Audit model automatic generation method and system based on large language model technology

Abstract

The invention provides an auditing model automatic generation method and system based on a large language model technology, and the method comprises the steps: constructing an auditing knowledge vector database according to obtained target auditing data; based on a large language model technology, performing deep learning on the data features of the target audit data in the audit knowledge vector database to obtain an audit model corresponding to the target audit data; matching an audit input problem input by the target object with target audit data in the audit knowledge vector database, and determining a target type of audit data corresponding to the audit input problem in the target audit data; according to the target type and an audit model corresponding to the target audit data, a target audit model is determined, and the target audit model is a model corresponding to the audit input problem; the auditing knowledge vector database is combined with the large language model technology, and the auditing model is automatically generated by using the large language model technology, so that the timeliness and error rate of auditing are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning technology, specifically relating to a method and system for automatically generating audit models based on large language model technology. Background Technology

[0002] With the continuous development and application of internet technology, the application level and popularity of information technology are increasing, and digital auditing has become an important direction for the future development of auditing. Regardless of the industry, to achieve high-quality development, it is necessary to explore digital auditing while considering how to make it more effective and faster to implement.

[0003] Taking power grid companies as an example, digital auditing in power grid companies is based on a fundamental audit database established by various business systems. Through the collection, transformation, cleaning, analysis, and verification of data required for audit inspections, more user-friendly and accurate intermediate tables are created. These intermediate tables are primarily implemented using SQL (Structured Query Language) and advanced Excel analysis tools. Multi-dimensional audit data within these intermediate tables is then mined and analyzed, and audit models are constructed using audit analysis techniques. This reveals hidden trends and potential anomalies behind big data, enabling a comprehensive understanding of the data, highlighting key areas of focus, effectively collecting audit evidence, and ultimately achieving the goals of the audit task efficiently. With the rapid development of the power industry, the auditing work of power grid companies has become increasingly important.

[0004] In the massive auditing work of projects, it is necessary to collect and audit a large amount of audit data. The existing processing methods mostly rely on manual processing, which leads to low timeliness and high error rate in the auditing work. Summary of the Invention

[0005] To overcome the shortcomings of the existing technology, this invention proposes an automatic audit model generation method based on large language model technology, comprising:

[0006] Based on the acquired target audit data, construct an audit knowledge vector database;

[0007] Based on the large language model technology, deep learning is performed on the data features of the target audit data in the audit knowledge vector database to obtain the audit model corresponding to the target audit data.

[0008] The audit input question input by the target object is matched with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question; and the target audit model is determined according to the target type and the audit model corresponding to the target audit data, wherein the target audit model is the model corresponding to the audit input question.

[0009] Furthermore, the process of performing deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology to obtain the audit model corresponding to the target audit data includes:

[0010] Based on the large language model technology, feature extraction is performed on the target audit data in the audit knowledge vector database to obtain the data features of the target audit data;

[0011] Using the question features in the data features as input and the answer features in the data features as output, a large language model is trained to predict the corresponding answer features based on the input question features. When the trained large language model reaches the convergence condition, the audit model corresponding to the target audit data is obtained.

[0012] Furthermore, both the question features and the answer features include at least one of the following features: statistical features, text features, and time series features;

[0013] The answer features include an audit question prompt field and / or an audit question answer field.

[0014] Furthermore, after performing deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology to obtain the audit model corresponding to the target audit data, the method further includes:

[0015] Select the first audit model to be evaluated from the audit models corresponding to the target audit data; and obtain test data for evaluating the first audit model.

[0016] Based on the large language model technology, feature extraction is performed on the test data to obtain the data features of the test data;

[0017] The first audit model is used to perform audit analysis on the problem features in the test data to obtain the predicted answers corresponding to the problem features in the test data.

[0018] Based on the answer characteristics and predicted answers in the test data, the first audit model is evaluated, and the evaluation result of the first audit model is determined.

[0019] Based on the evaluation results of the first audit model, the first audit model is optimized.

