Method and device for auditing based on large language model

Abstract

The invention discloses a method and a device for auditing based on a large language model. The method comprises the steps that audit data are processed, an audit knowledge base is constructed, and the audit knowledge base comprises common concepts and rules in audit tasks; combining the auditing knowledge base with a large language model by utilizing a LangChain framework, and performing automatic retrieval and analysis on auditing materials by utilizing the large language model based on the auditing knowledge base to generate a related auditing report and a data matching result; and performing articulation detection on the financial statement in the audit report and the data matching result by using the large language model, so that the audit report and the data matching result are accurate in logic and finance. The technical problems of high labor cost and long consumed time of auditing work in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and device for auditing based on a large language model. Background Art

[0002] With the digital transformation of the global economy, the audit industry is undergoing profound changes. The rapid development of digital technology, combined with changes in market demand and regulatory environment, has driven new models and processes in the industry. One of the biggest challenges facing auditors is managing and ensuring the quality and security of data. With the surge in the amount of corporate data, auditors must effectively process and analyze large amounts of data while ensuring data security and compliance. As a result, the timeliness and accuracy of audit reports are affected.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present invention provide a method and device for auditing based on a large language model, so as to at least solve the technical problems of high labor cost and long time consumption of auditing work in the prior art.

[0005] According to one aspect of an embodiment of the present invention, a method for auditing based on a large language model is provided, comprising: constructing an audit knowledge base by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks; combining the audit knowledge base with a large language model using a LangChain framework, and based on the audit knowledge base, using the large language model to automatically retrieve and analyze audit materials to generate relevant audit reports and data matching results; using the large language model to perform cross-reference detection on the financial statements in the audit report and the data matching results, so that the audit report and the data matching results are logically and financially accurate.

[0006] According to another aspect of an embodiment of the present invention, there is also provided an apparatus for auditing based on a large language model, comprising: a construction module, configured to construct an audit knowledge base by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks; a generation module, configured to combine the audit knowledge base and the large language model using the LangChain framework, and based on the audit knowledge base, use the large language model to automatically retrieve and analyze audit materials to generate relevant audit reports and data matching results; a detection module, configured to use the large language model to perform cross-reference detection on the financial statements in the audit report and the data matching results, so that the audit report and the data matching results are logically and financially accurate.

[0007] In an embodiment of the present invention, an audit knowledge base is constructed by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks; the audit knowledge base and the large language model are combined using the LangChain framework, and based on the audit knowledge base, the large language model is used to automatically retrieve and analyze audit materials to generate relevant audit reports and data matching results; the large language model is used to perform cross-reference relationship detection on the financial statements in the audit report and data matching results, so that the audit report and data matching results are logically and financially accurate. Through the above scheme, the technical problems of high labor cost and long time consumption of audit work in the prior art are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0009] Figure 1 is a flow chart of a method for auditing based on a large language model according to an embodiment of the present invention;

[0010] Figure 2 A complex cross-check relationship diagram between various parts of a financial statement according to an embodiment of the present invention;

[0011] Figure 3 is a flowchart of another method for auditing based on a large language model according to an embodiment of the present invention;

[0012] Figure 4 is a schematic diagram of a LangChain-based agent according to an embodiment of the present invention;

[0013] Figure 5 is a diagram of the working principle of an intelligent agent according to an embodiment of the present invention;

[0014] Figure 6 is an example diagram of the arrangement of a cross-reference relationship detection agent according to an embodiment of the present invention;

[0015] Figure 7 is a flow chart of an agent performing cross-reference relationship detection according to an embodiment of the present invention;

[0016] Figure 8 is a schematic diagram of a cross-reference relationship detection system according to an embodiment of the present invention;

[0017] Fig. 9 is a schematic diagram of the structure of an apparatus for auditing based on a large language model according to an embodiment of the present invention; DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. 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 creative work should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] According to an embodiment of the present invention, a method embodiment for auditing based on a large language model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0021] Figure 1 is a method for auditing based on a large language model according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0022] Step S102, constructing an audit knowledge base by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks.

[0023] For example, unstructured financial statements are obtained as the audit data, and the unstructured financial statements are processed and converted into text data; the text content in the text data is divided into multiple independent text blocks, and the multiple independent text blocks are vectorized to establish a vector library as an audit knowledge base.

[0024] Step S104, using the LangChain framework to combine the audit knowledge base and the large language model, based on the audit knowledge base, using the large language model to automatically retrieve and analyze the audit materials, and generate relevant audit reports and data matching results.

