An AI-based enterprise intelligent accounting full-process automatic processing method and system

Abstract

This invention discloses an AI-based automated method and system for the entire enterprise intelligent expense reimbursement process, comprising: identifying acquired expense receipts to generate receipt feature data; performing semantic analysis on extracted expense background information to obtain business elements; determining the mapping relationship between business elements and receipt feature data, calculating the semantic correlation degree based on the mapping relationship, and generating an electronic expense reimbursement form; performing compliance verification on the electronic expense reimbursement form, performing anomaly detection on the electronic expense reimbursement form, and calculating a compliance confidence score; dynamically adjusting a floating score threshold based on the semantic correlation degree, and executing a payment instruction when the compliance confidence score is greater than the floating score threshold. Compared with existing technologies, this invention overcomes the shortcomings of low efficiency in manual verification, difficulty in identifying isolated information, and rigidity in risk control rules and standards, improves the automation level of the entire expense reimbursement process, enhances the accuracy of correlation verification between receipt features and business elements, and improves the adaptability of compliance decisions.

Description

Technical Field

[0001] This invention relates to the fields of information technology and artificial intelligence, and more specifically, to an AI-based method and system for automating the entire process of intelligent expense reimbursement for enterprises. Background Technology

[0002] With the deepening evolution of global economic digital transformation and the advancement of enterprise management towards hyper-automation, the processing efficiency, risk control accuracy, and business closed-loop capability of intelligent expense reimbursement systems have become key indicators for measuring the core competitiveness of enterprise operations. As the core hub for realizing intelligent financial and tax management, the in-depth extraction of invoice information, the contextual alignment of business logic, and the intelligent measurement of compliance risks directly determine the overall effectiveness of the entire expense reimbursement chain in terms of cost reduction, efficiency improvement, and financial governance. Among these, invoice feature extraction based on machine vision is widely adopted due to its high degree of automation and wide applicability; intelligent processing solutions combining multimodal semantic parsing and adaptive threshold adjustment, with their advantages of deep logical understanding and strong decision-making flexibility, have become an ideal solution for achieving full-process, highly reliable, and contactless automated expense reimbursement.

[0003] While existing technologies include methods for processing reimbursement information by scanning invoices and supplementing them with manual verification, thereby achieving digital recording of reimbursement information, they still suffer from several problems. These include low overall processing efficiency due to the heavy reliance on manual input and visual verification throughout the entire process; difficulty in identifying fraudulent reimbursements due to the lack of logical mapping between invoice feature data and business background information; and inconsistent standards due to rigid and singular risk control audit rules, which fail to balance decision-making flexibility with risk control accuracy.

[0004] Therefore, how to provide an AI-based automated processing method for the entire enterprise intelligent expense reimbursement process that can overcome the shortcomings of low efficiency of manual verification, difficulty in identifying isolated information, and rigidity of risk control rules and standards, improve the automation level of the entire expense reimbursement process, enhance the accuracy of correlation verification between invoice features and business elements, and improve the adaptability of compliance decisions has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides an AI-based automated processing method for the entire enterprise intelligent expense reimbursement process. This method overcomes the shortcomings of low efficiency in manual verification, difficulty in identifying isolated information, and rigid risk control rules and standards. It improves the automation level of the entire expense reimbursement process, enhances the accuracy of correlation verification between invoice features and business elements, and improves the adaptability of compliance decisions.

[0006] The first technical solution provided by this invention is as follows: This invention provides an AI-based method for automating the entire process of intelligent enterprise expense reimbursement, comprising the following steps: S1 acquiring expense receipts, identifying the receipts, and generating receipt feature data; S2 extracting expense background information, performing semantic parsing on the expense background information, and obtaining business elements; S3 determining the mapping relationship between the business elements and the receipt feature data, calculating the semantic correlation degree based on the mapping relationship, and generating an electronic expense reimbursement form; S4 performing compliance verification on the electronic expense reimbursement form, and performing anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculating a compliance confidence score; S5 dynamically adjusting the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree, and executing an automatic payment instruction when the compliance confidence score is greater than the floating score threshold.

[0007] Furthermore, in a preferred embodiment of the present invention, the step of identifying the reimbursement receipt and generating receipt feature data includes: A deep learning model is used to encode and decode the reimbursement receipts, and to locate and extract key fields from the receipts. The global layout features of the reimbursement receipts are extracted based on the deep learning model, and the receipt type is determined based on the global layout features. Based on the type of invoice, locate the anti-counterfeiting feature area in the reimbursement invoice, extract the anti-counterfeiting verification features within the anti-counterfeiting feature area to verify authenticity, and generate authenticity feature labels. The key fields of the invoice, the type of invoice, and the authenticity feature labels are standardized and mapped to generate invoice feature data.

[0008] Furthermore, in a preferred embodiment of the present invention, the step of using a deep learning model to perform feature encoding and decoding on the reimbursement receipts includes: Based on the deep learning model, the reimbursement voucher is subjected to layer-by-layer convolution and downsampling processing to extract image features at different resolution levels and generate visual feature maps. A two-dimensional pixel coordinate system is constructed with the upper left corner of the reimbursement receipt as the origin, and text region detection is performed based on the visual feature map to determine the bounding box coordinates of each text line in the two-dimensional pixel coordinate system. The bounding box coordinates are normalized and mapped to position embedding vectors, and the association weight matrix between the visual feature map and the position embedding vectors is calculated. The visual feature map is weighted and concatenated based on the association weight matrix to generate a multimodal fusion feature vector. The multimodal fusion feature vector is then input into a sequence labeling decoder to extract the key fields of the invoice.

[0009] Furthermore, in a preferred embodiment of the present invention, semantic parsing is performed on the reimbursement background information to obtain business elements, including: Obtain the reimbursement background information, which includes the reimbursement reason and itinerary planning data submitted by the user; The reimbursement reason description text is segmented and part-of-speech tagged, and named entity recognition is performed in combination with the trip planning data to extract candidate semantic entities; The candidate semantic entities and the invoice feature data are mapped to the same high-dimensional feature space, and the structural dependency relationship between the candidate semantic entities and the invoice feature data is constructed. Based on the structural dependency relationship, the interaction attention weight between the candidate semantic entity and the invoice feature data is calculated, and the candidate semantic entity is filtered according to the interaction attention weight to generate business elements.

[0010] Furthermore, in a preferred embodiment of the present invention, the electronic expense reimbursement form undergoes compliance verification, and the electronic expense reimbursement form that passes the compliance verification is subjected to anomaly detection, and a compliance confidence score is calculated, including: Build a financial policy and rule engine and collect historical compliance benchmark distributions; The values ​​of each field in the electronic expense report are parsed and input into the financial policy rule engine, which includes a rule decision tree and a priority list. Based on the priority list, traverse the branch nodes of the rule decision tree from top to bottom and output the logical verification result; The electronic expense reports are filtered based on the logical verification results, and the filtered electronic expense reports are subjected to anomaly detection to uncover the hidden risk characteristics of the electronic expense reports. Based on the implicit risk characteristics, a feature vector is constructed, and the outlier deviation metric of the feature vector based on the historical compliance benchmark distribution is calculated. The outlier deviation metric is then normalized and mapped to generate the compliance confidence score.

[0011] Furthermore, in a preferred embodiment of the present invention, constructing a financial policy rule engine includes: Obtain the enterprise financial management system document, perform semantic dependency analysis on the enterprise financial management system document, and obtain rule logic tuples, which represent the relationship between reimbursement business elements and corresponding financial control standards. The rule logic tuples are mapped to standardized logic verification operators, and the logic verification operators are instantiated as rule nodes; Based on the hierarchical structure and business inclusion relationships of the clauses and chapters in the enterprise financial management system document, the logical dependencies between each rule node are determined. The rule nodes are connected according to the logical dependencies to generate the initial rule logical topology, and the rule logical topology is subjected to logical loop detection and conflict verification to generate the financial policy rule engine.

