Intelligent auxiliary system and method for enterprise financial reimbursement based on artificial intelligence technology

Abstract

The invention provides an enterprise financial reimbursement auxiliary system and method based on an artificial intelligence technology. The system comprises a user operation interface, an administrator operation interface, a standard system database, a large language model, a financial knowledge question and answer module, a receipt uploading and recognition module, a paper report generation module, an online form filling guide module and the like. Through an innovative financial knowledge enhanced contrast learning framework and an intelligent question and answer and verification mechanism, the accuracy of the large language model in the financial field is remarkably improved. The efficiency and accuracy of reimbursement processing are improved, and a sustainable, optimized and highly intelligent financial management solution is provided for enterprises.

Description

Technical Field

[0001] The present invention belongs to the technical field of enterprise financial reimbursement management, and relates to an enterprise financial reimbursement auxiliary system and method based on artificial intelligence technology. Background Art

[0002] Financial expense reimbursement systems play a vital role in business operations. First, they significantly reduce errors associated with manual processing by automating processes. Second, they provide accurate financial data, helping companies with cost control and budget management. Furthermore, financial expense reimbursement systems enhance transparency and compliance, ensuring that every expense is properly reviewed and approved, reducing the risk of fraud. Finally, they bring transparency to the expense reimbursement process, thereby improving overall operational efficiency.

[0003] However, users face many pain points and challenges in the existing financial reimbursement system. These pain points mainly include the following aspects: 1. High communication and learning costs due to changes in reimbursement standards and regulations: Corporate financial reimbursement standards and regulations frequently change based on policy adjustments and internal management needs. Every change requires notification to employees, who then need to relearn the new reimbursement rules and procedures. This frequent change not only increases employee learning costs but also increases communication costs between the finance department and employees, often requiring repeated explanation and clarification of new regulations.

[0004] 2. Reimbursement delays caused by changes in the reimbursement process and rework: The reimbursement process is complex and ever-changing, making it easy for employees to make mistakes when submitting reimbursement applications. For example, filling out incorrect information or omitting essential information can result in applications being rejected and requiring resubmission. This rework not only increases employee workload but also prolongs processing time, leading to delays in reimbursement payments and impacting employee cash flow.

[0005] 3. Parallel online form filling and offline paper receipt signing and review: Currently, many companies' reimbursement systems still operate in parallel with online and offline systems. Employees must complete reimbursement application forms online and simultaneously prepare paper receipts for signing and review. This dual-track system results in a significant amount of duplication of effort. Employees must not only enter reimbursement information into the system but also prepare paper documents for offline processing, increasing workload and the potential for errors.

[0006] 4. Duplicate reimbursement information: Due to the duplication of online and offline systems, employees often need to repeatedly enter the same information when filling out reimbursement applications. This is not only time-consuming and labor-intensive, but also prone to data inconsistencies and errors. When reviewing reimbursement documents, financial personnel also need to verify the consistency of information in the online system and offline paper documents, further increasing the complexity and time cost of the review.

[0007] 5. Poor user experience: Many existing financial expense reimbursement systems lack user-friendly interfaces, are cumbersome, and unintuitive, often confusing users and requiring significant time to familiarize themselves with and operate the system. The systems lack intelligent and automated features, failing to effectively reduce the user burden or provide personalized guidance and recommendations tailored to their needs.

[0008] Existing AI-based financial reimbursement systems still have some limitations. For example, Patent No. 202410600551.5, "Financial Reimbursement Method Based on Human-Computer Interaction," and Patent No. 202410170390.0, "Method for Managing a RAG Natural Language Intelligent Knowledge Base," both utilize natural language models and large language models to process user input and generate responses. While these methods have made some progress in improving user experience and streamlining operational processes, they are prone to errors when processing precise financial data involving monetary values.

[0009] While large language models offer powerful natural language processing capabilities, they also carry the risk of generating "hallucinations"—generating seemingly plausible but inaccurate information. In areas such as financial reimbursement, which require extreme accuracy, relying solely on large language models can lead to incorrect data processing or inappropriate decision recommendations. Furthermore, large language models can be susceptible to misleading inputs, making them ineffective in identifying and preventing potential fraud. Summary of the Invention

[0010] To build a financial reimbursement system that provides intelligent and personalized services while ensuring data processing accuracy and reliability, this system will utilize innovative methods such as multi-source data cross-validation, a professional financial rules engine, and a human-machine collaborative review mechanism. This system will maximize the accuracy and compliance of financial data while improving reimbursement efficiency and user experience, providing enterprises with a secure, efficient, and intelligent financial reimbursement solution. Therefore, this invention proposes an enterprise financial reimbursement assistance system and method based on artificial intelligence technology.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is as follows: In one aspect, the present invention provides an intelligent assistance system for enterprise financial reimbursement based on artificial intelligence technology, comprising: The user operation interface module provides a user interaction interface; the user interaction interface includes a first user interface for the user to input documents and basic user information, and a second user interface for providing answer information to the user, identifying basic user information, and providing prompt messages to the user; The administrator operation interface module provides administrators with functions including authority control, standard system database update, historical reimbursement data summary and financial risk warning; The standard system database module stores the company's financial reimbursement rules, automatically updates and pushes updated financial reimbursement rules; The large language model module has a built-in financial management large language model. Through multimodal optimization of the financial management large language model, it provides support for reimbursement process management and user interaction. The financial knowledge question-and-answer module, based on a large language model, provides interpretation and question-and-answer functions; The document upload and recognition module uploads user-entered documents including invoices and tickets, automatically recognizes the document content through OCR technology, and fills the document content into the corresponding reimbursement application form; Paper report generation module, which automatically generates paper reports after users upload information; The online form filling guidance module guides users to correctly fill in reimbursement information in the financial system.