[0025] Based on the user's instruction, a query vector is generated, and a search is performed in the vector library based on the query vector to locate the most relevant text as the relevant audit report and data matching result. The search may include: extracting balance sheet data in the vector library based on the query vector; extracting income statement data in the vector library based on the query vector; and extracting cash flow statement data in the vector library based on the query vector.

[0026] Step S106, using the large language model to perform cross-reference detection on the audit report and the financial statements in the data matching result, so that the audit report and the data matching result are logically and financially accurate.

[0027] Based on the preset audit requirements, the relevant parameters of the cross-check detection are set, wherein the relevant parameters of the cross-check detection include cross-check rules and concerned data paragraphs; based on the relevant parameters of the cross-check detection, the large language model is used to perform cross-check relationship detection on the audit report and the financial statements in the data matching results. For example, the structure and content of the financial statements in the audit report and the data matching results are deeply analyzed using the large language model to obtain analysis results; using chain thinking reasoning technology, the large language model is used to deepen the data relationship interpretation of the financial statements in the audit report and the data matching results to obtain interpretation results; based on the analysis results and the interpretation results, a cross-check relationship detection is performed, and a correction plan is proposed or the data is directly adjusted.

[0028] Among them, the cross-checking relationship detection includes at least one of the following: checking the cross-checking relationship within the table based on the balance sheet data, the income statement data, and the cash flow statement data; checking the cross-checking relationship between statements based on the balance sheet data, the income statement data, and the cash flow statement data; checking the cross-checking relationship between the current period and the previous period statements based on the balance sheet data, the income statement data, and the cash flow statement data.

[0029] In the audit work of accounting firms, the accuracy of financial statements and the effectiveness of cross-checking relationships are crucial. Traditional manual cross-checking verification methods are inefficient and prone to errors. Artificial intelligence technology, especially the application of LLMs, is driving the audit industry towards intelligent development. In this application, LLMs are applied to the intelligent detection of cross-checking relationships in financial statements, aiming to improve the efficiency and accuracy of audit work. Combined with the audit knowledge base and RAG technology, LLMs can automatically identify and verify cross-checking relationships in financial statements, greatly improving the quality and efficiency of audits. The audit knowledge base provides background information and data support, and RAG technology enhances the analytical capabilities of LLMs. This application demonstrates how to use these technologies to achieve intelligent cross-checking relationship detection, bringing revolutionary improvements to audit work and ushering in a new era of intelligent auditing.

[0030] Financial statements are the core of corporate decision-making and market trust, and they reflect the economic activities and financial health of the company in detail. The balance sheet, income statement and cash flow statement are independent and cross-checked, and together they build a complete financial information network. For example, the net profit of the income statement is linked to the change in owners' equity of the balance sheet, and the increase or decrease in cash of the cash flow statement must correspond to the cash balance of the balance sheet. This internal consistency is not only a reflection of the company's historical and future financial performance, but also serves the company's internal control and strategic planning, and is crucial to supporting the decision-making of investors and market participants.

[0031] In the context of globalization, financial statements follow international accounting standards, ensuring that global investors and regulators use this common language for communication and supervision. Accurate financial statements reflect the financial standardization and transparency of an enterprise and are a symbol of corporate professional ethics. The preparation of statements and cross-reference testing are the basis of audit work, ensuring the authenticity and fairness of financial data, and are the key to winning market trust and maintaining fair trading order.

[0032] Cross-checking ensures the accuracy and completeness of financial statements in expressing the financial status of an enterprise, and is a guarantee of the reliability of financial information. The balance sheet shows the financial status of an enterprise at a specific moment. Its cross-checking relationship, such as the total assets being equal to the total liabilities plus owner's equity, directly reflects financial health. The income statement and cash flow statement dynamically describe the business activities of an enterprise from the two dimensions of profitability and cash liquidity, forming a strict logical chain to ensure the consistency and continuity of data between statements. The cross-checking relationship under specific conditions is more detailed, revealing the truth of the economic activities behind the statement data. For example, the net increase in cash and cash equivalents in the cash flow statement is equal to the change in monetary funds in the balance sheet, which is only true when there is no change in cash equivalents. Such cross-checking analysis helps to discover potential operating problems or audit risks.

