A large model-based travel expense reimbursement auditing system and method

Abstract

This invention discloses a travel expense reimbursement review system and method based on a large model, comprising: acquiring travel intention information and travel policy documents; performing atomized segmentation and sequence mapping on the travel policy documents to generate an atomic rule sequence; based on the travel intention information and the atomic rule sequence, performing intent recognition and rule routing through a large model to generate a local rule index set; concatenating rule sub-documents according to the local rule index set; calling a strategy simulation toolset to perform compliance deduction on the travel intention information to generate a travel budget blueprint; after the travel is completed, acquiring the actual reimbursement voucher set and extracting structured voucher elements, performing review intent recognition and evidence chain analysis through a large model; dynamically scheduling a deterministic rule execution toolset for calculation and logic verification; and integrating the results to generate a travel audit report. This invention achieves efficient and accurate understanding of long documents and deterministic financial logic execution, improving review efficiency and reliability.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and specifically to a travel expense reimbursement review system and method based on a large model. Background Technology

[0002] Travel expense reimbursement is a crucial aspect of corporate financial management. Its review process typically involves understanding unstructured travel management policy documents and assessing the compliance of various receipts and vouchers. Currently, the industry generally employs manual review or automated systems based on fixed rules. Manual review is inefficient and easily influenced by the subjective experience of reviewers, making it difficult to ensure consistent standards. While traditional automated systems improve efficiency, their rules are usually hard-coded. When corporate travel policies change, the system requires significant investment in code-level modifications and maintenance, resulting in insufficient flexibility. In recent years, with the development of artificial intelligence technology, solutions using large language models for semantic understanding to assist in review have emerged. However, these solutions face challenges when processing long policy documents, including high computational resource consumption, limited accuracy in understanding long texts, and insufficient reliability when performing numerical calculations and logical judgments requiring strict determinism. Therefore, constructing an intelligent review system that can efficiently and accurately understand complex policy semantics while ensuring rigorous, traceable, and easily maintainable review logic has become a key technical problem that needs to be solved in this field. Summary of the Invention

[0003] In view of the above problems, the present invention provides a travel expense reimbursement review system and method based on a large model. By combining the atomic indexing of policy documents with the intent routing of the large model and calling deterministic rule tools to perform calculations, the system solves the efficiency and reliability problems of the large model in long document processing and precise financial logic execution.

[0004] To achieve the above objectives, in a first aspect, this application provides a travel expense reimbursement review method based on a large model, comprising:

[0005] Obtain travel intention information and travel policy documents. Travel intention information includes destination, time, job level and reason for travel, while travel policy documents include unstructured travel expense management policy texts.

[0006] Atomize and number the travel policy document to generate an atomic rule sequence consisting of a unique number and the corresponding original text.

[0007] Based on travel intention information and atomic rule sequences, a large model is used for intent recognition and rule routing to generate a set of local rule indexes for travel intention information.

[0008] Based on the local rule index set, rule sub-documents associated with travel intention information are physically concatenated from the atomic rule sequence;

[0009] Call the strategy simulation toolset that is semantically bound to the rule sub-document to perform compliance simulation and multi-scheme simulation on travel intention information, and generate a travel budget blueprint that includes compliance cost caps and recommended strategies;

[0010] After the business trip is completed, an actual reimbursement voucher set consisting of multiple receipts is obtained, and a multimodal key information extraction model is used to extract structured voucher elements from the actual reimbursement voucher set.

[0011] Based on the elements of structured vouchers, a large model is used to identify audit intent and analyze the chain of evidence, resulting in conclusions from the chain of evidence analysis.

[0012] In addition, the deterministic rule execution toolkit is dynamically scheduled to perform calculations and logical verifications to obtain the execution results of the deterministic rule execution toolkit;

[0013] By integrating the results of the travel budget blueprint, the implementation of the deterministic rule enforcement toolset, and the conclusions of the evidence chain analysis, a travel audit report is generated that includes pre-planning comparisons and violation tracing.

[0014] Furthermore, the travel policy document is atomically segmented and numbered, generating an atomic rule sequence consisting of a unique number and its corresponding original text, including:

[0015] The document parsing engine is used to parse the format of travel policy documents and identify chapter titles, clause numbers, paragraph boundaries, and table structures in the travel policy documents.

[0016] Based on the identified hierarchical relationship of chapter titles and the logical order of clause numbers, a hierarchical semantic tree of travel policy documents is constructed, wherein each leaf node of the hierarchical semantic tree corresponds to a text fragment with independent and complete semantics.

[0017] Using each leaf node of the hierarchical semantic tree as the smallest segmentation unit, a physical segmentation operation is performed to deconstruct the travel policy document into multiple atomic text units;

[0018] Assign a globally unique and incrementing serialization identifier to each atomic text unit;

[0019] Establish a mapping relationship between serialization identifiers and the original text content of the corresponding atomic text units, sort and store all mapping relationships according to serialization identifiers, and generate an atomic rule sequence.

[0020] Furthermore, based on travel intention information and atomic rule sequences, a large model is used for intent recognition and rule routing to generate a set of local rule indices for travel intention information, including:

[0021] The travel intention information, along with all serialization identifiers in the atomic rule sequence and the corresponding original text summary, are input into the large model.

[0022] Construct a rule classification prompt template, which is used to instruct the large model to classify the terms in the atomic rule sequence into multiple predefined cost dimension categories based on the destination and reason included in the travel intention information;

[0023] The large model is processed based on rule-based classification prompt templates to output structured classification results. In the structured classification results, each cost dimension category is associated with one or more serialized identifiers.

[0024] Extract all associated serialization identifiers from the structured classification results to form a local rule index set.

[0025] Furthermore, based on the local rule index set, rule sub-documents associated with travel intention information are physically concatenated from the atomic rule sequence, including:

[0026] Based on each serialization identifier contained in the local rule index set, traverse and query the atomic rule sequence to obtain the original text content that completely matches each serialization identifier;

[0027] According to the numerical order of the serialization identifier, all the original text content obtained is concatenated end to end to generate the initial concatenated text;

[0028] Based on the relationship between the cost dimension category and the serialization identifier defined in the structured classification results, the text segments in the initial concatenated text are split and classified into the corresponding cost dimension category;

[0029] All text paragraphs categorized under the same cost dimension are sorted and merged according to the numerical value of their corresponding serialization identifiers to form rule sub-documents corresponding to the cost dimension category.

[0030] Furthermore, the strategy simulation toolset, semantically bound to the rule sub-documents, is invoked to perform compliance simulations and multi-scenario simulations on travel intention information, generating a travel budget blueprint that includes compliance cost caps and recommended strategies, including:

[0031] Input travel intention information and rule sub-documents into the large model;

[0032] The large model is used to analyze the intent and identify one or more target cost dimensions that need to be simulated, including:

[0033] Analyze the destination, time, and reason for travel intention information, and match the clauses about transportation, accommodation, and allowances in the rule sub-document;

[0034] Identify one or more target cost dimensions that require cost projection;

[0035] For each identified target cost dimension, the large model selects and calls one or more corresponding strategy simulation tools from the pre-registered strategy simulation tool set based on the constraints in the rule sub-document; these are referred to as target strategy simulation tools.

[0036] Based on travel intention information and rule sub-documents, a parameter set that meets the input format requirements of the target strategy simulation tool is generated, and the target strategy simulation tool is called.

[0037] Each invoked target strategy simulation tool performs compliance simulation calculations. Based on the set of input parameters and the benchmark data obtained from accessing external real-time data sources, it follows the calculation logic in the rule sub-document to calculate the upper limit of compliance costs for that target cost dimension.

[0038] For the target cost dimension where there are multiple optional execution options, the target strategy simulation tool also performs a multi-scenario simulation step, including:

[0039] Based on different scheme parameters, multiple compliance simulation calculation steps are executed in parallel or serially to generate multiple scheme simulation results and their corresponding estimated costs and compliance status.

[0040] The large model receives and integrates the compliance cost caps for all target cost dimensions and the simulation results of all solutions, and performs formatting, filling, and strategy recommendation sorting according to the predefined blueprint template to generate a travel budget blueprint.

[0041] Furthermore, after the travel activity is completed, a set of actual reimbursement vouchers consisting of multiple receipts is obtained, and a multimodal key information extraction model is used to extract structured voucher elements from the actual reimbursement voucher set, including:

[0042] Receive the actual reimbursement voucher set uploaded by the user. The actual reimbursement voucher set contains various types of vouchers, including image format and scanned document format.

[0043] Each receipt in the actual reimbursement voucher set is input into the multimodal key information extraction model;

[0044] The multimodal key information extraction model identifies the preset ticket category to which each ticket belongs based on the image features and layout of the ticket.

