A reimbursement auditing method, system and storage medium

Abstract

This application provides a method, system, and storage medium for expense reimbursement review, including parsing expense reimbursement application data packets to extract reimbursement reason information, invoice information, and supporting material information; constructing a triplet data structure containing business entities, invoice entities, and physical evidence; performing semantic vectorization and spatiotemporal attribute alignment operations on the business entities, invoice entities, and physical evidence based on the triplet data structure to generate a multidimensional feature vector set; performing cross-validation operations on the content consistency between invoice entities and business entities, the evidentiary consistency between invoice entities and physical evidence, and the spatiotemporal logical consistency between business entities and physical evidence based on the multidimensional feature vector set to generate a consistency verification result; and generating a review result in natural language format based on the consistency verification result. This invention can accurately identify hidden violations such as genuine invoices for fake purposes and invoices issued in other locations, significantly improving the accuracy and automation level of the review.

Description

Technical Field

[0001] This invention relates to the fields of financial intelligence and artificial intelligence technology, and more specifically, to a reimbursement review method, system, and storage medium. Background Technology

[0002] In corporate financial management, expense reimbursement approval is a crucial aspect of internal control. Current technology typically relies solely on the national tax authority's interface to verify the authenticity of invoice elements such as codes, numbers, and verification codes.

[0003] With the widespread use of electronic invoices, it has become easier to forge documents using Photoshop and to repeatedly print and reimburse expenses. At the same time, employees use genuine business invoices to obtain expenses (such as using restaurant invoices from business trips to reimburse meals for personal expenses) or ask friends to collect invoices for reimbursement. The authenticity of these transactions is difficult to verify through the tax bureau's interface, resulting in a large number of cases of false reporting with genuine invoices going undetected.

[0004] Therefore, there is an urgent need for an intelligent auditing solution that can penetrate the surface of invoices and deeply verify the consistency of business operations. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a reimbursement review method, system and storage medium to accurately identify hidden violations such as genuine invoices for fraudulent purposes and invoices issued in other locations, thereby improving the accuracy of the review.

[0006] Firstly, a reimbursement review method is provided, including: Upon receiving a reimbursement application data packet awaiting review, the reimbursement application data packet is parsed to extract information on the reason for reimbursement, invoice information, and supporting materials. Based on the reimbursement reason information, invoice information, and supporting material information, a triplet data structure is constructed that includes the business entity, the invoice entity, and the physical evidence entity. Based on the triplet data structure, semantic vectorization and spatiotemporal attribute alignment operations are performed on business entities, invoice entities, and physical entities to generate a multi-dimensional feature vector set. Based on a multidimensional feature vector set, cross-validation operations are performed on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence to generate consistency verification results. Generate audit results in natural language format based on the consistency verification results.

[0007] Optionally, the reimbursement application data packet is parsed to extract information on the reason for reimbursement, invoice information, and supporting documentation, including: Based on image enhancement algorithms, image files in the reimbursement application data package are denoised and corrected to generate standardized image data. Based on optical character recognition technology and document layout analysis technology, key region segmentation and text extraction operations are performed on standardized image data and text files to generate raw text data streams. Based on predefined data cleaning rules, the original text data stream is formatted and noise filtered to generate structured reason fields, invoice fields, and supporting material fields.

[0008] Optionally, based on the cause information, invoice information, and supporting material information, a triplet data structure is constructed, including the business entity, the invoice entity, and the physical evidence entity, including: Using a named entity recognition model, business entities are constructed by extracting business time, business location, business matter, and expected target from structured reason fields. Using a named entity recognition model, invoice code, invoice date, invoice location, product name, and amount are extracted from structured invoice fields to construct invoice entities; Using a named entity recognition model, consumption details, transaction location, transaction time, check-in records, or shopping lists are extracted from structured supporting material fields to construct physical evidence entities; Based on the pre-defined relationship graph, business entities, invoice entities, and physical evidence entities are logically linked to form a triplet data structure.

[0009] Optionally, based on the triplet data structure, semantic vectorization and spatiotemporal attribute alignment operations are performed on the business entity, the invoice entity, and the physical entity to generate a multi-dimensional feature vector set, including: Based on a pre-trained semantic encoding model, semantic encoding operations are performed on the description of business items in business entities, the description of commodity names in invoice entities, and the description of consumption details in physical entities to generate semantic feature vectors. Based on the standard timeline and geographic coordinate system, the business time and location of business entities, the invoicing time and location of bill entities, and the transaction time and location of physical evidence entities are mapped to a unified spatiotemporal dimension to generate spatiotemporal aligned feature vectors. The semantic feature vector and the spatiotemporally aligned feature vector are fused and concatenated to generate a multidimensional feature vector set.