[0012] Furthermore, in a preferred embodiment of the present invention, dynamically adjusting the floating score threshold used to evaluate the compliance confidence score based on the semantic relevance includes: Collect historical basic compliance judgment thresholds and obtain preset semantic association adjustment weights; The trust deduction value is calculated based on the semantic association degree and the semantic association adjustment weight. The formula for calculating the trust deduction value is as follows: ; Wherein, V represents the trust deduction value, S represents the semantic association degree, and K represents the semantic association adjustment weight; The floating score threshold is calculated based on the historical compliance judgment threshold and the trust deduction value.

[0013] Furthermore, in a preferred embodiment of the present invention, obtaining a preset semantic association adjustment weight includes: Extract historical closed expense reimbursement form samples, obtain the semantic association degree value of each historical closed expense reimbursement form sample, and the historical compliance verification status corresponding to the semantic association degree value; The historical compliance verification status is binarized, with compliance status mapped to the value 1 and violation status mapped to the value 0. A first feature vector is constructed based on the semantic association degree value, and a second feature vector is constructed based on the binarized historical compliance verification status. Calculate the covariance between the first eigenvector and the second eigenvector, and calculate the standard deviation of the first eigenvector and the second eigenvector respectively; The covariance is normalized based on the standard deviation to obtain the positive correlation coefficient; Collect the preset maximum allowable fluctuation value, and perform a weighted mapping between the positive correlation coefficient and the maximum allowable fluctuation value to obtain the semantic association adjustment weight.

[0014] The present invention provides a second technical solution as follows: This invention also provides an AI-based enterprise intelligent expense reimbursement system with fully automated processing, including: The invoice feature extraction module acquires reimbursement invoices, identifies the reimbursement invoices, and generates invoice feature data. The business element parsing module extracts reimbursement background information, performs semantic parsing on the reimbursement background information, and obtains business elements. The semantic association modeling module determines the mapping relationship between the business elements and the invoice feature data, calculates the semantic association degree based on the mapping relationship, and generates an electronic expense reimbursement form. The compliance risk quantification module performs compliance verification on the electronic expense reimbursement form, and performs anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculates the compliance confidence score; The dynamic decision-making payment module dynamically adjusts the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree. When the compliance confidence score is greater than the floating score threshold, an automatic payment instruction is executed.

[0015] Furthermore, in a preferred embodiment of the present invention, the invoice feature extraction module includes: The key field identification unit uses a deep learning model to perform feature encoding and decoding on the reimbursement vouchers, and locates and extracts the key fields of the vouchers. The invoice type identification unit extracts the global layout features of the reimbursement invoices based on the deep learning model, and determines the invoice type of the reimbursement invoices based on the global layout features; The document authenticity verification unit locates the anti-counterfeiting feature area in the reimbursement document based on the document type, extracts the anti-counterfeiting verification features within the anti-counterfeiting feature area to verify authenticity, and generates an authenticity feature label. The feature data normalization unit standardizes and maps the key fields of the invoice, the type of invoice, and the authenticity feature label to generate invoice feature data.

[0016] This invention provides an AI-based automated processing method for the entire enterprise intelligent expense reimbursement process. It overcomes the shortcomings of low efficiency in manual verification, difficulty in identifying isolated information, and rigid risk control rules and standards. This method improves the automation level of the entire expense reimbursement process, enhances the accuracy of the correlation verification between invoice features and business elements, and improves the adaptability of compliance decisions. The AI-based automated processing method for the entire enterprise intelligent expense reimbursement process includes: S1 acquiring expense invoices, identifying the invoices, and generating invoice feature data; S2 extracting expense background information, performing semantic parsing on the expense background information, and obtaining business elements; S3 determining the mapping relationship between the business elements and the invoice feature data, calculating the semantic correlation degree based on the mapping relationship, and generating an electronic expense reimbursement form; S4 performing compliance verification on the electronic expense reimbursement form, and performing anomaly detection on the electronic expense reimbursement forms that pass the compliance verification, calculating a compliance confidence score; S5 dynamically adjusting the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree, and executing an automatic payment instruction when the compliance confidence score is greater than the floating score threshold. This application establishes a unified physical access layer for multi-source information by constructing a unified scheduling center and communicating with heterogeneous intelligent agents and their capability description data. Based on this, it performs semantic normalization on the capability description data of heterogeneous intelligent agents and constructs a capability knowledge graph. This eliminates protocol and standard differences between heterogeneous interfaces from a semantic perspective, achieving unified measurement and seamless compatibility of heterogeneous resources at the logical level. Based on the capability knowledge graph, it deconstructs and logically maps received abstract task requests level by level, generating atomic task sequences containing capability constraint tags. This transforms complex business logic into standardized sub-tasks that are machine-understandable and decomposable, solving the problem of business logic not being automatically decomposable. By collecting the running status of each heterogeneous intelligent agent in real time and performing bidirectional closed-loop adaptive matching with the capability constraint tags in the atomic task sequences, the scheduling decision process is coupled with the node load status in real time, thereby outputting a global task allocation scheme. This overcomes the drawbacks of decision-making disconnect from load and resource allocation imbalance. Finally, based on the global task allocation scheme, it drives the collaborative execution of heterogeneous intelligent agents, ensuring the reliability and efficiency of cross-system collaboration, and achieving highly automated processing of the heterogeneous intelligent agent hub and optimal allocation of system resources. Compared with existing technologies, this invention can overcome the shortcomings of low efficiency of manual verification, difficulty in identifying isolated information, and rigidity of risk control rules and standards. It can improve the automation level of the entire reimbursement process, enhance the accuracy of correlation verification between invoice features and business elements, and improve the adaptability of compliance decisions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the steps of the AI-based enterprise intelligent expense reimbursement fully automated processing method provided in this embodiment of the invention; Figure 2 A flowchart illustrating the steps for acquiring business elements provided in this embodiment of the invention; Figure 3 A logical framework diagram for calculating compliance confidence scores provided in embodiments of the present invention; Figure 4 A logical framework diagram of an AI-based enterprise intelligent expense reimbursement fully automated processing system provided in this embodiment of the invention; Figure 5 This is a comparison chart of experimental results for the invoice recognition and anti-counterfeiting verification methods in this embodiment of the invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly set on the other component; when a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to the other component.

[0021] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "first", "second", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0023] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0024] like Figures 1 to 5 As shown in the embodiments of the present invention, the AI-based enterprise intelligent expense reimbursement process automation method can overcome the shortcomings of low efficiency of manual verification, difficulty in identifying isolated information, and rigidity of risk control rules and standards. It can improve the automation level of the entire expense reimbursement process, enhance the accuracy of correlation verification between invoice features and business elements, and improve the adaptability of compliance decisions.