[0045] The multimodal key information extraction model extracts textual, visual, and spatial information from each ticket simultaneously, and then fuses and encodes these information to generate a unified ticket feature representation.

[0046] The multimodal key information extraction model is based on the feature representation of invoices. It uses pre-trained sequence labeling heads or region detection heads to extract key information entities of predefined categories from each invoice. The key information entities include at least personnel identity, timestamp, location, amount, reason and invoice type.

[0047] The multimodal key information extraction model is based on all extracted key information entities. Through entity disambiguation and time-space logical reasoning, it establishes semantic relationships between different documents and constructs structured voucher elements with business trip events as the core. The structured voucher elements include a set of related entities and an entity relationship graph.

[0048] Furthermore, based on structured credential elements, a large-scale model is used to identify audit intent and analyze the chain of evidence, yielding conclusions including:

[0049] Input the structured voucher elements into the large model;

[0050] The large model analyzes the document type and reason in the structured document elements to determine the categories of expenses involved in this audit and the compliance assertions that need to be verified.

[0051] Based on the entity relationship graph in the structured voucher elements, the association paths between personnel identity, timestamp, location and voucher type are extracted to construct an evidence chain topology network describing the entire business trip process;

[0052] Traverse the evidence chain topology network, check for logical omissions or evidence breaks based on the mandatory clauses regarding the completeness of the documents in the rule sub-documents, and generate completeness detection results;

[0053] The entities and relationships in the evidence chain topology are matched with the conditional statements in the rule sub-documents to verify whether there are logical contradictions in the travel behavior reflected by the actual reimbursement voucher set, and a consistency verification result is generated.

[0054] Integrate the integrity test results with the consistency verification results to form a chain of evidence analysis conclusion.

[0055] Furthermore, the deterministic rule execution toolkit is dynamically scheduled to perform computation and logical verification to obtain the execution results of the deterministic rule execution toolkit, including:

[0056] Based on the cost dimension categories determined in the audit intent identification step and the compliance assertions that need to be verified, select one or more target deterministic rule enforcement tools whose function descriptions match from the pre-registered deterministic rule enforcement tool set;

[0057] Based on structured credential elements and rule sub-documents, parameter data that meets the input interface requirements of the target deterministic rule execution tool is extracted and constructed;

[0058] Each target deterministic rule execution tool is invoked sequentially or in parallel. Each target deterministic rule execution tool performs black-box rule computation, including:

[0059] Based on the deterministic business logic and calculation code encapsulated within the target deterministic rule execution tool, the input parameter data is processed to independently generate the rule execution result corresponding to the tool. The rule execution result includes at least one of compliance status, calculated amount, and violation identifier.

[0060] Collect and integrate the rule execution results output by all invoked target deterministic rule execution tools to form the execution results of the deterministic rule execution tool set.

[0061] Furthermore, by integrating the execution results and evidence chain analysis conclusions of the travel budget blueprint, the deterministic rule enforcement toolset, and other relevant data, a travel audit report is generated that includes pre-planning comparisons and violation tracing, including:

[0062] The compliant expense caps and recommended execution strategies for each expense dimension in the travel budget blueprint are matched and compared with the actual audit results in the execution results of the deterministic rule execution toolset, and the deviation data is calculated.

[0063] For anomalies in the biased data or conflicts in the chain of evidence analysis, combine the original text of the corresponding specific clauses in the rule sub-document to locate the rule basis and missing evidence that caused the anomaly or conflict.

[0064] Based on the predefined audit report template, the travel budget blueprint, actual audit results, deviation data, results of the root cause tracing steps for violations, and the original text of the referenced rule sub-document clauses are organized and filled in according to the cost dimension.

[0065] Add natural language explanations based on evidence chain topology and rule logic to each conclusion in the audit report template to generate the final version of the travel audit report.

[0066] In a second aspect, the present invention also provides a travel expense reimbursement review system based on a large model, applicable to the method described in the first aspect. The system includes an information acquisition module, a rule preprocessing module, an intelligent routing module, a strategy deduction module, a voucher parsing module, an review execution engine, and a report generation module. The information acquisition module is used to acquire travel intention information and travel policy documents. The travel intention information includes destination, time, job level, and reason for travel, while the travel policy documents contain unstructured travel expense management system text. The rule preprocessing module is used to perform atomic segmentation and sequence mapping on the travel policy documents, generating an atomic rule sequence composed of a unique sequence number and the corresponding original text. The intelligent routing module is used to perform intent recognition and rule routing based on the travel intention information and the atomic rule sequence through a large model, generating a local rule index set for the travel intention information. The system also includes a module for physically splicing the relevant rules from the atomic rule sequence according to the local rule index set. The system comprises several modules: a rule sub-document associated with travel intention information; a strategy deduction module that calls a strategy simulation toolset semantically bound to the rule sub-document to perform compliance deduction and multi-scheme simulation on travel intention information, generating a travel budget blueprint that includes compliance cost limits and recommended strategies; a voucher parsing module that, after travel is completed, obtains a set of actual reimbursement vouchers consisting of multiple receipts and extracts structured voucher elements from the actual reimbursement voucher set using a multimodal key information extraction model; an audit execution engine that, based on structured voucher elements, performs audit intent identification and evidence chain analysis through a large model to obtain evidence chain analysis conclusions; and a module that dynamically schedules a deterministic rule execution toolset for calculation and logical verification to obtain the execution results of the deterministic rule execution toolset; and a report generation module that integrates the travel budget blueprint, the execution results of the deterministic rule execution toolset, and the evidence chain analysis conclusions to generate a travel audit report that includes pre-planning comparison and violation tracing.

[0067] Unlike existing technologies, the above technical solution provides a travel expense reimbursement review system and method based on a large model, including: acquiring travel intention information and travel policy documents; atomically segmenting and mapping the travel policy documents to generate atomic rule sequences; based on the travel intention information and atomic rule sequences, performing intent recognition and rule routing through a large model to generate a set of local rule indexes; concatenating rule sub-documents based on the set of local rule indexes; calling a strategy simulation toolset to perform compliance simulation of the travel intention information and generate a travel budget blueprint; after the travel is completed, acquiring the actual reimbursement voucher set and extracting structured voucher elements, performing audit intent recognition and evidence chain analysis through a large model; dynamically scheduling a deterministic rule execution toolset for calculation and logic verification; and integrating the results to generate a travel audit report. This invention achieves efficient and accurate understanding of long documents and deterministic financial logic execution, improving audit efficiency and reliability.

[0068] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0069] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0070] In the accompanying drawings of the instruction manual:

[0071] Figure 1 This is a schematic diagram illustrating steps S101 to S108 of the method described in the specific implementation embodiment;

[0072] Figure 2 This is a schematic diagram illustrating steps S201 to S205 of the method described in a specific implementation.

[0073] Figure 3 This is a schematic diagram illustrating steps S301 to S304 of the method described in a specific implementation.

[0074] Figure 4 This is a schematic diagram illustrating steps S401 to S404 of the method described in a specific embodiment;

[0075] Figure 5 This is a schematic diagram of the audit system described in a specific implementation.

[0076] The reference numerals used in the above figures are explained as follows:

[0077] 1. Audit system;

[0078] 11. Information Acquisition Module;

[0079] 12. Rule preprocessing module;

[0080] 13. Intelligent routing module;

[0081] 14. Strategy Deduction Module;

[0082] 15. Voucher parsing module;

[0083] 16. Audit execution engine;

[0084] 17. Report generation module. Detailed Implementation

[0085] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0086] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0087] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0088] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0089] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0090] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0091] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0092] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0093] The processor described in the embodiments of this application can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0094] The computer program involved in the embodiments can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical, or chemical structures that can achieve the same or equivalent functions as the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types, or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiments can be centrally stored in a single medium, or distributed and stored in multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device, or can be connected to the device involved in the embodiments as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0095] Please see Figure 1 In a first aspect, this embodiment provides a travel expense reimbursement review method based on a large model, including:

[0096] S101. Obtain travel intention information and travel policy documents. Travel intention information includes destination, time, job level and reason for travel. Travel policy documents include unstructured travel expense management system text.

[0097] S102. Perform atomic segmentation and sequence mapping on the travel policy document to generate an atomic rule sequence consisting of a unique sequence number and the corresponding original text.

[0098] S103. Based on travel intention information and atomic rule sequences, intent recognition and rule routing are performed through a large model to generate a set of local rule indexes for travel intention information;

[0099] S104. Based on the local rule index set, physically concatenate the rule sub-documents associated with the travel intention information from the atomic rule sequence;

[0100] S105. Call the strategy simulation toolset that is semantically bound to the rule sub-document to perform compliance simulation and multi-scheme simulation on travel intention information, and generate a travel budget blueprint that includes compliance cost cap and recommended strategy.

