Financial shared single auditing method and system based on AI large model technology

Abstract

This invention relates to the field of intelligent auditing technology, specifically to a financial shared auditing method and system based on AI large-scale model technology. The method includes the following steps: collecting document information and attachment fields; extracting key elements and establishing binding relationships; screening similar fields for consistency comparison; marking credibility results; executing judgments according to rules to generate an evidence chain; summarizing judgment records to determine the audit status; and outputting audit conclusion records. In this invention, by extracting key fields from various attachments in reimbursement documents, data traceability and field credibility are enhanced. Anomalies are identified through field type comparison, improving the consistency judgment capability of multi-source data. Comparison results are marked with credibility levels, supporting hierarchical processing of rule judgments. Field values, judgment logic, parameters, and results form a binding chain, enhancing the verifiability of judgments. The audit conclusion, combined with field identifiers and rule results, is output, enhancing the accuracy, compliance, and review support of automatic auditing.

Description

Technical Field

[0001] This invention relates to the field of intelligent auditing technology, and in particular to a financial shared auditing method and system based on AI big data model technology. Background Technology

[0002] The field of intelligent auditing technology involves data processing and business judgment methods for auditing and reviewing business in corporate financial management activities. Its core aspects include the collection, identification, verification, and compliance judgment of expense reimbursement documents, accounting vouchers, and related attachments. Typically based on a financial shared service model, it unifies the processing of multi-source and multi-type financial data, completing the review process through established business rules and audit procedures. Overall, it belongs to the methodology system of using information technology to replace or assist manual review in accounting and auditing. Traditional financial shared service review refers to judging the matching relationship between amount, time, and item based on preset review rules during the expense reimbursement review process. This method mainly targets structured or semi-structured invoice information. For non-invoice attachments such as hotel bills, itineraries, and meal receipts, the review is usually completed through fixed field identification and rule comparison. Related technical aspects focus on text extraction, element matching, and rule verification of attachment content, mainly relying on character recognition results and rule bases for review processing.

[0003] Traditional financial shared service document review processes rely primarily on character recognition for expense attachments. This leads to a lack of unique identification of fields across multiple sources, especially when multiple unstructured attachments are present, blurring the field attribution relationships and affecting the accuracy of subsequent judgments. In terms of comparison logic, existing methods often employ simple rules for matching numerical values ​​and times between fields, lacking a comprehensive verification mechanism for semantic relationships between fields, making it difficult to identify anomalies across attachments. For non-standard documents such as hotel bills and itineraries, the reliance on fixed field recognition results in omissions and misjudgments when fields are deformed or have abnormal formats, impacting the scope of the review. The current document review process lacks a mechanism for expressing and tracing the credibility of fields, making the determination of whether a field passes or fails difficult to interpret and lacking a chain of evidence structure, leaving review conclusions without a basis for verification. In practice, discrepancies between invoice amounts and the main reimbursement form, or contradictions in personnel information among attachments, will be missed due to the lack of fine-grained comparison and verification mechanisms, creating audit loopholes and increasing the exposure of corporate financial risks. Summary of the Invention

[0004] To address the technical problems existing in the current technology, embodiments of the present invention provide a financial shared auditing method based on AI large model technology, comprising the following steps: S1: Collect accounting computerization information, invoices and non-invoice attachments, extract date, amount, personnel, item, and quantity fields from hotel bills, itineraries, and meal receipts, record field attachment identifiers, establish the binding relationship between fields and documents, and form a set of document review fields; S2: Filter fields of the same type from the document review field set, determine whether the date is consistent, whether the amount is consistent, whether the personnel are consistent, and whether the quantity is matched in pairs, and record the comparison results to construct a field consistency set; S3: Based on the classification and comparison results of the field consistency set, mark consistent fields as passed, and inconsistent or uncompared fields as pending verification. Establish a correspondence between the marking results and the fields to form a credibility labeling result. S4: Call the credibility labeling result, apply the corresponding rules according to the expense type, perform limit judgment on the amount field, period judgment on the date field, and subject consistency judgment on the personnel field, and bind the field value, rule parameters and judgment result to form a rule judgment evidence chain; S5: Based on the rules, determine the evidence chain, collect all judgment records of the document, determine the document review status, and associate the status with the field identifier, rule number, and credibility result, and output the review conclusion record of the document that has passed and been reviewed.

[0005] As a further aspect of the present invention, the document review field set includes a master document information field, invoice attachment field entries, non-invoice attachment field entries, attachment identifier mapping information, and a field-document binding relationship table; the field consistency set includes date consistency results, amount consistency results, personnel consistency results, quantity matching results, and field comparison result records; the credibility labeling results include passed labeling items, unverified labeling items, and a field-label correspondence table; the rule judgment evidence chain includes field value snapshots, rule parameter sets, rule judgment conclusions, rule-field association items, and evidence node sequences; the document review conclusion record includes document review status codes, field identifier lists, rule number lists, credibility result summary items, passed conclusion information, and review conclusion information.

[0006] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Obtain accounting computerization information and the corresponding invoice and non-invoice attachments, call the expense type, applicant, reimbursement date and attachment identifier in the main order field, classify the attachment content according to the attachment identifier, separate and record invoice and non-invoice image data, and establish unified mapping data to generate attachment classification mapping data. S102: Based on the attachment classification mapping data, extract the date, amount, personnel, items and quantity fields from the hotel bill, itinerary and meal receipt images, call the text in the image recognition results, combine the field content structure features, record the attachment identifier corresponding to the field, and generate field location identifier information; S103: Based on the field location identification information, construct the binding structure between field content and attachments, call the attachment identification content, organize the correspondence between field attributes and field values, assign the content in the field set to the attachment number, uniformly classify it into the document field structure table, and generate the document review field set.