[0010] Optionally, based on a multi-dimensional feature vector set, cross-validation operations are performed on the content consistency between the bill entity and the business entity, the evidentiary consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence, including: Calculate the similarity between the semantic feature vector of the commodity name in the bill entity and the semantic feature vector of the expected subject in the business entity. If the similarity is lower than the preset threshold, the content consistency verification is deemed to have failed. Call the third-party map API interface to calculate the spatial distance between the invoice issuance location in the physical document and the transaction location in the physical document; if the spatial distance between the two exceeds the preset range, the evidence consistency verification is deemed to have failed. Verify whether the business time period in the business entity covers the transaction time in the physical entity. If the transaction time exceeds the business time period, the spatiotemporal logic consistency verification is deemed to have failed.

[0011] Optionally, an audit result in natural language format is generated based on the consistency check result, including: Based on a pre-defined risk scoring model, the total risk score of the verification failure item in the consistency verification result is calculated. Based on the range of the total risk score, the reimbursement application data to be reviewed is marked as high, medium, or low risk level; In response to being marked as high risk, an automatic rejection operation is performed; and a large language model is invoked to generate a rejection reason description in natural language format, and the rejection reason description and risk level are used together as the review result; In response to being marked as medium risk, the reimbursement application data package to be reviewed is transferred to the manual review queue, and auxiliary review information is generated, which includes highlighted suspicious features and targeted questioning scripts. The targeted questioning scripts are automatically generated by the big language model and are used to guide the manual reviewers to conduct secondary verification or request users to provide supplementary evidence. In response to being marked as low risk, the status of the pending reimbursement application data is updated to approved, and the financial payment process is triggered.

[0012] Optionally, based on a preset risk scoring model, the total risk score for verification failure items in the consistency verification results is calculated, including: Parse the consistency verification results, extract the spatiotemporal attribute conflict markers, business category semantic deviation markers, and transaction logic contradiction markers contained in the verification failure items, and map them to the corresponding basic violation scores; The system calls an external knowledge base interface to verify the business scope of the invoice issuer in the invoice information; and performs geographic reverse encoding on the location entities in the reason information, invoice information and supporting material information; if the seller's business scope does not match or the distance after geographic reverse encoding exceeds the physical feasibility threshold, the basic violation score is multiplied by the first weight to generate an enhanced risk score. Based on the invoice sequence and amount distribution characteristics in the reimbursement application data to be reviewed, abnormal behavior is identified. Abnormal behavior includes abnormal invoice number continuity, abnormal reimbursement amount splitting, and the distribution of the first digit of the amount violating Benfold's law. If the identified abnormal behavior is related to the verification failure item in terms of time or subject, the association penalty score with the second weight is added to the enhanced risk score. The risk score for the verification failure item is obtained by weighted summation of the enhanced risk score and the associated penalty score.

[0013] Optionally, the method also includes: In response to manual review and correction, the corrected data content is obtained and a negative sample set is constructed. Based on the negative sample set, the named entity recognition model is fine-tuned and updated.

[0014] Secondly, a reimbursement approval system is provided, including: The parsing unit is used to respond to the receipt of a reimbursement application data packet to be reviewed, parse the reimbursement application data packet, and extract the reimbursement reason information, invoice information, and supporting material information; The construction unit is used to build a triplet data structure containing business entities, invoice entities, and physical evidence entities based on reimbursement reason information, invoice information, and supporting material information. The processing unit is used to perform semantic vectorization and spatiotemporal attribute alignment operations on business entities, invoice entities and physical entities according to the triplet data structure, and generate a multi-dimensional feature vector set. The verification unit is used to perform cross-verification operations on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence based on a multi-dimensional feature vector set, and generate consistency verification results. The generation unit is used to generate audit results in natural language format based on the consistency verification results.

[0015] Thirdly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements any of the methods of the first aspect.

[0016] This invention provides a reimbursement review method, system, and storage medium. Upon receiving a reimbursement application data packet to be reviewed, the method parses the data packet to extract reimbursement reason information, invoice information, and supporting material information. Based on these information, a triplet data structure is constructed, comprising a business entity, a document entity, and a physical evidence. According to the triplet data structure, semantic vectorization and spatiotemporal attribute alignment are performed on the business entity, document entity, and physical evidence to generate a multidimensional feature vector set. Based on this multidimensional feature vector set, cross-validation is performed on the content consistency between the document entity and the business entity, the evidentiary consistency between the document entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence to generate a consistency verification result. A review result in natural language format is generated based on the consistency verification result. This invention, by constructing a triplet structure of business, document, and physical evidence and performing semantic and spatiotemporal multidimensional vector alignment, solves the problems of traditional reimbursement review that only verify authenticity without verifying the authenticity of the business, resulting in isolated documents and a lack of correlation verification. It can accurately identify hidden violations such as genuine tickets being used for fraudulent purposes and tickets being purchased from other locations, significantly improving the accuracy and automation level of the review process.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a reimbursement review method provided by an embodiment of the present invention is shown; Figure 2 A flowchart of another expense reimbursement review method provided by an embodiment of the present invention is shown; Figure 3 A flowchart of yet another expense reimbursement review method provided by this invention is shown; Figure 4 This diagram illustrates the structure of a reimbursement review system provided in an embodiment of the present invention. Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] This invention provides a method for expense reimbursement review, such as... Figure 1 As shown, it includes the following steps: Step S101: In response to receiving the reimbursement application data packet to be reviewed, the reimbursement application data packet is parsed to extract the reimbursement reason information, invoice information and supporting material information.