[0101] S106. After the business trip is completed, obtain the actual reimbursement voucher set consisting of multiple receipts, and use the multimodal key information extraction model to extract structured voucher elements from the actual reimbursement voucher set.

[0102] S107. Based on the elements of structured vouchers, a large model is used to identify audit intent and analyze the chain of evidence to obtain the conclusion of the chain of evidence analysis.

[0103] In addition, the deterministic rule execution toolkit is dynamically scheduled to perform calculations and logical verifications to obtain the execution results of the deterministic rule execution toolkit;

[0104] S108. Integrate the execution results and evidence chain analysis conclusions of the travel budget blueprint and the set of deterministic rule enforcement tools to generate a travel audit report that includes pre-planning comparison and violation tracing.

[0105] In step S101, the travel intention information is submitted by the user through the front-end interface during the travel planning stage, and its structured fields are used for subsequent accurate matching of policy rules. Travel policy documents are usually internal documents issued by enterprises or institutions regarding travel expense standards. They have various formats and the content is unstructured text. Directly processing the full text by a large model is inefficient and easily affected by interference.

[0106] In step S102, atomic segmentation aims to decompose lengthy travel policy documents into multiple semantically independent text fragments, each fragment corresponding to a specific cost provision or constraint. Sequence mapping assigns a globally unique identifier to each segmented text fragment, thereby generating an atomic rule sequence. This atomic rule sequence is essentially an index database, transforming complex natural language text into discrete units that can be quickly located and retrieved using sequence numbers, laying the foundation for subsequent precise rule extraction.

[0107] In step S103, intent recognition refers to the large model understanding the core expense types (such as transportation, accommodation, and meal allowances) involved in the current business trip based on key fields such as destination and reason in the travel intent information. Rule routing, based on this intent and combined with an atomic rule sequence, determines which rule clauses corresponding to the sequence numbers are relevant to the current travel intent. During this process, the large model does not output the original rule text, but only the sequence numbers of these relevant rules, forming a local rule index set. This step enables rapid and accurate filtering of relevant subsets from massive rules, greatly reducing the processing burden on the model.

[0108] In step S104, physical splicing is a deterministic, procedural operation. The system directly accesses the atomic rule sequence based on the sequence numbers provided in the local rule index set, extracts the corresponding complete original text according to the sequence numbers, and combines these texts in logical order to generate rule sub-documents. This process is entirely controlled by code, ensuring complete consistency between the spliced ​​rule text and the original policy clauses, avoiding potential deviations or information omissions that may occur in the model during text generation or summarization.

[0109] In step S105, the strategy simulation toolset is a series of predefined executable program modules that encapsulate specific cost calculation or compliance judgment logic. The strategy simulation toolset is semantically bound to the rule sub-document, allowing the system to automatically select the appropriate tools based on the sub-document content. Compliance simulation refers to using the selected tools, combined with travel intention information (such as job level, destination, and time) and specific numerical standards in the rule sub-document, to simulate and calculate the compliance ceiling for various expenses. Multi-scenario simulation can deduce multiple compliance options and their costs within the scope permitted by the rules (such as different modes of transportation). The final generated travel budget blueprint provides users with clear cost expectations and strategic recommendations.

[0110] In step S106, the actual reimbursement voucher set includes various receipts generated after the trip, such as transportation tickets and accommodation invoices, usually in image or PDF format. The multimodal key information extraction model can simultaneously process the visual layout and text content of the receipts, automatically extracting key information such as personnel, time, location, amount, and receipt type, and structuring it into a machine-readable data format to form structured voucher elements. This replaces traditional manual data entry or simple OCR recognition, achieving a deep understanding of complex receipt information.

[0111] In step S107, the audit intent identification refers to the large model determining the key expense items and compliance points that need to be focused on in this reimbursement audit based on the extracted structured voucher elements. Evidence chain analysis involves the large model performing logical reasoning on the structured voucher elements to check for temporal, spatial, or logical contradictions between different voucher information, and whether the voucher set completely supports the reimbursement reason. Dynamic scheduling of the deterministic rule execution toolset refers to the system automatically calling corresponding tools that encapsulate deterministic calculation logic (such as addition, subtraction, multiplication, division, and conditional judgments) according to the audit intent to perform precise numerical verification and rule compliance judgment on the voucher elements. This step combines the semantic analysis capabilities of the large model with the precise calculation capabilities of deterministic tools.

[0112] In step S108, the integration process compares the pre-planned travel budget blueprint with the actual results of the post-audit to reveal discrepancies. Violation tracing involves locating violations based on the evidence chain analysis and the verification results of deterministic tools, and linking them back to specific clauses in the rule sub-document. The final generated travel audit report not only provides the audit conclusions but also clearly demonstrates the comparison between pre- and post-audit details and the basis for violations, achieving transparency and traceability in the audit process.

[0113] This embodiment transforms unstructured policies into indexable rule sequences through atomic segmentation and sequence mapping. It utilizes a large model for efficient intent recognition and rule routing, and employs a "model indexing, program concatenation" approach to accurately generate rule sub-documents. Furthermore, it leverages a policy simulation toolset to perform pre-event compliance simulations and budget planning. In the post-event review phase, it combines multimodal information extraction, large model evidence chain analysis, and dynamic scheduling of deterministic rule tools to conduct in-depth reviews of expense reimbursement vouchers. Finally, it integrates all process information to generate an audit report. This embodiment combines the semantic understanding advantages of large-scale language with the precise execution capabilities of deterministic procedures. While ensuring the rigor and accuracy of financial audits, it significantly improves the processing efficiency and intelligence level of long, unstructured travel policy documents, solving the problems of poor flexibility in traditional automated systems and the unreliable computation and high cost of processing long documents when relying solely on large model reviews.

[0114] Please see Figure 2 In some embodiments, the travel policy document is atomically segmented and mapped with serial numbers to generate an atomic rule sequence consisting of a unique serial number and the corresponding original text, including:

[0115] S201. Use a document parsing engine to parse the format of the travel policy document and identify the chapter titles, clause numbers, paragraph boundaries and table structures in the travel policy document.

[0116] S202. Based on the identified hierarchical relationship of chapter titles and the logical order of clause numbers, construct a hierarchical semantic tree for travel policy documents, wherein each leaf node of the hierarchical semantic tree corresponds to a text fragment with independent and complete semantics.

[0117] S203. Using each leaf node of the hierarchical semantic tree as the smallest segmentation unit, perform physical segmentation operations to deconstruct the travel policy document into multiple atomic text units.

[0118] S204. Assign a globally unique and incrementing serialization identifier to each atomic text unit;

[0119] S205. Establish a mapping relationship between serialization identifiers and the original text content of the corresponding atomic text units, sort and store all mapping relationships according to serialization identifiers, and generate an atomic rule sequence.

[0120] In step S201, the document parsing engine processes travel policy documents in different formats, such as PDF, Word, or HTML files. The document parsing engine extracts text flow, font styles, position coordinates, and layout information from the document by calling the corresponding format parsing library. Based on this information, the document parsing engine identifies and marks headings representing chapter structures (usually with specific fonts, sizes, or bold styles), clause numbers with fixed patterns, natural paragraphs bounded by line breaks or indentation, and table areas composed of cells, thereby completing the initial structured parsing of the document's unstructured content.

[0121] In step S202, the process of constructing a hierarchical semantic tree is based on the parsing results of step S201. The system analyzes and identifies the nesting relationships between chapter titles (such as first-level headings and second-level headings) and the continuity and hierarchy of clause numbers, organizing these structural elements into a tree-like data structure. In this hierarchical semantic tree, the root node represents the entire document, intermediate nodes represent chapters or major clauses, and each leaf node corresponds to an indivisible, semantically complete text fragment, such as an independent clause item, a complete paragraph, or the overall description of a table. This tree structure clearly expresses the logical organization of the policy document.

[0122] In step S203, the physical segmentation operation is performed based on the leaf node information of the hierarchical semantic tree. The system extracts the corresponding text content directly from the document's text stream or parsed data blocks based on the start and end positions or index range of the text content associated with each leaf node in the original document, using program instructions. Each extracted complete text block corresponding to a leaf node constitutes an atomic text unit. This operation ensures that the segmentation boundaries are strictly aligned with the original logical structure of the document.

[0123] In step S204, the serialization identifier is assigned following a simple incremental rule. The system may assign a globally unique integer ID to each unit according to the order in which the atomic text units are segmented, or according to the preorder traversal order of their corresponding leaf nodes in the hierarchical semantic tree, for example, starting from 1 and incrementing sequentially. This serialization identifier will serve as the unique key for locating and referencing the specific rule-based text segment in all subsequent processes.