[0025] This invention provides an AI-based method for automating the entire process of intelligent enterprise expense reimbursement, specifically including: S1 acquiring expense receipts, identifying the receipts, and generating receipt feature data; S2 extracting expense background information, performing semantic parsing on the expense background information, and obtaining business elements; S3 determining the mapping relationship between the business elements and the receipt feature data, calculating the semantic correlation degree based on the mapping relationship, and generating an electronic expense reimbursement form; S4 performing compliance verification on the electronic expense reimbursement form, and performing anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculating a compliance confidence score; S5 dynamically adjusting the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree, and executing an automatic payment instruction when the compliance confidence score is greater than the floating score threshold. This application establishes a unified physical access layer for multi-source information by constructing a unified scheduling center and communicating with heterogeneous intelligent agents and their capability description data. Based on this, it performs semantic normalization on the capability description data of heterogeneous intelligent agents and constructs a capability knowledge graph. This eliminates protocol and standard differences between heterogeneous interfaces from a semantic perspective, achieving unified measurement and seamless compatibility of heterogeneous resources at the logical level. Based on the capability knowledge graph, it deconstructs and logically maps received abstract task requests level by level, generating atomic task sequences containing capability constraint tags. This transforms complex business logic into standardized sub-tasks that are machine-understandable and decomposable, solving the problem of business logic not being automatically decomposable. By collecting the running status of each heterogeneous intelligent agent in real time and performing bidirectional closed-loop adaptive matching with the capability constraint tags in the atomic task sequences, the scheduling decision process is coupled with the node load status in real time, thereby outputting a global task allocation scheme. This overcomes the drawbacks of decision-making disconnect from load and resource allocation imbalance. Finally, based on the global task allocation scheme, it drives the collaborative execution of heterogeneous intelligent agents, ensuring the reliability and efficiency of cross-system collaboration, and achieving highly automated processing of the heterogeneous intelligent agent hub and optimal allocation of system resources. Compared with existing technologies, this invention can overcome the shortcomings of low efficiency of manual verification, difficulty in identifying isolated information, and rigidity of risk control rules and standards. It can improve the automation level of the entire reimbursement process, enhance the accuracy of correlation verification between invoice features and business elements, and improve the adaptability of compliance decisions.

[0026] It should be noted that the semantic relevance is calculated in this embodiment as follows: First, the reimbursement reason text and the invoice product name text are converted into high-dimensional semantic vectors. Then, the cosine similarity between these two vectors is calculated. This calculation involves dividing the dot product of the two vectors by the product of their respective moduli, resulting in a value between -1 and +1. The closer this value is to +1, the more consistent the semantic direction of the reimbursement reason description with the invoice content; this value is defined as the text semantic score. Next, the absolute time difference between the trip planning time and the actual invoice issuance time is calculated. Then, this time difference is used as the independent variable and substituted into a pre-set Gaussian function model for calculation. The model includes a preset time tolerance factor (e.g., set to the standard deviation of 24 hours), causing the calculated score to decrease according to a normal distribution curve as the time difference increases. This process generates a time-related score, ensuring that spending near the travel time window receives a high score, while the score decreases the further the deviation. Finally, the absolute value of the difference between the budgeted amount and the actual amount is calculated, and then divided by the budgeted amount to obtain the deviation ratio. Subsequently, the deviation ratio is subtracted from the value (if the deviation ratio is greater than one, it is rounded down to one) to obtain the confidence score. This score intuitively reflects whether the actual spending is strictly controlled within the budget. At the same time, to address the subjectivity of weight allocation for each dimension, the "weighting factor" in this embodiment is not manually set, but automatically generated based on the statistical correlation of historical data. The system selects a sample set of audited and confirmed compliant data from the company's historical database. It calculates the maximum information coefficient between the text semantic score, time correlation score, and monetary confidence score, and the final compliance result. The maximum information coefficient captures complex linear and non-linear correlations between variables. The system normalizes the calculated three correlation coefficients, meaning the weight of a particular item equals its correlation coefficient divided by the sum of the three correlation coefficients. This weighting factor objectively reflects which dimension contributes most to the compliance determination under the current business model of the company. Finally, the system multiplies the text semantic score by its corresponding weight, adds the time correlation score multiplied by its corresponding weight, and adds the monetary confidence score multiplied by its corresponding weight. To eliminate potential negative correlation interference in text matching, the text score is linearly rectified (negative values ​​are reduced to zero) before weighting. The final accumulated value... The following describes in detail the steps of the AI-based enterprise intelligent expense reimbursement automation method with specific embodiments.

[0027] Specifically, in a specific embodiment of the present invention, the step of identifying reimbursement receipts and generating receipt feature data includes: using a deep learning model to encode and decode the features of the reimbursement receipts, locating and extracting key fields of the receipts; extracting the global layout features of the reimbursement receipts according to the deep learning model, and determining the receipt type based on the global layout features; locating the anti-counterfeiting feature area in the reimbursement receipts based on the receipt type, extracting the anti-counterfeiting verification features within the anti-counterfeiting feature area for authenticity verification, and generating authenticity feature labels; and standardizing and mapping the key fields of the receipts, the receipt type, and the authenticity feature labels to generate receipt feature data.

[0028] In a specific embodiment of the present invention, firstly, an image of the reimbursement receipt to be processed is acquired. A deep learning model is used to encode and decode the features of the reimbursement receipt. This step is mainly a preliminary processing step, where a feature extraction network is used to convert the image into a high-dimensional feature map, and the text regions on the receipt are initially located and parsed to obtain the key fields of the receipt. On this basis, the core processing logic shifts to the extraction of global layout features of the reimbursement receipt. Specifically, a global context aggregation layer, such as a global average pooling layer or a spatial pyramid pooling layer, is connected to the end of the feature extraction layer of the deep learning model to compress the two-dimensional feature map into a one-dimensional global feature vector containing the layout structure, color distribution, and texture style of the entire receipt. Based on this global layout feature, it is input into a fully connected classification network. This network is trained on a large number of receipts of different formats and can output the confidence probability of each receipt type, thereby accurately determining whether the reimbursement receipt belongs to a VAT special invoice, a fixed-amount invoice, an electronic air transport ticket itinerary, or a taxi receipt. After determining the receipt type, the system enters the crucial authenticity verification stage. The system uses a pre-set anti-counterfeiting feature positioning template for the invoice type index. This template records the relative coordinates of the anti-counterfeiting areas specified by national standards for various types of invoices, such as the QR code area in the upper left corner of a VAT invoice, the supervisory seal area in the middle, or the anti-counterfeiting fiber area in a specific location. Based on these relative coordinates, the anti-counterfeiting feature area image is accurately cropped from the original high-resolution image, and the anti-counterfeiting verification features within the area are extracted. This includes frequency domain texture analysis of the anti-counterfeiting watermark, detection of the edge sharpness of the supervisory seal, and decoding verification of the QR code information. The extracted features are input into a binary classification verifier to generate authenticity feature labels indicating whether the invoice is genuine or counterfeit. Finally, a standardized mapping is performed. The system calls the built-in field mapping rule library to convert the original field names extracted in the previous steps into system standard metadata key names. For example, "buyer" is uniformly mapped to "buyer_name", the date format is uniformly converted to a timestamp, and the currency symbol is removed from the amount value. Finally, the standardized key fields, the determined invoice type code, and the authenticity feature labels are structurally encapsulated to generate invoice feature data that can be directly called by the subsequent risk control system.

[0029] To verify the effectiveness of the above-mentioned methods for classifying invoices, verifying anti-counterfeiting measures, and standardizing data, such as... Figure 5 As shown, the experiment constructed a test dataset containing 300,000 real corporate expense reimbursement receipts, covering four major scenarios: catering, accommodation, transportation, and office procurement. This dataset includes 2,000 highly realistic forged receipts created by professionals to test the authenticity verification performance. The experiment was conducted on a server equipped with an NVIDIA A100 GPU, and the deep learning framework used was PyTorch. 1.10; In the experiment of invoice type recognition, by comparing the classification effect of using only local text features and introducing global layout features, the results show that the method based on global layout features improves the accuracy of invoice type recognition from 94.5% to 99.92%, especially in distinguishing between VAT ordinary invoices and special invoices with similar appearances, the error rate is reduced by 90%; In the anti-counterfeiting verification stage, for common features in counterfeit invoices such as blurred supervisory seals and broken anti-counterfeiting background patterns, the method based on specific area cropping and texture analysis proposed in this embodiment achieves a recall rate of 97.8% and a precision rate of 99.5%, which is significantly better than the traditional global image binary classification method; In the data standardization test, through the pre-set mapping rule library, the system successfully achieved 100% normalization of heterogeneous invoice fields from different provinces and industries, effectively solving the problem of financial system input failure caused by inconsistent field names; Comprehensive experiments show that this technical solution, while ensuring basic text recognition functions, greatly improves the security and standardization of reimbursement invoice processing by strengthening layout analysis and anti-counterfeiting verification.