[0007] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the document review field set, filter field groups with consistent field types, call the field type identifier and the corresponding attachment identifier, and assign the date, amount, personnel, and quantity fields to their respective groups to generate a field type grouping sequence; S202: Group the sequence according to the field type, perform consistency judgment on the field values ​​in the field group, judge the time content for the date field, the numerical content for the amount field, the identification content for the personnel field, and the pair matching status for the quantity field, summarize all judgment results and record the consistency status to obtain the field comparison mark; S203: Based on the field comparison markers, organize the comparison status corresponding to the field group, call the field type and identifier content to establish a combination structure of field items and consistency status, output the data structure format in a unified manner according to the field type, and generate a field consistency set.

[0008] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the field consistency set, filter the comparison status records corresponding to the fields, call the field identifier and consistency status mark, distinguish and organize the fields that participated in the comparison and those that did not participate in the comparison, and uniformly collect the field comparison status into the field-level record item to form a status list that can be used for classification and processing, and generate a summary value of the field comparison status. S302: Based on the field comparison status summary value, determine the corresponding comparison status type of the field, assign a pass mark to fields with consistent status, assign a pending verification mark to fields with inconsistent status and fields not participating in the comparison, and organize the mark results and field identifiers to form a field-level identifier record structure to obtain the field status identifier value. S303: Based on the field status identifier value, bind the field identifier result with the original field content, call the field identifier, field type and annotation status to combine and organize, output the field and annotation relationship record structure in a unified manner, establish the field credible annotation data structure, and generate credible annotation results.

[0009] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Call the credibility labeling result to obtain the field identifier, field type and labeling status content, filter the corresponding field set according to the expense type record item, organize the correspondence between the amount field, date field and personnel field, establish a mapping structure between expense type and field type, form a list of fields that can be called by rules, and generate rule field mapping values; S402: Based on the rule field mapping value, call the corresponding rule parameter according to the expense type, determine whether the amount field meets the limit threshold, determine whether the date field falls within the cycle benchmark range, determine whether the personnel field is consistent with the main identifier, and associate and record each judgment result with the field value and rule parameter to obtain the rule judgment record; S403: Based on the rule judgment record, organize the correspondence between field values, rule parameters and judgment status, call the field identifier to collect multiple judgment records by field dimension, establish the field judgment process association structure, form a continuous record chain from field to rule, and generate a rule judgment evidence chain.

[0010] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the rule-based evidence chain, collect all rule-based records corresponding to the current expense reimbursement document, call the document identifier and rule number, collect and organize the judgment status associated with the fields, summarize the scattered records to the document dimension, and form a set of rule records that can be used for status determination, and generate a document rule summary value. S502: Based on the document rule summary value, determine the overall document review status, call the rule to determine the status and credibility labeling results, classify the rule records by status, determine whether the document is in the pass or review status based on the classification results, and record the status to the document-level data structure to obtain the document review status. S503: Based on the document review status, the review status is associated with the field identifier, rule number and credibility result. The document identifier is called to combine multiple types of associated information in a unified manner to form a structured review conclusion record. Traceable document processing result data is output to generate the review conclusion record.

[0011] As a further aspect of the present invention, the field location identification information is obtained by extracting the text content and image area coordinates from the attached images of hotel bills, itineraries, and meal receipts using image recognition technology. Combined with predefined field templates, the identified fields are located and named, and the field text content and its corresponding attachment identifiers are recorded for subsequent binding of fields to documents and consistency judgment processing.

[0012] As a further aspect of the present invention, the document review status is determined based on the collected rules and the credibility labeling results of the records and fields. A preset status judgment logic model is used to classify and judge the status. If all rules judge the record as passed and the credibility score of the associated field is higher than the set threshold, the document is marked as passed; otherwise, the document is marked as reviewed and recorded in the document-level data structure for subsequent traceability.

