Method and device for acquiring audit data based on audit data robot
Through the audit data robot automatically identifying audit requirements and obtaining data from the database, the problem of inefficient auditing is solved, efficient data retrieval and intelligent analysis are realized, and the skills and decision-making level of auditors are improved.
Patent Information
- Application Number
- CN202411312209.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-19
- Publication Date
- 2025-01-17
AI Technical Summary
In the audit of corporate financial statements, there are problems such as high digital and intelligent skills requirements, weak understanding of personalized audit requirements, and the level of intelligent analysis and decision-making of audit results need to be improved, resulting in low audit efficiency.
Using an audit data robot-based method, natural language instructions are obtained through a code generator, and the data acquisition intention is automatically identified and converted into query code. The database connector and code executor are used to retrieve audit data from the database, identify audit draft requirements and generate drafts. , use large language models for intelligent analysis.
It improves audit efficiency, reduces the mastery of SQL technology, simplifies the data acquisition process, and improves the efficiency of data retrieval and the intelligent analysis capabilities of audit results.
Smart Images
Figure CN119322784A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and device for obtaining audit data based on an audit data robot. Background Art
[0002] Although the field of corporate financial statement auditing is constantly using computer technology and big data technology to improve audit efficiency, with the rapid development of large language models and generative artificial intelligence technology, there are still problems in corporate financial statement auditing, such as high requirements for auditors' digital skills, weak ability to understand personalized audit needs, and the level of intelligent analysis and decision-making of audit results needs to be improved.
[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 obtaining audit data based on an audit data robot, so as to at least solve the technical problem of low audit efficiency in the prior art.
[0005] According to one aspect of an embodiment of the present invention, a method for acquiring audit data based on an audit data robot is provided, comprising: obtaining, through a code generator of the audit data robot, natural language instructions for indicating the semantics of the current audit requirement, automatically identifying data acquisition intentions corresponding to the semantics of the current audit requirement according to the natural language instructions, and converting the identified data acquisition intentions into query codes for querying corresponding audit data from a database; connecting a code executor and the database through a database connector of the audit data robot; and retrieving the corresponding audit data from the database based on the query code using the database connector through the code executor of the audit data robot.
[0006] According to another aspect of an embodiment of the present invention, there is also provided a device for acquiring audit data, namely, an audit data robot, comprising: a code generator, configured to acquire natural language instructions for indicating the semantics of the current audit requirement, automatically identify data acquisition intentions corresponding to the current audit requirement semantics according to the natural language instructions, and convert the identified data acquisition intentions into query codes for querying corresponding audit data from a database; a database connector, configured to connect a code executor and the database; and a code executor, configured to retrieve the corresponding audit data from the database using the database connector based on the query code.
[0007] In the embodiment of the present invention, the data acquisition intention corresponding to the current audit demand semantics is identified, and audit data is acquired from the database according to the identified data acquisition intention; the working paper production intention corresponding to the audit working paper demand semantics is identified, the corresponding working paper production tool is selected according to the identified working paper production intention, and the audit working paper is generated using the working paper production tool; the data analysis intention corresponding to the audit analysis demand semantics is identified, and the audit working paper is intelligently analyzed based on the identified data analysis intention using historical audit experience. Through the above scheme, the technical problem of low audit efficiency in the prior art is 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 flowchart of a method for obtaining audit data based on an audit data robot according to an embodiment of the present invention;
[0010] Figure 2 is an architectural diagram of an audit data robot according to an embodiment of the present invention;
[0011] Figure 3 is a structural diagram of an audit data acquisition model according to an embodiment of the present invention;
[0012] Figure 4 is a flowchart of another method for obtaining audit data based on an audit data robot according to an embodiment of the present invention;
[0013] Figure 5 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION
[0014] 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.
[0015] 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.
[0016] The present application embodiment provides a method for obtaining audit data based on an audit data robot. The audit data robot is used to execute the method for obtaining audit data based on an audit data robot, such as Figure 1 As shown, the method includes:
[0017] Step S102, through the code generator of the audit data robot, obtain natural language instructions for indicating the current audit requirement semantics, automatically identify the data acquisition intention corresponding to the current audit requirement semantics according to the natural language instructions, and convert the identified data acquisition intention into a query code for querying the corresponding audit data from the database.