[0124] In step S205, the mapping relationship is established by creating key-value pairs. The system associates each serialization identifier with the original text content of its corresponding atomic text unit, forming a record. All records are arranged according to the numerical order of the serialization identifiers and stored in a linear data structure such as a list or array, thus generating an atomic rule sequence. This sequence is essentially an ordered rule text library that can be directly accessed through an index (sequence number).

[0125] This embodiment achieves automatic structured parsing of policy documents through a document parsing engine and constructs a hierarchical semantic tree based on the parsed logical structure, thereby intelligently determining the precise boundaries of atomic segmentation. Physical segmentation at the leaf node level ensures the semantic integrity of each rule unit. By assigning globally unique serialization identifiers and establishing an ordered mapping storage, the final generated atomic rule sequence provides an efficient, accurate, and absolutely reliable data foundation for subsequent rule retrieval, routing, and concatenation. This embodiment refines the specific steps of rule preprocessing, automating and standardizing the process of transforming unstructured documents into a structured, indexable knowledge base, which is a key prerequisite for supporting the efficient and accurate rule processing of the entire solution.

[0126] Please see Figure 3 In some embodiments, based on travel intention information and atomic rule sequences, intent recognition and rule routing are performed through a large model to generate a local rule index set for travel intention information, including:

[0127] S301. Input the travel intention information, along with all serialization identifiers and corresponding original text summaries in the atomic rule sequence, into the large model.

[0128] S302. Construct a rule classification prompt template. The rule classification prompt template is used to instruct the large model to classify the terms in the atomic rule sequence into multiple predefined cost dimension categories based on the destination and reason contained in the travel intention information.

[0129] S303. The large model is processed based on the rule-based classification prompt template to output structured classification results. In the structured classification results, each cost dimension category is associated with one or more serialization identifiers.

[0130] S304. Extract all associated serialization identifiers from the structured classification results to form a local rule index set.

[0131] In step S301, the generation of the original text summary aims to compress the verbose content of each text unit in the atomic rule sequence while preserving its core semantics. The system can generate the corresponding original text summary by extracting the first sentence or key sentence of each atomic text unit, or by applying an automatic text summarization algorithm (such as key sentence extraction based on TF-IDF or TextRank). Inputting travel intention information, all serialization identifiers, and their corresponding original text summaries into the large model, instead of inputting the complete original rule text, can significantly reduce the length of the input sequence, thereby reducing the computational resources (number of tokens) required for processing by the large model, and helping the model to focus more on the categorical semantics of the terms rather than detailed numerical values.

[0132] In step S302, the rule classification prompt template is a pre-designed natural language instruction template used to guide a large model to complete a specific classification task. The rule classification prompt template includes a clear description of the classification task, references to the destination and reason fields in travel intention information, a list of cost dimension categories that the model is required to classify (e.g., intercity transportation costs, accommodation costs, meal allowances, intra-city transportation costs, etc.), and strictly defined output format requirements (e.g., specifying JSON format for output). By constructing and using this template, the system transforms the open semantic understanding task into a structured and controllable classification problem.

[0133] In step S303, after receiving the prompt template filled with specific travel intention information and rule summaries, the large model, based on its pre-trained language understanding and reasoning capabilities, analyzes the semantic relevance between each rule summary and the travel intention, and assigns it to one or more of the most matching cost dimension categories. After processing, the large model outputs structured classification results, which are typically presented in a machine-readable structured data format, such as a JSON object, where each key is a cost dimension category and the corresponding value is an array of serialized identifiers associated with that category.

[0134] In step S304, the extraction operation is performed by parsing the structured classification results (such as JSON objects). The system program iterates through the serialized identifier array under each cost dimension category key in the results, collects all identifiers into a set, and performs deduplication to ensure that each identifier appears only once in the final set. The resulting deduplicated identifier set constitutes the local rule index set. The local rule index set is essentially a list of numeric pointers pointing to relevant clauses in the atomic rule sequence.

[0135] This embodiment optimizes model input by generating rule summaries and guides the large model to perform efficient semantic classification using a constructed rule classification prompt template. The large model only needs to output the associated cost category and sequence number identifier, without generating or restating the rule text, which greatly reduces the model's output length and computational cost. The final extracted local rule index set accurately filters out all rule clauses related to the current travel intention, providing a precise index basis for the subsequent physical splicing of rule sub-documents.

[0136] Please see Figure 4 In some embodiments, rule sub-documents associated with travel intention information are physically concatenated from the atomic rule sequence based on a local rule index set, including:

[0137] S401. Based on each serialization identifier contained in the local rule index set, traverse and query the atomic rule sequence to obtain the original text content that completely matches each serialization identifier.

[0138] S402. According to the numerical order of the serialization identifier, concatenate all the original text content obtained to generate the initial concatenated text.

[0139] S403. Based on the relationship between the cost dimension category and the serialization identifier defined in the structured classification results, the text segments in the initial concatenated text are split and classified into the corresponding cost dimension category.

[0140] S404. All text paragraphs categorized under the same cost dimension are sorted and merged according to the numerical value of their corresponding serialization identifiers to form rule sub-documents corresponding to the cost dimension category.

[0141] In step S401, the atomic rule sequence is typically organized into a data structure that can be directly accessed through serialization identifiers, such as a dictionary or hash table with serialization identifiers as keys and original text content as values. By traversing each serialization identifier in the local rule index set and using it as a key to directly query the data structure, the original text content that perfectly matches it can be accurately obtained, ensuring that the target clause is accurately extracted from the massive rule base, and that the extracted text is exactly the same as the original policy clause.

[0142] In step S402, all the original text content obtained in step S401 is sorted in ascending order according to the numerical value of its corresponding serialization identifier, and then joined together sequentially. Preferably, line breaks or specific paragraph separators are inserted between adjacent text blocks to maintain readability, thereby generating the initial concatenated text. The initial concatenated text is a continuous text containing all relevant rule clauses, but not yet categorized by fee topic.

[0143] In step S403, the system parses the initial concatenated text and identifies the serialization identifier associated with each text segment (corresponding to a block of original text content). Based on the mapping relationship recorded in the structured classification results (i.e., which cost dimension category each serialization identifier belongs to), the system allocates or copies the content of the text segment to the corresponding cost dimension category container. For example, all text segments corresponding to serialization identifiers classified as "accommodation fee" will be placed under the "accommodation fee" category.

[0144] In step S404, all text paragraphs already categorized under the same cost dimension are sorted again based on the serialization identifier values ​​corresponding to each paragraph to maintain the original narrative order of the policy clauses. After sorting, these paragraphs are merged into a coherent text, forming a rule sub-document corresponding to that cost dimension category. Each rule sub-document independently encapsulates all relevant review rules for a specific cost type (such as transportation or accommodation).

[0145] This embodiment retrieves accurate rule texts through efficient key-value queries and intelligently classifies and reorganizes the clauses based on structured classification results. This process transforms the abstract index output by the model into a concrete, complete set of rule texts organized by cost dimension. The generated rule sub-documents are not only absolutely accurate in content (due to physical splicing) but also have a clear structure, directly corresponding to the cost categories relevant to subsequent review and calculation, providing accurate and reliable textual basis for strategy simulation and rule execution.

[0146] In some embodiments, a policy simulation toolset semantically bound to the rule sub-document is invoked to perform compliance extrapolation and multi-scenario simulation on travel intention information, generating a travel budget blueprint that includes compliance cost caps and recommended strategies, including:

[0147] Input travel intention information and rule sub-documents into the large model;

[0148] The large model is used to analyze the intent and identify one or more target cost dimensions that need to be simulated, including:

[0149] Analyze the destination, time, and reason for travel intention information, and match the clauses about transportation, accommodation, and allowances in the rule sub-document;

[0150] Identify one or more target cost dimensions that require cost projection;

[0151] For each identified target cost dimension, the large model selects and calls one or more corresponding strategy simulation tools from the pre-registered strategy simulation tool set based on the constraints in the rule sub-document; these are referred to as target strategy simulation tools.

[0152] Based on travel intention information and rule sub-documents, a parameter set that meets the input format requirements of the target strategy simulation tool is generated, and the target strategy simulation tool is called.

[0153] Each invoked target strategy simulation tool performs compliance simulation calculations. Based on the set of input parameters and the benchmark data obtained from accessing external real-time data sources, it follows the calculation logic in the rule sub-document to calculate the upper limit of compliance costs for that target cost dimension.