[0013] A financial shared auditing system based on AI large-scale model technology includes: The field collection module is used to execute S1: collect computerized accounting information for expenses, invoice attachments and non-invoice attachments. In hotel bills, itineraries and meal receipts, it identifies the date, amount, personnel, items and quantity fields, records the attachment identifiers corresponding to the fields, builds the binding relationship between fields and documents, and forms a set of audit fields. The field consistency module is used to execute S2: filter field groups with the same field type from the document review field set, determine whether the time of the date field is the same, determine whether the value of the amount field is consistent, determine whether the identifier of the personnel field is consistent, determine whether the value of the quantity field is matched in pairs, and record the field comparison results to form a field consistency set. The trust labeling module is used to perform S3: based on the field consistency set, classify the comparison results of each field, mark the fields that match the comparison as passed, mark the fields that do not match the comparison or did not participate in the comparison as pending verification, and establish a correspondence between the labeling results and the fields themselves to form a trust labeling result; The rule verification module is used to execute S4: call the credibility labeling result, call the corresponding rule according to the expense type, perform limit condition judgment on the amount field, perform period range judgment on the date field, perform subject consistency judgment on the personnel field, and bind the field value, rule parameter and judgment result during the rule execution process to form a rule judgment evidence chain; The status output module is used to execute S5: based on the rules, determine the evidence chain, collect all rule judgment records of the current expense reimbursement document, determine the document's review status, associate the status with field identifiers, rule numbers, and credibility results, and output the review conclusion record of passing and reviewing.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by extracting key fields from various attachments in expense reimbursement documents and recording their sources, the traceability and credibility of data are enhanced. Anomalies are identified through field type comparison, improving the consistency judgment capability of multi-source data. The comparison results are labeled with credibility levels, supporting hierarchical processing of rule judgments. Field values, judgment logic, parameters, and results form a binding chain, improving the verifiability of judgments. The review conclusion is combined with field identifiers and rule results for aggregated output, enhancing the accuracy, compliance, and review support of automatic review. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a financial shared auditing method based on AI large model technology, including the following steps: S1: Collect computerized accounting information for expenses, invoice attachments and non-invoice attachments. In hotel bills, itineraries and meal receipts, identify the date, amount, personnel, items and quantity fields, record the attachment identifiers corresponding to the fields, build the binding relationship between fields and documents, and form a set of document review fields. Expense reimbursement forms refer to data forms generated and submitted in the financial shared service system for the review of expense reimbursements. They consist of expense items, expense types, structured information of the person making the reimbursement, and associated invoice attachments and non-invoice attachments. Master order information refers to the structured field information bound to the expense reimbursement document, which is used to represent at least the expense item content, expense type identifier, reimbursement personnel identifier, and application time business elements; Non-invoice attachments refer to image or text vouchers that are related to expense reimbursement but do not have the attributes of tax invoices, and originate from hotel bills, itineraries, and meal receipts submitted by the person seeking reimbursement. The document review field set refers to the set of field information identified from the invoice attachments and non-invoice attachments corresponding to the expense reimbursement document, which is used to record the field values ​​and field types. S2: Filter field groups with the same field type from the document review field set, determine whether the time of the date field is the same, whether the value of the amount field is consistent, whether the identifier of the personnel field is consistent, and whether the value of the quantity field is matched in pairs, and record the comparison results of the fields to form a field consistency set; The field consistency set refers to the result set formed after performing consistency judgment on fields of the same type in different attachments under the same expense reimbursement document. It is used to characterize the judgment status of whether the fields are consistent. S3: Based on the field consistency set, classify the comparison results of each field, mark the fields that match the comparison as passed, mark the fields that do not match the comparison or did not participate in the comparison as pending verification, and establish a correspondence between the marking results and the fields themselves to form the credibility labeling results; The credibility labeling result refers to the field credibility status identifier formed after classifying the fields according to the field consistency set. It is used to characterize whether the field has passed the consistency verification or is in a state of pending verification. S4: Call the credibility labeling results, call the corresponding rules according to the expense type, perform limit condition judgment on the amount field, perform period range judgment on the date field, and perform subject consistency judgment on the personnel field. During the rule execution process, bind the field values, rule parameters and judgment results to form a rule judgment evidence chain. Rule parameters refer to the benchmark information used to make judgments during the expense reimbursement review process. They are derived from the expense management system or review rule configuration content corresponding to the expense type. The rule-based judgment evidence chain refers to a structured record set formed by associating the rule execution order with the field information, field values, rule parameters, and judgment results involved in the expense reimbursement review process. S5: Based on the rule-based evidence chain, collect all rule-based records for the current expense reimbursement document, determine the document's review status, and associate the status with field identifiers, rule numbers, and credibility results, outputting the review conclusion record of "passed" and "reviewed". The review status refers to the review result identifier formed after the expense reimbursement document has completed the rule judgment, which is used to represent the current review conclusion of the document in the review process. The review conclusion record refers to a structured record used to store the review results of expense reimbursement documents and the corresponding judgment basis, which is used to support query, approval and review operations.

[0023] The document review field set includes the master document information field, invoice attachment field entries, non-invoice attachment field entries, attachment identifier mapping information, and a field-document binding relationship table; the field consistency set includes date consistency results, amount consistency results, personnel consistency results, quantity matching results, and field comparison result records; the credibility labeling results include passed labeling items, pending verification labeling items, and a field-labeling correspondence table; the rule judgment evidence chain includes field value snapshots, rule parameter sets, rule judgment conclusions, rule-field association items, and evidence node sequences; the document review conclusion record includes the document review status code, field identifier list, rule number list, credibility result summary item, passed conclusion information, and review conclusion information.

[0024] Please see Figure 2 The specific steps of S1 are as follows: S101: Obtain accounting computerization information and the corresponding invoice and non-invoice attachments, call the expense type, applicant, reimbursement date and attachment identifier in the main order field, classify the attachment content according to the attachment identifier, separate and record invoice and non-invoice image data, and establish unified mapping data to generate attachment classification mapping data. For VAT invoice images in uploaded attachments, the existing VAT invoice recognition and management module is invoked to parse the VAT invoice information in the attachment images, and invoice verification and deduplication operations are completed simultaneously. The VAT invoice recognition results and verification conclusions are directly incorporated into the reimbursement master data as structured data and do not enter the subsequent attachment parsing process. After VAT invoice processing is completed, the remaining attachment images that are not recognized as VAT invoices are uniformly classified as non-invoice attachments, and these attachment images are pushed to the attachment parsing module of this invention for further processing. To achieve traceable management of attachment categories and field sources, a classification mapping relationship is established for all attachments based on attachment identification information. Attachments that have completed VAT invoice recognition are marked as invoices, and the remaining attachments are marked as non-invoices, and attachment classification mapping data is generated. For example, attachments identified as "F001" are mapped to the invoice category, and attachments identified as "F002" and "F003" are mapped to the non-invoice category, forming the following mapping structure: {"F001" → "Invoice Category", "F002" → "Non-Invoice Category", "F003" → "Non-Invoice Category"}. This attachment classification mapping data is used in subsequent field extraction processes to bind and manage the attachments from which fields originate, ensuring the uniqueness and traceability of field identification results in multi-attachment scenarios.