[0020] Further, the step of evaluating the first audit model based on the answer features in the data features of the test data and the predicted answer, and determining the evaluation result of the first audit model, includes:

[0021] Based on the answer features in the data features of the test data and the predicted answer, the accuracy of the first audit model is evaluated, and the accuracy evaluation result of the first audit model is determined.

[0022] And / or; based on the answer features in the data features of the test data and the predicted answer, determine the precision and recall evaluation results of the first audit model.

[0023] Furthermore, optimizing the first audit model based on the evaluation results of the first audit model includes:

[0024] Based on the evaluation results of the first audit model, an optimization algorithm is determined; the optimization algorithm includes at least one of the following: parameter adjustment algorithm, network structure optimization algorithm, and regularization algorithm;

[0025] Based on the evaluation results and the optimization algorithm, the parameters and / or network structure of the first audit model are adjusted.

[0026] Furthermore, the step of constructing an audit knowledge vector database based on the acquired target audit data includes:

[0027] Based on the acquired target audit data, an audit data corpus is formed; the target audit data includes at least one of the following types: structured report data, unstructured contract text data, business document data, power load data, equipment operating status data, and user electricity consumption data;

[0028] The target audit data in the audit data corpus is vectorized to obtain an audit knowledge vector database.

[0029] Furthermore, the step of forming an audit data corpus based on the acquired target audit data includes:

[0030] Based on the obtained target audit data, the target audit data is preprocessed; the preprocessing includes at least one of the following processing methods: data cleaning, data integration, data transformation, data reduction, missing value handling, outlier detection and handling, and text segmentation;

[0031] An audit data corpus is formed based on the preprocessed target audit data.

[0032] Furthermore, after determining the target audit model based on the target type and the audit model corresponding to the target audit data, the process further includes:

[0033] Based on the audit input questions, a question-and-answer prompt is generated;

[0034] The target audit model is used to perform audit analysis on the question-and-answer request prompt to obtain the answer corresponding to the audit input question.

[0035] Based on the same inventive concept, this invention also provides an automatic audit model generation system based on large language model technology, comprising:

[0036] The database construction module is used to build an audit knowledge vector database based on the acquired target audit data;

[0037] The model generation module is used to perform deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology, so as to obtain the audit model corresponding to the target audit data.

[0038] The model determination module is used to match the audit input question input by the target object with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question in the target audit data; and to determine the target audit model according to the target type and the audit model corresponding to the target audit data, wherein the target audit model is the model corresponding to the audit input question.

[0039] Based on the same inventive concept, the present invention also provides a computer device, comprising: one or more processors;

[0040] Memory, used to store one or more programs;

[0041] When the one or more programs are executed by the one or more processors, an automatic audit model generation method based on large language model technology, as described above, is implemented.

[0042] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the aforementioned method for automatically generating an audit model based on a large language model technology.

[0043] Compared with the closest existing technology, the present invention has the following beneficial effects:

[0044] This invention provides a method and system for automatically generating audit models based on large language modeling technology. The method includes: constructing an audit knowledge vector database based on acquired target audit data; performing deep learning on the data features of the target audit data in the audit knowledge vector database based on large language modeling technology to obtain an audit model corresponding to the target audit data; matching the audit input question input by the target object with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question; and determining the target audit model based on the target type and the audit model corresponding to the target audit data, wherein the target audit model is the model corresponding to the audit input question. This invention combines an audit knowledge vector database with large language modeling technology to automatically generate audit models, reducing manual intervention and errors, improving the accuracy and efficiency of audit models, thereby improving the timeliness and error rate of audit work, and providing more efficient, accurate, and reliable audit support. Attached Figure Description

[0045] Figure 1 A schematic diagram of the process for automatically generating an audit model based on large language model technology provided by this invention;

[0046] Figure 2 A schematic diagram of an example process for an automatic audit model generation method based on large language model technology provided by the present invention;

[0047] Figure 3 This invention provides a schematic diagram of the structure of an automatic audit model generation system based on large language model technology. Detailed Implementation

[0048] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Example 1:

[0050] This invention provides a method for automatically generating audit models based on large language model technology, such as... Figure 1 As shown, it includes:

[0051] S1. Construct an audit knowledge vector database based on the obtained target audit data;

[0052] S2. Based on the large language model technology, deep learning is performed on the data features of the target audit data in the audit knowledge vector database to obtain the audit model corresponding to the target audit data.

