Accounting business processing method and system based on multi-agent

Abstract

The invention discloses a multi-agent-based accounting business processing method and system. The multi-agent-based accounting business processing method and system comprises the following steps: acquiring a correct invoice number, amount, category, responsible department, principal and the like through the technologies of image acquisition, denoising, enhancement, tilt correction, feature extraction, character correction and the like of an initial voucher; automatically inputting the identified related information into an account book according to the classification of each accounting subject, and generating a total account book and a detail account book; regularly adopting a computer-aided auditing technology, randomly generating a random number according to an importance (such as involved amount or business type risk level) proportion, automatically extracting an account for auditing, and realizing unified calling and checking of an original voucher, a general account book and a detail account book; through integration of deep learning, knowledge graph and other technologies, more complex prediction is realized by using big data, a potential financial trend is identified, detailed risk assessment is carried out, and decision support is provided. The query and scheduling service of the whole business process is realized by using a large language model and adopting a voice interaction mode, unit financial data is preprocessed, and an RAG containing financial statements and related financial knowledge is constructed with related knowledge such as related legal policies, tax policies, financial analysis and bank settlement and is used for retrieval and query, and the financial statement and the related financial knowledge are stored in a database. All modules and execution processes of financial software are fused, and intelligent query and execution based on a large model are achieved.

Description

Technical Field

[0001] The present application relates to the field of intelligent accounting business processing, and specifically to a multi-agent-based accounting business processing method and system. Background Art

[0002] Traditional accounting processes are complex, requiring manual or system-based entry of transaction vouchers, which are then entered into ledgers, including the general ledger and various subsidiary ledgers. Financial statements, such as the balance sheet and income statement, are compiled from these ledgers at quarterly and annual ends. Furthermore, tax payables must be calculated and designated personnel assigned to file tax returns. Accounts must also be regularly reconciled to ensure data accuracy in preparation for internal and external audits of a company's financial statements and operating performance. These processes have numerous drawbacks, including time-consuming and error-prone manual data entry and processing, which can be prone to errors when dealing with large amounts of data. Information silos between different departments or systems are often lacking, hindering decision-making efficiency. Manual accounts are difficult to track, resulting in complex and time-consuming audit processes. Financial reporting and data analysis are challenging, hindering timely reflection of a company's financial status and limiting the ability to identify and address potential financial risks. Furthermore, traditional accounting systems often lack real-time data updates and analysis, resulting in delayed financial decisions and hindering a company's responsiveness. Finally, traditional accounting systems lack data analysis and forecasting capabilities, preventing them from providing deep financial insights. Therefore, a technical solution is desired to address or at least mitigate these shortcomings of existing technologies.

[0003] Large language models (LLMs) are natural language processing models with a large number of parameters trained using deep learning techniques. These models are typically capable of understanding and generating natural language text and performing various language-related tasks, such as text translation, text summarization, sentiment analysis, question-answering systems, and text generation. Large language models have extensive applications in natural language processing (NLP) and are a hot topic in current artificial intelligence research and application. Their development and application are continuously pushing the boundaries of human-computer interaction and automated processing of language information. Multi-agent technology, commonly referred to as multi-agent systems (MAS), is a computing system involving multiple intelligent agents that interact, communicate, collaborate, or compete in a shared environment to complete tasks or solve problems. Each agent is autonomous, capable of perceiving the environment, making decisions, and executing actions. Multi-agent systems can solve complex problems that are difficult for single agents or single-layer systems to handle, and they exhibit characteristics such as autonomy, distribution, complexity, self-organization, and real-time performance.

[0004] In order to overcome the shortcomings of traditional accounting business processes and turn to automated, digital, and intelligent accounting business processing solutions, this patent proposes an accounting business processing method and system based on multi-agents. The system can reduce manual operations through automated and intelligent processes, improve work efficiency, and reduce error rates. At the same time, it can update and process financial data in real time, integrate data from different sources, eliminate information silos, and use data mining and artificial intelligence algorithms to identify complex patterns and trends, thereby providing accurate financial forecasts and analysis, providing enterprises with in-depth financial insights and forecasting capabilities, and supporting better decision-making. Finally, by using large models and intelligent agents, it can process and understand text information such as financial reports and compliance documents, and provide richer analysis and task execution. Summary of the Invention

[0005] The purpose of the present invention is to provide an accounting business processing method and system based on multi-agent to at least solve the above problems.