[0154] For the target cost dimension where there are multiple optional execution options, the target strategy simulation tool also performs a multi-scenario simulation step, including:

[0155] Based on different scheme parameters, multiple compliance simulation calculation steps are executed in parallel or serially to generate multiple scheme simulation results and their corresponding estimated costs and compliance status.

[0156] The large model receives and integrates the compliance cost caps for all target cost dimensions and the simulation results of all solutions, and performs formatting, filling, and strategy recommendation sorting according to the predefined blueprint template to generate a travel budget blueprint.

[0157] In this embodiment, the strategy simulation toolset consists of multiple independent computation modules. Each module registers with the central scheduler during the system initialization phase through a pre-registration mechanism. The pre-registration process includes each module providing its functional description text, specifying the cost dimension category it processes, the required input parameter data structure, and the output format. The functional description text allows the large model to understand its purpose without accessing the module's internal code.

[0158] When large models perform inference intent parsing, natural language understanding techniques, such as attention-based sequence labeling or intent classification models, can be used to analyze the semantic relationship between travel intention information and rule sub-documents. By identifying key entities (such as transportation type and accommodation keywords) and their constraint context in the rule sub-documents, the model abstracts and outputs a list of target cost dimensions that need to be simulated.

[0159] For the identified target cost dimension, the large model performs semantic matching between the computational logic features described in the rule sub-document and the functional descriptions of each tool in the pre-registered strategy simulation toolkit, thereby selecting and invoking the target strategy simulation tool. Subsequently, the large model extracts key parameters from travel intention information and rule sub-documents, and performs structured assembly according to the input pattern defined by the target strategy simulation tool to generate a parameter set.

[0160] When the target strategy simulation tool performs compliance simulation calculations, its internal calculation logic can be based on rule engine-based judgment, numerical calculation based on calculation formulas, or limit matching based on table lookup. The calculation process strictly follows the input parameter set and the processing rules embedded in the tool that are semantically consistent with the rule sub-documents. During the calculation process, the tool can access external data interfaces to obtain benchmark data such as real-time ticket prices and exchange rates to ensure the timeliness of the simulation results. For cost dimensions with multiple alternatives, multi-option simulation is achieved by the tool receiving different combinations of option parameters and executing the above calculation process separately, ultimately outputting the estimated cost and compliance status of each option.

[0161] A predefined blueprint template is a structured data schema or document framework that defines the fields a travel budget blueprint should include, such as compliance limits for each expense dimension, a list of recommended options, and dimensions for comparing options. After integrating all the projection results, the large model populates the template with data and sorts the options according to preset optimization objectives (such as cost minimization and time minimization) to generate the final travel budget blueprint.

[0162] This embodiment achieves tool discoverability and standardized invocation through a pre-registration mechanism, precisely schedules tools using the semantic parsing capabilities of large models, and allows the tools to perform deterministic compliant calculations and multi-scheme simulations. By decoupling the understanding of natural language strategies from the execution of coded rules, it ensures both the flexibility to handle complex semantics and the absolute accuracy and verifiability of financial calculations. The generated budget blueprint provides users with quantitative and multi-choice decision-making basis.

[0163] In some embodiments, after the travel activity is completed, an actual reimbursement voucher set consisting of multiple receipts is obtained, and structured voucher elements are extracted from the actual reimbursement voucher set using a multimodal key information extraction model, including:

[0164] Receive the actual reimbursement voucher set uploaded by the user. The actual reimbursement voucher set contains various types of vouchers, including image format and scanned document format.

[0165] Each receipt in the actual reimbursement voucher set is input into the multimodal key information extraction model;

[0166] The multimodal key information extraction model identifies the preset ticket category to which each ticket belongs based on the image features and layout of the ticket.

[0167] The multimodal key information extraction model extracts textual, visual, and spatial information from each ticket simultaneously, and then fuses and encodes these information to generate a unified ticket feature representation.

[0168] The multimodal key information extraction model is based on the feature representation of invoices. It uses pre-trained sequence labeling heads or region detection heads to extract key information entities of predefined categories from each invoice. The key information entities include at least personnel identity, timestamp, location, amount, reason and invoice type.

[0169] The multimodal key information extraction model is based on all extracted key information entities. Through entity disambiguation and time-space logical reasoning, it establishes semantic relationships between different documents and constructs structured voucher elements with business trip events as the core. The structured voucher elements include a set of related entities and an entity relationship graph.

[0170] In this embodiment, the multimodal key information extraction model is a pre-trained deep learning model. Its training process includes: performing self-supervised or supervised pre-training on a large-scale general image-text pair dataset to enable the model to learn the association between basic visual features and linguistic features; and performing supervised fine-tuning on a specially constructed invoice dataset, which contains a large number of image samples that have been accurately labeled with invoice categories, text regions, and key information entity locations. The model parameters are optimized by minimizing the classification and detection loss functions to adapt it to specific tasks in the invoice domain.

[0171] Predefined ticket categories refer to the set of ticket types defined before model deployment, such as transportation tickets, accommodation invoices, and restaurant invoices. The multimodal key information extraction model predicts the category of the input ticket image based on its image features and layout. Image features are extracted using a convolutional neural network, encoding color, texture, and shape information; layout features are obtained by analyzing the spatial distribution of text lines, table structure, or the positional relationships of specific areas.

[0172] For each ticket, the multimodal key information extraction model simultaneously processes information from multiple modalities. Textual information is extracted from the image by integrating an optical character recognition engine to extract the original character sequence and its position. Visual information refers to the deep semantic features extracted from the image, typically output by the backbone of a convolutional neural network. Spatial location information records the coordinate bounding boxes of each identified text block or visual element in the original image. The multimodal key information extraction model uses a feature fusion module, such as employing a cross-modal attention mechanism, to interact and integrate text feature vectors, visual feature vectors, and positional encodings to generate a unified, high-dimensional ticket feature representation.

[0173] Based on this invoice feature representation, the multimodal key information extraction model uses pre-trained sequence labeling heads or region detection heads to extract key information entities. Sequence labeling heads are suitable for processing text sequences extracted by OCR. They perform context encoding on the sequence using a recurrent neural network or Transformer encoder, followed by a conditional random field layer for sequence labeling, assigning each word its corresponding entity category. Region detection heads, based on an object detection architecture, predict bounding boxes containing entity information and their category labels on the image feature map. Predefined categories of key information entities include personnel identity, timestamps, locations, amounts, reasons for transactions, and invoice types; these categories are explicitly defined in the annotation system during model training.

[0174] After extracting all key information entities from a single ticket, the multimodal key information extraction model further performs semantic association across tickets. Entity disambiguation aims to address the inconsistency in the representation of the same entity across different tickets. This can be achieved by calculating the edit distance between entity names or the cosine similarity of semantic embedding vectors, clustering and normalizing entities pointing to the same real-world object. Time-space logical reasoning analyzes the timestamps and location information extracted from different tickets, establishing semantic associations between tickets based on the logical continuity of time sequence and geographical location. For example, tickets that are consecutive in time and connected in location are associated as the same trip. The final structured voucher element is a graph-structured data, where nodes are key information entities and edges represent relationships between entities, thus forming an evidence network centered on the business trip event.

[0175] This embodiment achieves end-to-end deep understanding and structured information extraction from heterogeneous document images through a pre-trained multimodal key information extraction model. The model integrates textual, visual, and spatial information, improving entity extraction accuracy under complex layouts and noise interference. Further cross-document entity disambiguation and logical reasoning integrate discrete document information into a coherent, semantically interconnected chain of evidence, providing high-quality, computable structured input for subsequent automated review, fundamentally changing the traditional processing methods that rely on manual input or simple OCR.

[0176] In some embodiments, based on structured credential elements, a large model is used to identify audit intent and analyze the chain of evidence to obtain chain of evidence analysis conclusions, including:

[0177] Input the structured voucher elements into the large model;

[0178] The large model analyzes the document type and reason in the structured document elements to determine the categories of expenses involved in this audit and the compliance assertions that need to be verified.

[0179] Based on the entity relationship graph in the structured voucher elements, the association paths between personnel identity, timestamp, location and voucher type are extracted to construct an evidence chain topology network describing the entire business trip process;

[0180] Traverse the evidence chain topology network, check for logical omissions or evidence breaks based on the mandatory clauses regarding the completeness of the documents in the rule sub-documents, and generate completeness detection results;

[0181] The entities and relationships in the evidence chain topology are matched with the conditional statements in the rule sub-documents to verify whether there are logical contradictions in the travel behavior reflected by the actual reimbursement voucher set, and a consistency verification result is generated.

[0182] Integrate the integrity test results with the consistency verification results to form a chain of evidence analysis conclusion.