[0022] In this embodiment of the invention, the reimbursement application data packet typically contains unstructured or semi-structured data submitted by the user. Upon receiving the data packet, a preliminary unpacking and classification operation is performed to distinguish between text-based data (such as a summary of the reimbursement form) and image-based data (such as photos of invoices, screenshots of itineraries, meeting attendance sheets, etc.). A basic data parser then separates the mixed data stream into information about the reason for the expense (i.e., the user's reimbursement application form), document information (i.e., electronic data of invoices or receipts), and supporting material information (such as contracts, receipts, bank slips, etc.).

[0023] Step S102: Based on the reimbursement reason information, invoice information, and supporting material information, construct a triplet data structure containing the business entity, the invoice entity, and the physical evidence entity.

[0024] Specifically, a business entity represents the business activity described in the reimbursement application, a document entity represents the voucher for financial payment, and a physical document entity represents the physical evidence of the business transaction.

[0025] By constructing triples, these three types of entities are linked together through specific relationships to form a logical closed loop. For example, "travel application" is treated as a business entity, "flight itinerary" as a document entity, and "boarding pass" as a physical document entity, forming an approval unit through association.

[0026] Step S103: Based on the triplet data structure, perform semantic vectorization and spatiotemporal attribute alignment operations on the business entity, the invoice entity, and the physical entity to generate a multidimensional feature vector set.

[0027] In order to perform a computer-computable comparison, the above entities need to be converted into mathematical vectors.

[0028] Semantic vectorization uses pre-trained models to transform natural language descriptions into high-dimensional vectors to capture their deeper meanings; spatiotemporal attribute alignment unifies timestamps and geographical locations involved in different entities into a standard coordinate system and time axis.

[0029] The multidimensional feature vector set generated by combining these two elements contains both the semantic features of the entity and precise spatiotemporal features, providing a unified computational basis for subsequent cross-validation.

[0030] Step S104: Based on the multi-dimensional feature vector set, perform cross-validation operations on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence, and generate consistency verification results.

[0031] This can be explained as the core step in this embodiment. The content consistency verification checks whether the content of the invoice falls within the scope of the business application (e.g., the business is "business trip," and the invoice is "restaurant"). Evidence consistency verification: whether the invoice and physical evidence point to the same transaction (e.g., the distance between the invoice location and the map location); The spatiotemporal logic consistency verification checks whether the time and location of the business transaction are consistent with the physical evidence logic (e.g., whether the business time covers the consumption time).

[0032] By calculating the distance or similarity between multidimensional feature vector sets, a pass or fail verification result is generated.

[0033] Step S105: Generate audit results in natural language format based on the consistency verification results.

[0034] For example, if the spatiotemporal logic is inconsistent, a specific description will be generated stating that "the consumption location shown on the receipt does not match the business trip location," thereby improving the transparency and explainability of the audit.

[0035] This invention addresses the problems of traditional expense reimbursement audits, which only verify authenticity without verifying the genuineness of the business transaction, and suffer from isolated documents and lack of correlation verification, by constructing a triple structure of business transactions, invoices, and physical evidence, and performing semantic and spatiotemporal multidimensional vector alignment. It can accurately identify hidden violations such as genuine invoices for fraudulent purposes and invoices issued in other locations, significantly improving the accuracy and automation level of the audit.

[0036] Based on the above embodiments, the reimbursement application data packet is parsed to extract reimbursement reason information, invoice information, and supporting material information, including: Step S101A: Based on the image enhancement algorithm, perform noise reduction and correction processing on the image files in the reimbursement application data packet to generate standardized image data.

[0037] It should be noted that user-uploaded photos of invoices taken with mobile phones often have issues such as creases, shadows, and tilt. In this embodiment, generative adversarial networks or traditional image processing algorithms are applied for document denoising.

[0038] For example, firstly, the Canny edge detection algorithm or a deep learning detection model is used to identify the four corner points of the invoice in the original image. Assume the pixel coordinates of these four corner points are... , , , These coordinates constitute the source matrix, the source image coordinate matrix. .

[0039] Then according to Calculate a 3×3 homography transformation matrix from the four points in the matrix. This matrix contains the mathematical parameters needed to stretch a trapezoid (a slanted invoice) into a rectangle (a standard invoice).

[0040] Then the matrix The following mapping formula is used to calculate the positions of all pixels in the target plane, applied to the source image, to correct the image:

[0041] in, Represents the target image coordinate matrix; Represents the transformation matrix; This represents the coordinate matrix of the source image.

[0042] This step ensures the accuracy of subsequent OCR recognition, generating clear, well-formed, standardized image data.

[0043] Step S101B: Based on optical character recognition technology and document layout analysis technology, perform key region segmentation and text extraction operations on standardized image data and text files to generate the original text data stream.