[0033] The consistency of financial statements requires that the key values ​​of the current report be consistent with the data of the previous report to prevent the manipulation of financial data. The cross-check relationship between the main table and the appendix further refines the verification of data and ensures the progressive and accurate correspondence of the report data. Figure 2 It shows the complex cross-checking relationship between the various parts of the financial statements, like a sophisticated network that closely connects different financial data. With the application of financial data automation tools and intelligent technology, this network will become more compact and efficient, providing stronger support for the company's financial management and external investment decisions.

[0034] In the traditional practice of financial auditing and report preparation, manual cross-reference verification is a critical process that ensures that the data in the financial statements is not only accurate but also logically consistent. By manually comparing and verifying the data in the report, auditors can confirm the accuracy of the data and discover and correct potential errors or inconsistencies. Although this process plays a key role in ensuring the reliability of the report, it also faces a series of challenges, especially in terms of efficiency and accuracy when dealing with large amounts of complex data.

[0035] Manual cross-checking relationship verification requires auditors to have a deep understanding of the composition and cross-checking logic of financial statements and to accurately perform data comparison and analysis. This process usually involves the following steps:

[0036] (1) Data collation and preparation: Collect financial statements and related accounting documents and organize them into a format that can be analyzed.

[0037] (2) Item-by-item comparison: Based on the cross-checking relationship between financial statements, check the consistency of data one by one, such as whether the net profit in the income statement matches the change in retained earnings in the balance sheet.

[0038] (3) Analysis and Adjustment: Conduct in-depth analysis of any data inconsistencies or anomalies found, determine their causes, and make appropriate adjustment recommendations.

[0039] Although manual cross-reference verification has its place in traditional audits, it faces increasing challenges as enterprises grow in size and transactions become more complex:

[0040] (1) Inefficiency: Manually processing large amounts of financial data is time-consuming and inefficient, and it is difficult to meet rapidly changing business needs and reporting cycles.

[0041] (2) Accuracy risk: As the number of manual operations increases, the possibility of error increases, especially when dealing with complex or subtle cross-references.

[0042] (3) High professional knowledge requirements: Accurately performing manual cross-reference verification requires deep financial and accounting knowledge and a thorough understanding of business details, which places high demands on auditors.

[0043] (4) Adaptability issues: As financial standards and business practices continue to change, traditional manual verification methods may find it difficult to adapt to new report formats and cross-references.

[0044] In response to the problems in the manual cross-checking verification process, the industry is exploring and adopting new technologies and methods to improve efficiency and accuracy. For example, automated tools and financial software can help automatically identify data inconsistencies and potential errors, while artificial intelligence and machine learning technologies can provide deeper data analysis and insights. Through the application of these technologies, the burden of manual verification can be significantly reduced, and the efficiency and quality of audits can be improved.

[0045] In addition, auditors are also gradually adopting advanced analytical techniques, such as data mining and predictive analysis, to detect abnormal patterns and potential fraud in financial statements. These technologies can help auditors gain a deeper understanding of the economic activities behind financial statements, thereby improving the depth and breadth of audits.

[0046] In summary, manual cross-checking relationship verification is a traditional method to ensure the accuracy and completeness of financial statements. However, in the face of the challenges of the modern financial environment, it is necessary to combine new technologies and methods to optimize this process. In this way, the efficiency and adaptability of audit work can be improved while ensuring the quality of the statements. With the development and application of technology, future audit work will rely more on advanced tools and methods to cope with the increasingly complex financial environment.

[0047] This application also provides another method for auditing based on a large language model, such as Figure 3 As shown, the method comprises the following steps:

[0048] Step S302, constructing an intelligent agent.

[0049] In modern auditing practice, ensuring the accuracy and consistency of financial statements is a core task, but the efficiency and accuracy of traditional manual cross-checking verification are often questioned. Auditors need to be proficient in accounting knowledge and have a high level of attention and analytical ability. With the rapid development of artificial intelligence, especially the application of LLMs in NLP, the possibility of using these technologies to optimize the cross-checking detection process of financial statements has emerged.

[0050] LLMs, such as GPT-4 and its ilk, are trained to understand and generate human language. These models have demonstrated the ability to solve complex problems by analyzing large amounts of text data. In particular, in the financial field, these models can deeply analyze the relationship between accounting entries, discover anomalies and inconsistencies in the data, and thus provide automated solutions for cross-checking. The introduction of the LangChain framework provides an ideal foundation for the deployment of intelligent solutions based on LLMs. The framework combines the language understanding and domain-specific knowledge of LLMs, allowing developers to modularly build agents that focus on accounting and financial tasks. These agents can thoroughly analyze the content of financial statements, perform cross-checking verification, and make adjustment suggestions or corrections when necessary, greatly improving the efficiency of cross-checking and reducing the error rate of traditional auditing methods. Figure 4 Demonstrated LangChain-based agent.