[0012] Furthermore, the construction method of the user operation interface module and the administrator operation interface module is: Design the user interface of the user operation interface module based on the user interaction interface and the functions of providing answer information to users, identifying basic user information and providing prompt messages to users; Design the system interface of the administrator operation interface module based on functions including authority control, standard system database update, historical reimbursement data aggregation and financial risk warning; Based on the front-end framework and back-end technology stack, develop front-end pages and back-end services respectively to implement permission management, document upload, standard system update, data aggregation and analysis functions; Conduct functional, performance, and security testing, collect feedback from users and administrators, and optimize system functionality and user experience.

[0013] Furthermore, the construction method of the standard system database module is: Collect internal and external financial reimbursement rules of the enterprise; Digitize the rules. If they are paper documents, use OCR technology to convert them into editable electronic text; Clean and pre-process the text data obtained after digitization to remove noise and redundant information and unify the format; Classify and label the cleaned and pre-processed data, and store them in a structured manner according to categories including different reimbursement items and policy terms; Select a database management system corresponding to a relational database or a non-relational database, and design the database table structure and index; Import the classified and labeled data into the database, create interfaces and tools for interacting with the database, and complete the construction of the standard system database module.

[0014] Furthermore, the financial management language model built into the large language model module is trained based on the data in the standard system database module, and the specific method is as follows: Extract financial reimbursement rules from the standard system database module; Preprocess financial reimbursement rules and construct financial professional word vectors ,in Indicates the Vector representation of financial terms; Design special financial task reminder templates , used to generate financial reimbursement related question-answer key-value pair data; Generate a large-scale training dataset based on question-answer key-value pairs ,in It's a problem. is the corresponding standard answer; Introduce the financial reimbursement rule mechanism into the large language model module and define a financial reimbursement rule function , the financial reimbursement rule function accepts input And returns the financial rule vector related to it. The specific formula is: ; in It is the encoding output when the financial management language model does not introduce financial reimbursement rules. is the equilibrium parameter; The training process uses a dynamic batch size strategy to automatically adjust the number of samples in each batch based on the complexity of the financial reimbursement rules, and controls the output of the financial management large language model through a loss function. When new financial reimbursement rules are detected, the incremental learning process is automatically triggered.

[0015] Furthermore, the loss function in the training process of the financial management large language model is: ; in, The cross entropy loss ensures that the large language model for financial management can generate the correct answer; is a comparative loss used to distinguish correct from incorrect financial information. is the consistency loss, and is the weight coefficient; The calculation of is as follows: ; in It's a problem and answers The similarity score of is the correct answer, is the wrong answer; The calculation of is as follows: ; After each training batch, the model performance is evaluated using any financial scenario test set.

[0016] Furthermore, the specific method for evaluating the performance of the financial management large language model using any financial scenario test set is: Introducing financial risk scoring indicators, comprehensively considering the performance of the financial management large language model in terms of accuracy, compliance, and robustness, and introducing long-term financial risk scoring trends; Financial risk scoring indicators The formula is:

[0017] in is the weight coefficient; Accuracy evaluation The specific calculation formula is: ; in, For a real example, is a true negative example, For a false positive example, For false negative examples, the accuracy of model training is calculated based on the prediction results of the financial management large language model on the test set; Compliance Evaluation The specific calculation formula is: ; in is the number of rule conflicts, is the total number of test cases. This indicator measures the degree to which the financial management language model complies with the financial reimbursement rules; Robustness evaluation The specific calculation formula is: ; in is the number of errors in the perturbation input, is the total number of perturbation inputs. During the robustness test, the test set is perturbed to evaluate the performance of the financial management large language model on the perturbation inputs. Long-term financial risk score trends The calculation formula is: ; in is the current FRS score, is the initial baseline score; The three indicators of financial risk scoring and the long-term financial risk score trend are calculated. Users set different weights for each item and calculate the comprehensive score. The comprehensive score is used to evaluate the performance of the financial management large language model in a single scenario.

[0018] Furthermore, the method for constructing the financial knowledge question-answering module based on the financial management large language model built into the large language model module is as follows: Use reinforcement learning and human feedback techniques to train large language models and define model policies ,in are model parameters, is the current state, is the model’s answer; Modeling human feedback as a reward function , to maximize the expected reward, maximize the expected reward The specific formula is: ; Update model parameters through policy gradient method: ; in is the learning rate, is the model parameter vector; A multi-round dialogue mechanism is introduced to enhance the model's ability to understand context and capture key information through the attention mechanism. The specific formula is: ; in 、 、 are query, key, and value matrices respectively, is the transpose of the key matrix, is the dimension of the key.