[0044] Specifically, OCR (Optical Character Recognition) technology is used to convert image pixels into a character stream. Combined with document layout analysis technology, it can identify different areas in the image, such as titles, tables, and seals. For example, it can distinguish between invoice code areas and amount areas, rather than simply extracting them line by line.

[0045] For plain text files, word segmentation and paragraph splitting are performed directly. This process transforms unstructured visual or textual information into a raw text data stream that can be processed by a computer.

[0046] Step S101C: Based on predefined data cleaning rules, perform format unification and noise filtering on the original text data stream to generate structured reason fields, invoice fields, and supporting material fields.

[0047] For example, the raw text data stream may contain a large number of irrelevant characters (such as headers, footers, and advertising slogans).

[0048] This step removes noise based on predefined cleaning rules (such as regular expression matching and stop word filtering). Then, using field mapping techniques, the cleaned data is categorized into standard structured fields. For example, the extracted "2023-10-01" is mapped to the "Invoice Date" field, and "¥1000.00" is mapped to the "Amount" field.

[0049] This implementation ensures the purity and standardization of the input data, providing a high-quality data source for subsequent entity construction.

[0050] This embodiment addresses the challenges of inconsistent data source quality and diverse formats in reimbursement scenarios by introducing GAN image enhancement, edge detection and correction, and multimodal document parsing technologies. It enables intelligent full-element collection of heterogeneous documents, providing a high-quality, standardized data foundation for subsequent entity extraction and logical verification, significantly reducing the misjudgment rate caused by blurry images or format errors.

[0051] Based on the above embodiments, a triplet data structure is constructed based on the cause information, invoice information, and supporting material information, including the business entity, the invoice entity, and the physical evidence entity, such as... Figure 2 As shown, it includes: Step S102A: Using the named entity recognition model, extract the business time, business location, business matter, and expected target from the structured reason field to construct the business entity.

[0052] In this embodiment of the invention, a trained Named Entity Recognition (NER) model is used to perform in-depth analysis of the reimbursement reason.

[0053] For example, for the reason "purchasing round-trip airfare to attend the Shanghai AI Developer Conference", the model identifies "Shanghai" as the business location, "AI Developer Conference" as the business event, "2023-11-01 to 2023-11-03" as the business period, and "airfare" as the expected object. These attributes are encapsulated into a business entity object, representing the business background of this expense reimbursement.

[0054] Step S102B: Using the named entity recognition model, extract the invoice code, invoice time, invoice location, product name and amount from the structured invoice fields to construct the invoice entity.

[0055] Specifically, for invoice fields, the NER model is fine-tuned to identify specific financial terms.

[0056] For example, from an electronic invoice, we can extract "123456789" as the invoice code, "2023-11-02" as the invoice date, "XX Road, Pudong New Area, Shanghai" as the invoice location, and "Domestic Air Transport Electronic Ticket Itinerary" as the product name. These attributes constitute the physical document, representing proof of the outflow of funds.

[0057] Step S102C: Using a named entity recognition model, extract consumption details, transaction location, transaction time, check-in records or shopping lists from the structured supporting material fields to construct physical evidence entities.

[0058] Following the previous example, the NER model was also used to extract key information from supporting materials (such as boarding passes, hotel bills, and payment screenshots).

[0059] For example, "2023-11-02 08:00" can be extracted from the boarding pass as the transaction time, and "Pudong T2" as the transaction location. This information constitutes physical evidence used to prove that the transaction actually occurred.

[0060] Step S102D: Based on the preset relationship graph, perform logical linking operations on the business entity, the invoice entity, and the physical entity to form a triplet data structure.

[0061] This step connects the three independent entities mentioned above through association.

[0062] For example, establish a link between "the business entity contains the bill entity" and "the bill entity corresponds to the physical certificate entity," forming a complete triple structure.

[0063] This structured storage method clearly shows which invoice was used for which transaction, and what physical evidence supports it, providing a topological foundation for subsequent vector calculations.

[0064] Based on the above embodiments, and according to the triplet data structure, semantic vectorization and spatiotemporal attribute alignment operations are performed on business entities, invoice entities, and physical evidence entities to generate a multi-dimensional feature vector set, including: Step S103A: Based on the pre-trained semantic encoding model, perform semantic encoding operations on the business item description in the business entity, the commodity name description in the invoice entity, and the consumption detail description in the physical entity to generate semantic feature vectors.

[0065] For example, this embodiment uses the Sentence-BERT (SBERT) model to encode the text. For instance, the "description of the reason" (such as Text1) in the expense report and the "name of goods" (such as Text2) in the invoice are encoded into 768-dimensional dense vectors v1 and v2, respectively.

[0066] Specifically, the model can be trained using a contrastive learning strategy. First, a triplet sample is constructed: anchor (e.g., the definition of expense items), positive sample (e.g., compliant invoice details), and negative sample (e.g., non-compliant invoice details).