[0051] Step S304, auditing based on the intelligent agent.

[0052] In the process of implementing the cross-reference relationship detection task, the workflow is divided into stages such as understanding and perception, logical reasoning and planning, and execution and adjustment. In the understanding and perception stage, the intelligent agent deeply analyzes the structure and content of the financial statements; in the logical reasoning stage, it uses reasoning techniques such as chain thinking to deepen the interpretation of data relationships; in the execution and adjustment stage, based on the analysis results of the first two stages, it proposes correction plans or directly adjusts the data. The design of the intelligent agent workflow needs to consider planning and coordination capabilities. Through iterative reflection, tools such as searching knowledge bases, generating codes, and performing mathematical operations are used until results that meet user needs are generated, which reflects the deep thinking and self-improvement capabilities of the intelligent agent when solving complex audit tasks. Figure 5 Demonstrates how the agent works.

[0053] Referring to the successful application of ProtAgents in the field of protein design, this application provides a set of specialized agents to perform cross-reference detection tasks in different parts of financial statements. For example, one agent processes the balance sheet, and another processes the income statement and cash flow statement. These agents can analyze data independently or work together to ensure the consistency and accuracy of the overall data. Figure 6 Demonstrate an example of cross-reference relationship detection agent orchestration.

[0054] With the continuous optimization of agent technology and the continuous accumulation of data, they can self-learn and adapt, and continuously improve the strategy and efficiency of cross-check relationship detection. This self-evolution ability indicates that agents based on LLMs and LangChain framework will play an increasingly important role in the future financial auditing field.

[0055] This application also provides an automated cross-reference relationship detection system, and its framework design and workflow are described in detail below.

[0056] In the process of automating financial auditing, the design and implementation of the cross-reference detection agent aims to solve the efficiency and accuracy problems of traditional manual methods. To achieve this goal, the agent integrates a series of steps and components to enable it to process financial statements in various formats and adapt to the specific needs of different industries and enterprises. The following is an optimized content summary to present the entire design and workflow more clearly and logically.

[0057] In the design of the agent, flexibility and scalability are the core principles to ensure that it can adapt to the unique needs of different industries and enterprises. Personalized customization allows the agent to accurately adapt to the changing financial environment and audit standards. In order to reduce the complexity of operation, the design team emphasized the intuitiveness and simplicity of the user interface, so that non-professional users can easily complete the entire process of cross-reference relationship detection.

[0058] In the actual operation process, the user first uploads the required financial statements through an interface customized for ease of use. The agent supports multiple file formats, such as PDF and Excel, to ensure the compatibility of different data sources. The user sets the relevant parameters for cross-check detection according to specific audit requirements, including cross-check rules and focus data paragraphs. After the parameters are set, the user submits the financial statements and parameters, and the agent immediately starts the automated cross-check relationship detection and is responsible for performing subsequent data processing and analysis. The smoothness of this process not only greatly improves audit efficiency, but also greatly reduces the technical requirements for users, representing a new chapter in financial audit automation.

[0059] The agent workflow is at the core of the automated audit design, covering the entire process from uploading financial statements to cross-reference detection. First, the uploaded unstructured financial statements are processed by the agent and converted into text data, using advanced text extraction technology to ensure accurate information extraction and data authenticity. Next, the agent divides the text content into independent text blocks and vectorizes them, creating a local vector library for subsequent analysis and comparison.

[0060] Based on the user's instructions, the agent generates a query vector and searches the local vector library to locate the text block that is most relevant to the query. This step is a crucial part of the agent's workflow, which ensures the directionality and efficiency of cross-reference detection. Subsequently, the agent uses LLMs to perform in-depth cross-reference analysis on the selected text block and outputs the analysis report in natural language, allowing users to easily understand the cross-reference results.

[0061] Next, the method for constructing the local vector library will be described in detail.