[0030] Specifically, in a specific embodiment of the present invention, the steps of using a deep learning model to perform feature encoding and decoding on the reimbursement receipt include: performing layer-by-layer convolution and downsampling processing on the reimbursement receipt based on the deep learning model to extract image features at different resolution levels and generate a visual feature map; constructing a two-dimensional pixel coordinate system with the upper left corner of the reimbursement receipt as the origin, and performing text region detection based on the visual feature map to determine the bounding box coordinates of each text line in the two-dimensional pixel coordinate system; normalizing the bounding box coordinates and mapping them to position embedding vectors, and calculating the association weight matrix between the visual feature map and the position embedding vectors; weighting and concatenating the visual feature map based on the association weight matrix to generate a multimodal fusion feature vector, and inputting the multimodal fusion feature vector into a sequence label decoder to extract the key fields of the receipt.

[0031] In this embodiment of the invention, a deep neural network integrating spatial awareness is constructed, specifically including four sub-steps: multi-scale visual feature extraction, positional encoding embedding, multimodal attention fusion, and sequence decoding. First, layer-by-layer convolution and downsampling are performed. The preprocessed ticket image is input into a backbone architecture based on a feature pyramid network, such as ResNet-101. This backbone contains four stages of convolutional modules. Each stage performs downsampling through convolution or pooling operations with a stride of 2, extracting features at a resolution equal to the original. Figure 1 Feature maps at resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 are used. These feature maps at different resolution levels not only contain low-level texture edge information but also high-level semantic abstraction information. The system upsamples the deep features and then performs lateral concatenation and pixel-by-pixel addition with the shallow features, ultimately outputting a visual feature map V that integrates multi-scale information. Secondly, a two-dimensional pixel coordinate system is constructed and position embeddings are generated. The origin (0,0) is set at the top left corner of the ticket image, with the positive X-axis pointing horizontally to the right and the positive Y-axis pointing vertically downwards. A lightweight text detection branch is used to predict the four-point bounding box coordinates (x, y) of each text line on the visual feature map. min ,y min ,x max ,y max The coordinate values ​​are normalized by dividing the image width and height, ensuring they fall within the [0,1] interval. The normalized coordinates are then input into a multilayer perceptron or sinusoidal position encoder, mapping them to a fixed-dimensional high-dimensional vector, the position embedding vector P. This vector explicitly represents the spatial layout of the text field on the ticket page. Next, the association weight matrix is ​​calculated, generating multimodal fusion features. This is the core step in implementing the attention mechanism. The system transforms the visual feature map V into a query vector and the position embedding vector P into a key vector. The similarity between the two is calculated through dot product operations, and then normalized using Softmax to obtain the association weight matrix A. Each element in this matrix... A i，j Represents the first i The visual feature point pair of the first jThe model focuses on the attention level of each location region. Based on the weight matrix A, the visual feature map V is weighted and summed, enabling the model to dynamically focus on key areas of the invoice according to location cues (e.g., when the location embedding indicates "bottom right corner", the model automatically suppresses noise in the top left corner visual features). Then, the weighted features are concatenated with the original visual features in the channel dimension, and after 1×1 convolution dimensionality reduction, a multimodal fusion feature vector M containing rich visual information and sensitive to spatial location is generated. Finally, sequence labeling and decoding are performed. The multimodal fusion feature vector M is input into the sequence labeling decoder, which adopts a bidirectional long short-term memory network combined with a conditional random field architecture, or a Transformer Decoder architecture. The global probability transition matrix of the CRF layer is used to constrain the legality of the label sequence (e.g., the "amount" label cannot be directly followed by the "date" label). The feature sequence is classified frame by frame, and the entity label corresponding to each character is output, such as B-TotalAmount, I-TotalAmount, thereby accurately extracting the key fields of the reimbursement invoice.

[0032] It should be noted that, in this embodiment, the deep learning model preferably adopts an end-to-end heterogeneous network structure of ResNet-101 backbone network + FPN feature pyramid + Transformer decoder + CRF conditional random field; wherein, ResNet-101 is used as encoder to extract features at resolutions of the original... Figure 4The network consists of four feature maps ranging from 1 / 1 to 1 / 32. The FPN module is responsible for mapping the number of channels in the feature maps at different scales to a unified 256-dimensional dimension and performing multi-scale fusion. The Transformer decoder is configured with a 6-layer stacked structure, with each layer containing 8 multi-head attention mechanism components. The hidden layer dimension of the feedforward neural network is set to 1024, and the dropout ratio is set to 0.1. The CRF layer is placed at the end of the network to learn the transition probability matrix between labels, and to perform global constraints and path optimization on the entity label sequence of the decoded output. Regarding the training dataset parameters, a dedicated hybrid dataset containing 500,000 invoice images was constructed. This includes 150,000 real de-identified enterprise invoices with manual entity-level annotation, and 350,000 synthetic augmented data generated through random background replacement, motion blur simulation, and random perspective transformation (rotation angle ±15 degrees) to address the problem of insufficient long-tail invoice samples in the real data. For key hyperparameter configuration, model training was deployed on a server cluster equipped with four NVIDIA A100 GPUs, and the deep learning framework used was PyTorch 1.10. Input images were uniformly preprocessed to a resolution of 640 x 640 pixels and mean normalized. The optimizer used was AdamW, with a weight decay coefficient set to 0.01%. The learning rate was adjusted using a "linear warm-up combined with cosine annealing" strategy, with an initial learning rate of 5 x 10^-4 and a warm-up period of 5 epochs, followed by a cosine decay to 1 x 10^-6. The loss function consisted of a weighted average of the Smooth L1 loss for text detection, the cross-entropy loss for character recognition, and the negative log-likelihood loss for sequence labeling, with a weight ratio of 1:1:0.5. The global batch size was set to 128, the total number of iterations was 100 epochs, and a random weight averaging strategy was used in the last 20 epochs to improve the model's generalization accuracy.

[0033] To verify the effectiveness of the aforementioned encoding and decoding method based on association weight matrix and multimodal fusion features, a dedicated dataset containing 200,000 complex-format reimbursement invoices was constructed. This dataset covers various types, including VAT invoices, itinerary slips, and fixed-amount invoices, and specifically includes 50,000 challenging samples with interference features such as folding, occlusion, and text misalignment. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The experiment was conducted on a server equipped with four NVIDIA A100 graphics processors, using PyTorch. 1.13 Deep learning framework was used to build the model. During training, the batch size was set to 32, the initial learning rate was set to 1e-4, and a cosine annealing strategy was used for dynamic adjustment. The loss function consisted of a weighted average of text detection loss, position regression loss, and sequence decoding loss. To quantify the technical contribution of the "position embedding and association weight matrix," an ablation comparison group was set up in the experiment. The baseline model used only pure visual features for decoding, while the model in this application fully incorporated the position embedding and attention weighting modules. Experimental results showed that on the test set, the overall F1 score of the model in this application for extracting key fields of invoices reached 99.6%, compared to 97.2% for the baseline model. The accuracy was improved by 2.4 percentage points. In particular, in the recognition of the easily confused "capitalized amount" and "lowercase amount" fields, thanks to the effective use of the correlation weight matrix for page position information, the model can correct according to the spatial rule that "capitalized amount is usually located above or to the left of lowercase amount", which reduces the recognition error rate of this type of field by 45%. In addition, when processing invoice images with a tilt angle of more than 15 degrees, the average intersection-union ratio (IoU) of the field localization of the model in this application remains above 0.92, which verifies the robustness of multimodal fusion features in complex spatial transformation scenarios and proves the significant effect of this technical solution in improving the accuracy of automated enterprise expense reporting.