[0019] Furthermore, the financial knowledge question-answering module based on the financial management large language model in the large language model module also includes a financial reimbursement rule engine, specifically: Define a set of financial reimbursement rules , each financial reimbursement rule There is an associated validation function , the return value is , the initial answer generated by the large language model It will be verified by the financial reimbursement rules engine. The specific formula is: ; when When , the answer is considered valid; when When the answer is judged as invalid, the system will generate correction suggestions based on the violated financial reimbursement rules and require the model to regenerate the answer; A dynamic learning loop is also implemented, defining an update function ,in is the new training data, including manual intervention and correction. The specific formula of the update function is: ; This function updates the model parameters at the same time and rule sets ; Finally, through continuous performance monitoring, define a comprehensive performance indicator : ; in is the weight coefficient, which is adjusted according to specific business needs. The definitions of each indicator are as follows: Accuracy evaluation : ; in is the number of correct answers, is the total number of questions, and the correct answers are determined based on human review or automated test sets, which measures the ability of the model to provide accurate information; Compliance Evaluation : ; in The number of violations of financial reimbursement rules, is the total number of answers, and the financial reimbursement rule violations are determined by the financial reimbursement rule engine, which measures the extent to which the model complies with the latest financial reimbursement rules; Responsiveness evaluation : ; in, is the average time it takes for the large language model for financial management to generate an answer, measured in seconds. For normalization, its reciprocal is taken, which measures the system's ability to quickly respond to user queries. The value range of each indicator is ,in Indicates the best performance, by adjusting the weight 、 、 , can emphasize different aspects of performance according to needs; In addition, define a dynamic optimization threshold : ; in It is the performance benchmark for the initial deployment of a large language model for financial management. It is a coefficient greater than 1 and is used to set goals for continuous improvement; when When the financial management language model optimization process is triggered, it includes retraining the financial management language model and updating the financial reimbursement rules.

[0020] In another aspect, the present invention provides an intelligent assistance method for enterprise financial reimbursement based on artificial intelligence technology, comprising: The administrator selects data through the administrator operation interface module to build the standard system database module; The financial management language model built into the language model module is trained using the data stored in the standard system database module; The financial management large language model built into the large language model module provides technical support for the financial knowledge question and answer module, document upload and recognition module, paper report generation module, and online form filling guidance module; Users interact through the user operation interface module, the financial knowledge question and answer module, the document upload and recognition module, the paper report generation module and the online form filling guidance module to complete the reimbursement process.

[0021] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides an enterprise financial reimbursement assistance system and method based on artificial intelligence technology, which has the following beneficial effects: the system effectively improves the accuracy and efficiency of reimbursement processing through an innovative financial knowledge-enhanced comparative learning framework and an intelligent question-answering and verification mechanism. Combined with a large language model and a professional financial rule engine, the system realizes intelligent and personalized reimbursement process management. Multi-source data cross-validation and dynamic learning cycles ensure continuous optimization of system performance, especially when dealing with complex financial policies. The system's multi-round dialogue capability significantly improves the user experience and makes the reimbursement process more convenient. Its adaptive learning ability enables it to quickly adapt to the company's specific needs and policy changes, reducing maintenance costs. Through real-time performance monitoring and optimization, the system always maintains efficient operation, providing enterprises with an intelligent and sustainable financial management solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of the module architecture of the system of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, the present invention aims to provide an enterprise financial reimbursement assistance system and method based on artificial intelligence technology, which solves the pain points of the existing system through the following technical means.

[0025] An enterprise financial reimbursement assistance system based on artificial intelligence technology, comprising: The user interface module provides a simple and intuitive user interface. Users only need to enter questions or documents and basic user information in the first user interface, and the system will provide answers, automatically identify and fill in the necessary data. The system also provides interactive content such as providing answers, identifying basic user information, and providing prompts to users in the second user interface.

[0026] The administrator operation interface module provides a simple administrator interface, supports functions such as permission control, updating of standard system database, summary of historical reimbursement data and financial risk warning, and simplifies the administrator's operation process.

[0027] The Standard System Database module stores the company's financial reimbursement rules. The system automatically updates and pushes the latest financial reimbursement rules to ensure users always use the latest rules. These financial reimbursement rules include the company's internal financial policies, systems, and standards, as well as relevant external laws and regulations.

[0028] The large language model module, a core module, has a built-in financial management large language model. Through multimodal optimization of the financial management large language model, it supports intelligent reimbursement process management and user interaction, thereby improving the intelligence and accuracy of the system.

[0029] The financial knowledge question and answer module, based on a large language model, provides intelligent interpretation and question and answer functions to help users quickly understand and adapt to new financial reimbursement rules, reducing communication and learning costs.

[0030] In the document upload and recognition module, users can upload invoices, tickets and other document information. The system automatically recognizes the document content through OCR technology and fills it into the corresponding reimbursement application form, reducing manual input errors.

[0031] Paper report generation module: after the user uploads the information, the system automatically generates a paper report for the user to sign and review offline.

[0032] In the online form filling guidance module, the system generates detailed online form filling guidance to guide users to correctly fill in the reimbursement information in the financial system to ensure a smooth reimbursement process.