[0025] S102: Based on the attachment classification mapping data, extract the date, amount, personnel, items and quantity fields from the hotel bill, itinerary and meal receipt images, call the text in the image recognition results, combine the field content structure features, record the attachment identifier corresponding to the field, and generate field location identifier information; Based on the attachment classification mapping data, key information is extracted from non-invoice attachments sequentially according to their attachment identifiers. For each attachment image, image recognition and text parsing capabilities are used to obtain the text content. Furthermore, the information extraction capabilities of a large model are invoked, combining the semantics and contextual structure of fields in the expense reimbursement scenario to perform structured parsing of the attachment text, extracting key fields including date, amount, personnel, matter, and quantity. During field extraction, no preset templates or fixed coordinate positioning rules are relied upon; instead, fields are identified based on their semantic features, contextual relationships, and text structure, establishing a correspondence between each field and its source attachment identifier. To ensure the traceability of field results, evidence citation information supporting the field value is recorded simultaneously with the generation of the field result. This evidence citation information includes at least one of the following: field-hitting fragment, keyword triggering basis, and the text paragraph or line block index where the field is located, forming field positioning identification information. For example, for the itinerary image with attachment identifier "F002", information such as "travel date", "departure / destination", "fare amount", and "passenger" is extracted from the recognized text, and the extracted date field, amount field, and person field are bound to "F002" respectively; for the meal receipt image with attachment identifier "F003", fields such as "consumption date", "consumption amount", "number of diners / quantity", and "consumption items" are extracted, and the field results are bound to "F003" to form field location identification information, which is used for subsequent field source binding and consistency comparison processing.

[0026] S103: Based on the field location identifier information, construct the binding structure between field content and attachments, call the attachment identifier content, organize the correspondence between field attributes and field values, and uniformly classify the content and the associated attachment number in the field set into the document field structure table to generate the document review field set; The field coordinates of the retrieved record are bound to the attachment number information for field content binding. Using "520.00" as the field value, "Amount" as the field type, and "F001" as the source attachment number, the record structure is constructed as follows: {Field Type: "Amount", Field Value: "520.00", Attachment: "F001"}. The fields "Date", "Personnel", "Item", and "Quantity" are then organized into the structure and added to the field set in the same way, forming the following field structure table: Field values ​​correspond one-to-one with field types, and are uniquely bound to the attachment number. For example, the field "Item" has coordinates [154, 212] in image "F001" and its value is "Accommodation". The field content and attribution information are then merged and uniformly included in the field structure table, forming the preliminary review field set.

[0027] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the document review field set, filter field groups with consistent field types, call the field type identifier and the identifier of the attachment to which they belong, and classify the date, amount, personnel, and quantity fields into their corresponding groups to generate a field type grouping sequence; Based on the document review field set, the system first reads the field type identifier, field value, field source attachment number, and the field's sort index in the attachment from the field structure table. The field type is then used as the first-level filtering condition for an equality matching operation. The field type set is set as {Amount, Date, Personnel, Quantity}. When a field type string exactly matches an item in the set, the record is written to the corresponding field type cache. For example, when reading the field record {Field Type = Amount, Field Value = 520.00, Attachment = F001, Serial Number = 1}, it is written to the amount field cache. Subsequently, {Field Type = Amount, Field Value = 120.00, Attachment = F002, Serial Number = 1} and {Field Type = Amount, Field Value = 89.50, Attachment = F003, Serial Number = 1} are read and... The records are sequentially written to the same cache area to complete the construction of the amount field group. Then, the records in the amount field group are sorted in ascending order according to the attachment number to fix the field arrangement order as F001, F002, F003, thus completing the order unification process. The same steps are used to filter the date field, reading the field values ​​"2023-12-24", "2023-12-24", and "2023-12-25", which are respectively assigned to the date field cache area and sorted by attachment number. The personnel field reads "Zhang San", "Zhang San", and "Zhang San" and completes the aggregation. The quantity field reads "1", "1", and "1" and completes the aggregation. Finally, a field type grouping sequence structure is generated. This structure uses the field type as the key and the ordered field record list as the value to form a stable field grouping sequence, which is used for subsequent group-by-group consistency judgment.

[0028] S202: Group the sequence according to field type, perform consistency judgment on the field values ​​in the field group, judge the time content for date field, the numerical content for amount field, the identification content for personnel field, and the pair matching status for quantity field, summarize all judgment results and record the consistency status to obtain the field comparison mark; After obtaining the field type grouping sequence, consistency checks are performed on the field values ​​within each field group one by one. First, the three field values ​​520.00, 120.00, and 89.50 in the amount field group are read. The consistency check interval is set to ±10.00 yuan. Pairwise difference calculation is performed. The difference between 520.00 and 120.00 is calculated to be 400.00. It is then checked whether 400.00 is less than or equal to 10.00. The result is no, and the comparison result is recorded as inconsistent. Next, the difference between 120.00 and 89.50 is calculated to be 30.50, which is also greater than 10.00, and is recorded as inconsistent. The overall consistency flag for the amount field group is set to 0. Then, the date field group is read, and the three date values ​​are converted to... In the timestamp format, 2023-12-24 corresponds to timestamp T1, and 2023-12-25 corresponds to timestamp T2. The maximum time difference is calculated as |T2-T1|=1 day. The allowed interval for date consistency is set to ≤1 day. The judgment result is that the condition is met, and the record date field group is marked as 1 for consistency. In the personnel field group, all three field values ​​are "Zhang San". The length and character encoding of the three strings are compared character by character. No difference is found, and the record is marked as 1 for consistency. In the quantity field group, all three field values ​​are the value 1. The value equality judgment is performed. The three values ​​are exactly the same, and the record is marked as 1 for consistency. The judgment results of each field group are summarized to form a field comparison mark set, where amount=0, date=1, personnel=1, and quantity=1.