[0053] S3. Match the audit input question input by the target object with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question in the target audit data; and determine the target audit model based on the target type and the audit model corresponding to the target audit data, wherein the target audit model is the model corresponding to the audit input question.

[0054] This invention combines an audit knowledge vector database with large language model technology to automatically generate audit models, reducing manual intervention and errors, and improving the accuracy and efficiency of audit models. This improves the timeliness and error rate of audit work, and provides more efficient, accurate and reliable audit support.

[0055] The target audit data in S1 above may include a set of audit data, specifically including but not limited to historical and / or real-time data related to the audit collected from multiple data sources. For example, target audit data includes, but is not limited to, at least one of the following types: structured report data, unstructured contract text data, business document data, power load data, equipment operating status data, and user electricity consumption data. During the data collection phase, the comprehensiveness and accuracy of the collected audit data can be determined through reasonable collection strategies and tools. For example, audit data can be collected from various data sources through specific data collection tools or interfaces. The aforementioned interface may be an API (Application Programming Interface). API callers can be implemented using various programming languages, including Java, Python, PHP (Hypertext Preprocessor), or other programming languages.

[0056] The audit knowledge vector database mentioned in S1 above can quickly process large-scale data to improve the efficiency of auditing work. For example, the audit knowledge vector database could be the Faiss vector library (an open-source vector database). The Faiss vector library can not only quickly process large-scale data, but also perform similarity searches in high-dimensional spaces, offering the benefits of high efficiency and scalability.

[0057] In one example, in S1, an audit data corpus can be formed based on the acquired target audit data. The target audit data in the audit data corpus is then vectorized to obtain an audit knowledge vector database. The audit data corpus is used to construct the audit knowledge vector database. The audit data corpus can be a large-scale electronic text library obtained by sampling and processing the target audit data. During the automated generation of audit models, the audit data corpus can provide a text library or dataset for training the audit models.

[0058] To reduce errors and biases in the data analysis and modeling process, in this embodiment of the invention, forming an audit data corpus based on the acquired target audit data may include: preprocessing the target audit data; the preprocessing includes at least one of the following processing methods: data cleaning, data integration, data transformation, data reduction, missing value handling, outlier detection and handling, and text segmentation; and forming an audit data corpus based on the preprocessed target audit data.

[0059] Specifically, the preprocessing process involves several steps: data cleaning filters and processes invalid, missing, abnormal, duplicate, or other problematic data from the original data; data integration merges and integrates data from different audit data sources after cleaning; data transformation performs feature extraction and data transformation on the integrated data to convert it into a modeling data format; and data reduction unifies and standardizes the transformed data and categorizes it. Data reduction can obtain a streamlined dataset from the original dataset, thus reducing its size while preserving its complete characteristics, thereby improving work efficiency.

[0060] In one implementation, step S2 above may include: extracting features from the target audit data in the audit knowledge vector database using large language modeling technology to obtain data features of the target audit data; using question features from the data features as input and answer features from the data features as output, training a large language model to predict corresponding answer features based on the input question features; and obtaining the audit model corresponding to the target audit data when the trained large language model reaches convergence. This step utilizes large language modeling technology and advanced algorithms such as deep learning to automatically generate audit models, thereby better handling large-scale data and improving the accuracy and efficiency of the audit models. Furthermore, large modeling technology can extract richer and more accurate features, further improving the performance of the audit models. It can be understood that if the target audit data includes multiple types of audit data, this implementation can automatically generate corresponding audit models for each type of audit data to meet the work requirements of various audit scenarios.

