform a query in the storage unit containing the expense audit results and invoice audit results to obtain the data query response.

[0051] The query language can be a database language, such as the database language corresponding to the storage unit. The data query response includes at least one of the following, but is not limited to: user identifier, time range, anomaly type, and corresponding audit result.

[0052] For example, a query language can be used to query the storage units of storage fee audit results and invoice audit results to obtain an initial query response; then, a preset third language model can be used to perform data analysis on the initial query response and data query statement to obtain a data query response based on a preset display format. The third language model can be the same as the first or second language model, but is not limited to this.

[0053] Specifically, the initial query response and data query statement can be filled into a preset prompt word template. The prompt word template can instruct the third language model to generate a data query response in a preset display format. The preset display format includes at least one display format, such as statistical charts (bar charts, pie charts, or line charts, etc.), tables, text (a piece of natural language), images, and files (Word files or PDF files), etc.

[0054] For example, the vLLM framework can be used to support the real-time response of the intelligent question-answering module corresponding to the above steps, thereby improving the overall throughput efficiency and user experience of the system. vLLM is a high-performance inference engine for deploying and inferring large language models, possessing high concurrency and low latency dynamic batch processing capabilities.

[0055] For example, users can input queries in natural language through the front-end interface, such as: whether employee 1 had any abnormal car usage in July of this year, which employees had duplicate bookings, why the invoice was deemed non-compliant, and what the company's reimbursement standards are for business entertainment expenses. Then, the electronic device can use an intent recognition module built on a large language model (such as Qwen-32B or other large models) to determine which type of query the above statement belongs to: abnormal data query (i.e., financial audit query type, such as: querying which financial data or which employee's financial data is abnormal), financial knowledge Q&A (such as: reimbursement policies, tax regulations), or non-financial related (the query content is outside the scope of financial audit). For non-financial related types, standardized guiding prompts can be generated as a data query response, such as: Please ask questions related to financial audit. For financial knowledge Q&A types, the query can be directly submitted to the model and combined with the system's built-in company policies / knowledge documents to generate a natural language answer. For abnormal data query types, the system can automatically invoke the parsing module based on prompt words in the large language model (which can adapt to new abnormal types or new question formats without requiring a large amount of labeled data) to extract the following key information: employee name (orderUserName), time range (orderTime), and abnormal type (warnType, including ride-hailing abnormality, food delivery abnormality, invoice discrepancy, amount abnormality, duplicate reimbursement, etc.). Then, based on Text2SQL, the above key information is mapped to an SQL query statement (or the query statement corresponding to other data call interfaces), and the system queries the abnormal database or audit logs. Additionally, for result explanation questions (e.g., why is this invoice non-compliant), the system can invoke the abnormal logs and model inference chain to output explanatory text with causal analysis.

[0056] Furthermore, electronic devices can accurately and comprehensively respond to queries through context tracking in multi-turn dialogues containing multiple user-input queries—that is, continuous question-and-answer sessions based on historical dialogues (managed by a history dialogue manager). A multi-turn dialogue can be structured as follows: User: Please check for any recent unusual activity from Li Si; System: Li Si had two abnormal food delivery transactions in July 2025; User: What is the reason? System: The food delivery order was submitted at 21:34 on July 13, but the clock-out time for the same day was 20:30, which is outside the reasonable time range.

[0057] In actual testing, through testing in an internal enterprise simulation environment, the technical solution of this application performed as follows: the accuracy rate of intent recognition reached 96.6%, the accuracy rate of structured parameter extraction for abnormal query tasks reached 94.1%, the consistency rate of generated answers for abnormal explanation questions reached 87.3%, and the user satisfaction score reached 4.6 / 5. In addition, compared with the fixed field query method, the user learning cost was reduced by about 70%.