[0018] First, an example selector is used to select the example most relevant to the natural language instruction based on the natural language instruction. For example, the natural language instruction is converted into a query embedding vector Vquery, and the sample embedding vector Vs corresponding to each sample in the sample library is obtained; the cosine similarity between the query embedding vector Vquery and the sample embedding vector Vs is calculated, and the sample with the highest semantic similarity to the current audit requirement text is selected based on the cosine similarity to obtain the most relevant example. For another example, the semantic similarity example selector and the embedded vector database integrated in the LangChain framework are used to select the most relevant example based on the natural language instruction.
[0019] Then, the corresponding data acquisition intention is automatically identified based on the natural language instruction and the most relevant example. The audit data robot includes a large language model. The natural language instruction and the most relevant example are input into the large language model of the audit data robot, and the data acquisition intention will be output. Among them, the audit data robot is trained by the following method: creating an audit data acquisition sample library, wherein the audit data acquisition sample library includes two types of data acquisition tasks, and the two types of data acquisition tasks include simple audit data acquisition tasks and complex audit data acquisition tasks; using the audit data acquisition sample library to train the audit data robot. Among them, the simple audit data acquisition task is used to inquire about the value of a specific subject in the current period or the same period of the previous year; the complex audit data acquisition task is used to inquire about the calculation of fields in the table.
[0020] Step S104, connecting the code executor and the database through the database connector of the audit data robot.
[0021] First, ensure that the database connector establishes a secure and stable connection with the target database. The connection process includes database address configuration, identity authentication, and compatibility check with the database management system. By implementing security protocols (such as SSL / TLS), ensure security and data integrity during data transmission.
[0022] Next, the database connector will parse the query code and generate the corresponding database query request based on the generated query code. The format and content of the query request will be adjusted according to the type of the target database (such as relational database, NoSQL database, etc.) to ensure that the query code can be executed correctly and obtain the required audit data.
[0023] During this process, the database connector also performs the following specific operations: 1) Query optimization. Analyze the SQL statements or other database query languages in the query code to determine the optimal execution path to improve query efficiency and response speed. For large databases or complex queries. 2) Concurrency management. If multiple query requests are executed at the same time, the database connector needs to manage concurrent access to avoid database locking or resource competition to ensure that each query request can be executed smoothly. 3) Error handling. When executing the query code, if you encounter database connection failures, query syntax errors, or data retrieval anomalies, the database connector needs to capture and handle these errors. At this point, the system can choose to retry the connection, record logs for audit tracking, or feedback error information to the user and suggest possible solutions. 4) Data format conversion: Convert the format of the original data retrieved from the database to match the system requirements or the user's expected format. For example, convert table data in a relational database to JSON format, or convert document data in a NoSQL database to a tabular form.
[0024] Step S106, through the code executor of the audit data robot, based on the query code, use the database connector to retrieve the corresponding audit data from the database.
[0025] First, the code executor receives the query code passed by the database connector. This query code has been generated and optimized by the previous step to ensure that it can run efficiently in the target database.
[0026] Next, the code executor starts to execute the query code and sends a specific query request to the database. Depending on the complexity of the query and the amount of data, the code executor will dynamically adjust the execution strategy. For example, for simple query requests, the code executor can directly initiate a single request to the database and return the result immediately; for complex queries or join operations involving multiple data tables, the code executor may execute the query in batches, or gradually merge and process the query results to optimize performance and memory usage.
[0027] During the data retrieval process, the code executor also monitors the query execution status in real time. If any problems occur during the query process, such as data timeout, query failure, or database connection interruption, the code executor will immediately initiate a predetermined error handling mechanism, for example, including re-initiating the query request, adjusting the query strategy, or sending a warning to the system administrator for further action.
[0028] When the query is successfully completed, the code executor will perform preliminary processing and formatting on the audit data retrieved from the database to ensure that the data meets the requirements of subsequent analysis or presentation. This processing may include steps such as data cleaning, deduplication, type conversion, or calculation of derived indicators.
[0029] Finally, the processed audit data will be stored in a designated storage location or directly passed to the next functional module for more in-depth analysis or presentation. In this application, the code executor is not only a tool for executing queries, but also ensures the accuracy and efficiency of data retrieval and the overall stability and security of the system.