[0067] The loss function uses triplet loss:

[0068] in, Represents the anchor vector; Represents a positive sample vector; Represents a negative sample vector; This represents the minimum distance difference that must be maintained between positive and negative samples.

[0069] By minimizing this loss function, the model learns that in the semantic space, compliant invoice description vectors will be close to their corresponding expense item vectors, while being far away from non-compliant description vectors.

[0070] By applying this loss function to train the model, a highly discriminative semantic vector space can be constructed. In subsequent reviews, even if the product name on the invoice (e.g., "accommodation service fee") does not perfectly match the reimbursement reason (e.g., "business trip"), the model can still identify that they are close in the vector space (belonging to a positive sample relationship), thus determining that the content is consistent. Conversely, for invoices with semantically mismatched information (e.g., "shopping card" corresponding to "travel"), the model will identify a significant vector distance, thereby accurately intercepting irregular reimbursements.

[0071] Step S103B: Based on the standard time axis and geographic coordinate system, map the business time and location of the business entity, the invoice time and location of the invoice entity, and the transaction time and location of the physical entity to a unified spatiotemporal dimension to generate a spatiotemporal aligned feature vector.

[0072] It should be noted that the spatiotemporal information formats differ among different entities (e.g., "Shanghai" vs. "Shanghai"). These will be uniformly mapped to a standard spatiotemporal dimension: time will be uniformly represented by a UTC+8 timestamp, and location by WGS-84 or GCJ-02 geographic coordinates.

[0073] For example, both "Pudong New Area" and "Pudong" are mapped to the longitude and latitude coordinates (121.49, 31.23). The generated spatio-temporal alignment feature vectors can quantify the spatio-temporal distance between entities.

[0074] Step S103C: Perform a fusion and splicing operation on the semantic feature vectors and the spatio-temporal alignment feature vectors to generate a multi-dimensional feature vector set.

[0075] In an embodiment of the present invention, the semantic vectors generated in step S103A and the spatio-temporal vectors generated in step S103B are spliced or weighted and fused.

[0076] For example, a high-dimensional vector containing [semantic dimension 1...n, longitude, latitude, timestamp] is formed. This multi-dimensional feature vector set not only retains the semantic connotation of the entity but also accurately records its spatio-temporal position, providing a complete data basis for subsequent multi-dimensional consistency verification.

[0077] This embodiment solves the problems of weak semantic understanding and rigid rules in traditional reimbursement systems by introducing an entity extraction model and semantic encoding technology. The model trained by contrastive learning can understand that "Japanese food set meal" belongs to the semantic category of "business entertainment", achieving a qualitative change from keyword matching to semantic similarity matching, and greatly improving the audit accuracy for non-standard commodity names and complex business scenarios.

[0078] Based on the above embodiment, based on the multi-dimensional feature vector set, cross-verification operations are performed on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence entity, and the spatio-temporal logical consistency between the business entity and the physical evidence entity, as Figure 3 shown, including: Step S104A: Calculate the similarity between the semantic feature vector of the commodity name in the bill entity and the semantic feature vector of the expected subject matter in the business entity. If the similarity is lower than a preset threshold, it is determined that the content consistency verification fails.

[0079] Exemplarily, the cosine similarity algorithm can be used to calculate the semantic vector distance between a bill (such as a "dining invoice") and a business (such as a "technical seminar").

[0080] If the similarity is lower than the set threshold (such as 0.3), it means that the bill content does not match the business reason (for example, the business is "office supplies procurement" but the bill is for "consumption in an entertainment venue"), that is, it is determined that the content consistency verification fails and is marked as abnormal.

[0081] Step S104B: Call a third-party map API interface to calculate the spatial distance between the invoicing location in the bill entity and the transaction location in the physical evidence entity; if the spatial distance between the two exceeds a preset range, it is determined that the evidence consistency verification fails.

[0082] For example, the system calls the API of Gaode or Baidu Maps to calculate the straight-line distance or commuting distance between the merchant's registered address on the invoice and physical evidence (such as payment location). If the distance exceeds a preset range (such as 50 kilometers), it indicates that the invoice may have been fraudulently issued in a different location or borrowed, and the evidence consistency verification fails.

[0083] Step S104C: Verify whether the business time period in the business entity covers the transaction time in the physical entity. If the transaction time exceeds the business time period, the spatiotemporal logic consistency verification is deemed to have failed.

[0084] In this step, the time of the evidence transaction is verified. Does it fall within the business start time? and business end time Between. If or If so, it is determined that the spatiotemporal logic is inconsistent.

[0085] For example, the business entity specifies the time period as "November 1, 2023 to November 3, 2023", while the physical entity (such as a hotel bill) specifies the time period as "November 5, 2023". If the transaction time exceeds the business time period, the spatiotemporal logic consistency verification is deemed to have failed, preventing subsequent supplementary reporting or false reimbursement.

[0086] Based on the above embodiments, a natural language format audit result is generated according to the consistency verification result, including: Step S105A: Based on the preset risk scoring model, calculate the total risk score of the verification failure items in the consistency verification results.