[0062] First, unstructured financial statements are obtained as audit data, and they are preprocessed and converted into text data; then the text data is divided into multiple independent text blocks, and these text blocks are vectorized to generate corresponding text block vectors. Then, based on the key factors in the financial field, relevant keywords are extracted from the audit knowledge base, and these keywords are combined with the text block vectors to generate financial element feature vectors containing contextual dependencies. Deep hidden layer features are extracted layer by layer through a multi-layer neural network. The leading layer captures local dependencies, the middle layer extracts long-distance dependency features, and the top layer obtains overall financial information through a global attention mechanism. Subsequently, based on the dynamic attention mechanism, the deep hidden layer features are fused with the text block vector to generate a fused feature vector containing financial element keywords, which is stored in the vector library to construct an audit knowledge base based on financial elements. Finally, the text blocks are classified by a classifier, and they are reorganized or corrected according to the classification results to ensure the accuracy and consistency of data processing, thereby improving the efficiency and precision of audit data processing. Among them, financial elements refer to the basic components of financial statements that reflect the economic activities of an enterprise, which are used to describe the financial status, operating results and cash flow of an enterprise. Common financial elements include assets, liabilities, owner's equity, income, expenses and profits. These elements help analyze the financial health of a company and assess its profitability, debt repayment ability and operating efficiency. Financial elements provide a classification framework in financial reports to help companies and stakeholders clearly understand and evaluate the performance of their economic activities.

[0063] After integrating the text block classification method, combined with the multi-scale feature mapping and speech classification methods, the text blocks can be classified in the following steps: First, after converting the unstructured financial statements into text data, multiple independent text blocks in the text data are linearly transformed based on the index and multi-scale feature mapping of the current network layer. Specifically, the text block features output by the previous layer (i.e., the input of the current layer) are linearly mapped by using the weight matrix of the current layer, and the linear transformation results of each text block are generated by combining the bias vector of the current scale. Subsequently, the linear transformation results are nonlinearly mapped using the activation function to extract feature information from different levels and scales to generate nonlinear mapping results. This multi-scale feature mapping method can capture local and global information in the text block, so as to better understand its pattern. Then, in the last layer of the network, based on the weight matrix of the last layer and the output features of the previous layer, the high-level features are mapped to multiple classification category spaces to generate the category scores corresponding to each text block. Through the Softmax activation function, these category scores are converted into probability values ​​and a probability distribution is generated. The Softmax function exponentially normalizes the category scores and ensures that the sum of the probabilities of all categories is 1, thereby generating a corresponding probability distribution for each text block. In order to improve the flexibility and accuracy of classification, the correlation weight between categories is introduced, and the temperature parameter is combined to adjust the smoothness of the output. The temperature parameter can control the degree of exploration of different categories by the model during the training process, and the model decision is made more accurate by lowering the temperature during the inference stage. Finally, the generated category probability distribution is classified by maximizing the likelihood method. Specifically, the category with the highest probability value is selected as the final classification result. This process ensures that the model can make the best classification decision for the text block based on the generated fusion features, and then make accurate classification judgments for the input text block, thereby improving the efficiency and accuracy of audit data processing.

[0064] This process not only significantly improves the efficiency and accuracy of cross-check detection, but also simplifies the user's operation steps. With the assistance of the agent, users can quickly start complex cross-check relationship analysis without having to deeply understand complex financial rules. In this way, the agent not only solves the time efficiency and accuracy problems in traditional auditing, but also is an easy-to-use tool for non-professional users.

[0065] With the help of intelligent agent technology, the detection process of financial statement cross-checking relationships can be significantly automated and efficient. The automation effect of the intelligent agent is reflected in its evolving learning and adaptation mechanism. Through machine learning technology, it continuously improves its own cross-checking detection strategy, significantly improving detection efficiency and accuracy.

[0066] Figure 7It provides an intuitive view of how the agent can efficiently perform the task of cross-reference detection. This process not only speeds up the audit workflow, but also improves the accuracy of corporate financial reports. With the further development of agent technology, its application is expected to go beyond the scope of cross-reference detection and expand to a wider range of audit fields such as financial risk assessment and compliance inspection, contributing new strength to the automation and intelligence of financial auditing.

[0067] In order to accurately verify the core cross-checking relationship in the balance sheet, especially the cross-checking relationship of current assets, this application provides a cross-checking relationship detection system, such as Figure 8 As shown in the figure. This system uses intelligent agents to perform cross-reference detection of financial statements, significantly improving the efficiency of the audit process and the accuracy of financial reports. By automating these complex cross-references, the system not only optimizes the audit workflow, but also strengthens the credibility of corporate financial statements, providing solid and reliable data support for key stakeholders such as investors, management and creditors.

[0068] Intelligent agents not only optimize the workflow of traditional financial auditing, but also open up a new development path for the analysis and verification of financial statements. As intelligent agent technology continues to improve and its application scope continues to expand, its role in the automation and intelligentization of financial auditing is expected to become more significant.

