Automatic auditing method for engineering project based on large model

Abstract

The invention provides an automatic auditing method for an engineering project based on a large model, which automatically extracts knowledge rules and auditing relationships in a policy file according to a preset prompt library through an AI large model which is finely adjusted based on engineering auditing domain knowledge, and combines an auditing unit set which contains an auditing calculation process and calculation description and is introduced by the invention, so that the auditing efficiency is improved. According to the technical scheme, a scheduling algorithm model based on an auditing calculation unit is generated by understanding and using an AI large model, so that an engineering project auditing doubtful point is obtained, and end-to-end automatic auditing from a policy file end to an auditing decision end is realized. Particularly, the accuracy of automatically generating the auditing ternary pair and the auditing model is improved by applying the circulation reverse verification technology of the AI large model twice, and the reliability of the auditing result is improved while automatic auditing of the engineering project is achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence applications, and in particular to an automatic audit method for engineering projects based on a large model. Background Art

[0002] With the call of the digital age and the promotion of technological progress, the volume of corporate audit data is growing exponentially. Coupled with the intensified market competition and the increase in corporate compliance requirements, the internal audit work of large enterprises is facing severe challenges. In this context, the transformation of digital automated auditing is imminent. Among them, automatic auditing of engineering projects is an important application scenario in the field of digital auditing, which aims to automatically and accurately extract engineering audit doubts from existing engineering project data, and assist auditors in further investigation, verification and evidence collection.

[0003] At present, in the field of digital auditing, the two representative technologies are knowledge graph technology and AI big model technology. Both have certain applications in specific business scenarios such as audit knowledge question and answer, audit knowledge recommendation, audit abnormality warning, audit report generation, and intelligent audit system construction. However, none of the existing technical routes can fully meet the actual needs of automatic auditing of engineering projects. Specifically, although the knowledge graph can query, analyze and recommend based on a clear semantic structure, it lacks the ability to learn independently, cannot generate new content, and has a weak understanding of natural language text; in addition, it performs poorly in terms of versatility, especially in the field of engineering project auditing. It cannot understand all the knowledge in policy documents like auditors, and usually requires the collaboration of auditors to complete specific audit work. In terms of AI big models, it has a rich knowledge reserve, has the ability to follow instructions, can handle multiple tasks, and automatically generate diverse data, but its decision-making process is black box and difficult to explain, and depends on verification by manual or other systems; in addition, since the audit work involves massive data in the digital platform, the generation function of the AI ​​big model cannot fully realize the connection with the existing audit database, resulting in the inability to directly generate audit decisions from the reasoning results of the AI ​​big model. In summary, the two mainstream technical applications in the current digital auditing field have some drawbacks and cannot truly realize the autonomous auditing process from the data segment to the decision-making end. Summary of the invention

[0004] In order to make up for the deficiencies of the prior art, the present invention provides an automatic audit method for engineering projects based on a large model.

[0005] The present invention is implemented by the following technical solution: an automatic audit method for engineering projects based on a large model, specifically comprising the following steps: Step S1: Based on the internal digital platform of the enterprise, create an audit unit set S for audit work to extract specific audit data respectively; Step S2: Select the document policies used in the project audit as the audit knowledge data, perform preprocessing, and use the obtained data set to fine-tune the AI ​​model; The document policies used in project audits are selected as audit knowledge data, including four categories: audit policies, laws and regulations, internal enterprise regulations, and project audit cases. The collected audit knowledge data is processed through standard preprocessing steps including textualization, cleaning, deduplication, and denoising. The preprocessed text data is converted into a text format suitable for AI large model training to obtain the fine-tuning training data set T. Next, the LORA algorithm is used to fine-tune the AI ​​large model. That is, the inherent low-rank characteristics of the large model are used to simulate the effect of full parameter fine-tuning by adding low-rank matrices to achieve efficient model adaptive adjustment, ensuring that the professionalism in the field of engineering auditing is enhanced while retaining general intelligence. Step S3: Take the text data of the audit policy of the target project as the input of the AI ​​big model, use the prompt word technology to extract the audit entities and audit relations, and form a candidate audit triple set AT_CAND; obtain the specific audit policy text data of the target project, and perform necessary formatting and preprocessing including: noise removal, text standardization, word segmentation, and stop word removal, and input the preprocessed audit policy text data into the AI ​​big model; then, use the prompt word technology to let the AI ​​big model automatically extract a series of audit entities and attached audit relations; Step S4: Send the obtained triples in the candidate audit triple set AT_CAND to the AI ​​big model for reverse verification, remove inaccurate audit triples, and retain the audit triples that have passed the verification to form a set AT_VALID; Step S5: Based on the information of each audit unit in the audit unit set S, the audit triples of the set AT_VALID are used to prompt the AI ​​large model to obtain the candidate audit model set AM_CAND; Step S6: Send the audit models in the candidate audit model set AM_CAND to the AI ​​big model for reverse verification, remove inaccurate audit models, and retain the audit models that have passed the verification to form a set AM_VALID; Step S7, run the audit model code in the audit model set AM_VALID to find the audit doubts of the engineering project; for the audit model set AM_VALID formed in step S6, run the audit model codes it contains one by one on the internal digital platform of the enterprise, and determine the specific audit doubts of the engineering project based on the returned results.