[0006] One aspect of the present invention is a multi-agent-based accounting business system, characterized by comprising:

[0007] Acquisition module, collects and obtains information;

[0008] Capture, denoise, enhance, tilt correct, extract features, recognize text, and repair the image of the initial voucher to obtain the correct invoice number, amount, category, responsible department, and person in charge;

[0009] Identification module, identification and entry into account books;

[0010] Use OCR technology to recognize text, numbers, and symbols in electronic images, categorize the recognized data by income, output, bank receipts, and expenses, and automatically enter them into the ledger to generate a general ledger and detailed ledgers.

[0011] Computational module, risk assessment and decision support;

[0012] Computer-assisted auditing technology is used to enable unified verification of original documents, general ledgers, and detailed ledgers. By integrating deep learning, knowledge graphs, and other technologies, big data is used to achieve more complex forecasts, identify potential financial trends, conduct detailed risk assessments, and provide decision support.

[0013] Generate modules, query and dispatch services;

[0014] Generate preset financial data reports from the financial system's dedicated financial data, and comprehensively utilize large language models to implement query and scheduling services for the entire business process through voice interaction.

[0015] Optionally, the text recognition and text correction technology includes: moving to a specific embodiment

[0016] Text correction is based on the results of fuzzy text prediction using the BERT model. Text correction is performed based on contextual semantic information. Each word s in the initially predicted text T is masked in turn, and the most appropriate word for s's position is predicted based on the context of s. A tolerance threshold k is set. If the original word s appears in the top three prediction results, it is considered that the position is not a typo. Otherwise, it is a typo and is corrected according to the maximum probability.

[0017] One aspect of the present invention provides a multi-agent-based accounting business processing method, characterized in that it is based on the accounting business system according to claim 1;

[0018] When collecting and obtaining information:

[0019] For unclear text recognition, Densenet121 is used as the encoder and Transformer is used as the decoder to predict the unclear text;

[0020] For text correction, the BERT model is used based on the results of fuzzy text prediction, and text correction is performed according to contextual semantic information.

[0021] One aspect of the present invention provides a multi-agent-based accounting business processing method, characterized in that the risk assessment and decision support includes:

[0022] Random numbers are randomly generated according to the amount and risk level, and accounts are automatically selected for audit. Sampling is done from high risk, medium risk, and low risk in a ratio of 6:4:2. The randomness of the sampling is determined by the linear congruential method. i+1 =(ar i +C)(mod M);

[0023] Among them, r i is the seed, a, c, M are magic numbers to ensure random sampling;

[0024] If any problems arise during the audit, the account problem, original vouchers, general ledger, detailed ledger and responsible persons will be fed back to the relevant department leaders through the system feedback mechanism so that appropriate measures can be taken.

[0025] One aspect of the present invention provides a multi-agent-based accounting business processing method, characterized in that the risk assessment and decision support further includes:

[0026] Utilize the long short-term memory (LSTM) network to split the training and test data, perform data normalization, train on historical data, capture the time dependency in the data, conduct risk forecasting and investment analysis, integrate the relevant experience and knowledge of managers and experts, use reinforcement learning to learn from experience, continuously optimize its decision-making strategy, gradually approach the optimal strategy, and provide timely business feedback and decision support.

[0027] One aspect of the present invention provides a multi-agent-based accounting business processing method, wherein the query and scheduling service includes:

[0028] Import the processed financial data into the database and vectorize it using the dual encoder model;

[0029] Regularly collect and process relevant legal policies, tax policies, financial knowledge, bank settlement and other related knowledge, and organize this complex information into a timely updated structured knowledge graph through entity, relationship and attribute extraction technology;

[0030] Build a Graph-RAG containing financial statements and related financial knowledge for retrieval and query, integrating various modules and execution processes of financial software to implement intelligent query and execution based on large models;

[0031] By integrating the language interaction module and realizing a large interactive intelligent body model, it can conduct continuous voice dialogues and multiple rounds of questions and answers on relevant knowledge. The information database is updated regularly to provide comprehensive and accurate real-time query services. According to different query requirements, it can generate customized information summaries and provide personalized information services.

[0032] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0033] 1. Multi-agent-based accounting business processing methods and systems can, on the one hand, automate repetitive accounting tasks, reducing manual operations and allowing financial personnel to focus on more complex and value-added tasks. This improves work efficiency and data accuracy, reduces reliance on manpower, and lowers the company's operating costs. Furthermore, by introducing AI-powered analytical technologies, using big data analytics and machine learning algorithms, they can deeply mine and analyze massive amounts of financial data, providing more accurate and comprehensive data support for a company's strategic decision-making. This data-driven decision-making support helps companies respond quickly to fierce market competition, seize opportunities, and mitigate risks.