[0061] The aforementioned question and answer features each include at least one of the following: statistical features, textual features, and time-series features. Question features may include an audit question field, and answer features may include an audit question hint field and / or an audit question answer field. These various feature types can be used to describe audit-related attributes in the audit data. The purpose of feature extraction from the audit data in the audit knowledge vector database is to extract audit-related information from the raw data, providing input information for subsequent training of the large language model. For example, the feature extraction process may include: using large language model techniques, such as Transformer (a deep learning model architecture) and LSTM (Long Short-Term Memory), to extract features from the preprocessed data in the audit knowledge vector database.

[0062] The process of training a large language model can include: training an audit model using large language model technology based on extracted data features (such as the question and answer features mentioned above). Through training, the audit model can learn audit rules, patterns, and trends in the data, providing support for subsequent audit applications. During the model training phase, a model architecture and optimization algorithm are selected, and model parameters are adjusted according to actual needs. For example, the selected model can be a large language model chosen based on requirements, such as, but not limited to, OpenAI's GPT series models (Generative Pre-Trained) (a deep learning model for text generation based on available internet data). Convergence conditions during model training can include, but are limited to, an iteration threshold and / or a loss function value threshold. For example, convergence is considered achieved when the number of iterations of the large language model training reaches the iteration threshold. Similarly, convergence is considered achieved when the loss function value between the expected output answer features and the answer features predicted by the large language model reaches the loss function value threshold.

[0063] To ensure the accuracy and reliability of the generated audit model, after S2 above, a first audit model to be evaluated can be selected from the audit models corresponding to the target audit data; test data for evaluating the first audit model can be obtained; feature extraction can be performed on the test data based on large language model technology to obtain the data features of the test data; audit analysis processing can be performed on the problem features in the data features of the test data using the first audit model to obtain the predicted answers corresponding to the problem features in the data features of the test data; the first audit model can be evaluated based on the answer features and predicted answers in the data features of the test data to determine the evaluation result of the first audit model; and the first audit model can be optimized based on the evaluation result. After the audit model is trained, by evaluating the audit model and then optimizing and adjusting the audit model based on the evaluation results, the performance of the audit model can be improved, ensuring its good performance in practical applications.

[0064] When selecting the first audit model to be evaluated, it can be selected sequentially from the audit models corresponding to the target audit data in the order of generation, or arbitrarily, or in other selection methods. That is, in this embodiment of the invention, there is no restriction on the method of selecting the first audit model.

[0065] When evaluating the first audit model, test data is used to evaluate the trained audit model. For example, one or more metrics such as accuracy, precision, recall, and F1 score (generally the harmonic mean of precision and recall) can be used to measure the performance of the audit model. For instance, when evaluating the first audit model based on answer features and predicted answers in the test data, and determining the evaluation result of the first audit model, the accuracy of the first audit model can be evaluated based on the answer features and predicted answers in the test data to determine the accuracy evaluation result; and / or, the precision and recall of the first audit model can be evaluated based on the answer features and predicted answers in the test data to determine the precision and recall evaluation results.

[0066] Assuming the evaluation metrics include accuracy, precision, and recall, the specific steps for evaluating an audit model may include:

[0067] 1. Calculate the accuracy of the output results: Compare the output results of the audit model (such as the predicted answer above) with the actual results (such as the answer features above) to obtain the accuracy of the output results. This accuracy can be used to evaluate the overall performance of the audit model; the higher the accuracy, the better the performance of the first audit model.

[0068] 2. Calculate the precision and recall of the output results: Precision represents the proportion of true positive samples among the positive samples output by the audit model; recall represents the proportion of true positive samples among the samples output as positive samples by the audit model. By considering these two indicators simultaneously, the performance of the audit model can be evaluated more comprehensively.

[0069] 3. Plotting the ROC curve: The ROC (Receiver Operating Characteristic Curve) curve considers both precision and recall of the auditing model, helping to select the optimal classification threshold. The horizontal axis of the ROC curve is FPR (False Positive Rate), and the vertical axis is TPR (True Positive Rate). The better the auditing model's performance, the closer the ROC curve is to the upper left corner. Calculating the AUC (Area Under Curve): The AUC value is the area under the ROC curve; a larger AUC value indicates better auditing model performance.