[0006] As a preferred solution, step S1 specifically includes the following steps: for the project audit, create an audit unit set S, which includes multiple audit units with parameter lists, each audit unit is used to automatically extract eight types of original audit data from the database of the digital platform, including funding, progress, contract, quality, personnel information, project information, equipment information, and material. The created audit unit AU i It can be expressed as AU i ={Unit DescriptionAU info , parameter list AU param , executable operation behavior AU opt}, where the audit unit describes the AU info Includes description of audit unit functions, parameters, and return types, which are used to be recognized, understood, and used by the AI ​​big model; parameter list AU param Used to specify the specific parameters of the operation performed by the audit unit; executable operation behavior AU opt It is a sequence of operations that can be performed by the audit digital platform to extract specific audit data based on a list of parameters.

[0007] As a preferred solution, fine-tuning the AI ​​large model using the LORA algorithm in step S2 specifically includes the following steps: Step S2-1, keep the parameter weight matrix W of the pre-trained AI large model unchanged; Step S2-2: Introduce low-rank weight matrix A m,p , B p,n , where m, p, and n represent the rows of matrix A, the columns of matrix A, and the columns of matrix B, respectively. The product of these two matrices is C m,n It will be used to simulate the parameter weight changes of full parameter fine-tuning; in particular, p is much smaller than m, which means that most of the information in the matrix can be represented by a few main components, which helps to reduce the complexity of model parameter learning.

[0008] Step S2-3, based on the training data set T, the low-rank weight matrix A is conditioned by the gradient descent method. m,p , B p,n After training, the trained weight matrix A is obtained after the number of iterations. * m,p , B * p,n ; Step S2-4, merge weights; * m,p , B * p,n Multiply them together to get the weight update matrix C of the AI ​​large model parameters * m,n , and update the matrix C* m,n Merge into the parameters of the pre-trained model to obtain the fine-tuned parameter weight matrix W * m,n ; The specific formula for merging weights is: W * m,n =W m,n +C * m,n .

[0009] Furthermore, the number of iterations epoch in step S2-4 is set to 200-500.

[0010] As a preferred solution, step S3 specifically includes the following steps: designing an audit relationship prompt vocabulary P={prompt1, prompt2,...} to prompt the AI ​​big model to automatically identify and extract relevant audit entities and their associated audit relationships to generate structured audit triples.

[0011] As a preferred solution, step S4 specifically includes the following steps: the element AT_CAND in the candidate audit triple set AT_CAND formed in step S3 is i , and then send them to the AI ​​big model for reverse verification through the prompt word technology, asking whether the triple accurately reflects the rule facts in the policy document, and let the big model return a result from the set {accurate, inaccurate}; if the result is "accurate", AT_CAND i Add the verified triples to the set AT_VALID; if it returns "inaccurate", let it return a similar revised audit triple and verify it again. If the above steps are repeated more than 3 times, discard the triple; otherwise, add the verified triples to the set AT_VALID.