[0058] In the above embodiments, intent recognition, semantic parsing using large language models, and structured query chains using database languages ​​can be integrated into the financial system, forming a closed loop from natural language to executable tasks. This supports queries across multiple scenarios, including anomaly queries, anomaly explanations, and knowledge-based question answering. It possesses strong scalability, portability, and interactive diversity, as well as high efficiency in anomaly localization and root cause analysis, enhancing system usability and interactivity. Simultaneously, it exhibits high automation and intelligence, reducing manual auditing workload while improving the efficiency and accuracy of anomaly identification. This effectively ensures the compliance of various financial expense reimbursements, reduces the risk of financial manpower investment and irregular expenditures, and provides solid technical support for enterprises to build an efficient, intelligent, and scalable financial risk control system, achieving the core goal of "cost reduction and efficiency improvement." Furthermore, an interpretable output strategy is adopted to clarify the reasoning process and judgment basis corresponding to the query response, enhancing system trustworthiness.

[0059] Based on the above embodiments, the system architecture and modules involved in this application are described below: In one embodiment, such as Figure 2 As shown, Figure 2This is a functional layered architecture diagram of the intelligent auditing system involved in the technical solution of this application. Based on a bottom-up, capability abstraction, and layered decoupling design approach, the intelligent auditing system is divided into a data layer, a service layer, an AI capability layer, and an application layer. The data layer aggregates structured data (such as travel orders, various expense reimbursement documents, attendance records, organizational information, employee information, etc.) and unstructured data (such as electronic invoices, image receipts, uploaded attachments, etc.) required for enterprise financial auditing, and constructs a high-quality data foundation for auditing through data integration, cleaning, and alignment. The service layer provides underlying resource scheduling and model operation support, covering large model loading and runtime environments (such as AI chip support like Ascend 910B), vLLM inference engines, rule and task scheduling services, model call interfaces, etc., providing a high-concurrency, low-latency, and highly available operational foundation for AI capabilities. The AI ​​capability layer integrates core capabilities of large models, such as semantic understanding, multimodal analysis, text generation, prompt word engineering, intent recognition, and Text2SQL-based structured queries, providing semantic and inference capabilities support for tasks such as anomaly detection, invoice comparison, and intelligent question answering. At the application layer, it can provide users with functions such as visual audit dashboards, intelligent financial assistants, abnormal behavior identification, invoice compliance review, and business knowledge Q&A, building an end-to-end closed-loop audit and interactive application scenario that covers a variety of financial expense types.

[0060] In one embodiment, on top of the aforementioned layered architecture, such as Figure 3 As shown, Figure 3This intelligent audit execution process, designed for enterprise financial expense scenarios, comprises a closed-loop workflow from raw data access to intelligent auditing and user interaction feedback. It can be executed by a multi-source data access and preprocessing module, an intelligent audit processing module, an abnormal audit result storage module, and an intelligent question-and-answer and interaction module. The multi-source data access and preprocessing module can uniformly access and preprocess various types of financial expense-related data within the enterprise. This data includes structured data such as expense reimbursement forms, expense details, employee and organizational information, budget indicators, attendance records, and expense type configurations, as well as unstructured data such as electronic invoices, screenshots of invoices, PDF invoices, and voucher attachments (images or documents). Through data integration, cleaning, standardization, and field alignment, a high-quality data foundation is built to ensure consistent, complete, and parsable data input for subsequent model processing. The intelligent audit processing module leverages the semantic understanding and reasoning capabilities of multimodal large-scale models and large-scale language models to automatically identify and intelligently judge various abnormal financial expense behaviors. Specifically, this includes, but is not limited to: auditing abnormal travel behavior, auditing the compliance of invoices, verifying the consistency between the user and the person claiming reimbursement, auditing high-frequency rescheduling / cancellation behavior, auditing fraudulent / proxy reimbursement behavior, auditing duplicate booking behavior, auditing abnormal car-hailing consumption behavior, and auditing abnormal food delivery consumption behavior. During the audit process, structural logic rules and semantic reasoning mechanisms can be integrated to ensure coverage of complex audit scenarios with strong contextual relevance and difficult-to-enumerate rules. The abnormal audit result storage module can uniformly store various abnormal behaviors identified by the audit module and their associated details, providing data support for subsequent analysis, query responses, and upstream and downstream system interface calls. The intelligent question-answering and interaction module can build natural language interaction capabilities based on a large language model. Through user intent recognition, structured query generation, text generation, and rule tracing capabilities, it can intelligently ask questions about audit results, obtain abnormal explanations, and conduct context-based multi-turn dialogues, achieving a more convenient and efficient human-computer interaction experience.