[0183] In this embodiment, audit intent identification refers to the large model performing semantic understanding on the input structured voucher elements to determine the core focus of the audit task. The large model parses the voucher type and reason fields in the structured voucher elements and maps them to predefined expense dimension categories using natural language reasoning technology. For example, if a voucher set containing accommodation invoices and transportation tickets is identified, it determines that accommodation fees and intercity transportation fees are involved. At the same time, based on the combination of voucher type and reason, the large model derives compliance assertions that need to be verified, such as "accommodation fees do not exceed the standard limit for the corresponding job level and destination" or "transportation ticket time matches the business trip approval time".

[0184] The construction of the evidence chain topology network is based on the entity relationship graph in the structured credential elements. The large model extracts the associated paths centered on the business trip event from this graph. For example, it connects the "personnel identity" node to the "ticket" node through the "holding" relationship, and then forms a spatiotemporal sequence through the "timestamp" and "location" nodes. By analyzing the connectivity and temporal relationships of these paths, the large model constructs an evidence chain topology network describing the entire business trip process. This network presents the logical and spatiotemporal connections between all evidence entities in a graph structure.

[0185] Integrity checks are performed on the evidence chain topology network. The large model traverses the network and checks for structural deficiencies based on mandatory clauses regarding the completeness of receipts in the rule sub-documents (e.g., "reimbursement of accommodation expenses must be accompanied by a corresponding accommodation invoice"). Specifically, the large model translates the rule clauses into constraints on the existence of network nodes and edges, checking whether all necessary receipt type nodes exist and whether there are complete association paths between key entities (such as time and location). If deficiencies are found, the specific missing points are recorded and integrity check results are generated.

[0186] Consistency checks focus on logical inconsistencies within the evidence chain topology and between it and rule sub-documents. The large model matches and compares entity attributes (such as monetary values ​​and time points) in the network with conditional statements in the rule sub-documents. For example, it checks whether the time sequence of transportation tickets conforms to the logic of a business trip, or whether the reimbursement amount exceeds the limit specified in the rule sub-document. This process involves logical reasoning and numerical comparison, generating consistency check results by identifying attribute conflicts or violations of rule constraints.

[0187] The conclusion of the evidence chain analysis is formed by integrating the results of integrity checks and consistency verifications. This conclusion is a structured output that clearly indicates whether the evidence chain is complete, whether there are logical contradictions, and categorizes and locates the problems found, providing direct semantic-level judgment basis for subsequent review decisions.

[0188] This embodiment leverages the deep semantic understanding and reasoning capabilities of a large model to transform structured voucher data into a logically verifiable chain of evidence. By constructing an evidence chain topology network and performing systematic integrity checks and consistency verifications, it achieves automated and in-depth auditing of the business logic behind reimbursement vouchers. This goes beyond simple information extraction and matching, enabling the discovery of hidden logical errors and missing evidence, significantly improving the intelligence level and risk detection capabilities of the audit.

[0189] In some embodiments, a deterministic rule execution toolkit is dynamically scheduled to perform computation and logical verification to obtain the execution result of the deterministic rule execution toolkit, including:

[0190] Based on the cost dimension categories determined in the audit intent identification step and the compliance assertions that need to be verified, select one or more target deterministic rule enforcement tools whose function descriptions match from the pre-registered deterministic rule enforcement tool set;

[0191] Based on structured credential elements and rule sub-documents, parameter data that meets the input interface requirements of the target deterministic rule execution tool is extracted and constructed;

[0192] Each target deterministic rule execution tool is invoked sequentially or in parallel. Each target deterministic rule execution tool performs black-box rule computation, including:

[0193] Based on the deterministic business logic and calculation code encapsulated within the target deterministic rule execution tool, the input parameter data is processed to independently generate the rule execution result corresponding to the tool. The rule execution result includes at least one of compliance status, calculated amount, and violation identifier.

[0194] Collect and integrate the rule execution results output by all invoked target deterministic rule execution tools to form the execution results of the deterministic rule execution tool set.

[0195] In this embodiment, dynamic scheduling is the process by which the system automatically selects and activates corresponding computing resources based on real-time determined audit requirements. The selection logic for matching function descriptions is based on semantic similarity calculation or rule matching. The function description text provided by each deterministic rule execution tool during pre-registration clarifies the cost dimension it processes, the applicable compliance assertions, and the required input parameter structure. The system compares the cost dimension category and compliance assertions determined in the audit intent identification step with the function description of each tool in the toolset. By calculating the cosine similarity between text vectors or matching keywords, it selects one or more tools with similarity exceeding a threshold or the highest keyword matching degree as the target deterministic rule execution tool.

[0196] The construction of parameter data is a process of extracting and formatting data from multi-source information. Based on structured voucher elements, the system extracts entity attribute values ​​relevant to the target tool, such as personnel rank, invoice amount, and time interval from the voucher elements. Simultaneously, the system parses relevant calculation benchmarks or constraints from rule sub-documents, such as quota standards and coefficients in calculation formulas. Subsequently, the system assembles the extracted discrete data into specific data structures, such as dictionaries, JSON objects, or instances of specific classes, according to the parameter names, types, and order defined in the target deterministic rule execution tool input interface, ensuring that the data format is fully compatible with the tool's expected input.

[0197] Black-box rule calculation emphasizes the encapsulation and determinism of the tool's internal logic. Each target deterministic rule execution tool encapsulates the code logic that implements specific business rules, such as conditional statements, numerical calculation formulas (e.g., multiplication, summation), or lookup tables. When the tool is invoked, it receives pre-constructed parameter data and independently executes its internal code; the processing does not rely on real-time inference from a large model. The processing logic may include comparing the actual reimbursement amount with the standard limit, calculating the reimbursement amount based on the number of business trip days and the allowance standard, or verifying whether the invoice type complies with regulations. After execution, the tool generates its corresponding rule execution result, which is output in structured data format, explicitly including the compliance status (e.g., approved, disapproved), the calculated compliant amount, or a label identifying the specific type of violation.

[0198] The execution results of the deterministic rule enforcement toolset are generated by collecting and integrating the outputs of all invoked tools. The integration process may include categorizing the results of multiple tools by cost dimension, merging and calculating amounts, or summarizing a list of violation identifiers, ultimately forming a unified, structured result set covering all verified dimensions for use in subsequent report generation steps.

[0199] This embodiment achieves precise on-demand allocation of computing resources through a dynamic scheduling mechanism and intelligent tool selection through functional description matching. By constructing standardized parameters from structured data sources, the correctness of tool input is ensured; black-box rule computation delegates complex logic prone to uncertainty to deterministic code execution, fundamentally guaranteeing the absolute accuracy and repeatability of numerical calculations and logical judgments in the financial audit process. This embodiment achieves a complementary advantage between intelligent decision-making from large models and reliable execution of deterministic code.

[0200] In some embodiments, the results of the travel budget blueprint, the execution results of the deterministic rule enforcement toolset, and the conclusions of the evidence chain analysis are integrated to generate a travel audit report that includes pre-planning comparison and violation tracing, including:

[0201] The compliant expense caps and recommended execution strategies for each expense dimension in the travel budget blueprint are matched and compared with the actual audit results in the execution results of the deterministic rule execution toolset, and the deviation data is calculated.

[0202] For anomalies in the biased data or conflicts in the chain of evidence analysis, combine the original text of the corresponding specific clauses in the rule sub-document to locate the rule basis and missing evidence that caused the anomaly or conflict.

[0203] Based on the predefined audit report template, the travel budget blueprint, actual audit results, deviation data, results of the root cause tracing steps for violations, and the original text of the referenced rule sub-document clauses are organized and filled in according to the cost dimension.

[0204] Add natural language explanations based on evidence chain topology and rule logic to each conclusion in the audit report template to generate the final version of the travel audit report.

[0205] In this embodiment, item-by-item matching and numerical comparison are achieved through programmatic comparison. Specifically, the system aligns the compliant expense cap in the travel budget blueprint with the actual audit results (such as the actual approved amount) in the execution results of the deterministic rule enforcement toolset, based on the expense dimension category. For each dimension, deviation data is calculated, which is typically the difference between the actual result and the budget cap, or a value representing the proportion of overrun.

[0206] Root cause analysis of violations is an analytical and diagnostic process. For items identified as abnormal (e.g., exceeding limits) in the deviation data, or logical conflicts pointed out in the chain of evidence analysis, the system locates the source by combining the rule sub-documents. Specifically, based on the cost dimension and key attributes (e.g., job level, destination) associated with the abnormal item, the system retrieves the original text of the clauses containing the corresponding constraints in the rule sub-documents. Simultaneously, for chain of evidence conflicts, the system locates missing nodes or broken association paths based on the evidence chain topology. The output of this step is source-tracing information that clearly points to specific rule clauses and / or missing evidence points.