[0087] See the description of the following embodiments for details.

[0088] Step S105B: Based on the range of the total risk score, mark the reimbursement application data to be reviewed as high, medium, or low risk level.

[0089] Based on preset risk ranges (e.g., 0-10 points for low risk, 11-30 points for medium risk, and 31 points and above for high risk), reimbursement applications are automatically classified into different risk levels. This classification strategy determines the subsequent processing procedures.

[0090] Step S105C: In response to being marked as high risk, perform an automatic rejection operation; and call the large language model to generate a rejection reason description in natural language format, and use the rejection reason description and risk level together as the review result.

[0091] If the code is deemed high-risk, it will be automatically rejected. To improve user experience, a Large Language Model (LLM) is invoked to convert the validation failure code into natural language.

[0092] For example, a reason description such as "After review, the location of your reimbursed catering invoice is too far from the business trip location, which does not comply with the company's travel regulations" can be generated and sent to the user along with a "high risk" label.

[0093] Step S105D: In response to being marked as medium risk level, the reimbursement application data packet to be reviewed is transferred to the manual review queue, and auxiliary review information containing suspicious feature highlighting prompts and targeted questioning scripts is generated. The targeted questioning scripts are automatically generated by the large language model and are used to guide the manual reviewer to conduct secondary verification or request the user to provide supplementary evidence.

[0094] For medium-risk cases, instead of direct rejection, they are transferred to human reviewers. A large language model is used to automatically generate targeted questioning scripts, such as "Please confirm whether this expense is an excessive charge due to an emergency," and suspicious characteristics are highlighted (such as excessively large amounts or remote locations). This helps human reviewers quickly pinpoint problems and improves review efficiency.

[0095] Step S105E: In response to being marked as low risk, update the status of the pending reimbursement application data to approved and trigger the financial payment process.

[0096] If the risk level is determined to be low, it means that the reimbursement application complies with all compliance criteria. Its status is automatically updated to "Approved," and it seamlessly integrates with the financial system, triggering the subsequent payment process and achieving ultra-fast reimbursement without human intervention.

[0097] Based on the above embodiments, and using a preset risk scoring model, the total risk score for verification failure items in the consistency verification results is calculated as follows: Step S105A1: Parse the consistency verification results, extract the spatiotemporal attribute conflict markers, business category semantic deviation markers, and transaction logic contradiction markers contained in the verification failure items, and map them to the corresponding basic violation scores.

[0098] For example, the verification results are broken down in detail. If a "spatiotemporal attribute conflict" is found (e.g., a person is in Beijing but the ticket is in Shanghai), a base score of 5 points is assigned. If a "business category semantic deviation" is found (such as paying for meals when buying a computer), a base score of 10 points will be assigned. If a "contradictory transaction logic" is found (such as reimbursement before the meeting), a base score of 15 points is assigned. These base scores form the foundation for risk calculation.

[0099] Step S105A2: Call the external knowledge base interface to verify the business scope of the invoice issuer in the invoice information; and perform geographic reverse encoding on the location entities in the reason information, invoice information and supporting material information; if the seller's business scope does not match or the distance after geographic reverse encoding exceeds the physical feasibility threshold, then perform the first weight multiplication processing on the basic violation score to generate an enhanced risk score.

[0100] It should be noted that this step incorporates external data to enhance risk identification. For example, if the verification reveals that the merchant issuing the invoice has a business scope of "building materials" but the reimbursement item is "consulting fees," this constitutes a mismatch in business scope, and the risk score is multiplied by 2.

[0101] Alternatively, if the geocoding shows that the invoice is located in Antarctica while the business is located in Beijing, which is physically inaccessible, the risk score will be multiplied by 2.

[0102] This multiplication process can quickly identify obviously fake invoices.

[0103] Step S105A3: Based on the invoice sequence and amount distribution characteristics in the reimbursement application data to be reviewed, abnormal behavior is identified. Abnormal behavior includes abnormal invoice number continuity, abnormal reimbursement amount splitting, and the distribution of the first digit of the amount violating Benfold's law. If the identified abnormal behavior is related to the verification failure item in terms of time or subject, the association penalty score with the second weight is added to the enhanced risk score.

[0104] For example, by using a rules engine to analyze a user's historical expense reimbursement data, if it is found that a user submits multiple invoices with consecutive numbers but amounts slightly below the approval threshold (split reimbursement), or if the amount distribution violates the natural distribution law of numbers (Benford's Law), then it is marked as abnormal behavior.

[0105] If these abnormal behaviors occur within the same time period as the current verification failure or involve the same supplier, they are considered malicious behavior, and additional associated penalty points are added.

[0106] Step S105A4: Perform a weighted summation operation on the enhanced risk score and the associated penalty score to obtain the total risk score of the verification failure item.

[0107] Finally, the enhanced risk score after multiplication is weighted and summed with the accumulated associated penalty score.

[0108] This complex calculation logic ensures that risk scoring is not based solely on a single rule violation, but rather takes into account the external environment, historical behavior, and logical connections, making risk assessment more accurate and scientific.