[0034] 2. The accounting business processing method and system based on multi-intelligence adopts the multi-intelligence approach of "big model + control + execution" to realize query and scheduling services for the entire business process. Staff can automatically perform accounting tasks such as invoice processing, account reconciliation, and report analysis through voice interaction, greatly improving work efficiency, reducing the time and potential errors of manual operations, and enhancing the human-computer interaction experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of a multi-agent-based accounting business processing method and system of the present application. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the implementation of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below in conjunction with the drawings in the embodiments of this application. In the drawings, the embodiments described are part of the embodiments of this application, rather than all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be understood as limitations on this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The embodiments of this application are described in detail below in conjunction with the drawings.

[0037] like Figure 1 The illustrated multi-agent-based accounting business processing method and system includes the following:

[0038] First, a scanner captures an image of the invoice. Preprocessing the image involves grayscaling, denoising, enhancing, and tilt correction to improve image quality. Feature extraction then occurs to extract and analyze text features within the invoice image. Finally, the recognized text is matched against a pre-set template to extract key invoice information, such as the invoice number, date, and amount. This information is then automatically entered into the ledger based on classification, generating a general ledger and detailed ledgers.

[0039] For unclear text recognition, Densenet121 is used as the encoder and Transformer is used as the decoder to predict the unclear text;

[0040] Densenet121 is used as the encoder, and the obtained feature representation is used as the input of the decoder. First, a large-scale convolution is performed, followed by a pooling layer, followed by several consecutive sub-modules (Dense Block and TransitinLayer), and finally a pooling and full connection;

[0041] The Dense Block contains 6 [1*1conv, 3*3conv] Bottleneck structures, and the TransitinLayer is a convolution plus pooling layer used to integrate the learned features and reduce the size of the feature map. The batch size is 64 and the activation function is Swish.

[0042] The decoder uses a Transformer architecture, consisting of self-attention layers, encoder-decoder attention layers, and feedforward neural network layers. These components are connected via residual connections and layer normalization. The encoder-decoder attention layer uses two-dimensional attention to calculate the correlation between each pixel and all other pixels. Adaptive normalization (SN) is used for normalization, combining Batch Normalization (BN), Layer Normalization (LN), Instance Normalization (IN), and Group Normalization (GN). The weights are initialized, and the network adaptively selects the normalization method through learning.

[0043] Text correction uses the BERT model based on the results of ambiguous text prediction. This model captures contextual information in the text through bidirectional encoding and self-attention mechanisms to achieve text correction. The specific steps are as follows: First, pre-training is performed on a large amount of text data, primarily through two tasks: masked language modeling and next sentence prediction. Each character s in the initially predicted text T is masked in turn. The model relies on the context of s to predict the most appropriate character in its position. A tolerance threshold k is set. If the original character s appears in the top three predictions, it is considered not a typo at that position. Otherwise, it is a typo and is corrected according to the highest probability.

[0044] Computer-assisted auditing technology is used to uniformly retrieve and verify original vouchers, general ledgers, and detailed ledgers. First, random numbers are generated based on the amount and risk level, automatically selecting accounts for audit. Sampling is done in a 6:4:2 ratio from high-risk, medium-risk, and low-risk categories. The randomness of the sampling is guaranteed using the linear congruential method ri+1=(ari+C)(mod M), where ri is the seed and a, c, and M are magic numbers. Once the audit sample is determined, the intelligent system will uniformly retrieve the original vouchers, general ledger, and detailed ledgers and integrate them onto a single platform for comparison to verify authenticity and compliance. Any issues identified will be promptly notified to the relevant department heads.

[0045] The LSTM (Long Short-Term Memory) model is a specialized recurrent neural network (RNN) that first collects and processes historical financial data, including balance sheets, income statements, and cash flow statements. This data needs to be cleaned and standardized to facilitate model training. Feature data, such as financial ratios, cash flow, and debt levels, is also selected as training data. Using the LSTM network, the training and test data are split and normalized. Training on historical data captures temporal dependencies within the data, enabling risk forecasting and investment analysis. By integrating the relevant experience and knowledge of managers and experts, and leveraging reinforcement learning to learn from experience, the model continuously optimizes its decision-making strategies, gradually approaching the optimal strategy. Financial reports and management analysis are rapidly generated, providing timely business feedback and decision support.