[0207] The predefined audit report template is a structured document framework that defines the report's chapters, data table formats, and placeholders for conclusions. The system categorizes the travel budget blueprint, actual audit results, deviation data, results of the root cause analysis steps, and the original text of retrieved rule sub-documents by cost dimension and then populates them into the corresponding positions in the template. For example, it creates independent analysis paragraphs for each cost dimension and embeds comparison tables, deviation values, reasons for violations, and referenced rule texts.

[0208] Adding natural language explanations based on the evidence chain topology and rule logic to each conclusion in the audit report template is typically done by a large model. The large model generates a coherent and easily understandable text description based on the populated template content, the structure of the evidence chain topology, and the logic of the rule sub-documents. This explains the causes of the deviations, the basis for the violations, and the completeness of the evidence chain, ultimately generating a complete travel audit report containing data, analysis, and explanations.

[0209] This embodiment deeply integrates pre-audit, in-process audit, and post-audit analysis through systematic data comparison, precise rule and evidence tracing, and structured template filling. Combined with natural language interpretation generated by the large model, the final audit report not only presents the audit results but also clearly reveals the business logic, rule basis, and evidence chain behind the results, achieving transparency, traceability, and high automation of the audit process.

[0210] Please see Figure 5In a second aspect, this embodiment also provides a travel expense reimbursement review system 1 based on a large model, applicable to the method described in the first aspect. The system includes an information acquisition module 11, a rule preprocessing module 12, an intelligent routing module 13, a strategy deduction module 14, a voucher parsing module 15, an review execution engine 16, and a report generation module 17. The information acquisition module 11 is used to acquire travel intention information and travel policy documents. The travel intention information includes destination, time, job level, and reason for travel, and the travel policy documents include unstructured travel expense management system text. The rule preprocessing module 12 is used to perform atomic segmentation and sequence mapping on the travel policy documents to generate an atomic rule sequence consisting of a unique sequence number and the corresponding original text. The intelligent routing module 13 is used to perform intent recognition and rule routing through a large model based on the travel intention information and the atomic rule sequence to generate a local rule index set for the travel intention information; and is used to extract the relevant information from the atomic rule sequence based on the local rule index set. The system physically assembles rule sub-documents associated with travel intention information; the strategy deduction module 14 calls a strategy simulation toolset semantically bound to the rule sub-documents to perform compliance deduction and multi-scheme simulation on travel intention information, generating a travel budget blueprint that includes compliance cost limits and recommended strategies; the voucher parsing module 15 obtains an actual reimbursement voucher set consisting of multiple receipts after the travel activity is completed, and extracts structured voucher elements from the actual reimbursement voucher set using a multimodal key information extraction model; the audit execution engine 16 performs audit intent identification and evidence chain analysis based on structured voucher elements through a large model to obtain evidence chain analysis conclusions; and dynamically schedules a deterministic rule execution toolset for calculation and logical verification to obtain the execution results of the deterministic rule execution toolset; the report generation module 17 integrates the travel budget blueprint, the execution results of the deterministic rule execution toolset, and the evidence chain analysis conclusions to generate a travel audit report that includes pre-planning comparison and violation tracing.

[0211] In this embodiment, the travel expense reimbursement audit system 1, through the collaborative work of its various modules, sequentially performs structured preprocessing of policy documents, intelligent rule routing and sub-document generation based on a large model, pre-budget deduction, post-reimbursement voucher parsing and multi-dimensional auditing, and finally generates an audit report. Specifically, the information acquisition module 11 and the rule preprocessing module 12 provide a structured data foundation for subsequent processing; the intelligent routing module 13 and the strategy deduction module 14 realize intelligent pre-planning, ensuring the accuracy and efficiency of rule extraction; the voucher parsing module 15 and the audit execution engine 16 collaborate to complete in-depth post-reimbursement auditing; and the report generation module 17 outputs conclusions based on comprehensive information from the entire process. This system combines the semantic understanding advantages of a large language model with the reliable execution of deterministic procedures, significantly improving the automation and intelligence level of processing long documents, complex invoices, and variable rules while ensuring the rigor of financial auditing.

[0212] By adopting the above technical solutions, this invention differs from existing technologies and has the following beneficial effects: It transforms unstructured travel policy documents into indexable atomic rule sequences through atomic segmentation and sequence mapping. A large model is used for intent recognition and rule routing, outputting only a local rule index set. Then, precise rule sub-documents are generated through programmatic physical concatenation, significantly reducing the computational overhead and semantic interference of processing long documents by the large model. By calling a strategy simulation toolset semantically bound to the rule sub-documents, pre-compliance deduction and multi-scheme simulation are performed to generate a travel budget blueprint, achieving intelligent expense planning. In the post-audit stage, a multimodal key information extraction model is used to deeply analyze invoices and construct structured voucher elements. Combined with evidence chain analysis from the large model and a dynamically scheduled deterministic rule execution toolset, logical consistency verification and precise numerical calculations are completed, ensuring the rigor of the audit process and the credibility of the results. Finally, the entire process information is integrated to generate an audit report that includes pre- and post-audit comparisons and violation tracing. This invention combines the semantic understanding and flexible reasoning capabilities of large models with the reliable execution capabilities of deterministic program tools, effectively solving the problems of difficulty in updating and maintaining rules in traditional automated systems, as well as the low efficiency of long document processing, unreliable numerical calculations, and lack of rigorous logical judgments when relying solely on large models for review. It achieves efficient, accurate, and traceable intelligent review of complex travel policies and multi-source heterogeneous invoices.

[0213] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A travel expense reimbursement review method based on a large model, characterized in that, include: Obtain travel intention information and travel policy documents. The travel intention information includes destination, time, job level and reason for travel, and the travel policy documents include unstructured travel expense management system text. The travel policy document is atomically segmented and mapped with serial numbers to generate an atomic rule sequence consisting of a unique serial number and the corresponding original text. Based on the travel intention information and the atomic rule sequence, intent recognition and rule routing are performed through a large model to generate a local rule index set for the travel intention information; Based on the local rule index set, rule sub-documents associated with the travel intention information are physically concatenated from the atomic rule sequence; The strategy simulation toolset, which is semantically bound to the rule sub-document, is invoked to perform compliance simulation and multi-scheme simulation on the travel intention information, and a travel budget blueprint including the compliance cost limit and recommended strategy is generated. After the business trip is completed, an actual reimbursement voucher set consisting of multiple receipts is obtained, and a multimodal key information extraction model is used to extract structured voucher elements from the actual reimbursement voucher set. Based on the structured credential elements, the audit intent is identified and the chain of evidence is analyzed using the large model to obtain the chain of evidence analysis conclusions. In addition, the deterministic rule execution toolkit is dynamically scheduled to perform calculations and logical verifications to obtain the execution results of the deterministic rule execution toolkit; By integrating the travel budget blueprint, the execution results of the deterministic rule enforcement toolset, and the conclusions of the evidence chain analysis, a travel audit report is generated that includes pre-planning comparison and violation tracing.

2. The travel expense reimbursement review method based on a large model according to claim 1, characterized in that, The travel policy document is atomically segmented and numbered to generate an atomic rule sequence consisting of a unique number and its corresponding original text, including: The document parsing engine is used to parse the format of the travel policy document and identify the chapter titles, clause numbers, paragraph boundaries and table structures in the travel policy document. Based on the identified hierarchical relationship of the chapter titles and the logical order of the clause numbers, a hierarchical semantic tree of the travel policy document is constructed, wherein each leaf node of the hierarchical semantic tree corresponds to a text fragment with independent and complete semantics. Using each leaf node of the hierarchical semantic tree as the smallest segmentation unit, a physical segmentation operation is performed to deconstruct the travel policy document into multiple atomic text units. Assign a globally unique and incremental serialization identifier to each of the atomic text units; Establish a mapping relationship between the serialization identifier and the original text content of the corresponding atomic text unit, sort and store all mapping relationships according to the serialization identifier, and generate the atomic rule sequence.

3. The travel expense reimbursement review method based on a large model according to claim 2, characterized in that, Based on the travel intention information and the atomic rule sequence, intent recognition and rule routing are performed through a large model to generate a local rule index set for the travel intention information, including: The travel intention information, along with all the serialization identifiers and corresponding original text summaries in the atomic rule sequence, are input into the large model. Construct a rule classification prompt template, which is used to instruct the large model to classify the terms in the atomic rule sequence into multiple predefined cost dimension categories based on the destination and reason contained in the travel intention information; The large model processes the data based on the rule classification prompt template and outputs a structured classification result. In the structured classification result, each cost dimension category is associated with one or more serialization identifiers. Extract all associated serialization identifiers from the structured classification results to form the local rule index set.