[0029] S203: Based on the field comparison markers, organize the comparison status corresponding to the field group, call the field type and the identifier content to establish the combination structure of field items and consistency status, output the data structure format in a unified manner according to the field type, and generate a field consistency set; After the field comparison markers are generated, the field types are structurally bound to their corresponding consistency markers. Each record in the field comparison marker set is read. For example, when reading {field type = amount, marker = 0}, it is associated with all fields in the amount field group, and the overall comparison status of the amount field item is recorded as inconsistent. Then, {field type = date, marker = 1} is read, and the overall status of the date field item is marked as consistent. The same processing is performed on the personnel field and the quantity field. During the processing, the one-to-one correspondence between the field type and the consistency marker is maintained, and the field values ​​are not modified. Only the consistency status attribute is appended, and finally, a field consistency set structure is formed. Each field type in this structure corresponds to a clear consistency status value, which is used for subsequent field-level status summary processing.

[0030] Please see Figure 4 The specific steps of S3 are as follows: S301: Based on the field consistency set, filter the corresponding comparison status records of the fields, call the field identifier and consistency status mark, distinguish and organize the fields that participated in the comparison and those that did not participate in the comparison, and uniformly collect the field comparison status into the field-level record items to form a status list that can be used for classification and processing, and generate a summary value of the field comparison status. Based on the field consistency set, the comparison status at the field level is reorganized. First, the field type and status value in the field consistency set are read, and the status value is expanded to the field-level record item. For example, the amount field group contains three field records: F001-amount, F002-amount, and F003-amount. Since the consistency of the amount field is marked as 0, the comparison status of all three fields is recorded as inconsistent. The consistency of the date field group records F001-date, F002-date, and F003-date is marked as 1. The personnel field and the quantity field are expanded in the same way. If there are fields that have not entered the field consistency set, their status is assigned as not participating. Finally, a field comparison status summary value structure is formed. This structure uses the field identifier as an index to record its corresponding comparison status value, providing a direct basis for subsequent status labeling.

[0031] S302: Based on the summary value of field comparison status, determine the corresponding comparison status type of the field, assign a pass mark to fields with consistent status, assign a pending verification mark to fields with inconsistent status and fields not participating in the comparison, organize the mark results and field identifiers one by one to form a field-level identifier record structure, and obtain the field status identifier value. Based on the summary value of field comparison status, the field status identifier assignment operation is performed. The field identifier and its corresponding comparison status value are read one by one. When the comparison status value is equal to 1, the field status is marked as passed. When the comparison status value is equal to 0 or equal to not participated, the field status is marked as pending verification. For example, the comparison status value of field F001-Amount is 0, so its status is marked as pending verification. The comparison status value of field F001-Date is 1, so its status is marked as passed. In this process, the field values ​​are not modified. Only the status attribute is appended to the field record. Finally, a set of field status identifier values ​​is formed. This set completely records the current status of each field.

[0032] S303: Based on the field status identifier value, bind the field identifier result with the original field content, call the field identifier, field type and annotation status to combine and organize, output the field and annotation relationship record structure in a unified manner, establish the field credible annotation data structure, and generate credible annotation results; The field status identifier value is bound to the original field content. The field identifier, field type, field value, and status label information are read and integrated into a unified field trusted label record. For example, a structured record {Field Identifier=F001-Amount, Field Value=520.00, Field Type=Amount, Label Status=Pending Verification} is generated. After performing the same operation on all fields, a set of field trusted label data structures is formed. Each record in this set retains a stable association between the field source attachment number, field value, and label status.

[0033] Please see Figure 5 The specific steps of S4 are as follows: S401: Call the credibility labeling results, obtain the field identifier, field type and labeling status content, filter the corresponding field set according to the expense type record item, organize the correspondence between the amount field, date field and personnel field, establish a mapping structure between expense type and field type, form a list of fields that can be called by rules, and generate rule field mapping values; First, extract field identifiers, field types, and corresponding status annotation information from the credibility annotation results. Let the field credibility status set be {"F001-Amount"-"Pending Verification", "F001-Date"-"Passed", "F001-Personnel"-"Passed", "F002-Amount"-"Passed"}, etc. Then, call the expense type record item "Travel Expenses". This expense type is predefined in the system with a bound field type set of {Amount, Date, Personnel}. Based on the expense type, call the field type mapping list, filtering fields whose types belong to the above three categories, removing irrelevant fields such as quantity, and retaining the field types "Amount", "Date", and "Personnel". Perform a mapping structure generation operation on each field. For example, for the field F001-Amount, the record structure is {Expense Type: "Travel Expenses", Field Type: "Amount", Field Identifier: "F001-Amount"}. Then, perform the following operations on fields such as "F001-Date" and "F001-Personnel". Similarly, the above structure is aggregated into a field call list set for rule parameter input preparation. Fields in the field call list must be marked as "Passed" or "Pending Verification". Fields with a status of "Not Participated" will be marked as "Rule Not Applicable" and excluded from the field list. Based on the mapping relationship, the following structure record is established: [Travel Expenses → Amount → F001-Amount, Travel Expenses → Date → F001-Date, Travel Expenses → Personnel → F001-Personnel]. At the same time, a rule field mapping value table is established. This table is used to call the rule parameter template and trigger the field matching process. All fields in the mapping value table must exist in the trusted annotation structure. If a field does not appear in the annotation result, it will be marked as "Field Missing" and a prompt message will be generated. The final mapping value is in the following format: {Expense Type = Travel Expenses, Field Set = [Amount: F001-Amount, Date: F001-Date, Personnel: F001-Personnel]}, which serves as the input for stage S402.