[0070] The above methods can be used to evaluate the performance of audit models, and then the best audit model can be selected for subsequent application. In actual audit tasks, matching evaluation metrics and methods can also be selected based on the specific task and dataset.

[0071] After evaluating the first audit model, it can be optimized and adjusted. For example, based on the evaluation results, a series of optimization algorithms and techniques can be used to optimize the audit model. Model optimization is an important step in improving model performance, requiring continuous adjustment and improvement of the model's parameters and structure based on the evaluation results. For instance, when optimizing the first audit model based on its evaluation results: the optimization algorithm can be determined based on the evaluation results; and the parameters and / or network structure of the first audit model can be adjusted based on the evaluation results and the optimization algorithm. The optimization algorithm includes at least one of the following: parameter adjustment algorithm, network structure optimization algorithm, and regularization algorithm, which can improve the accuracy and stability of the audit model.

[0072] The optimization and adjustment steps described above may specifically include: First, preparing audit input questions; inputting the audit input questions (specifically, problem features from the data characteristics of the audit input questions) into the trained audit model, using the audit model to analyze the data, and analyzing the output results of the audit model; identifying problems in the output results of the audit model based on the auditor's experience and judgment; analyzing the identified problems to determine their causes and impacts; improving and optimizing the audit model based on the identified problems; and re-evaluating the improved audit model to determine its accuracy and generalization ability, until the output results of the audit model meet the design requirements, thus obtaining the optimized audit model.

[0073] The optimization and adjustment of the above audit model can be mainly achieved by adjusting the audit knowledge vector database to fine-tune the audit model, including:

[0074] a) Adjust search thresholds and sorting rules:

[0075] Adjust the retrieval parameters of the audit knowledge vector database to ensure a higher degree of matching or relevance between the retrieval results and the model output. For example, use a hybrid ranking algorithm that comprehensively considers the probability score generated by the audit model and the similarity score of the retrieval results.

[0076] b) Feedback loops and reinforcement learning:

[0077] Collect user feedback on the audit model's output, and further improve the audit model or retrieval strategy based on the feedback. Furthermore, through reinforcement learning, the audit model adjusts its behavioral strategies based on user satisfaction or business metrics, including when and how to rely on audit knowledge vector database retrieval results.

[0078] c) Integrating multiple resources:

[0079] Key information is extracted and integrated from the search results of the audit knowledge vector database and directly injected into the output of the audit model, or used to guide the audit model to generate more accurate answers again.

[0080] After optimizing and adjusting the audit model, it can be applied to actual audit work. By inputting new data, the audit model can automatically perform audit analysis and provide audit results. In practical applications, it is necessary to ensure the stability and reliability of the model and make necessary adjustments and optimizations according to actual needs.

[0081] In summary, the method for automatically generating audit models based on large language model technology mainly includes data collection, preprocessing, feature extraction, model training, evaluation, optimization, and application. These steps are interconnected, forming a complete automated audit process that provides enterprises with efficient and accurate audit support. In practical applications, it is necessary to select appropriate implementation methods and technical solutions based on specific needs and scenarios.

[0082] When applying the audit model, the target audit model corresponding to the audit input question, the audit knowledge vector database, the interaction layer Prompt, external audit knowledge, and external audit tools can be integrated to construct a private large language audit model oriented towards audit data. For example, the construction steps of the private large language audit model include: building a local audit knowledge vector database and storing it in the Faiss vector database; integrating the Large Language Model (LLM), the audit knowledge vector database, the interaction layer Prompt, external audit knowledge, and external audit tools to construct a private large language audit model application oriented towards audit data; where the large language model includes OpenAI's GPT series models; subsequently, fine-tuning the model based on the data in the audit knowledge vector database; and finally, building a user interface based on the interaction layer Prompt for inputting audit input questions.

[0083] The audit input question in S3 above can be user input or input from an electronic device with question input functionality. In other words, the target object can be a user or an electronic device with question input functionality. In S3, after obtaining the audit input question, a corresponding target audit model can be determined for that audit input question to ensure that the target object completes the required audit work based on the target audit model. It is understood that the matching tools and matching rules used in this embodiment of the invention when matching the audit input question in S3 with the target audit data in the audit knowledge vector database are not limited.