[0109] This embodiment constructs a fine-grained, multi-dimensional logical consistency reasoning engine and a dynamic risk scoring model. By introducing an external knowledge base for business scope verification and utilizing statistical methods (such as Benford's law) to detect abnormal amounts, it achieves precise targeting of genuine invoices with fabricated information and fraudulent transactions.

[0110] Based on the above embodiments, the method further includes: Step S106: In response to the manual review and correction operation, obtain the corrected data content and form a negative sample set.

[0111] Specifically, when a human reviewer intervenes and corrects the system's judgment (e.g., the system judges it as high-risk, but the human judges it as pass, or vice versa), the system records this human-machine difference. The corrected data (such as entities that were originally misidentified or logic that was originally judged incorrectly) is extracted and used to construct a negative sample set. These samples represent boundary cases that the current model cannot handle.

[0112] Step S107: Based on the negative sample set, perform parameter fine-tuning and update operations on the named entity recognition model.

[0113] In this embodiment of the invention, the named entity recognition model used in step S102 is incrementally trained or fine-tuned using the constructed negative sample set. The model parameters are updated through the backpropagation algorithm, enabling the model to learn the manually corrected logic.

[0114] For example, if the model frequently misidentifies "meeting tea break" as "personal catering", after fine-tuning with negative samples, the model will be able to correctly identify it next time.

[0115] This closed-loop feedback mechanism allows the expense reimbursement approval system to continuously evolve over time, resulting in increasingly higher accuracy.

[0116] Based on the same inventive concept, embodiments of the present invention provide a reimbursement approval system, such as... Figure 4 As shown, it includes: The parsing unit 401 is used to parse the reimbursement application data packet in response to receiving the reimbursement application data packet to be reviewed, and extract the reimbursement reason information, invoice information and supporting material information. Construction unit 402 is used to construct a triplet data structure containing business entity, invoice entity and physical entity based on reimbursement reason information, invoice information and supporting material information; Processing unit 403 is used to perform semantic vectorization processing and spatiotemporal attribute alignment operations on business entities, invoice entities and physical entities according to the triplet data structure, and generate a multi-dimensional feature vector set. Verification unit 404 is used to perform cross-verification operations on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence based on a multi-dimensional feature vector set, and generate consistency verification results. The generation unit 405 is used to generate audit results in natural language format based on the consistency verification results.

[0117] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 5 As shown, it includes a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504.

[0118] Memory 503 is used to store computer programs; The processor 501 is used to implement the steps of the reimbursement review method when executing the program stored in the memory 503.

[0119] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0120] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0121] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0122] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0123] In another embodiment of the present invention, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements the steps of any of the above-described reimbursement review methods. Specific implementation details can be found in the method embodiments, and will not be repeated here.

[0124] The reimbursement review system provided in this embodiment of the invention can be specific hardware on a device or software or firmware installed on the device. The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiments. For the sake of brevity, any parts not mentioned in the system embodiments can be referred to the corresponding content in the aforementioned method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, units, and processes described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0125] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces. Indirect couplings or communication connections between systems or units may be electrical, mechanical, or other forms.

[0126] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0128] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0130] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for expense reimbursement review, characterized in that, include: In response to receiving a reimbursement application data packet to be reviewed, the reimbursement application data packet is parsed to extract information on the reason for reimbursement, invoice information, and supporting materials information; Based on the reimbursement reason information, the invoice information, and the supporting material information, a triplet data structure is constructed that includes the business entity, the invoice entity, and the physical evidence entity. Based on the triplet data structure, semantic vectorization and spatiotemporal attribute alignment operations are performed on the business entity, the invoice entity, and the physical entity to generate a multidimensional feature vector set. Based on the multidimensional feature vector set, cross-validation operations are performed on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence to generate consistency verification results. The audit results are generated in natural language format based on the consistency verification results.

2. The method according to claim 1, characterized in that, The process of parsing the reimbursement application data packet extracts information on the reason for reimbursement, invoice information, and supporting documentation, including: Based on the image enhancement algorithm, the image files in the reimbursement application data package are denoised and corrected to generate standardized image data; Based on optical character recognition technology and document layout analysis technology, key region segmentation and text extraction operations are performed on the standardized image data and text files to generate the original text data stream. Based on predefined data cleaning rules, the original text data stream is formatted and noise filtered to generate structured reason fields, invoice fields, and supporting material fields.

3. The method according to claim 2, characterized in that, The construction of a triplet data structure based on the cause information, the invoice information, and the supporting material information, comprising a business entity, a document entity, and a physical evidence entity, includes: Using a named entity recognition model, the business time, business location, business matter, and expected target are extracted from the structured reason field to construct the business entity; Using the named entity recognition model, the invoice code, invoice date, invoice location, product name, and amount are extracted from the structured invoice fields to construct the invoice entity; Using the named entity recognition model, consumption details, transaction location, transaction time, check-in records or shopping lists are extracted from the structured supporting material fields to construct the physical evidence entity; Based on the preset relationship graph, the business entity, the bill entity, and the physical entity are logically linked to form a triplet data structure.