[0061] The technical solution of this application can realize steps from multi-source data fusion, intelligent audit processing, abnormal result storage to intelligent interactive feedback, deeply integrating the capabilities of large models with multi-source financial data. Based on large language models and multimodal large models, it integrates structured and unstructured financial data to achieve intelligent, automated and interpretable financial data audit.

[0062] Figure 4 This is a schematic diagram of a financial data auditing device 400 provided in an embodiment of this application. Figure 4 As shown, the financial data auditing device 400 includes: a data acquisition module 410, a data auditing module 420, and a data query module 430.

[0063] In one embodiment, the data acquisition module 410 is used to acquire the target user's expense financial data and invoice financial data; the data audit module 420 is used to perform data audit on the expense financial data through a preset first large language model to obtain the expense audit result; and to perform data audit on the invoice financial data through a preset multimodal large model to obtain the invoice audit result; the data query module 430 is used to acquire the input data query statement, and generate a data query response corresponding to the data query statement based on the data query statement, the expense audit result, and the invoice audit result.

[0064] For example, the data audit module 420 is specifically used to: add expense and financial data to a preset prompt word template to obtain expense and financial prompt words, the prompt word template being used to instruct the first language model to output expense audit results; and through the first language model, perform semantic reasoning on the expense and financial prompt words based on at least one expense audit dimension to obtain expense audit results; wherein, at least one expense audit dimension includes at least one of the following: region dimension, time dimension, expense type dimension, expense item dimension, budget control dimension, payment method dimension, personnel role dimension, and expense correlation dimension; the expense audit results include the audit results under at least one expense audit dimension, the anomaly type, and the data on the reasons for generating the corresponding audit results.

[0065] For example, the data acquisition module 410 is specifically used to: acquire user expense data of the target user within a preset time period, the user expense data including at least one of the following: expense order data, attendance data; aggregate the user expense data to obtain aggregated expense data; and format the aggregated expense data to obtain expense financial data in natural language form.

[0066] For example, the data audit module 420 is specifically used to: extract financial data from bills through a multimodal large model to obtain structured bill field information; and compare the structured bill field information with the recorded bill expense data to obtain the bill audit result.

[0067] For example, the data acquisition module 410 is specifically used to: acquire user ticket data of at least one modality in at least one format of the target user; and perform at least one of the following processes on the user ticket data: clarity detection, noise reduction, tilt correction, and resolution enhancement to obtain ticket financial data.

[0068] For example, the data query module 430 is used to: extract information from the data query statement through a preset second language model to obtain key query information; convert the key query information into query language; and use the query language to query in the storage unit storing the cost audit results and invoice audit results to obtain a data query response.

[0069] For example, the data query module 430 is specifically used to: use the query language to query in the storage unit storing the cost audit results and the invoice audit results to obtain an initial query response; and perform data analysis on the initial query response and the data query statement through a preset third language model to obtain a data query response based on a preset display format.

[0070] It should be understood that the embodiments at the device level are similar to the embodiments in the above-described methods, and their content and effects can be referred to the content and effects in the above-described methods, which will not be repeated here. Specifically, Figure 4 The apparatus 400 shown can execute the above-described method embodiments, and the aforementioned and other operations and / or functions of each module in the apparatus 400 are respectively for implementing the corresponding processes in the above-described methods. For the sake of brevity, they will not be described in detail here.

[0071] The apparatus 400 of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0072] Figure 5 This is a schematic diagram of an electronic device 500 provided in an embodiment of this application.

[0073] like Figure 5 As shown, the electronic device 500 may include: The system includes a memory 510 and a processor 520. The memory 510 stores computer programs and transfers the program code to the processor 520. In other words, the processor 520 can retrieve and run the computer program from the memory 510 to implement the methods described in the embodiments of this application.

[0074] For example, the processor 520 can be used to execute the above-described method embodiments according to instructions in the computer program.

[0075] In some embodiments of this application, the processor 520 may include, but is not limited to: General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0076] In some embodiments of this application, the memory 510 includes, but is not limited to: Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0077] In some embodiments of this application, the computer program may be divided into one or more modules, which are stored in the memory 510 and executed by the processor 520 to perform the method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0078] like Figure 5 As shown, the electronic device may further include: Transceiver 530, which can be connected to processor 520 or memory 510.