[0030] The present application also provides an audit data robot. The overall framework of the enterprise financial statement audit determines that the audit data robot will mainly realize the functions of connecting to the database, extracting data in natural language, and performing substantive testing in the audit procedure. Figure 2 As shown in the figure, auditors can provide natural language instructions to the audit data robot built on the big language model according to their own data acquisition and substantive testing needs. The audit data robot will automatically identify the required intent, such as "profit statement", "total operating income" and other more prominent words. According to the results of intent recognition, the relevant information of the data table corresponding to the intent is obtained, and the big language model is used to convert the intent into query code. The results are extracted from the corresponding data table in the database through automatic code execution, and finally the output results are sorted according to the required output format and passed to the auditor.
[0031] The functional modules implemented by the audit data robot in the corporate financial statement audit process mainly include two aspects: audit data and substantive audit testing. Since both the acquisition of audit data and substantive testing involve the acquisition of the audited company's financial data, and the substantive audit testing is to verify the acquired data. Therefore, the automation of substantive testing will also be completed by the audit data robot. The audit data robot's function of acquiring audit data is mainly achieved by using a large language model to implement Text-to-SQL, that is, the method of converting text into database commands, accurately converting the auditor's needs into database commands that can be executed in the database, and executing relevant commands on the connected database platform to obtain relevant data. In addition, regarding the audit data robot's function involving substantive testing in the audit procedure, the data in the database will be verified step by step through the Chain of Thought (CoT) method, and the relevant results will be output in the end.
[0032] 1) Construction of audit data acquisition model based on Text-to-SQL
[0033] The large language model not only demonstrates powerful capabilities in text understanding and generation, but also demonstrates excellent performance in code writing. The application of the large language model in the field of Text-to-SQL allows users to query the database through natural language without having to master the professional knowledge of the database language, greatly reducing the threshold for database queries and improving the efficiency of data retrieval. Applying Text-to-SQL technology to the field of audit data acquisition can convert auditors' natural language queries into specific database commands, greatly simplifying the data acquisition process and reducing auditors' mastery of SQL technology.
[0034] This application uses the Text-to-SQL technology of the large language model, which uses the large language model to understand the auditor's query intent and capture key information in user needs, and accurately converts it into a query statement that can be executed by the database. The formula for generating database commands is as follows: Y = f(Q, S|θ).
[0035] Where Y represents the generated database command, Q represents the audit data-related questions raised by the auditor, and S represents the structural information of the tables, columns, and relationships between tables in the audit database. This information describes the audit data structure in detail and is crucial for generating accurate database query statements. f(·|θ) is the core transformation mechanism of the large language model. It uses the model parameter θ to map the natural language queries and database structure information raised by the user into specific database commands.
[0036] The audit data robot includes an audit data acquisition model. The audit data acquisition model is as follows: Figure 3The code generator contains three main core components, namely the code generator, the code executor, and the database connector. The code generator is the entry point of the model and is responsible for converting the natural language queries passed by the auditors into executable commands for the database. The powerful natural language processing capabilities of the large language model are used to understand the intent and requirements in the query and capture key information. Through the sophisticated algorithms of the large language model, the code generator implements the structural and semantic analysis of natural language queries, maps them to the corresponding database structure, and generates accurate database commands. This process involves complex semantic understanding and language-to-code conversion to ensure that the generated database query statements not only conform to the database query syntax, but also accurately reflect the user's query intent. The code generator passes the generated code to the code executor, runs the generated database commands in the code executor, and retrieves data from the database. It is used as a bridge between the model and the database, which not only ensures the correct execution of the database statements, but also handles any technical problems that may be encountered during the execution process, such as query optimization. The code executor is also responsible for returning the query results in a standardized format, which facilitates the subsequent processing and analysis of the data results by the large language model. The large language model uses its own language generation capabilities to process the output results of this executor in natural language, answer the acquired data in natural language, or process the data results according to user needs, such as output in the form of tables, lists, etc. to meet the user's needs for data viewing. The database connector is responsible for establishing and maintaining the connection between the model and the target database to ensure the secure transmission of data, including the transmission of audit data query statements and audit data query results. At the same time, an important part of the database connector is to transmit the information of the data table in the database to the audit data acquisition model, so that the large language model can grasp the relevant information of the data it owns. The database connector needs to establish a stable and reliable connection with different database environments and configurations, so that the audit data acquisition model can seamlessly access the company's financial data and support complex and efficient data queries.