[0033] The financial system is the internal financial system of the user's company and includes a general ledger module, accounts receivable module, accounts payable module, cash management module, fixed asset accounting module, salary accounting module, cost module, financial planning module, financial analysis module, financial decision module, report generation module, budget management module, etc. The artificial intelligence-based enterprise financial reimbursement intelligent assistance system provided in this application is an auxiliary tool for the financial reimbursement function in the financial system.

[0034] Each module in the above-mentioned enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology is constructed by the following method.

[0035] 1. Data preprocessing: Extract relevant data from the company's ERP, financial, and human resources systems. Use Python's pandas library to clean and structure the data, unifying the data from different sources into a unified format.

[0036] 2. Financial Knowledge Graph Construction: Use the Neo4j graph database to store the financial knowledge graph. Use natural language processing to extract entities and relationships from corporate financial system documents and construct a knowledge graph containing information such as reimbursement categories, amount limits, and approval processes.

[0037] 3. Large Language Model Fine-tuning: Based on the pre-trained ChatGLM model, we fine-tune it using a company-specific financial dataset to generate a large language model for financial management. This model uses Hugging Face's Transformers library to fine-tune the model, focusing on optimizing its understanding of financial terminology and policy interpretation.

[0038] 4. Define a financial rules engine: This is implemented using the Drools rules engine. Internal financial policies, systems, and standards, as well as external laws and regulations, are encoded and constructed into financial reimbursement rules. These rules are then integrated with the output of the financial management language model to ensure that the generated answers comply with the financial reimbursement rules.

[0039] 5. Multimodal input processing: Use OpenCV to process uploaded bill images and integrate the Tesseract OCR engine for text recognition.

[0040] 6. User Interface: The React framework is used to develop the web front-end, providing an intuitive user interface. The Django REST framework is used to build the API back-end, ensuring efficient communication between the front-end and back-end.

[0041] 7. Performance Monitoring and Optimization: Use Prometheus to collect system performance metrics in real time and combine it with Grafana for visualization. Set up automated scripts to trigger the retraining of the financial management language model and the update of financial reimbursement rules when the performance metrics of the financial management language model fall below the preset threshold.

[0042] The AI-powered enterprise financial reimbursement intelligent assistance system is deployed in Docker containers, orchestrated using Kubernetes to ensure high availability and scalability. This modular and containerized design allows the system to flexibly meet the needs of diverse enterprises while ensuring ease of deployment and maintenance.

[0043] After the overall framework of the enterprise financial reimbursement system based on artificial intelligence technology is completed, each specific module will be constructed and improved.

[0044] The construction method of the user operation interface module and the administrator operation interface module is: First, conduct a needs analysis to determine the core functions. Design the user interface of the user operation interface module based on the user interaction interface and the functions of providing answers to users, identifying users' basic information, and providing prompt messages to users.

[0045] The system interface of the administrator operation interface module is designed based on functions such as authority control, update of standard system database, summary of historical reimbursement data and financial risk warning.

[0046] Next, select the appropriate front-end framework and back-end technology stack, developing the front-end pages and back-end services, respectively, to implement functions such as permission management, document upload, standard system updates, and data aggregation and analysis. Front-end frameworks can include React, Vue, and Angular, while back-end technology stacks can include Node.js, Django, and Spring Boot. To ensure data security and privacy, strict access control and logging are implemented to ensure system security.

[0047] Finally, conduct comprehensive functional, performance, and security testing, collect feedback from users and administrators, and continuously optimize the system's functionality and user experience.

[0048] The construction method of the standard system database module is as follows: First, collect the company's internal and external financial reimbursement rules, including policy documents, internal corporate systems, laws, government regulations, etc. These documents can exist in electronic documents, paper documents, etc.

[0049] Next, the financial reimbursement rules are digitized. If they are paper documents, they are converted into editable electronic text using OCR technology.

[0050] Then, the digitized text data is cleaned and preprocessed to remove noise and redundant information, unify the format, and ensure the consistency and accuracy of the data.

[0051] Afterwards, the cleaned and pre-processed data will be classified and labeled, and stored in a structured manner according to different reimbursement items, policy terms and other categories.

[0052] Next, design and implement the database storage architecture, select an appropriate database management system (RDB), such as a relational or non-relational database, and design the database table structure and indexes based on actual needs. Import the classified and labeled data into the database to ensure data integrity and queryability.

[0053] Finally, we developed database interaction interfaces and tools to facilitate data query and update functionality, ensuring that the standard system database reflects the latest financial reimbursement rules. Through these steps, we built a structured, maintainable, and scalable standard system database module, providing reliable data support for subsequent financial management.

[0054] The method for training the financial management big language model built into the big language model module based on the standard system database module is as follows.

[0055] This method uses an innovative financial knowledge-enhanced contrastive learning framework to incorporate financial knowledge into the financial management language model, combining the advantages of professional financial knowledge and the language model. The specific process is as follows: First, extract financial reimbursement rules from the standard system database module, including financial reimbursement regulations, policies, and historical records. Preprocess these financial reimbursement rules and construct financial professional word vectors. ,in Indicates the A vector representation of financial terms.

[0056] Next, design a special financial task reminder template , used to generate a large number of financial reimbursement-related question-answer key-value pairs. These templates are the basis for generating training data. For example: Q1: According to the travel expenses management regulations, what is the maximum limit for domestic air tickets? a1: According to the latest travel expense management regulations, the maximum reimbursement limit for domestic air tickets is the economy class fare. Special circumstances require approval from the department head.