[0079] The processor 520 can control the transceiver 530 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 530 may include a transmitter and a receiver. The transceiver 530 may further include antennas, and the number of antennas may be one or more.

[0080] It should be understood that the various components in the electronic device are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0081] This application also provides a computer storage medium storing a computer program thereon, which, when executed by a computer, enables the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0082] When implemented using software, it can be implemented entirely or partially as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, the computer can perform all or part of the corresponding processes in the methods of the embodiments of this application, producing the functions achievable by the methods of the embodiments of this application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital video discs (DVDs)), or semiconductor media (e.g., solid state disks (SSDs)).

[0083] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems, apparatuses, or modules may be electrical, mechanical, or other forms.

[0085] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

Claims

1. A method for auditing financial data, characterized in that, include: Obtain the target user's expense and invoice financial data; The expense and financial data are audited using a pre-defined first language model to obtain the expense audit results. The financial data of the bills is audited using a pre-set multimodal large model to obtain the bill audit results; Obtain the input data query statement, and generate a data query response corresponding to the data query statement based on the data query statement, the expense audit result, and the invoice audit result.

2. The method according to claim 1, characterized in that, The process of auditing the expense and financial data using a pre-defined first language model to obtain the expense audit results includes: The expense and financial data are added to a preset prompt word template to obtain expense and financial prompt words. The prompt word template is used to instruct the first language model to output the expense audit results. Using the first large language model, semantic reasoning is performed on the expense financial prompt words based on at least one expense audit dimension to obtain the expense audit result; The at least one cost audit dimension includes at least one of the following: region dimension, time dimension, cost type dimension, cost item dimension, budget control dimension, payment method dimension, personnel role dimension, and cost relevance dimension; The cost audit results include audit results under at least one cost audit dimension, anomaly types, and data on the reasons for generating the corresponding audit results.

3. The method according to claim 2, characterized in that, The acquisition of the target user's financial data includes: Obtain user expense data of the target user within a preset time period, wherein the user expense data includes at least one of the following: expense order data and attendance data; The user fee data is aggregated to obtain aggregated fee data; The aggregated cost data is formatted to obtain cost financial data in natural language form.

4. The method according to claim 1, characterized in that, The process of auditing the financial data of the bills using a pre-set multimodal large model to obtain the bill audit results includes: The financial data of the bills is extracted using the multimodal large model to obtain structured bill field information; The structured invoice field information is compared with the recorded invoice fee data to obtain the invoice audit result.

5. The method according to claim 4, characterized in that, Obtaining the target user's invoice financial data includes: Obtain user ticket data of at least one modality in at least one format for the target user; The user ticket data is processed by at least one of the following methods: clarity detection, noise reduction, tilt correction, and resolution enhancement, to obtain the ticket financial data.

6. The method according to any one of claims 1-5, characterized in that, The step of generating a data query response corresponding to the data query statement based on the data query statement, the expense audit result, and the invoice audit result includes: The data query statement is extracted using a pre-defined second language model to obtain key query information; Convert the key query information into query language; The query language is used to perform a query in the storage unit that stores the cost audit results and the invoice audit results to obtain the data query response.

7. The method according to claim 6, characterized in that, The step of using the query language to query the storage unit storing the expense audit results and the invoice audit results to obtain the data query response includes: The query language is used to perform a query in the storage unit that stores the cost audit results and the invoice audit results to obtain an initial query response; The initial query response and the data query statement are analyzed using a preset third language model to obtain a data query response based on a preset display format.

8. A financial data auditing device, characterized in that, include: The data acquisition module is used to acquire the target user's expense and invoice financial data. The data audit module is used to audit the expense and financial data using a preset first language model to obtain the expense audit results. The financial data of the bills is audited using a pre-set multimodal large model to obtain the bill audit results; The data query module is used to obtain the input data query statement, and generate a data query response corresponding to the data query statement based on the data query statement, the expense audit result and the invoice audit result.

9. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the method of any one of claims 1-7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

Images