[0034] Specifically, such as Figure 3 As shown, in a specific embodiment of the present invention, semantic parsing is performed on the reimbursement background information to obtain business elements, including: S21 obtaining reimbursement background information, which includes the reimbursement reason submitted by the user and the trip planning data; S22 performing word segmentation and part-of-speech tagging on the reimbursement reason description text, and performing named entity recognition in combination with the trip planning data to extract candidate semantic entities; S23 mapping the candidate semantic entities and the invoice feature data to the same high-dimensional feature space, and constructing a structural dependency relationship between the candidate semantic entities and the invoice feature data; S24 calculating the interaction attention weight between the candidate semantic entities and the invoice feature data based on the structural dependency relationship, and filtering the candidate semantic entities according to the interaction attention weight to generate business elements.

[0035] In this embodiment of the invention, the reimbursement background information is first obtained, specifically including unstructured reimbursement reason text input by employees and structured travel planning data. To accurately understand financial terminology, a dedicated language model is constructed. The construction process is as follows: First, the pre-trained weights of a general BERT model are loaded. This model consists of 12 stacked Transformer encoder layers and uses a WordPiece tokenizer to process the input text, converting the reimbursement reason text into a combined vector sequence containing token embeddings, position embeddings, and paragraph embeddings. Based on this, a specific output layer for sequence labeling tasks is added to the top-level output of the BERT model, such as a Conditional Random Field (CRF) layer or a linear classification layer, to predict the entity label of each token in the financial context. To adapt the model to the reimbursement scenario, a pre-training to fine-tuning mode is adopted. A dedicated dataset is constructed by collecting millions of text descriptions from historical reimbursement documents of the enterprise, and the model is fully optimized. Parameters are fine-tuned to accurately identify specific business entities such as business banquets and customer visits, thereby extracting candidate semantic entities. Subsequently, to verify the authenticity of these entities, a multimodal feature interaction space is constructed, mapping the extracted candidate semantic entity vectors and invoice feature data to the same high-dimensional vector space. A heterogeneous graph structure is built using a graph attention network (GAT), where semantic entities and invoice features serve as graph nodes, and their temporal overlap and semantic similarity serve as edge weights. Based on this graph structure, the interaction attention weights between nodes are calculated, i.e., the degree of attention paid by invoice feature nodes to semantic entity nodes. If a catering entity has a high attention weight with an invoice node that matches the amount in the same time period, the entity is deemed valid. Finally, based on the calculated interaction attention weights, candidate semantic entities are threshold-filtered, eliminating isolated entities with weights below a preset value, generating business elements corroborated by invoices, completing the logical loop from subjective description to objective fact.

[0036] Specifically, in this embodiment of the invention, compliance verification is performed on the electronic expense reimbursement form, and anomaly detection is performed on the electronic expense reimbursement forms that pass the compliance verification to calculate a compliance confidence score. This includes: constructing a financial policy rule engine and collecting historical compliance benchmark distributions; parsing the values ​​of each field of the electronic expense reimbursement form and inputting each field value into the financial policy rule engine, which includes a rule decision tree and a priority list; traversing the branch nodes of the rule decision tree from top to bottom according to the priority list and outputting logical verification results; filtering the electronic expense reimbursement forms according to the logical verification results and performing anomaly detection on the filtered electronic expense reimbursement forms to uncover the implicit risk characteristics of the electronic expense reimbursement forms; constructing a feature vector based on the implicit risk characteristics, calculating the outlier deviation metric of the feature vector based on the historical compliance benchmark distribution, normalizing and mapping the outlier deviation metric, and generating the compliance confidence score.

[0037] In a specific embodiment of the present invention, a pre-built financial policy rule engine and historical compliance benchmark distribution data are first loaded. The historical compliance benchmark distribution is a statistical feature model generated based on all audited and confirmed compliant expense reimbursement documents from the past three fiscal years, including the mean, variance, and probability density distribution of various expenses in a high-dimensional space. Subsequently, the system starts a parsing program to read the values ​​of various fields in the electronic expense reimbursement form, such as reimbursement amount, expense type, occurrence time, number of participants, and invoice tax rate, and passes these field values ​​as input parameters to the financial policy rule engine. The rule engine internally maintains a set of dynamic rule decisions. The system employs a tree-like structure and a priority list. The priority list defines the execution order. The system traverses the branch nodes of the rule decision tree from top to bottom according to the priority list. For example, when traversing the travel expense decision tree, the root node determines the city level, the left child node determines whether accommodation is included, and the leaf nodes output logical judgments indicating whether the expense exceeds the limit or is compliant. The final output is a logical verification result containing a pass, rejection, or warning status. Based on this logical verification result, an initial screening is performed. Electronic expense reports with a pass or warning status are sent to the deep anomaly detection module. This module aims to uncover hidden risk characteristics in electronic expense reports, which are often difficult to detect using simple If-Else rules. Therefore, the system... A deep neural network model based on a variational autoencoder (VAE) is constructed. First, discrete fields (such as department ID) of the electronic expense report are embedded and encoded using embedding. Continuous fields (such as amount) are normalized using Z-scores, and the resulting concatenation generates a high-dimensional feature vector. This feature vector is then input into the encoder network of the VAE model. The encoder consists of three fully connected layers with 128, 64, and 32 nodes respectively, using ReLU activation to compress the input into a latent variable distribution. The latent variables are then reconstructed into an output vector through a decoder network. The system calculates the Euclidean distance between the input feature vector and the reconstructed output vector. Alternatively, the cosine distance can be used, which represents the outlier deviation of the current expense report relative to the historical compliance baseline distribution. This is because the reconstruction error of compliant samples is extremely small, while outlier samples, due to their non-compliance with historical distribution patterns, will have a high reconstruction error. Finally, a normalization mapping is performed, and the upper limit of the preset deviation threshold is read. This threshold is determined based on the reconstruction error quantile of 99.9% of the samples in the training set. The outlier deviation measure is mapped to a value between 0 and 100 using an inverse sigmoid function or a linear scaling formula, generating a compliance confidence score. This can effectively identify complex fraudulent behaviors that comply with the rules but are statistically abnormal, such as splitting large amounts into smaller ones or frequent expense reports.

[0038] To verify the effectiveness of the dual verification method based on rule-based decision trees and variational autoencoders, a dataset containing 1 million real historical expense reimbursement records from enterprises was constructed. The training set consisted of 980,000 unlabeled historical records (assuming most were compliant), and the test set consisted of 20,000 records, including 500 manually injected complex violation samples (such as frequent cash withdrawals within compliant limits and unusually large dining expenses outside of working hours). The experimental model was built on the PyTorch 1.12 framework, with an NVIDIA Tesla T4 GPU. The input layer of the variational autoencoder network was set to 64 dimensions, including 12 statistical features such as amount, time interval, and average departmental reimbursement amount, as well as embeddings. The loss function consisted of a weighted sum of reconstruction error loss (MSE) and KL divergence loss, with weight coefficients set to 1.0 and 0.001, respectively. The optimizer was Adam, with an initial learning rate of 1e-3. With a size of 256, the training iterations were performed for 100 epochs until the loss converged to the order of 1e-4. Experimental results show that during the rule engine filtering stage, it can 100% intercept hard violations such as excessive document amounts and incorrect invoice headers. In the subsequent anomaly detection stage, for those "hidden risk" samples that bypass the rule engine, the outlier detection method based on reconstruction error proposed in this embodiment achieved an AUC of 0.96, significantly better than the 0.89 of the traditional isolated forest algorithm. In particular, for fraud patterns of "breaking down large amounts into smaller ones," since the individual amounts are normal but the overall feature vector deviates from the baseline distribution in the latent space, the compliance confidence score calculated by the model is on average lower than 0.3, successfully triggering a manual review warning. Ablation experiments further demonstrate that, in the normalization mapping step, introducing a dynamic threshold based on historical data quantiles reduces the false alarm rate of the system by 15% compared to a fixed threshold, verifying the dual advantages of this technical solution in ensuring financial security and improving audit efficiency.