[0057] Q2: When on a business trip in Beijing, what is the upper limit of daily accommodation expenses? a2: According to current regulations, the upper limit for accommodation expenses on business trips in Beijing is RMB 500 per night. Exceeding this limit requires special application and approval.

[0058] Q3: What documents are required for reimbursement of transportation expenses? A3: To claim transportation expenses, you must provide the following documents: 1. Original receipts (e.g., air tickets, train tickets, taxi tickets, etc.); 2. Electronic ticket itinerary (for air tickets); 3. Reimbursement application form. If you drive, you must also provide toll and gas receipts.

[0059] These question-answer pairs constitute the training dataset used to fine-tune the large language model so that it can accurately answer questions related to financial reimbursement. Using these templates and information from the standard system database, a large-scale training dataset is generated based on question-answer key-value pairs. ,in It's a problem. is the corresponding standard answer.

[0060] Then, the financial reimbursement rule mechanism is introduced into the financial management language model. Define a financial reimbursement rule function , the specific formula is: ; in For input No. The judgment function of the financial reimbursement rules, is its corresponding weight. For example, You can check whether the reimbursement amount exceeds the limit. The validity of the invoice can be verified, ….

[0061] Financial reimbursement rule function Accepting Input And return the financial reimbursement rule vector related to it. Integrated into the encoder of the large language model for financial management: ; in It is the encoding output when the financial management language model does not introduce financial reimbursement rules. is a balance parameter. In this embodiment .

[0062] During training, we use Optimizer, with learning rate set to , the batch size is 32. Based on this, we define the loss function of FACL: ; in: The cross entropy loss ensures that the large language model for financial management can generate the correct answer; It is a comparative loss used to distinguish correct from incorrect financial information; It is the consistency loss, which ensures that the output of the large language model of financial management is consistent with the financial rules; and is the weight coefficient.

[0063] The calculation of is as follows: ; in It's a problem and answers The similarity score of Is the correct answer, that is, a positive sample; is a wrong answer, that is, a negative sample.

[0064] The calculation of is as follows: ; The training process uses a dynamic batch size strategy to automatically adjust the number of samples in each batch based on the complexity of the financial rules. At the same time, it introduces progressive financial knowledge distillation technology to transfer professional financial knowledge into the model in stages.

[0065] After each training batch, the performance of the financial management large language model is evaluated using a specific financial scenario test set. A financial risk scoring indicator is introduced to comprehensively consider the model's performance in terms of accuracy, compliance, and robustness. The formula for the financial risk scoring indicator is:

[0066] in is the weight coefficient, which is adjusted according to specific business needs. The definitions of each indicator are as follows: 1. For accuracy evaluation: ; in, For a real example, is a true negative example, For a false positive example, is a false negative. Accuracy Calculated based on the prediction results of the financial management large language model on the test set, this measures the ability of the financial management large language model to make correct decisions in various financial reimbursement scenarios.

[0067] 2. For compliance evaluation: ; in is the number of rule conflicts, is the total number of test cases. This metric measures the degree to which the financial management language model adheres to financial rules and policies. Rule violations may include exceeding reimbursement limits or missing necessary documents.

[0068] 3. For robustness evaluation: ; in is the number of errors in the perturbation input, is the total number of perturbed inputs. To test the robustness of the model, we perturb the test set, such as adding noise or changing the decimal point of the amount, and then evaluate the model's performance on these perturbed inputs. This measures the resilience of the large language model for financial management to anomalous or potentially fraudulent inputs.

[0069] The value range of each indicator is , where 1 indicates the best performance. By adjusting the weights , , You can emphasize different aspects of performance according to the needs of the enterprise. For example, for enterprises that pay special attention to compliance, you can increase The value of .

[0070] In addition, the system also regularly calculates a long-term FRS trend to evaluate the performance of the financial management large language model over time: ; in is the current FRS score, is the initial benchmark score. This helps track the long-term improvement of the large language model for financial management. Finally, a dynamic update mechanism is established.

[0071] The three indicators of financial risk scoring and the long-term financial risk score trend are calculated. Users set different weights for each item and calculate the comprehensive score. The comprehensive score is used to evaluate the performance of the financial management large language model in a single scenario.

[0072] When new financial reimbursement rules are detected, the incremental learning process is automatically triggered to ensure that the model always reflects the latest financial reimbursement rules.

[0073] Through this innovative training method, we have built a large language model for financial management that can both understand natural language and strictly adhere to financial reimbursement rules. This model not only provides intelligent user interaction but also ensures high accuracy and reliability, effectively addressing the challenges faced by traditional large language models in the financial reimbursement field.

[0074] The financial knowledge question-answering module implements financial knowledge question-answering based on a large language model. The large language model uses a multi-stage intelligent question-answering and verification framework, combining reinforcement learning, human feedback, and a professional financial rules engine to ensure the accuracy and reliability of answers.

[0075] First, a large language model is trained using reinforcement learning and human feedback techniques. Define the model strategy ,in are model parameters, is the current state, including user questions and conversation history, is the model’s answer. Human feedback is modeled as a reward function Our goal is to maximize the expected reward, The specific formula is:

[0076] Update model parameters through policy gradient method:

[0077] in is the learning rate, is the model parameter Vector.