[0084] In one implementation, after S3 above, a question-and-answer request prompt can be generated based on the audit input question; the question-and-answer request prompt can be processed by audit analysis using the target audit model to obtain the answer corresponding to the audit input question.

[0085] Taking a specific implementation example, let's illustrate the application of the SQL (Structured Query Language) model to query the percentage of female newborns in 2023:

[0086] First, information on newborns in 2023 is collected from the `newborns` table, including birth date, gender, etc. The collected newborn information is preprocessed to form an audit corpus, which then constitutes an audit knowledge vector database. Next, large language modeling techniques, such as Transformer and LSTM, are used to extract features from the preprocessed data to extract the information to be collected. Based on the extracted information, the model is trained using the aforementioned model training process. The `count` function is used to calculate the number of female newborns and the total number of newborns, and then the newborn ratio is multiplied by 100 to obtain the percentage of female newborns. Figure 2 As shown, users can first input the audit input question: "Query the percentage of females among newborns in 2023". The audit knowledge vector database can perform a similarity search on it, and form a question-and-answer instruction through the retrieval fragments of the context knowledge base, which will send a question-and-answer prompt to the audit model. Then, the corresponding answer will be output based on the audit model.

[0087] The solution proposed in this invention is highly automated, accurate, scalable, real-time, and easy to maintain and use. Specifically, the entire process of generating the audit model requires no manual intervention; from data collection to practical application, everything is completed automatically, resulting in a high degree of automation and improved work efficiency. Furthermore, due to the use of large language model technology for training and optimization, the generated audit model has high accuracy and stability. Moreover, modules / functions can be added, removed, or adjusted according to actual needs during the audit model generation process, demonstrating strong scalability. Data can be collected and processed in real time during audit model generation, improving the real-time nature and effectiveness of audit work. The model / function structure is clear, with low coupling between modules / functions, making it easy to maintain and use. Therefore, this invention, based on an audit knowledge vector database built for audit knowledge (audit intermediate tables, audit business data, audit documents), and using large language model technology to acquire audit business data and generate audit models, can provide enterprises with efficient, accurate, and reliable audit support, and improve their risk control capabilities and operational efficiency.

[0088] The embodiments of this invention are mainly based on large model technology to automatically generate audit models, and have the following technical effects:

[0089] 1. High Precision and Efficiency: Large language model technology can handle massive amounts of data and complex patterns, improving the accuracy and efficiency of audit models. Compared with traditional machine learning algorithms, audit models based on large language model technology can better cope with large-scale data and complex scenarios, improving the accuracy and reliability of audits.

[0090] 2. Automation and Intelligence: Based on large language model technology, audit models are automatically generated, enabling automatic generation and optimization of audit models, reducing manual intervention and errors. Simultaneously, intelligent feature extraction and anomaly detection are performed using large language model technology, improving the intelligence level of the audit process.

[0091] 3. Explainability and Transparency: The audit model is automatically generated based on large language modeling technology, providing interpretability and helping companies understand and trust the model's decision-making process and results. This improves audit transparency and credibility, enabling companies to better understand and address potential risks.

[0092] 4. Large-scale data processing capability: Large language model technology can process large-scale data, including various types of data such as text, images, and time series. This enables audit models to cover a wider range of data sources, improving the comprehensiveness and accuracy of audits.

[0093] In summary, the solution for automatically generating audit models based on large language model technology has advantages such as high accuracy, high efficiency, automation, intelligence, interpretability, and large-scale data processing capabilities. These advantages help improve enterprises' risk control capabilities and operational efficiency, and provide enterprises with higher quality and more efficient audit services.

[0094] Example 2:

[0095] Based on the same inventive concept, this invention also provides an automatic audit model generation system based on large language model technology, such as... Figure 3 As shown, it includes:

[0096] The database construction module is used to build an audit knowledge vector database based on the acquired target audit data;

[0097] The model generation module is used to perform deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology, and obtain the audit model corresponding to the target audit data.