4. The travel expense reimbursement review method based on a large model according to claim 3, characterized in that, Based on the local rule index set, rule sub-documents associated with the travel intention information are physically concatenated from the atomic rule sequence, including: Based on each of the serialization identifiers contained in the local rule index set, the atomic rule sequence is traversed and queried to obtain the original text content that completely matches each of the serialization identifiers; According to the numerical order of the serialization identifier, all the obtained original text content is concatenated end to end to generate the initial concatenated text; Based on the relationship between the cost dimension category and the serialization identifier defined in the structured classification results, the text segments in the initial concatenated text are split and classified into the corresponding cost dimension category; All text paragraphs categorized under the same cost dimension category are sorted and merged according to the numerical value of their corresponding serialization identifiers to form the rule sub-documents corresponding to the cost dimension category.

5. The travel expense reimbursement review method based on a large model according to claim 1, characterized in that, The strategy simulation toolset, semantically bound to the rule sub-document, is invoked to perform compliance simulation and multi-scenario simulation on the travel intention information, generating a travel budget blueprint that includes compliance cost caps and recommended strategies, including: Input the travel intention information and the rule sub-document into the large model; The large model is used to analyze the inference intent and identify one or more target cost dimensions that need to be simulated, including: Analyze the destination, time, and reason for travel intention in the information, and match them with the clauses regarding transportation, accommodation, and subsidies in the rule sub-document; Identify one or more target cost dimensions that require cost projection; For each of the identified target cost dimensions, the large model selects and calls one or more corresponding strategy simulation tools from the pre-registered strategy simulation tool set according to the constraints in the rule sub-document, denoted as the target strategy simulation tool; Based on the travel intention information and the rule sub-document, a parameter set that conforms to the input format requirements of the target strategy simulation tool is generated, and the target strategy simulation tool is invoked. Each invoked target policy simulation tool performs compliance simulation calculations. Based on the input parameter set and benchmark data obtained from accessing external real-time data sources, it follows the calculation logic in the rule sub-document to calculate the upper limit of compliance costs for that target cost dimension. For the target cost dimension where there are multiple alternative execution schemes, the target policy simulation tool is also used to perform multi-scheme simulation steps, including: Based on different scheme parameters, multiple compliance simulation calculation steps are executed in parallel or serially to generate multiple scheme simulation results and their corresponding estimated costs and compliance status. The large model receives and integrates the compliance cost ceiling values ​​of all the target cost dimensions and the simulation results of all the schemes, and performs formatting, filling and strategy recommendation sorting according to the predefined blueprint template to generate the travel budget blueprint.

6. The travel expense reimbursement review method based on a large model according to claim 1, characterized in that, After the travel activity is completed, a set of actual reimbursement vouchers consisting of multiple receipts is obtained. A multimodal key information extraction model is then used to extract structured voucher elements from this set of actual reimbursement vouchers, including: Receive the actual reimbursement voucher set uploaded by the user, the actual reimbursement voucher set containing various types of vouchers in image format or scanned document format; Each receipt in the actual reimbursement voucher set is input into the multimodal key information extraction model; The multimodal key information extraction model identifies the preset ticket category to which each ticket belongs based on the image features and layout of the ticket. The multimodal key information extraction model simultaneously extracts textual information, visual information, and spatial location information from each ticket, and fuses and encodes the textual information, visual information, and spatial location information to generate a unified ticket feature representation. The multimodal key information extraction model, based on the invoice feature representation, uses a pre-trained sequence labeling head or region detection head to extract predefined categories of key information entities from each invoice. The key information entities include at least personnel identity, timestamp, location, amount, reason, and invoice type. The multimodal key information extraction model, based on all the extracted key information entities, establishes semantic relationships between different documents through entity disambiguation and time-space logical reasoning, and constructs the structured voucher elements with the business trip event as the core. The structured voucher elements include a set of associated entities and an entity relationship graph.

7. The travel expense reimbursement review method based on a large model according to claim 6, characterized in that, Based on the structured credential elements, the audit intent is identified and the chain of evidence is analyzed using the large model to obtain the chain of evidence analysis conclusions, including: Input the structured credential elements into the large model; The large model analyzes the bill type and the reason in the structured voucher elements to determine the category of expenses involved in this audit and the compliance assertions that need to be verified. Based on the entity relationship graph in the structured voucher elements, the association paths between the person's identity, the timestamp, the location, and the ticket type are extracted to construct an evidence chain topology network describing the entire business trip process; Traverse the evidence chain topology network, check for logical omissions or evidence breakpoints based on the mandatory clauses regarding the completeness of the invoices in the rule sub-documents, and generate a completeness check result; The entities and relationships in the evidence chain topology are matched with the conditional statements in the rule sub-document to verify whether there are logical contradictions in the travel behavior reflected by the actual reimbursement voucher set, and a consistency verification result is generated. The integrity detection results and the consistency verification results are integrated to form the evidence chain analysis conclusion.

8. The travel expense reimbursement review method based on a large model according to claim 3, characterized in that, The deterministic rule execution toolkit is dynamically scheduled to perform computation and logical verification, yielding the execution results of the toolkit, including: Based on the cost dimension category determined by the audit intent identification step and the compliance assertion that needs to be verified, one or more target deterministic rule execution tools with matching function descriptions are selected from the pre-registered deterministic rule execution tool set. Based on the structured credential elements and the rule sub-document, extract and construct parameter data that meets the input interface requirements of the target deterministic rule execution tool; Each of the target deterministic rule execution tools is invoked sequentially or in parallel, and each of the target deterministic rule execution tools performs black-box rule computation, including: Based on the deterministic business logic and calculation code encapsulated within the target deterministic rule execution tool, the input parameter data is processed to independently generate the rule execution result corresponding to the tool. The rule execution result includes at least one of compliance status, calculated amount, and violation identifier. Collect and integrate the rule execution results output by all invoked target deterministic rule execution tools to form the execution results of the deterministic rule execution tool set.

9. The travel expense reimbursement review method based on a large model according to claim 1, characterized in that, Integrating the aforementioned travel budget blueprint, the execution results of the aforementioned deterministic rule enforcement toolset, and the conclusions of the evidence chain analysis, a travel audit report is generated that includes pre-planning comparison and violation tracing, including: The compliant expense caps and recommended execution strategies for each expense dimension in the travel budget blueprint are matched and compared item by item with the actual audit results in the execution results of the deterministic rule execution toolset to calculate the deviation data. For the anomalies in the deviation data or the conflicts in the chain of evidence analysis conclusions, the rule basis and evidence missing points that lead to the anomalies or conflicts are located by combining the original text of the corresponding specific clauses in the rule sub-document. Based on the predefined audit report template, the travel budget blueprint, the actual audit results, the deviation data, the results of the root cause tracing steps, and the original text of the referenced rule sub-document clauses are organized and filled in according to the cost dimension. For each conclusion in the audit report template, add a natural language explanation based on the evidence chain topology network and rule logic to generate the final version of the travel audit report.

10. A travel expense reimbursement approval system based on a large model, characterized in that, The system applicable to the method of any one of claims 1 to 9 comprises: The information acquisition module is used to acquire travel intention information and travel policy documents. The travel intention information includes destination, time, job level and reason for travel, and the travel policy documents include unstructured travel expense management system text. The rule preprocessing module is used to perform atomic segmentation and sequence mapping on the travel policy document to generate an atomic rule sequence consisting of a unique sequence number and the corresponding original text. The intelligent routing module is used to perform intent recognition and rule routing through a large model based on the travel intention information and the atomic rule sequence, and generate a local rule index set for the travel intention information; and to physically splice out rule sub-documents associated with the travel intention information from the atomic rule sequence according to the local rule index set. The strategy deduction module is used to call the strategy simulation toolset that is semantically bound to the rule sub-document, perform compliance deduction and multi-scheme simulation on the travel intention information, and generate a travel budget blueprint that includes the compliance cost limit and recommended strategy. The voucher parsing module is used to obtain an actual reimbursement voucher set consisting of multiple receipts after the travel activity is completed, and to extract structured voucher elements from the actual reimbursement voucher set using a multimodal key information extraction model. The audit execution engine is used to identify audit intent and analyze the chain of evidence based on the structured credential elements through the large model to obtain the chain of evidence analysis conclusion; and to dynamically schedule the deterministic rule execution toolkit for calculation and logical verification to obtain the execution result of the deterministic rule execution toolkit. The report generation module is used to integrate the travel budget blueprint, the execution results of the deterministic rule enforcement toolset, and the evidence chain analysis conclusions to generate a travel audit report that includes pre-planning comparison and violation tracing.

Images