[0078] On this basis, a multi-round dialogue mechanism is introduced to enhance the model's ability to understand context. The attention mechanism is used to capture key information: ; in 、 、 are query, key, and value matrices respectively, is the transpose of the key matrix, is the dimension of the key.

[0079] To further improve accuracy, the financial management language model in the large language model module integrates a professional financial reimbursement rule engine. Define the financial reimbursement rule set , each rule There is an associated validation function , the return value is . Preliminary answers generated by the large language model for financial management Will be verified by the financial reimbursement rules engine: ; Only when The answer is considered valid only when , the system generates correction suggestions based on the violated financial reimbursement rules and requires the model to regenerate the answer.

[0080] The system also implements a dynamic learning loop and defines an update function ,in is the new training data, including manual intervention and correction. The update function is: ; This function updates the model parameters at the same time and rule sets .

[0081] Finally, through continuous performance monitoring, a comprehensive performance indicator is defined : ; in is the weight coefficient, which is adjusted according to specific business needs. The definitions of each indicator are as follows: 1. For accuracy evaluation: ; in is the number of correct answers, is the total number of questions, with correct answers determined based on human review or automated test sets. This measures the model's ability to provide accurate information.

[0082] 2. For compliance evaluation: ; in The number of violations of financial reimbursement rules, is the total number of answers. Financial reimbursement rule violations are determined by the financial reimbursement rules engine and may include exceeding reimbursement limits, citing outdated policies, etc. This measures the extent to which the model adheres to the latest financial reimbursement rules.

[0083] 3. For responsiveness evaluation: ; in, is the average time, in seconds, it takes for the model to generate an answer. For normalization, we take its inverse. This measures the system's ability to quickly respond to user queries.

[0084] The value range of each indicator is ,in Indicates the best performance. By adjusting the weight , which can emphasize different aspects of performance based on the needs of a specific enterprise.

[0085] In addition, define a dynamic optimization threshold : ; in It is the performance benchmark for the initial deployment of a large language model for financial management. is a coefficient greater than 1, and in this embodiment is set to 1.05, which is used to set the goal of continuous improvement. When the financial management language model optimization process is triggered, it includes retraining the financial management language model and updating the financial reimbursement rules.

[0086] The document upload and recognition module and the paper report generation module are both implemented based on the large language model module.

[0087] First, the user uploads documents, such as invoices and tickets, and the system uses OCR technology to scan and recognize text on the documents, and vectorizes the recognized text information.

[0088] Then, the large language model reviews and verifies the identified text content according to the financial reimbursement rules in the standard system database module to ensure that it complies with the requirements of the financial reimbursement rules and generates corresponding paper reports.

[0089] The large language model is responsible for intelligently formatting the audited data to ensure that the report format, entries, and data comply with financial reimbursement rules.

[0090] The final generated report can be printed directly for offline signature and review.

[0091] Example 2: The specific user usage scenarios are as follows.

[0092] This embodiment describes the complete process of a typical user using an enterprise financial reimbursement assistance system based on artificial intelligence technology.

[0093] Scenario: Zhang San is a sales manager who has just returned from a 3-day business trip and needs to be reimbursed for related expenses.

[0094] Step 1: Log in to the system Zhang San logs into the AI-powered enterprise financial reimbursement assistance system through the company's intranet. After the system recognizes his identity, it automatically loads the financial reimbursement rules relevant to his position and department.

[0095] Step 2: Submit a reimbursement request Zhang San clicks the "New Reimbursement" button, and the system pops up a smart dialog box: System: Hello, Manager Zhang San. What type of reimbursement is this? Zhang San: Business trip reimbursement.

[0096] System: I understand. What is the time and location of your business trip? Zhang San: From August 15th to 17th, I went to Shanghai on a business trip for 3 days.

[0097] System: Okay, based on your itinerary, you can claim reimbursement for transportation, accommodation, and daily meals. Have you prepared the relevant receipts? Step 3: Upload the ticket Zhang San: Yes, I'm ready.

[0098] System: Please upload your bill picture and I will automatically identify and classify it.

[0099] Zhang San uploaded photos of his flight tickets, hotel invoices, and several restaurant invoices. The system used OCR technology to quickly identify the invoice information and automatically filled in the corresponding reimbursement items.

[0100] Step 4: Smart Verification System: I have completed bill recognition. The total reimbursement amount for your business trip is RMB 3,800, including: - Round-trip air ticket: 2,000 yuan; -Hotel accommodation: 1,500 yuan, note: 500 yuan / night*3 nights; -Food and Beverage Fee: 300 yuan, Note: 100 yuan / day*3 days.

[0101] According to the company's latest travel reimbursement policy, these expenses are within the allowable range. Are there any other items that need to be reimbursed? Zhang San: That’s all.

[0102] Step 5: Generate Report System: OK, I've generated a detailed expense report based on the information and receipts you provided. You can review the report summary on the screen. If the information is correct, please click the "Submit" button.

[0103] Zhang San checks the report summary, confirms that it is correct, and clicks "Submit".