[0098] The model determination module is used to match the audit input question input by the target object with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question; and to determine the target audit model based on the target type and the audit model corresponding to the target audit data, wherein the target audit model is the model corresponding to the audit input question.

[0099] In one possible implementation, the model generation module is specifically used for:

[0100] Based on the large language model technology, feature extraction is performed on the target audit data in the audit knowledge vector database to obtain the data features of the target audit data. Using the question features in the data features as input and the answer features in the data features as output, the large language model is trained to predict the corresponding answer features based on the input question features. When the trained large language model reaches the convergence condition, the audit model corresponding to the target audit data is obtained.

[0101] In one possible implementation, both the question features and the answer features include at least one of the following features: statistical features, text features, and time series features; the answer features include an audit question prompt field and / or an audit question answer field.

[0102] In one possible implementation, the model generation module is also used for:

[0103] Select the first audit model to be evaluated from the audit models corresponding to the target audit data; and obtain the test data to evaluate the first audit model.

[0104] Based on the large language model technology, feature extraction is performed on the test data to obtain the data features of the test data;

[0105] The first audit model is used to perform audit analysis on the problem features in the test data to obtain the predicted answers corresponding to the problem features in the test data.

[0106] Based on the answer characteristics and predicted answers in the test data, the first audit model is evaluated, and the evaluation result of the first audit model is determined.

[0107] Based on the evaluation results of the first audit model, the first audit model is optimized.

[0108] In one possible implementation, the model generation module is specifically used for:

[0109] Based on the answer characteristics and predicted answers in the test data, the accuracy of the first audit model is evaluated, and the accuracy evaluation result of the first audit model is determined.

[0110] And / or; based on the answer features and predicted answers in the data features of the test data, determine the precision and recall evaluation results of the first audit model.

[0111] In one possible implementation, the model generation module is specifically used for:

[0112] Based on the evaluation results of the first audit model, an optimization algorithm is determined; the optimization algorithm includes at least one of the following: parameter adjustment algorithm, network structure optimization algorithm, and regularization algorithm;

[0113] Based on the evaluation results and optimization algorithm, the parameters and / or network structure of the first audit model are adjusted.

[0114] In one possible implementation, the database building module is specifically used for:

[0115] Based on the acquired target audit data, an audit data corpus is formed; the target audit data includes at least one of the following types: structured report data, unstructured contract text data, business document data, power load data, equipment operating status data, and user electricity consumption data;

[0116] The target audit data in the audit data corpus is vectorized to obtain an audit knowledge vector database.

[0117] In one possible implementation, the database building module is specifically used for:

[0118] Based on the obtained target audit data, the target audit data is preprocessed; the preprocessing includes at least one of the following methods: data cleaning, data integration, data transformation, data reduction, missing value handling, outlier detection and handling, and text segmentation;

[0119] An audit data corpus is formed based on the preprocessed target audit data.

[0120] In one possible implementation, it also includes:

[0121] The model application module is used to generate a question-and-answer prompt based on the audit input questions; the target audit model is then used to input the question-and-answer prompt into the target audit model for audit analysis and processing to obtain the answers corresponding to the audit input questions.

[0122] Example 3:

[0123] Based on the same inventive concept, this invention also provides a computer device, which includes a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement corresponding method flows or corresponding functions, thereby realizing the steps of the automatic audit model generation method based on large language model technology in the above embodiments.

[0124] Example 4:

[0125] Based on the same inventive concept, this invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the steps of the automatic audit model generation method based on large language model technology in the above embodiments.

[0126] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims of the present invention.

Claims

1. A method for automatically generating audit models based on large language model technology, characterized in that, include: Based on the acquired target audit data, construct an audit knowledge vector database; Based on the large language model technology, deep learning is performed on the data features of the target audit data in the audit knowledge vector database to obtain the audit model corresponding to the target audit data. The audit input question input by the target object is matched with the target audit data in the audit knowledge vector database to determine the target type of the audit data corresponding to the audit input question in the target audit data; And based on the target type and the audit model corresponding to the target audit data, a target audit model is determined, wherein the target audit model is the model corresponding to the audit input question.