[0046] First, the processed financial data is imported into a database and vectorized using a dual encoder model. Next, relevant knowledge, including legal policies, tax policies, financial knowledge, and bank settlements, is collected and processed. This unstructured text is combined with the collected financial data and other complex information using entity, relationship, and attribute extraction techniques, with each entity as a node and each relationship as an edge. This knowledge graph is reorganized into a knowledge graph, which is stored in Neo4j. A Graph-RAG containing financial statements and related financial knowledge is constructed, and an automatic Graph-RAG update mechanism is set up to maintain the timeliness and accuracy of the information as financial data and expertise are updated.

[0047] The Llama 3.1 large model is used as the model, integrated with the constructed Graph-RAG to enhance the large model generation effect. The language interaction module is integrated to realize the interactive intelligent large model. It can conduct voice dialogue and conduct multiple rounds of questions and answers on professional legal policies, tax policies, financial knowledge, bank settlement and other related knowledge, providing comprehensive and accurate real-time query services. According to different query requirements, it can generate customized information summaries and provide personalized information services. By integrating various modules and execution processes of financial software, each module is an intelligent agent capable of independent operation and analysis, forming a seamless connection between the large model and business modules, and realizing intelligent query and execution of the large model.

[0048] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the present invention. Therefore, such modifications and improvements, which do not depart from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. An accounting business system based on multi-agent, characterized by: include: Acquisition module, collects and obtains information; Capture, denoise, enhance, tilt correct, extract features, recognize text, and repair the image of the initial voucher to obtain the correct invoice number, amount, category, responsible department, and person in charge; Identification module, identification and entry into account books; Use OCR technology to recognize text, numbers, and symbols in electronic images, categorize the recognized data by income, output, bank receipts, and expenses, and automatically enter them into the ledger to generate a general ledger and detailed ledgers. Computational module, risk assessment and decision support; Computer-assisted auditing technology is used to enable unified verification of original documents, general ledgers, and detailed ledgers. By integrating deep learning, knowledge graphs, and other technologies, big data is used to achieve more complex forecasts, identify potential financial trends, conduct detailed risk assessments, and provide decision support. Generate modules, query and dispatch services; Generate preset financial data reports from the financial system's dedicated financial data, and comprehensively utilize large language models to implement query and scheduling services for the entire business process through voice interaction.

2. A multi-agent-based accounting business processing method, characterized in that: Based on the accounting business system according to claim 1; When collecting and obtaining information: For unclear text recognition, Densenet121 is used as the encoder and Transformer is used as the decoder to predict the unclear text; For text correction, the BERT model is used based on the results of fuzzy text prediction, and text correction is performed according to contextual semantic information.

3. The multi-agent-based accounting business processing method according to claim 2, characterized in that: The risk assessment and decision support include: Random numbers are randomly generated according to the amount and risk level, and accounts are automatically selected for audit. Sampling is done from high risk, medium risk, and low risk in a ratio of 6:4:

2. The randomness of the sampling is determined by the linear congruential method. i+1 =(ar i +C)(mod M); Among them, r i is the seed, a, c, M are magic numbers to ensure random sampling; If any problems arise during the audit, the account problem, original vouchers, general ledger, detailed ledger and responsible persons will be fed back to the relevant department leaders through the system feedback mechanism so that appropriate measures can be taken.

4. The multi-agent-based accounting business processing method according to claim 3, characterized in that: The risk assessment and decision support also includes: Utilize the long short-term memory (LSTM) network to split the training and test data, perform data normalization, train on historical data, capture the time dependency in the data, conduct risk forecasting and investment analysis, integrate the relevant experience and knowledge of managers and experts, use reinforcement learning to learn from experience, continuously optimize its decision-making strategy, gradually approach the optimal strategy, and provide timely business feedback and decision support.

5. The multi-agent-based accounting business processing method according to claim 2, characterized in that: The query and scheduling services include: Import the processed financial data into the database and perform vectorization using the dual encoder model; Regularly collect and process relevant legal policies, tax policies, financial knowledge, bank settlement and other related knowledge, and organize this complex information into a timely updated structured knowledge graph through entity, relationship and attribute extraction technology; Build a Graph-RAG containing financial statements and related financial knowledge for retrieval and query, integrating various modules and execution processes of financial software to implement intelligent query and execution based on large models; By integrating the language interaction module and realizing a large interactive intelligent body model, it can conduct continuous voice dialogues and multiple rounds of questions and answers on relevant knowledge. The information database is updated regularly to provide comprehensive and accurate real-time query services. According to different query requirements, it can generate customized information summaries and provide personalized information services.

Images