[0104] Step 6: Automatic Approval System: Your reimbursement request has been submitted successfully. According to company regulations, reimbursements under 5,000 RMB are automatically approved. Your request has passed preliminary review and will be sent directly to the Finance Department for processing. We expect your reimbursement to be processed within three business days.

[0105] Step 7: Status Tracking Zhang San: Thank you. How can I know the progress of the reimbursement? System: You can check your reimbursement status at any time in the system. I'll also send you updates via WeChat for Business. If the Finance Department requires additional documentation, I'll notify you immediately.

[0106] Zhang San: That’s great. Very convenient.

[0107] System: Do you have any other questions? Zhang San: No more, thank you.

[0108] System: OK, I wish you a pleasant work. If you have any questions later, feel free to ask me.

[0109] Throughout the entire process, the system leverages the financial management language model built into the large language model module to understand user intent, automatically processes invoices using optical character recognition (OCR), and intelligently verifies invoices based on the company's specific financial reimbursement rules. This significantly simplifies the reimbursement process, improves efficiency, and ensures accuracy and compliance.

[0110] Based on user input and document information, the online form filling guidance module is implemented using the financial management large language model built into the large language model module. First, the user uploads a document, such as an invoice or ticket, into the system and begins filling out the online reimbursement form. The system uses optical character recognition (OCR) technology to recognize the uploaded document information and convert it into structured data. This recognized data is then fed into the financial management large language model built into the large language model module. Next, the financial management large language model parses the user input and document information against the company's standard system database module, identifying key content and potential errors. As the user completes the online reimbursement form, the system analyzes the user's input in real time and provides dynamic form filling guidance based on the financial reimbursement rules in the standard system database. The financial management large language model generates personalized suggestions and error messages to ensure that the user follows the correct steps and standards when filling out the reimbursement form. For example, when the user enters the invoice amount, the financial management large language model checks whether it falls within the specified range and prompts for additional relevant information, such as the invoice number and date. Throughout the entire process, the system continuously updates and adjusts the guidance content based on the user's current input and the document information recognized by OCR, helping users avoid common errors and omissions. Ultimately, the system generates a complete reimbursement application form that complies with corporate financial reimbursement regulations, ensuring that users can submit their reimbursement applications smoothly.

[0111] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent assistance system for enterprise financial reimbursement based on artificial intelligence technology, characterized by: Includes: The user operation interface module provides a user interaction interface; the user interaction interface includes a first user interface for the user to input documents and basic user information, and a second user interface for providing answer information to the user, identifying basic user information, and providing prompt messages to the user; The administrator operation interface module provides administrators with functions including authority control, standard system database update, historical reimbursement data summary and financial risk warning; The standard system database module stores the company's financial reimbursement rules, automatically updates and pushes updated financial reimbursement rules; The large language model module has a built-in financial management large language model. Through multimodal optimization of the financial management large language model, it provides support for reimbursement process management and user interaction. The financial knowledge question-and-answer module, based on a large language model, provides interpretation and question-and-answer functions; The document upload and recognition module uploads user-entered documents including invoices and tickets, automatically recognizes the document content through OCR technology, and fills the document content into the corresponding reimbursement application form; Paper report generation module, which automatically generates paper reports after users upload information; The online form filling guidance module guides users to correctly fill in reimbursement information in the financial system.

2. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 1 is characterized in that: The construction method of the user operation interface module and the administrator operation interface module is: Design the user interface of the user operation interface module based on the user interaction interface and the functions of providing answer information to users, identifying basic user information and providing prompt messages to users; Design the system interface of the administrator operation interface module based on functions including authority control, standard system database update, historical reimbursement data aggregation and financial risk warning; Based on the front-end framework and back-end technology stack, develop front-end pages and back-end services respectively to implement permission management, document upload, standard system update, data aggregation and analysis functions; Conduct functional, performance, and security testing, collect feedback from users and administrators, and optimize system functionality and user experience.

3. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 1 is characterized in that: The construction method of the standard system database module is as follows: Collect internal and external financial reimbursement rules of the enterprise; Digitize the rules. If they are paper documents, use OCR technology to convert them into editable electronic text; Clean and pre-process the text data obtained after digitization to remove noise and redundant information and unify the format; Classify and label the cleaned and pre-processed data, and store them in a structured manner according to categories including different reimbursement items and policy terms; Select a database management system corresponding to a relational database or a non-relational database, and design the database table structure and index; Import the classified and labeled data into the database, create interfaces and tools for interacting with the database, and complete the construction of the standard system database module.

4. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 1 is characterized in that: The financial management language model built into the large language model module is trained based on the data in the standard system database module. The specific method is as follows: Extract financial reimbursement rules from the standard system database module; Preprocess financial reimbursement rules and construct financial professional word vectors ,in Indicates the Vector representation of financial terms; Design special financial task reminder templates , used to generate financial reimbursement related question-answer key-value pair data; Generate a large-scale training dataset based on question-answer key-value pairs ,in It's a problem. is the corresponding standard answer; Introduce the financial reimbursement rule mechanism into the large language model module and define a financial reimbursement rule function , the financial reimbursement rule function accepts input And returns the financial rule vector related to it. The specific formula is: ； in It is the encoding output when the financial management language model does not introduce financial reimbursement rules. λ is the equilibrium parameter; The training process uses a dynamic batch size strategy to automatically adjust the number of samples in each batch based on the complexity of the financial reimbursement rules, and controls the output of the financial management large language model through a loss function. When new financial reimbursement rules are detected, the incremental learning process is automatically triggered.

5. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 4 is characterized in that: The loss function in the training process of the financial management large language model is: ； in, The cross entropy loss ensures that the large language model for financial management can generate the correct answer; is a comparative loss used to distinguish correct from incorrect financial information. is the consistency loss, and is the weight coefficient; The calculation of is as follows: ； in It's a problem and answers The similarity score of is the correct answer, is the wrong answer; The calculation of is as follows: ； After each training batch, the model performance is evaluated using any financial scenario test set.

6. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 5 is characterized in that: The specific method for evaluating the performance of the financial management large language model using any financial scenario test set is: Introducing financial risk scoring indicators, comprehensively considering the performance of financial management large language models in terms of accuracy, compliance, and robustness, and introducing long-term financial risk scoring trends; Financial risk scoring indicators The formula is: ； in is the weight coefficient; Accuracy evaluation The specific calculation formula is: ； in, For a real example, is a true negative example, For a false positive example, For false negative examples, the accuracy of model training is calculated based on the prediction results of the financial management large language model on the test set; Compliance Evaluation The specific calculation formula is: ； in is the number of rule conflicts, is the total number of test cases. This indicator measures the degree to which the financial management language model complies with the financial reimbursement rules; Robustness evaluation The specific calculation formula is: ； in is the number of errors in the perturbation input, is the total number of perturbation inputs. During the robustness test, the test set is perturbed to evaluate the performance of the financial management large language model on the perturbation inputs. Long-term financial risk score trends The calculation formula is: ； in is the current FRS score, is the initial baseline score; The three indicators of financial risk scoring and the long-term financial risk score trend are calculated. Users set different weights for each item and calculate the comprehensive score. The comprehensive score results are used to evaluate the performance of the financial management large language model in a single scenario.

7. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 1 is characterized in that: The method for constructing the financial knowledge question-answering module based on the financial management large language model built into the large language model module is as follows: Use reinforcement learning and human feedback techniques to train large language models and define model policies ,in are model parameters, is the current state, is the model’s answer; Modeling human feedback as a reward function , to maximize the expected reward, maximize the expected reward The specific formula is: ； Update model parameters through policy gradient method: ； in is the learning rate, is the model parameter vector; A multi-round dialogue mechanism is introduced to enhance the model's ability to understand context and capture key information through the attention mechanism. The specific formula is: ； in 、 、 are query, key, and value matrices respectively, is the transpose of the key matrix, is the dimension of the key.

8. The enterprise financial reimbursement intelligent assistance system based on artificial intelligence technology according to claim 1 is characterized in that: The financial knowledge question-answering module based on the financial management large language model in the large language model module also includes a financial reimbursement rule engine, specifically: Define a set of financial reimbursement rules , each financial reimbursement rule There is an associated validation function , the return value is , the initial answer generated by the large language model It will be verified by the financial reimbursement rules engine. The specific formula is: ； when When , the answer is considered valid; when When the answer is judged as invalid, the system will generate correction suggestions based on the violated financial reimbursement rules and require the model to regenerate the answer; A dynamic learning loop is also implemented, defining an update function ,in is the new training data, including manual intervention and correction. The specific formula of the update function is: ； This function updates the model parameters at the same time and rule sets ; Finally, through continuous performance monitoring, define a comprehensive performance indicator P : ； in is the weight coefficient, which is adjusted according to specific business needs. The definitions of each indicator are as follows: Accuracy evaluation : ； in is the number of correct answers, is the total number of questions, and the correct answers are determined based on human review or automated test sets, which measures the ability of the model to provide accurate information; Compliance Evaluation : ； in The number of violations of financial reimbursement rules, is the total number of answers, and the financial reimbursement rule violations are determined by the financial reimbursement rule engine, which measures the extent to which the model complies with the latest financial reimbursement rules; Responsiveness evaluation : ； in, is the average time it takes for the large language model for financial management to generate an answer, measured in seconds. For normalization, its reciprocal is taken, which measures the system's ability to quickly respond to user queries. The value range of each indicator is ,in Indicates the best performance, by adjusting the weight 、 、 , can emphasize different aspects of performance according to needs; In addition, define a dynamic optimization threshold : ； in It is the performance benchmark for the initial deployment of a large language model for financial management. It is a coefficient greater than 1 and is used to set goals for continuous improvement; when When the financial management language model optimization process is triggered, it includes retraining the financial management language model and updating the financial reimbursement rules.

9. An intelligent auxiliary method for enterprise financial reimbursement based on artificial intelligence technology, characterized in that: Includes: The administrator selects data through the administrator operation interface module to build the standard system database module; The financial management language model built into the language model module is trained using the data stored in the standard system database module; The financial management large language model built into the large language model module provides technical support for the financial knowledge question and answer module, document upload and recognition module, paper report generation module, and online form filling guidance module; Users interact through the user operation interface module, the financial knowledge question and answer module, the document upload and recognition module, the paper report generation module and the online form filling guidance module to complete the reimbursement process.

Images