[0039] Specifically, in a specific embodiment of the present invention, constructing a financial policy rule engine includes: acquiring a corporate financial management system document; performing semantic dependency analysis on the corporate financial management system document to obtain rule logic tuples, whereby the rule logic tuples represent the association between reimbursement business elements and corresponding financial control standards; mapping the rule logic tuples to standardized logic verification operators and instantiating the logic verification operators into rule nodes; determining the logical dependency relationships between each rule node based on the clause and chapter hierarchy and business inclusion relationships in the corporate financial management system document; connecting each rule node according to the logical dependency relationships to generate an initial rule logic topology structure; and performing logical loop detection and conflict verification on the rule logic topology structure to generate the financial policy rule engine.

[0040] In a specific embodiment of this invention, the system first obtains the enterprise financial management system document, typically in PDF or Word format, through a document parsing interface. Using regular expressions combined with document layout analysis technology, it identifies the chapter numbers, clause numbers, and indentation levels to construct a document tree object reflecting the document's nested structure. Then, it performs semantic dependency analysis on the final clause text within the document tree, employing a dependency parsing model based on the Transformer architecture. This model first inputs the clause text into a BERT encoder to obtain context-dependent word vector representations. Then, it uses two multilayer perceptrons (MLPs) to calculate the score vectors between words as head nodes and modifier nodes, respectively. Finally, it uses a biaffine attention mechanism to calculate the probability distribution and relation labels of dependency arcs, thereby parsing the subject-verb-object structure and modification relations in the sentence. The system extracts rule-based logical tuples based on the parsed dependency tree structure. For example, it extracts (accommodation fee, less than or equal to, 800, first-tier city) as a tuple from the rule that accommodation fees in first-tier cities must not exceed 800 yuan. This tuple contains business elements, logical relations, thresholds, and preconditions. Then, it calls... Using a pre-built operator library containing standardized logical verification operators such as greater than, less than, contain, and belong, the extracted rule logic tuples are mapped to executable code functions, and these functions are instantiated as rule node objects in memory. To determine the execution order between nodes, the system establishes logical dependencies based on the previously constructed document tree hierarchy and business inclusion relationships. For example, all rule nodes under the Travel Expense Management Regulations section are set as child nodes of the Travel Control Node, and transportation expenses and accommodation expenses are set as parallel sibling nodes. Based on this, an initial rule logic topology structure, i.e., a Directed Acyclic Graph (DAG), is generated. Finally, strict logical loop detection and conflict verification are performed. The Tarjan algorithm is used to traverse the topology structure to detect the existence of infinite loop paths, and an interval overlap detection algorithm or a Satisfiability Modulus (SMT) solver is used to perform conflict verification on multiple rule nodes of the same business element. For example, it checks whether there are contradictory rules such as the reimbursement amount must be greater than 1000 and the reimbursement amount must be less than 500. After eliminating all logical errors, the final deployable financial policy rule engine is serialized, which can significantly reduce the complexity and error rate of manual rule configuration.

[0041] To verify the accuracy and robustness of the aforementioned automated financial policy rule engine construction method, the experiment collected real financial management system documents from 50 large and medium-sized enterprises in different industries, containing approximately 12,000 specific financial control clauses covering multiple business areas such as travel, procurement, and benefits. The dependency parsing model was fine-tuned based on the RoBERTa-Large pre-trained model. The training data included 5,000 manually annotated financial rule corpora, with annotations covering dependency arcs and logical tuples. Model training was conducted on a dual-processor NVIDIA A100 processor. The training was conducted on a GPU server with a batch size of 32, a learning rate of 1e-5, and gradient descent optimization using the cross-entropy loss function. The training period was 50 epochs. Experimental results show that the model achieves an accuracy of 94.5% in dependency arc connection and a 92.8% F1 score in extracting rule logic tuples on the semantic dependency analysis task of financial terms. In the topology construction and verification stage, the system successfully detected 15 logical conflicts (such as contradictions between old and new policy terms) and 3 logical infinite loops in historical documents. Performance comparison tests show that the automated construction method described in this embodiment reduces the rule engine configuration work, which originally required two weeks from professional consultants, to within one hour, and increases the rule coverage from 85% to over 98% with manual configuration. This effectively verifies the practical value and efficiency of this technical solution in the digital transformation of enterprise finance.

[0042] Specifically, in embodiments of the present invention, dynamically adjusting the floating score threshold used to assess compliance confidence scores based on the degree of semantic association includes: collecting historical basic compliance judgment thresholds and obtaining preset semantic association adjustment weights; calculating a trust deduction value based on the degree of semantic association and the semantic association adjustment weights, wherein the trust deduction value is calculated using the following formula: Where V represents the trust deduction value, S represents the semantic relevance, and K represents the semantic relevance adjustment weight; the floating score threshold is calculated based on the historical basic compliance judgment threshold combined with the trust deduction value.

[0043] In a specific embodiment of this invention, historical compliance judgment thresholds are first collected from the system configuration center or high-performance database. These thresholds are not arbitrarily set but are determined based on the confidence distribution curve of all historical compliant reimbursement documents in the anomaly detection model. Specifically, the working point corresponding to a 99% recall rate is selected as the baseline, for example, set to 0.85. Simultaneously, a preset semantic association adjustment weight is obtained. This weight is a hyperparameter characterizing the influence of semantic evidence on risk tolerance. Next, the trust deduction value is calculated. The system performs numerical calculations on the semantic association degree calculated in the previous steps and the obtained semantic association adjustment weight, specifically using a linear weighted product method. That is, the product of the semantic association degree value and the semantic association adjustment weight value is calculated to obtain the trust deduction value. This trust deduction value quantifies the risk score that can be exempted due to the high degree of consistency between the cause and the invoice. Subsequently, based on this trust deduction value... The system adjusts the historical compliance threshold by subtracting the trust deduction value from the historical compliance threshold, resulting in a dynamic floating score threshold. This allows for lenient review of documents with high semantic relevance. For example, if a reimbursement has extremely high semantic relevance, indicating a very genuine transaction background, even if its statistical characteristics are slightly abnormal, causing the compliance confidence score to be slightly below the baseline, the reduced floating score threshold will still allow it to pass the verification. Finally, the system compares the calculated compliance confidence score with the floating score threshold. When the compliance confidence score is greater than or equal to the floating score threshold, the electronic reimbursement form is deemed to meet dynamic compliance requirements. An automated processing instruction is immediately generated and passed. This instruction includes calling the enterprise-bank direct connection interface to generate a payment message or directly triggering a payment operation. This effectively solves the technical problems of high false negative rates and large workload of manual review caused by traditional rigid thresholds.