4. The method according to claim 1, characterized in that, The step involves performing semantic vectorization and spatiotemporal attribute alignment operations on the business entity, the invoice entity, and the physical entity based on the triplet data structure to generate a multidimensional feature vector set, including: Based on a pre-trained semantic coding model, semantic coding operations are performed on the business item descriptions in the business entity, the product name descriptions in the invoice entity, and the consumption detail descriptions in the physical entity to generate semantic feature vectors. Based on the standard timeline and geographic coordinate system, the business time and location of the business entity, the invoice issuance time and location of the invoice entity, and the transaction time and location of the physical entity are mapped to a unified spatiotemporal dimension to generate a spatiotemporal aligned feature vector. The semantic feature vector and the spatiotemporally aligned feature vector are fused and concatenated to generate the multidimensional feature vector set.

5. The method according to claim 1, characterized in that, The cross-validation operation based on the multi-dimensional feature vector set for the content consistency between the bill entity and the business entity, the evidentiary consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence includes: Calculate the similarity between the semantic feature vector of the product name in the bill entity and the semantic feature vector of the expected subject in the business entity. If the similarity is lower than a preset threshold, the content consistency verification is deemed to have failed. Call a third-party map API interface to calculate the spatial distance between the invoice issuance location in the invoice entity and the transaction location in the physical evidence entity; if the spatial distance between the two exceeds a preset range, the evidence consistency verification is deemed to have failed. Verify whether the business time period in the business entity covers the transaction time in the physical entity. If the transaction time exceeds the business time period, the spatiotemporal logic consistency verification is deemed to have failed.

6. The method according to claim 1, characterized in that, The step of generating a natural language format audit result based on the consistency verification result includes: Based on a preset risk scoring model, the total risk score of the verification failure item in the consistency verification result is calculated; Based on the range of the total risk score, the reimbursement application data to be reviewed is marked as high, medium, or low risk level; In response to being marked as high risk, an automatic rejection operation is performed; and a large language model is invoked to generate a rejection reason description in natural language format, and the rejection reason description and the risk level are used together as the review result; In response to being marked as medium risk level, the reimbursement application data packet to be reviewed is transferred to the manual review queue, and auxiliary review information containing suspicious feature highlighting prompts and targeted questioning scripts is generated. The targeted questioning scripts are automatically generated by the large language model and are used to guide the manual reviewer to conduct secondary verification or require the user to provide supplementary evidence. In response to being marked as low risk, the status of the pending reimbursement application data is updated to approved, and the financial payment process is triggered.

7. The method according to claim 6, characterized in that, The calculation of the total risk score for verification failure items in the consistency verification results based on the preset risk scoring model includes: The consistency verification results are analyzed, and the spatiotemporal attribute conflict markers, business category semantic deviation markers, and transaction logic contradiction markers contained in the verification failure items are extracted and mapped to the corresponding basic violation scores. The system calls an external knowledge base interface to verify the business scope of the invoice issuer in the invoice information; and performs geographic reverse encoding on the location entities in the reason information, the invoice information, and the supporting materials information; if the business scope of the seller does not match or the distance after geographic reverse encoding exceeds the physical feasibility threshold, the basic violation score is multiplied by the first weight to generate an enhanced risk score. Based on the invoice sequence and amount distribution characteristics in the reimbursement application data to be reviewed, abnormal behavior is identified. The abnormal behavior includes abnormal invoice number continuity, abnormal reimbursement amount splitting, and the distribution of the first digit of the amount violating Benfold's law. If the abnormal behavior is found to be related to the verification failure item in terms of time or subject, a second weighted association penalty score is added to the enhanced risk score. The enhanced risk score and the associated penalty score are weighted and summed to obtain the total risk score of the verification failure item.

8. The method according to claim 3, characterized in that, The method further includes: In response to manual review and correction, the corrected data content is obtained and a negative sample set is constructed. Based on the negative sample set, the named entity recognition model is fine-tuned and updated.

9. A reimbursement approval system, characterized in that, include: The parsing unit is used to parse the reimbursement application data packet in response to receiving the reimbursement application data packet to be reviewed, and extract the reimbursement reason information, invoice information and supporting material information. The construction unit is used to construct a triplet data structure containing a business entity, a document entity, and a physical evidence entity based on the reimbursement reason information, the invoice information, and the supporting material information. The processing unit is used to perform semantic vectorization processing and spatiotemporal attribute alignment operations on the business entity, the bill entity and the physical entity according to the triplet data structure, and generate a multi-dimensional feature vector set. The verification unit is used to perform cross-verification operations on the content consistency between the bill entity and the business entity, the evidence consistency between the bill entity and the physical evidence, and the spatiotemporal logical consistency between the business entity and the physical evidence based on the multidimensional feature vector set, and generate consistency verification results. The generation unit is used to generate audit results in natural language format based on the consistency verification results.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-8.

Images