[0037] For the audit data acquisition model built with Text-to-SQL technology using a large language model, this application will use the LangChain framework and the SQLDatabaseChain component in the LangChain library to implement it. According to the effects of natural language to database command query in combination with Text-to-SQL technology for various large language models, such as GPT3.5-Turbo, GPT-4, Claude-2, Llama, etc., it can be seen that GPT4 and GPT3.5-Turbo can achieve the best effect of generating database query commands. Therefore, this application selects GPT-4 as the large language model used in this application, and uses the LangChain framework to implement the API call to GPT-4 to call the large language model.
[0038] As an open source library, LangChain provides a wealth of tools and interfaces for building applications based on large language models, and is particularly suitable for processing complex natural language understanding tasks. With the help of LangChain, the audit data acquisition model provided by this application includes three main components: code generator, code executor, and database connector to realize the automation and intelligence of the audit data acquisition process. SQLDatabaseChain is a LangChain component specially designed to execute database queries. The code generator converts the auditor's natural language queries into precise database commands based on the language processing capabilities of LangChain and GPT-4's API. Subsequently, the SQLDatabaseChain component receives the database query statements generated by the code generator, executes these statements through the preset database connection, and retrieves data from the database safely and efficiently, thereby realizing the function of the code executor. At the same time, relying on SQLDatabaseChain to realize the function of database connection management, by configuring the database connection settings, including database address, authentication information, etc., to establish a secure database connection. Finally, the code generator, code executor, and database connector are integrated into a collaborative system to realize the audit data acquisition process of the process.
[0039] 2) Model optimization with a small number of sample prompts
[0040] The audit data acquisition model is built using the Text-to-SQL method. The GPT-4 used is still universal. There may be some subtle differences and problems in the output of the audit data set on the audit data platform and the auditors' questions about the audit data. At the same time, although the large language model has demonstrated excellent capabilities in zero-sample learning, it is still unable to complete more complex tasks when implementing zero-sample prompts. A small number of sample prompts can be used as a technology to achieve contextual learning, guiding the model to achieve better performance in the form of examples in the prompt.
[0041] This application will use a small number of sample prompts to optimize the audit data acquisition model. The core is to use limited, high-quality samples to guide the model to more accurately understand the audit-related natural language queries provided by auditors and convert them into database query commands. By obtaining a small number of carefully selected samples of audit-related data questions and answers, the performance of the audit data acquisition model is improved, so that the model can quickly meet the specific needs of the audit field without relying on large data sets for training.
[0042] According to the audit data acquisition tasks involved in the audit process of corporate financial statements, an audit data acquisition sample library is created. The audit data acquisition sample library mainly includes two aspects: commonly used simple audit data acquisition tasks and complex audit data acquisition tasks. The audit data acquisition model consists of four aspects: audit data acquisition input, audit code generation, audit code execution and output. When using audit data to obtain a small number of samples, it is also necessary to provide input (input), code generation (sql_cmd), code execution (result) and output (answer) in four aspects, so that the audit data acquisition model can better learn related examples. The specific examples of the audit data acquisition sample library are shown in Table 4.1 below. The table mainly takes profit and loss accounts as an example to illustrate the samples of some typical audit data acquisition tasks.
[0043] Table 4.1 Audit data acquisition task sample library types and examples
[0044]
[0045]
[0046] In common simple audit data acquisition tasks, it is usually necessary to inquire about the value of a specific account in the current period or the same period of the previous year, while complex audit data acquisition tasks involve calculations of fields within the table, etc. Therefore, this application will put complex query examples that may be used in the corporate financial statement audit process into the audit data acquisition sample library to improve the accuracy of the audit data model.