[0034] S402: Based on the rule field mapping value, call the corresponding rule parameters according to the expense type, determine whether the amount field meets the limit threshold, determine whether the date field falls within the period benchmark range, determine whether the personnel field is consistent with the main identifier, and associate each judgment result with the field value and rule parameters to obtain the rule judgment record; Based on the rule field mapping value, the rule parameter template bound to the expense type "Travel Expenses" is called. The template has the following parameters: Amount Limit Threshold = 1000.00 yuan, Date Tolerance Range = 30 days, Personnel Identifier Matching Value = "Zhang San". Fields in the field call list are read sequentially and field-by-field judgment is performed. The amount field F001 has a value of 520.00. The judgment logic is: mark as True when the field value is less than or equal to the threshold 1000.00, otherwise as False. The condition 520.00 ≤ 1000.00 is true, and the judgment status is marked as True. Next, F002 - Amount, has a value of 120.00, which is also less than the limit, so it is judged as True. The date field is then processed. F001 - Date Field has a value of "2023-12-24", and the reimbursement master form date is "2023". "-12-25" converts the two dates to T1=1703376000 and T2=1703462400 using timestamps. Execution |T1-T2|=86400 seconds, converted to 1 day, is less than the 30-day tolerance range, so it is judged as True. The personnel field F001-personnel value is "Zhang San". The personnel matching requirement is set to complete character matching and length matching. After comparison, it is confirmed to be consistent and marked as True. The above judgment operation is recorded as a rule judgment record array: [F001-amount-True, F001-date-True, F001-personnel-True]. Each record includes field values, judgment status, parameter thresholds, field identifiers and other meta-information, forming a rule judgment structure. For each field rule judgment, if the execution is successful, it records whether the field meets the parameter settings under the current expense type rule template.

[0035] S403: Based on rule-based judgment records, organize the correspondence between field values, rule parameters and judgment status, call field identifiers to aggregate multiple judgment records by field dimension, establish a field judgment process association structure, form a continuous record chain from field to rule, and generate a rule judgment evidence chain; Based on the rule-based judgment records, a field dimension aggregation operation is performed. All judgment items under the same field identifier are merged to construct a field judgment chain structure. Taking field F001-Amount as an example, the rule involved is the amount limit. The record judgment structure is {Field value: 520.00, Parameter: 1000.00, Status: True}. This structure is incorporated into the field judgment chain. If this field participates in multiple rule judgments (e.g., it also needs to judge the minimum amount or currency), each judgment is appended to the same chain. In the current example, the amount field only matches the limit rule, so the chain length is 1. The personnel field judgment chain structure is {Value: "Zhang San", Comparison target: "Zhang San", Status: True}. The date field judgment chain records have a time difference of 1 day and a tolerance of 30 days. The status is True. All field judgment chain structures are organized into an array structure, grouped and collected by field type to form a field judgment process association structure. The structure example is as follows: {field: "amount", chain: [(520.00≤1000.00=True)]}, {field: "personnel", chain: [(Zhang San==Zhang San=True)]}, {field: "date", chain: [(1≤30=True)]}. The judgment status in all chains retains Boolean records for subsequent logical operations. Finally, the fields are output to the set of continuous record chains of the rules for S501 to execute document-level status collection.

[0036] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the rule-based evidence chain, collect all rule-based records corresponding to the current expense reimbursement document, call the document identifier and rule number, collect and organize the judgment status associated with the fields, summarize the scattered records to the document dimension, and form a set of rule records that can be used for status determination, and generate a document rule summary value. The field judgment chain structure is aggregated at the document level. All field judgment records corresponding to the current document number "R20231225001" are read. The field identifier, rule number, and judgment status of the field chain record are called. The judgment status of all records in the chain structure is aggregated. The aggregation logic is set as follows: if all status values ​​under the field dimension are True, the field is marked as passed; otherwise, the field is pending. The status of field "F001-Amount" is True, the status of field "F001-Date" is True, and the status of field "F001-Personnel" is True. At the aggregation level, the fields are categorized into Amount, Date, and Personnel categories. Fields under each category are aggregated by attachment number to form aggregated records: { {Amount: [F001-520.00-True, F002-120.00-True]}, {Date: [F001-2023-12-24-True]}, {Personnel: [F001-Zhang San-True]}, then the judgment results are merged into the document dimension structure according to the field type dimension, forming a document rule record set array. The array is then structured and recorded in the following fields: {Document Number: "R20231225001", Rule Record Set: [[Amount: True], [Date: True], [Personnel: True]]}, forming a unified document-level view of rule judgment, which is used for the final judgment of subsequent document review status.

[0037] S502: Based on the document rule summary value, determine the overall document review status, call the rule judgment status and credibility labeling results, classify the rule records into statuses, determine whether the document is in the approved or reviewed status based on the classification results, and record the status to the document-level data structure to obtain the document review status. The document rule record set structure is read, and the document review status judgment rules are set as follows: If the judgment status of all fields is True, the document status is "Passed". If any field has a judgment status of False or is missing, the document status is set to "Reviewed". In the current document number "R20231225001", the status of the three fields of amount, date and personnel is True, and the corresponding judgment result is valid. The document review status is tentatively set as "Passed", but the field credibility label status is further judged. Among them, the F001-amount field has a status of "Pending Verification" in the credibility label. The credibility label influence factor weight is set to 0.2. When the field status is True but the credibility label is "Pending Verification", the single field score is reduced to 0.8.