[0044] Specifically, in a specific embodiment of the present invention, obtaining a preset semantic association adjustment weight includes: extracting historical closed reimbursement form samples, obtaining the semantic association degree value of each historical closed reimbursement form sample, and the historical compliance verification status corresponding to the semantic association degree value; binarizing the historical compliance verification status, mapping the compliance status to the value 1 and the violation status to the value 0; constructing a first feature vector based on the semantic association degree value, and constructing a second feature vector based on the binarized historical compliance verification status; calculating the covariance between the first feature vector and the second feature vector, and calculating the standard deviation of the first feature vector and the second feature vector respectively; normalizing the covariance based on the standard deviation to obtain a positive correlation coefficient; collecting a preset maximum allowable fluctuation value, and weighting the positive correlation coefficient and the maximum allowable fluctuation value to obtain the semantic association adjustment weight.

[0045] Specifically, in a specific embodiment of the present invention, firstly, historical expense reimbursement form samples that have been closed within the past year are extracted from the enterprise data warehouse. For example, 100,000 records that have completed financial audits are extracted. The semantic relevance value calculated by the NLP model during the audit process for each sample is obtained, as well as the final audit conclusion of the sample, i.e., the historical compliance verification status. Next, the data is cleaned and binarized, mapping the status of "approved" or "compliant" to the value 1, and mapping the status of "rejected," "violation," or "fraud" to the value 0, thereby transforming the qualitative compliance status into a quantitative numerical indicator. Based on the above-processed data, two numerical sequences of equal length are constructed, where the semantic relevance value constitutes the first feature vector, and the binarized compliance verification status constitutes the second feature vector. The system then calls a statistical analysis algorithm to calculate the covariance between the two feature vectors, which reflects the overall change in semantic score and compliance result. The system calculates the standard deviations of the first and second eigenvectors to measure the dispersion of their respective data. Based on the standard deviations, the covariance is normalized by dividing the covariance value by the product of the two standard deviations, thus obtaining the positive correlation coefficient, or Pearson correlation coefficient, which characterizes the linear correlation between the two. The closer this coefficient is to 1, the more statistically valid the hypothesis that the reasoning is clearer and more closely matches the invoice, and the higher the likelihood of compliance. Finally, the system collects a preset maximum allowable fluctuation value, which is an upper limit parameter set by the company's financial risk control committee based on the annual risk tolerance. For example, 0.2 indicates a maximum allowable threshold downward fluctuation of 20%. The system weights the calculated positive correlation coefficient with this maximum allowable fluctuation value, typically using multiplication, to obtain the final semantic association adjustment weight. This weight dynamically reflects the actual reference value of semantic quality for compliance judgment in the current business environment.

[0046] To verify the scientific validity and effectiveness of the aforementioned weight calculation method based on statistical correlation, the experiment selected a sample dataset of expense reimbursements from a specific business department of a large enterprise over the past two quarters, totaling 20,000 records. This dataset included various typical business types such as routine travel, entertainment expenses, and office procurement. The experiment first calculated the baseline correlation of this batch of data. The results showed a positive correlation coefficient of 0.78 between the degree of semantic association and the final compliance status, indicating that high-quality semantic descriptions are indeed a strong signal of compliance. The experiment set a maximum allowable fluctuation value of 0.15, meaning the threshold could be lowered by a maximum of 0.15. Based on the method in this embodiment, the semantic association adjustment weight was calculated to be 0.117. To compare the effects, the experiment set up three control groups: A Group A uses a fixed weight of 0.05, Group B uses a fixed weight of 0.15, and Group C uses a dynamic weight of 0.117 calculated in this embodiment. Test results show that when processing a batch of grayscale samples with high semantic matching but at the critical value of the amount, the automated pass rate of Group C is 18% higher than that of Group A, which greatly reduces unnecessary manual review. Compared with Group B, Group C reduces the miss rate of truly illegal samples by 4.5%. In addition, when the system periodically reruns the calculation process, the calculated correlation coefficient will fluctuate as fraud methods evolve, thereby automatically adjusting the weight. Experiments have shown that this dynamic adaptive mechanism enables the system to automatically tighten the weight when facing new illegal patterns, maintaining the long-term robustness of the risk control model.

[0047] This invention also provides an AI-based enterprise intelligent expense reimbursement fully automated processing system, such as... Figure 4 As shown, the system includes: a receipt feature extraction module, which acquires reimbursement receipts, identifies the receipts, and generates receipt feature data; a business element parsing module, which extracts reimbursement background information, performs semantic parsing on the background information, and acquires business elements; a semantic association modeling module, which determines the mapping relationship between the business elements and the receipt feature data, calculates the semantic association degree based on the mapping relationship, and generates an electronic reimbursement form; a compliance risk quantification module, which performs compliance verification on the electronic reimbursement form, performs anomaly detection on the electronic reimbursement form that passes the compliance verification, and calculates a compliance confidence score; and a dynamic decision payment module, which dynamically adjusts the floating score threshold used to evaluate the compliance confidence score based on the semantic association degree, and executes an automatic payment instruction when the compliance confidence score is greater than the floating score threshold.

[0048] Specifically, in a specific embodiment of the present invention, the invoice feature extraction module includes: a key field identification unit, which uses a deep learning model to perform feature encoding and decoding on the reimbursement invoice, locates and extracts key fields of the invoice; an invoice type identification unit, which extracts the global layout features of the reimbursement invoice according to the deep learning model, and determines the invoice type of the reimbursement invoice based on the global layout features; an invoice authenticity verification unit, which locates the anti-counterfeiting feature area in the reimbursement invoice based on the invoice type, extracts the anti-counterfeiting verification features in the anti-counterfeiting feature area for authenticity verification, and generates authenticity feature labels; and a feature data normalization unit, which standardizes and maps the key fields of the invoice, the invoice type and the authenticity feature labels to generate invoice feature data.

[0049] As described above, the AI-based enterprise intelligent expense reimbursement process automation method of this invention specifically includes: S1 acquiring expense receipts, identifying the expense receipts, and generating receipt feature data; S2 extracting expense background information, performing semantic parsing on the expense background information, and obtaining business elements; S3 determining the mapping relationship between the business elements and the receipt feature data, calculating the semantic association degree based on the mapping relationship, and generating an electronic expense reimbursement form; S4 performing compliance verification on the electronic expense reimbursement form, and performing anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculating a compliance confidence score; S5 dynamically adjusting the floating score threshold used to evaluate the compliance confidence score based on the semantic association degree, and executing an automatic payment instruction when the compliance confidence score is greater than the floating score threshold. This application establishes a unified physical access layer for multi-source information by constructing a unified scheduling center and communicating with heterogeneous intelligent agents and their capability description data. Based on this, it performs semantic normalization on the capability description data of heterogeneous intelligent agents and constructs a capability knowledge graph. This eliminates protocol and standard differences between heterogeneous interfaces from a semantic perspective, achieving unified measurement and seamless compatibility of heterogeneous resources at the logical level. Based on the capability knowledge graph, it deconstructs and logically maps received abstract task requests level by level, generating atomic task sequences containing capability constraint tags. This transforms complex business logic into standardized sub-tasks that are machine-understandable and decomposable, solving the problem of business logic not being automatically decomposable. By collecting the running status of each heterogeneous intelligent agent in real time and performing bidirectional closed-loop adaptive matching with the capability constraint tags in the atomic task sequences, the scheduling decision process is coupled with the node load status in real time, thereby outputting a global task allocation scheme. This overcomes the drawbacks of decision-making disconnect from load and resource allocation imbalance. Finally, based on the global task allocation scheme, it drives the collaborative execution of heterogeneous intelligent agents, ensuring the reliability and efficiency of cross-system collaboration, and achieving highly automated processing of the heterogeneous intelligent agent hub and optimal allocation of system resources. Compared with existing technologies, this invention can overcome the shortcomings of low efficiency of manual verification, difficulty in identifying isolated information, and rigidity of risk control rules and standards. It can improve the automation level of the entire reimbursement process, enhance the accuracy of correlation verification between invoice features and business elements, and improve the adaptability of compliance decisions.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fully automated enterprise intelligent expense reimbursement process based on AI, characterized in that, include: S1 acquires expense receipts, identifies the expense receipts, and generates receipt feature data; S2 extracts reimbursement background information, performs semantic parsing on the reimbursement background information, and obtains business elements; S3 determines the mapping relationship between the business elements and the invoice feature data, calculates the semantic association degree based on the mapping relationship, and generates an electronic expense reimbursement form; S4 performs compliance verification on the electronic expense reimbursement form, and performs anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculates a compliance confidence score; S5 dynamically adjusts the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree. When the compliance confidence score is greater than the floating score threshold, an automatic payment instruction is executed.

2. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 1, characterized in that, The step involves identifying the reimbursement receipts and generating receipt feature data, including: A deep learning model is used to encode and decode the reimbursement receipts, and to locate and extract key fields from the receipts. The global layout features of the reimbursement receipts are extracted based on the deep learning model, and the receipt type is determined based on the global layout features. Based on the type of invoice, locate the anti-counterfeiting feature area in the reimbursement invoice, extract the anti-counterfeiting verification features within the anti-counterfeiting feature area to verify authenticity, and generate authenticity feature labels. The key fields of the invoice, the type of invoice, and the authenticity feature labels are standardized and mapped to generate invoice feature data.

3. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 2, characterized in that, The steps involve using a deep learning model to perform feature encoding and decoding on the reimbursement receipts, including: Based on the deep learning model, the reimbursement voucher is subjected to layer-by-layer convolution and downsampling processing to extract image features at different resolution levels and generate visual feature maps. A two-dimensional pixel coordinate system is constructed with the upper left corner of the reimbursement receipt as the origin, and text region detection is performed based on the visual feature map to determine the bounding box coordinates of each text line in the two-dimensional pixel coordinate system. The bounding box coordinates are normalized and mapped to position embedding vectors, and the association weight matrix between the visual feature map and the position embedding vectors is calculated. The visual feature map is weighted and concatenated based on the association weight matrix to generate a multimodal fusion feature vector. The multimodal fusion feature vector is then input into a sequence labeling decoder to extract the key fields of the invoice.

4. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 1, characterized in that, Semantic parsing is performed on the reimbursement background information to obtain business elements, including: Obtain the reimbursement background information, which includes the reimbursement reason and itinerary planning data submitted by the user; The reimbursement reason description text is segmented and part-of-speech tagged, and named entity recognition is performed in combination with the trip planning data to extract candidate semantic entities; The candidate semantic entities and the invoice feature data are mapped to the same high-dimensional feature space, and the structural dependency relationship between the candidate semantic entities and the invoice feature data is constructed. Based on the structural dependency relationship, the interaction attention weight between the candidate semantic entity and the invoice feature data is calculated, and the candidate semantic entity is filtered according to the interaction attention weight to generate business elements.

5. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 1, characterized in that, The electronic expense reimbursement form undergoes compliance verification, and the electronic expense reimbursement forms that pass the compliance verification are subjected to anomaly detection. A compliance confidence score is calculated, including: Build a financial policy and rule engine and collect historical compliance benchmark distributions; The values ​​of each field in the electronic expense report are parsed and input into the financial policy rule engine, which includes a rule decision tree and a priority list. Based on the priority list, traverse the branch nodes of the rule decision tree from top to bottom and output the logical verification result; The electronic expense reports are filtered based on the logical verification results, and the filtered electronic expense reports are subjected to anomaly detection to uncover the hidden risk characteristics of the electronic expense reports. Based on the implicit risk characteristics, a feature vector is constructed, and the outlier deviation metric of the feature vector based on the historical compliance benchmark distribution is calculated. The outlier deviation metric is then normalized and mapped to generate the compliance confidence score.

6. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 5, characterized in that, Build a financial policy and rule engine, including: Obtain the enterprise financial management system document, perform semantic dependency analysis on the enterprise financial management system document, and obtain rule logic tuples, which represent the relationship between reimbursement business elements and corresponding financial control standards. The rule logic tuples are mapped to standardized logic verification operators, and the logic verification operators are instantiated as rule nodes; Based on the hierarchical structure and business inclusion relationships of the clauses and chapters in the enterprise financial management system document, the logical dependencies between each rule node are determined. The rule nodes are connected according to the logical dependencies to generate the initial rule logical topology, and the rule logical topology is subjected to logical loop detection and conflict verification to generate a financial policy rule engine.

7. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 1, characterized in that, The floating score threshold used to evaluate the compliance confidence score is dynamically adjusted based on the degree of semantic relevance, including: Collect historical basic compliance judgment thresholds and obtain preset semantic association adjustment weights; The trust deduction value is calculated based on the semantic association degree and the semantic association adjustment weight. The formula for calculating the trust deduction value is as follows: ； in, V This represents the trust deduction value. S Indicates the degree of semantic association. K This indicates the semantic association adjustment weight; The floating score threshold is calculated based on the historical compliance judgment threshold and the trust deduction value.

8. The AI-based enterprise intelligent expense reimbursement process automation method according to claim 7, characterized in that, Obtain the preset semantic association adjustment weights, including: Extract historical closed expense reimbursement form samples, obtain the semantic association degree value of each historical closed expense reimbursement form sample, and the historical compliance verification status corresponding to the semantic association degree value; The historical compliance verification status is binarized, with compliance status mapped to the value 1 and violation status mapped to the value 0. A first feature vector is constructed based on the semantic association degree value, and a second feature vector is constructed based on the binarized historical compliance verification status. Calculate the covariance between the first eigenvector and the second eigenvector, and calculate the standard deviation of the first eigenvector and the second eigenvector respectively; The covariance is normalized based on the standard deviation to obtain the positive correlation coefficient; Collect the preset maximum allowable fluctuation value, and perform a weighted mapping between the positive correlation coefficient and the maximum allowable fluctuation value to obtain the semantic association adjustment weight.

9. An AI-based enterprise intelligent expense reimbursement fully automated processing system, characterized in that, The system is used to execute the AI-based enterprise intelligent expense reimbursement full-process automated processing method according to any one of claims 1 to 8, including: The invoice feature extraction module acquires reimbursement invoices, identifies the reimbursement invoices, and generates invoice feature data. The business element parsing module extracts reimbursement background information, performs semantic parsing on the reimbursement background information, and obtains business elements. The semantic association modeling module determines the mapping relationship between the business elements and the invoice feature data, calculates the semantic association degree based on the mapping relationship, and generates an electronic expense reimbursement form. The compliance risk quantification module performs compliance verification on the electronic expense reimbursement form, and performs anomaly detection on the electronic expense reimbursement form that passes the compliance verification, and calculates the compliance confidence score; The dynamic decision-making payment module dynamically adjusts the floating score threshold used to evaluate the compliance confidence score based on the semantic correlation degree. When the compliance confidence score is greater than the floating score threshold, an automatic payment instruction is executed.

10. The AI-based enterprise intelligent expense reimbursement fully automated processing system according to claim 9, characterized in that, The invoice feature extraction module includes: The key field identification unit uses a deep learning model to perform feature encoding and decoding on the reimbursement receipts, and locates and extracts the key fields of the receipts. The invoice type identification unit extracts the global layout features of the reimbursement invoices based on the deep learning model, and determines the invoice type of the reimbursement invoices based on the global layout features; The document authenticity verification unit locates the anti-counterfeiting feature area in the reimbursement document based on the document type, extracts the anti-counterfeiting verification features within the anti-counterfeiting feature area to verify authenticity, and generates an authenticity feature label. The feature data normalization unit standardizes and maps the key fields of the invoice, the type of invoice, and the authenticity feature label to generate invoice feature data.

Images