[0047] Since the audit data acquisition sample library stores two types of data acquisition tasks, and involves data acquisition examples for multiple tables such as financial statements, chronological accounts, and account balance sheets, these examples are uniformly input into GPT-4 when conducting audit data acquisition work. On the one hand, the number of tokens in the API of GPT-4 will exceed the limit and cannot fully output the desired content. On the other hand, it aggravates the difficulty of GPT-4 in selecting examples, resulting in inaccurate output results after learning. Therefore, this application will introduce the example selector in LangChain to realize the selection of examples most relevant to the questions raised. The selected examples will be put into the entire audit data acquisition model as input together with the audit data acquisition requirements provided by the auditors to improve the accuracy of audit data acquisition. Therefore, this application will adopt the semantic similarity example selector integrated in the LangChain framework and the advanced technology of the embedded vector database Chroma to realize the precise selection of examples most relevant to the questions raised. The specific calculation formula for the most relevant example is as follows:
[0048]
[0049] Among them, S represents the sample set of the sample library, Vquery is the embedding vector converted from the query requirements provided by the auditor, and V S is the embedding vector corresponding to each sample in the sample library. By calculating the cosine similarity between the input query embedding vector and the embedding of each sample in the audit data acquisition sample library, the sample with the highest semantic similarity to the current audit requirement is selected.
[0050] The most relevant examples selected will be used as inputs together with the specific audit data acquisition requirements provided by the auditors and put into the entire audit data acquisition model. Combining the precise example selection of the semantic similarity example selector with Chroma's efficient embedding vector retrieval capabilities, this method greatly improves the accuracy of audit data acquisition. This example selection strategy based on semantic similarity ensures the high adaptability and accuracy of the model when processing specific audit queries, and provides an innovative and effective optimization approach for data acquisition in the audit field. At the same time, when selecting examples using semantic similarity, a specified number of examples can be selected by specifying parameters.
[0051] In some embodiments, the following method can also be used to calculate the most relevant example. First, the query embedding vector Vquery and the sample embedding vector V Si The cosine similarity of the jth dimension among the m total dimensions is calculated, and the cosine similarities of all the m total dimensions are weighted summed to obtain the comprehensive similarity between the embedding vector and the sample; then, based on the sample S i The embedding vector V Si The regularization term is calculated by the square of the L2 norm and the regularization coefficient; finally, the most relevant example is determined based on the comprehensive similarity and regularization. The specific calculation formula is as follows:
[0052]
[0053] Among them, W i is sample S i is a weight factor used to reflect the importance or quality of the sample. λ is the regularization term coefficient, which controls the strength of the regularization term to prevent overfitting. γj is used to control the nonlinear activation function of each embedding dimension. σ is used to control the activation function of the final example score. Vquery is the embedding vector converted from the query requirement text provided by the auditor. V Si is the embedding vector of the i-th sample in the sample library, m represents the number of dimensions of the embedding vector, Similarity(V query,j ,V Si,j ) represents the query embedding vector Vquery and the sample embedding vector V Si The cosine similarity in the jth dimension, a jis the weight coefficient of the j-th similarity calculation dimension.
[0054] The above scheme introduces multi-dimensional similarity calculation, sample weighting, nonlinear activation function and regularization term to comprehensively evaluate the similarity between audit requirements and each sample in the sample library, thereby achieving more accurate sample selection. This scheme can not only better capture complex multi-dimensional data relationships, but also effectively prevent model overfitting, improve the generalization ability and robustness of the model, and ultimately improve the accuracy and efficiency of audit data acquisition.
[0055] Finally, after completing the construction of the semantic similarity example selector, by giving the big language model relevant instructions, for example, by setting up a logo before the example, such as "Here are some examples for your reference:", it provides guidance for the big language model to convert the auditor's natural language query requirements into database instructions.
[0056] The present application embodiment provides another method for obtaining audit data based on an audit data robot, such as Figure 4 As shown, the method comprises the following steps:
[0057] Step S402, initialization of the audit data robot.
[0058] Before audit data acquisition begins, the audit data robot must be initialized. Initialization includes setting audit task parameters, such as audit scope, audit data source, data format requirements, etc. The robot generates the corresponding data acquisition strategy based on the set task parameters.
[0059] Step S404, connecting to the audit data source.
[0060] The audit data robot is connected to various audit data sources through a preset interface. The data sources may include but are not limited to financial systems, ERP systems, bank reconciliation systems, electronic invoice systems, etc. The robot can automatically log in to these systems and obtain the required data based on the pre-configured permissions and authentication information.
[0061] Step S406: data acquisition and processing.