[0038] S503: Based on the document review status, the review status is associated with field identifiers, rule numbers and credibility results. The document identifier is called to combine multiple types of associated information in a unified manner to form a structured review conclusion record, and traceable document processing result data is output to generate the review conclusion record. The document review status is integrated with field identifiers, rule numbers, and field credibility labels to form a structured review conclusion record. The record format is {Document Number: "R20231225001", Field Identifier: "F001-Amount", Rule Number: "R001-Limit Judgment", Status: "True", Credibility Label: "Pending Verification", Final Review Status: "Passed"}. The judgment record, parameter matching status, credibility status, and document status of each field are output together to form a traceable review conclusion data set, which is synchronously written to the "Review Result Log" table in the system database for external approval system calls or subsequent manual audit review. The structure format is standardized, and the fields can be used for log retrieval indexing, ultimately completing the closed loop of the review process and generating the review conclusion record.

[0039] Please see Figure 7 A financial shared auditing system based on AI big data model technology includes: The field collection module is used to execute S1: collect computerized accounting information for expenses, invoice attachments and non-invoice attachments. In hotel bills, itineraries and meal receipts, it identifies the date, amount, personnel, items and quantity fields, records the attachment identifiers corresponding to the fields, builds the binding relationship between fields and documents, and forms a set of audit fields. The field consistency module is used to execute S2: filter field groups with the same field type from the document review field set, determine whether the time of the date field is the same, whether the value of the amount field is consistent, whether the identifier of the personnel field is consistent, whether the value of the quantity field is paired and matched, and record the field comparison results to form a field consistency set; The trust labeling module is used to perform S3: based on the field consistency set, it classifies the comparison results of each field, marks the fields that match the comparison as passed, marks the fields that do not match the comparison or did not participate in the comparison as pending verification, and establishes a correspondence between the labeling results and the fields themselves to form trust labeling results; The rule verification module is used to execute S4: call the credibility labeling results, call the corresponding rules according to the expense type, perform limit condition judgment on the amount field, perform period range judgment on the date field, perform subject consistency judgment on the personnel field, and bind the field values, rule parameters and judgment results during the rule execution process to form a rule judgment evidence chain; The status output module is used to execute S5: based on rule-based evidence chain judgment, it collects all rule judgment records of the current expense reimbursement document, judges the document's review status, and associates the status with field identifiers, rule numbers, and credibility results, and outputs the review conclusion record of passing and reviewing.

[0040] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A financial shared service document review method based on AI large-scale model technology, characterized in that: Includes the following steps: S1: Collect accounting computerization information, invoices and non-invoice attachments, extract date, amount, personnel, item, and quantity fields from hotel bills, itineraries, and meal receipts, record field attachment identifiers, establish the binding relationship between fields and documents, and form a set of document review fields; S2: Filter fields of the same type from the document review field set, determine whether the date is consistent, whether the amount is consistent, whether the personnel are consistent, and whether the quantity is matched in pairs, and record the comparison results to construct a field consistency set; S3: Based on the classification and comparison results of the field consistency set, mark consistent fields as passed, and inconsistent or uncompared fields as pending verification. Establish a correspondence between the marking results and the fields to form a credibility labeling result. S4: Call the credibility labeling result, apply the corresponding rules according to the expense type, perform limit judgment on the amount field, period judgment on the date field, and subject consistency judgment on the personnel field, and bind the field value, rule parameters and judgment result to form a rule judgment evidence chain; S5: Based on the rules, determine the evidence chain, collect all judgment records of the document, determine the document review status, and associate the status with the field identifier, rule number, and credibility result, and output the review conclusion record of the document that has passed and been reviewed.

2. The financial shared auditing method based on AI large model technology according to claim 1, characterized in that, The document review field set includes the master document information field, invoice attachment field entries, non-invoice attachment field entries, attachment identifier mapping information, and a field-document binding relationship table; the field consistency set includes date consistency results, amount consistency results, personnel consistency results, quantity matching results, and field comparison result records; the credibility labeling results include passed labeling items, unverified labeling items, and a field-labeling correspondence table; the rule judgment evidence chain includes field value snapshots, rule parameter sets, rule judgment conclusions, rule-field association items, and evidence node sequences; the document review conclusion record includes the document review status code, field identifier list, rule number list, credibility result summary item, passed conclusion information, and review conclusion information.

3. The financial shared auditing method based on AI large model technology according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Obtain accounting computerization information and the corresponding invoice and non-invoice attachments, call the expense type, applicant, reimbursement date and attachment identifier in the main order field, classify the attachment content according to the attachment identifier, separate and record invoice and non-invoice image data, and establish unified mapping data to generate attachment classification mapping data. S102: Based on the attachment classification mapping data, extract the date, amount, personnel, items and quantity fields from the hotel bill, itinerary and meal receipt images, call the text in the image recognition results, combine the field content structure features, record the attachment identifier corresponding to the field, and generate field location identifier information; S103: Based on the field location identification information, construct the binding structure between field content and attachment, call the attachment identification content, organize the correspondence between field attributes and field values, and uniformly classify the content and the associated attachment number in the field set into the document field structure table to generate the document review field set.

4. The financial shared auditing method based on AI large model technology according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Based on the document review field set, filter field groups with consistent field types, call the field type identifier and the corresponding attachment identifier, and assign the date, amount, personnel, and quantity fields to their respective groups to generate a field type grouping sequence; S202: Group the sequence according to the field type, perform consistency judgment on the field values ​​in the field group, judge the time content for the date field, the numerical content for the amount field, the identification content for the personnel field, and the pair matching status for the quantity field, summarize all judgment results and record the consistency status to obtain the field comparison mark; S203: Based on the field comparison markers, organize the comparison status corresponding to the field group, call the field type and identifier content to establish a combination structure of field items and consistency status, output the data structure format in a unified manner according to the field type, and generate a field consistency set.