[0069] This application uses an agent developed using LLMs and LangChain technology to automate and improve the efficiency and accuracy of cross-checking relationship detection in financial statements. By analyzing the limitations of traditional manual cross-checking methods, it shows how the agent can understand and analyze financial data and automatically identify cross-checking relationships, solving the errors and inefficiencies of manual operations. This application also discusses the design principles and operating procedures of the agent, as well as its actual effect of improving automation in auditing work, demonstrating the innovative application and potential of LLMs and LangChain in the auditing field.

[0070] The present application also provides a device for auditing based on a large language model, such as Fig. 9 As shown, it includes: a construction module 92, which is configured to construct an audit knowledge base by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks; a generation module 94, which is configured to combine the audit knowledge base and the large language model using the LangChain framework, and based on the audit knowledge base, use the large language model to automatically retrieve and analyze audit materials to generate relevant audit reports and data matching results; a detection module 96, which is configured to use the large language model to perform cross-reference detection on the financial statements in the audit report and data matching results, so that the audit report and data matching results are logically and financially accurate.

[0071] It should be noted that the device for auditing based on a large language model provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device for auditing based on a large language model provided in the above embodiment and the method embodiment for auditing based on a large language model belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0072] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for auditing based on a large language model, characterized in that: include: By processing the audit data, an audit knowledge base is constructed, wherein the audit knowledge base includes common concepts and rules in audit tasks; The audit knowledge base and the large language model are combined using the LangChain framework. Based on the audit knowledge base, the large language model is used to automatically retrieve and analyze the audit materials to generate relevant audit reports and data matching results. The large language model is used to perform cross-reference detection on the audit report and the financial statements in the data matching results, so that the audit report and the data matching results are logically and financially accurate.

2. The method according to claim 1, characterized in that Build an audit knowledge base, including: Acquiring unstructured financial statements as the audit data, and processing and converting the unstructured financial statements into text data; The text content in the text data is divided into a plurality of independent text blocks, and the plurality of independent text blocks are vectorized to establish a vector library as the audit knowledge base.

3. The method according to claim 2, characterized in that The large language model is used to automatically retrieve and analyze audit materials, including: Generate a query vector based on the user's instructions; Searching the vector library based on the query vector to locate the most relevant text as the relevant audit report and data matching result; Among them, searching in the vector library based on the query vector includes at least one of the following: extracting balance sheet data in the vector library based on the query vector; extracting income statement data in the vector library based on the query vector; extracting cash flow statement data in the vector library based on the query vector.

4. The method according to claim 3, characterized in that Using the large language model to perform cross-reference detection on the audit report and the financial statements in the data matching result, including: Setting relevant parameters for cross-check detection based on preset audit requirements, wherein the relevant parameters for cross-check detection include cross-check rules and concerned data segments; Based on the relevant parameters of the cross-reference detection, the large language model is used to perform cross-reference detection on the audit report and the financial statements in the data matching results.

5. The method according to claim 4, characterized in that Using the large language model to perform cross-reference detection on the audit report and the financial statements in the data matching results, including: Using the large language model to conduct an in-depth analysis of the structure and content of the financial statements in the audit report and the data matching results to obtain analysis results; Using chain thinking reasoning technology, the large language model is used to deeply interpret the data relationship between the audit report and the financial statements in the data matching results to obtain the interpretation results; Based on the analysis results and the interpretation results, a cross-reference test is performed, and a correction plan is proposed or the data is directly adjusted.

6. The method according to claim 5, characterized in that Perform cross-reference relationship detection, including at least one of the following: Checking the internal cross-checking relationship of the table based on the balance sheet data, the income statement data, and the cash flow statement data; Based on the balance sheet data, the income statement data, and the cash flow statement data, checking the cross-checking relationship between the statements; Based on the balance sheet data, the income statement data, and the cash flow statement data, check the cross-checking relationship between the current period and the previous period statements.

7. A device for auditing based on a large language model, characterized in that: include: A construction module is configured to construct an audit knowledge base by processing audit data, wherein the audit knowledge base includes common concepts and rules in audit tasks; A generation module is configured to combine the audit knowledge base and the large language model using the LangChain framework, automatically retrieve and analyze audit materials based on the audit knowledge base using the large language model, and generate relevant audit reports and data matching results; The detection module is configured to use the large language model to perform cross-reference detection on the audit report and the financial statements in the data matching results, so that the audit report and the data matching results are logically and financially accurate.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images