[0062] The robot obtains the specified data from various audit data sources based on the audit task parameters. During the data acquisition process, the robot will preprocess the data. The preprocessing steps include: Data format conversion: converting data from different data sources into a unified format for subsequent processing and analysis. Data cleaning: removing redundant data, repairing missing data, and processing outliers to ensure data integrity and accuracy. Data classification and storage: the processed data is classified and stored in the specified audit database for subsequent analysis and reference.
[0063] Step S408: data review and verification.
[0064] After the data is acquired, the robot will perform the data review and verification steps. Through the preset audit rules and algorithms, the robot will conduct a preliminary audit of the acquired data to check the consistency, accuracy and compliance of the data. In this step, the robot can use historical audit data and industry standards for comparative analysis to identify potential audit risk points.
[0065] Step S410: Generate an audit report.
[0066] Based on the acquired data and audit results, the robot automatically generates a preliminary audit report. The audit report contains key data indicators, risk analysis results, and relevant audit recommendations. Auditors can conduct further audit work or make decisions based on the content of the report.
[0067] The method for obtaining audit data based on the audit data robot provided in this application can realize all the functions of the above-mentioned audit data robot, so it will not be repeated here.
[0068] An embodiment of the present application also provides a device for obtaining audit data, including: a code generator, configured to obtain natural language instructions for indicating the current audit requirement semantics, automatically identify data acquisition intentions corresponding to the current audit requirement semantics according to the natural language instructions, and convert the identified data acquisition intentions into query codes for querying corresponding audit data from a database; a database connector, configured to connect a code executor and the database; and the code executor, configured to retrieve the corresponding audit data from the database using the database connector based on the query code.
[0069] Figure 5 FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Figure 5 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.
[0070] like Figure 5 As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.
[0071] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.
[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 obtaining audit data based on an audit data robot, characterized in that: include: Obtaining, by means of a code generator of the audit data robot, a natural language instruction for indicating the semantics of the current audit requirement, automatically identifying a data acquisition intention corresponding to the semantics of the current audit requirement according to the natural language instruction, and converting the identified data acquisition intention into a query code for querying the corresponding audit data from a database; Connecting the code executor and the database via the database connector of the audit data robot; The code executor of the audit data robot retrieves the corresponding audit data from the database using the database connector based on the query code.
2. The method according to claim 1, characterized in that Automatically identifying the data acquisition intention corresponding to the semantics of the current audit requirement according to the natural language instruction, including: selecting, using an example selector, an example most relevant to the natural language instruction based on the natural language instruction; The corresponding data acquisition intention is automatically identified according to the natural language instruction and the most relevant example.
3. The method according to claim 2, characterized in that Selecting, using an example selector, an example most relevant to the natural language instruction based on the natural language instruction includes: Convert the natural language instruction into a query embedding vector, and obtain a sample embedding vector corresponding to each sample in the sample library; The cosine similarity between the query embedding vector and the sample embedding vector is calculated, and based on the cosine similarity, a sample with the highest semantic similarity to the current audit requirement text is selected to obtain the most relevant example.
4. The method according to claim 2, characterized in that: Utilizing an example selector to select the example most relevant to the natural language instruction based on the natural language instruction, including: utilizing a semantic similarity example selector and an embedded vector database integrated within a LangChain framework to select the most relevant example based on the natural language instruction.
5. The method according to claim 1, characterized in that The audit data robot is trained by: Creating an audit data acquisition sample library, wherein the audit data acquisition sample library includes two types of data acquisition tasks, the two types of data acquisition tasks including simple audit data acquisition tasks and complex audit data acquisition tasks; The audit data acquisition sample library is used to train the audit data robot.
6. The method according to claim 5, characterized in that The simple audit data acquisition task is used to inquire about the value of a specific subject in the current period or the same period of the previous year; the complex audit data acquisition task is used to inquire about the calculation of fields in the table.
7. A device for obtaining audit data, characterized in that: include: A code generator is configured to obtain a natural language instruction for indicating the semantics of the current audit requirement, automatically identify a data acquisition intention corresponding to the semantics of the current audit requirement according to the natural language instruction, and convert the identified data acquisition intention into a query code for querying corresponding audit data from a database; A database connector configured to connect the code executor and the database; A code executor is configured to retrieve the corresponding audit data from the database using the database connector based on the query code.
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.