5. The financial shared auditing method based on AI large model technology according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: Based on the field consistency set, filter the comparison status records corresponding to the fields, call the field identifier and consistency status mark, distinguish and organize the fields that participated in the comparison and those that did not participate in the comparison, and uniformly collect the field comparison status into the field-level record item to form a status list that can be used for classification and processing, and generate a summary value of the field comparison status. S302: Based on the field comparison status summary value, determine the corresponding comparison status type of the field, assign a pass mark to fields with consistent status, assign a pending verification mark to fields with inconsistent status and fields not participating in the comparison, and organize the mark results and field identifiers to form a field-level identifier record structure to obtain the field status identifier value. S303: Based on the field status identifier value, bind the field identifier result with the original field content, call the field identifier, field type and annotation status to combine and organize, output the field and annotation relationship record structure in a unified manner, establish the field credible annotation data structure, and generate credible annotation results.

6. The financial shared auditing method based on AI large model technology according to claim 5, characterized in that, The specific steps of S4 are as follows: S401: Call the credibility labeling result to obtain the field identifier, field type and labeling status content, filter the corresponding field set according to the expense type record item, organize the correspondence between the amount field, date field and personnel field, establish a mapping structure between expense type and field type, form a list of fields that can be called by rules, and generate rule field mapping values; S402: Based on the rule field mapping value, call the corresponding rule parameter according to the expense type, determine whether the amount field meets the limit threshold, determine whether the date field falls within the cycle benchmark range, determine whether the personnel field is consistent with the main identifier, and associate and record each judgment result with the field value and rule parameter to obtain the rule judgment record; S403: Based on the rule judgment record, organize the correspondence between field values, rule parameters and judgment status, call the field identifier to collect multiple judgment records by field dimension, establish the field judgment process association structure, form a continuous record chain from field to rule, and generate a rule judgment evidence chain.

7. The financial shared auditing method based on AI large model technology according to claim 6, characterized in that, The specific steps of S5 are as follows: S501: Based on the rule-based evidence chain, collect all rule-based records corresponding to the current expense reimbursement document, call the document identifier and rule number, collect and organize the judgment status associated with the fields, summarize the scattered records to the document dimension, and form a set of rule records that can be used for status determination, and generate a document rule summary value. S502: Based on the document rule summary value, determine the overall document review status, call the rule to determine the status and credibility labeling results, classify the rule records by status, determine whether the document is in the pass or review status based on the classification results, and record the status to the document-level data structure to obtain the document review status. S503: Based on the document review status, the review status is associated with the field identifier, rule number and credibility result. The document identifier is called to combine multiple types of associated information in a unified manner to form a structured review conclusion record. Traceable document processing result data is output to generate the review conclusion record.

8. The financial shared auditing method based on AI large model technology according to claim 3, characterized in that, The field location identification information is the text content extracted from the attached images of hotel bills, itineraries, and meal receipts using image recognition technology.

9. The financial shared auditing method based on AI large model technology according to claim 7, characterized in that, The document review status is determined based on the collected rules and the credibility labeling results of the records and fields. A preset status judgment logic model is used to classify and judge the status. If all rules judge the record as passed and the credibility score of the associated field is higher than the set threshold, the document is marked as passed; otherwise, the document is marked as reviewed and recorded in the document-level data structure for subsequent traceability. Based on the rules, the evidence chain is determined, all judgment records of the document are collected, the document review status is determined, and the status is associated with the field identifier, rule number, and credibility result. The review conclusion records of "pass" and "review" are output. Among them, if the credibility score of a field is higher than the set threshold, the field is marked as "pass"; otherwise, it is "to be verified". The threshold is set by the system according to historical data or business needs.

10. A financial shared auditing system based on AI large-scale model technology, characterized in that: The system is used to implement the financial shared auditing method based on AI large model technology as described in any one of claims 1-9, and the system includes: The field collection module is used to execute S1: collect computerized accounting information for expenses, invoice attachments and non-invoice attachments. In hotel bills, itineraries and meal receipts, it identifies the date, amount, personnel, items and quantity fields, records the attachment identifiers corresponding to the fields, builds the binding relationship between fields and documents, and forms a set of audit fields. The field consistency module is used to execute S2: filter field groups with the same field type from the document review field set, determine whether the time of the date field is the same, determine whether the value of the amount field is consistent, determine whether the identifier of the personnel field is consistent, determine whether the value of the quantity field is matched in pairs, and record the field comparison results to form a field consistency set. The trust labeling module is used to perform S3: based on the field consistency set, classify the comparison results of each field, mark the fields that match the comparison as passed, mark the fields that do not match the comparison or did not participate in the comparison as pending verification, and establish a correspondence between the labeling results and the fields themselves to form a trust labeling result; The rule verification module is used to execute S4: call the credibility labeling result, call the corresponding rule according to the expense type, perform limit condition judgment on the amount field, perform period range judgment on the date field, perform subject consistency judgment on the personnel field, and bind the field value, rule parameter and judgment result during the rule execution process to form a rule judgment evidence chain; The status output module is used to execute S5: based on the rules, determine the evidence chain, collect all rule judgment records of the current expense reimbursement document, determine the document's review status, associate the status with field identifiers, rule numbers, and credibility results, and output the review conclusion record of passing and reviewing.

Images