2. The method as described in claim 1, characterized in that, The method of performing deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology to obtain the audit model corresponding to the target audit data includes: Based on the large language model technology, feature extraction is performed on the target audit data in the audit knowledge vector database to obtain the data features of the target audit data; Using the question features in the data features as input and the answer features in the data features as output, a large language model is trained to predict the corresponding answer features based on the input question features. When the trained large language model reaches the convergence condition, the audit model corresponding to the target audit data is obtained.

3. The method as described in claim 2, characterized in that, Both the question features and the answer features include at least one of the following features: statistical features, text features, and time series features; The answer features include an audit question prompt field and / or an audit question answer field.

4. The method as described in claim 2 or 3, characterized in that, After obtaining the audit model corresponding to the target audit data by performing deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology, the method further includes: Select the first audit model to be evaluated from the audit models corresponding to the target audit data; and obtain test data for evaluating the first audit model. Based on the large language model technology, feature extraction is performed on the test data to obtain the data features of the test data; The first audit model is used to perform audit analysis on the problem features in the data features of the test data to obtain the predicted answer corresponding to the problem features in the data features of the test data. Based on the answer features in the data features of the test data and the predicted answer, the first audit model is evaluated, and the evaluation result of the first audit model is determined; Based on the evaluation results of the first audit model, the first audit model is optimized.

5. The method as described in claim 4, characterized in that, The step of evaluating the first audit model based on the answer features in the test data and the predicted answer, and determining the evaluation result of the first audit model, includes: Based on the answer features in the data features of the test data and the predicted answer, the accuracy of the first audit model is evaluated, and the accuracy evaluation result of the first audit model is determined. And / or; based on the answer features in the data features of the test data and the predicted answer, determine the precision and recall evaluation results of the first audit model.

6. The method as described in claim 4, characterized in that, The step of optimizing the first audit model based on the evaluation results of the first audit model includes: Based on the evaluation results of the first audit model, an optimization algorithm is determined; the optimization algorithm includes at least one of the following: parameter adjustment algorithm, network structure optimization algorithm, and regularization algorithm; Based on the evaluation results and the optimization algorithm, the parameters and / or network structure of the first audit model are adjusted.

7. The method as described in claim 1 or 2, characterized in that, The step of constructing an audit knowledge vector database based on the acquired target audit data includes: Based on the acquired target audit data, an audit data corpus is formed; the target audit data includes at least one of the following types: structured report data, unstructured contract text data, business document data, power load data, equipment operating status data, and user electricity consumption data; The target audit data in the audit data corpus is vectorized to obtain an audit knowledge vector database.

8. The method as described in claim 7, characterized in that, The process of forming an audit data corpus based on the acquired target audit data includes: Based on the obtained target audit data, the target audit data is preprocessed; the preprocessing includes at least one of the following processing methods: data cleaning, data integration, data transformation, data reduction, missing value handling, outlier detection and handling, and text segmentation; An audit data corpus is formed based on the preprocessed target audit data.

9. The method as described in claim 1, characterized in that, After determining the target audit model based on the target type and the audit model corresponding to the target audit data, the process further includes: Based on the audit input questions, a question-and-answer prompt is generated; The target audit model is used to perform audit analysis on the question-and-answer request prompt to obtain the answer corresponding to the audit input question.

10. An automatic audit model generation system based on large language model technology, characterized in that, include: The database construction module is used to build an audit knowledge vector database based on the acquired target audit data; The model generation module is used to perform deep learning on the data features of the target audit data in the audit knowledge vector database based on large language model technology, so as to obtain the audit model corresponding to the target audit data. The model determination module is used to match the audit input question input by the target object with the target audit data in the audit knowledge vector database, and determine the target type of the audit data in the target audit data that corresponds to the audit input question; And based on the target type and the audit model corresponding to the target audit data, a target audit model is determined, wherein the target audit model is the model corresponding to the audit input question.

Images