[0012] As a preferred solution, step S5 specifically includes the following steps: sending the information of each audit unit of the accounting unit set S to the AI ​​big model for learning; then, according to the audit triple AT_VALID of the set AT_VALID i Prompt the AI ​​model so that it can reasonably call the audit unit in the set S and verify the result according to the audit triple AT_VALID i Design executable audit code to form a candidate audit model AM_CAND that can detect audit doubts i , and added to the set AM_CAND.

[0013] As a preferred solution, step S6 specifically includes the following steps: the audit model AM_CAND in the candidate audit model set AM_CAND formed in step S5 is i, through the prompt word technology, it is sent to the AI ​​big model for reverse verification, asking whether the audit model accurately reflects the relationship described by the corresponding audit triple, and asking the big model to return a result from the set {accurate, inaccurate}; if it returns "accurate", AM_CAND i Add the verified audit model set AM_VALID; if it returns "inaccurate", let it return the revised audit model and verify it again; if the above steps are repeated more than 3 times, discard the audit model; otherwise, add the verified audit model to the set AM_VALID.

[0014] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects compared with the prior art: (1) The AI ​​big model based on the field of engineering project audit is used to automatically extract the audit rule ternary pairs from the policy documents, and combined with the audit unit set, the audit model that can be automatically run by the digital platform is automatically generated, and used to discover audit doubts, thereby realizing the "end-to-end" automated audit of engineering projects and improving the automation level of audit work. (2) By introducing the audit unit set containing the audit calculation process and calculation description, it can be understood, analyzed and used by the AI ​​big model, so that the generated results have the ability to connect with the existing digital platform, solving the problem that the AI ​​big model cannot connect with the digital audit platform database, and providing reference and reference for the existing automatic audit scheme based on the AI ​​big model and the knowledge graph. (3) In addition, by introducing the cyclic reverse verification step twice in the process of generating the AI ​​big model, the accuracy of the audit results is improved, the quality of the audit work is guaranteed, and the technical guarantee is provided for the actual promotion and application of the present invention. In summary, the method of the present invention has certain promotion and application value in the field of digital audit.

[0015] Additional aspects and advantages of the present invention will become apparent from the following description or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 It is the work flow chart of the present invention. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0018] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0019] Combine the following Figure 1 The automatic audit method for engineering projects based on a large model according to an embodiment of the present invention is specifically described.

[0020] like Figure 1 As shown, the present invention proposes an automatic audit method for engineering projects based on a large model, introduces an audit unit set, and realizes the docking of the generative model with the existing digital platform; introduces two-cycle reverse verification technology in the steps to ensure the accuracy of the audit model generated by the AI ​​large model; and realizes the "end-to-end" automatic audit of engineering projects from the "policy document end" to the "audit decision end". Specifically, it includes the following steps: Step S1: Based on the internal digital platform of the enterprise, create an audit unit set S for audit work to extract specific audit data respectively; the internal digital platform of the enterprise refers to an information technology-based software system that integrates data, applications, users, and businesses in the enterprise, and can run the audit model in the present invention and generate results.

[0021] For project audit, create an audit unit set S, which contains multiple audit units with parameter lists. Each audit unit is used to automatically extract eight types of original audit data from the database of the digital platform, including funding, progress, contract, quality, personnel information, project information, equipment information, and materials. These eight types of data cover the data involved in the project audit and can be used as the specific audit basis and evidence for the target project. Furthermore, the created audit unit AU i It can be expressed as AU i ={Unit DescriptionAU info , parameter list AU param , executable operation behavior AU opt}, where the audit unit describes the AU info Includes description of audit unit functions, parameters, and return types, which are used to be recognized, understood, and used by the AI ​​big model; parameter list AU param Used to specify the specific parameters of the operation performed by the audit unit; executable operation behavior AU opt It is a sequence of operations that can be performed by the audit digital platform to extract specific audit data based on a list of parameters.

[0022] Step S2: Select the document policies used in the project audit as the audit knowledge data, perform preprocessing, and use the obtained data set to fine-tune the AI ​​model; The document policies used in project audits are selected as audit knowledge data, including four categories: audit policies, laws and regulations, internal enterprise regulations, and project audit cases. The collected audit knowledge data is processed through standard preprocessing steps including textualization, cleaning, deduplication, and denoising. The preprocessed text data is converted into a text format suitable for AI large model training to obtain the fine-tuning training data set T. Next, the LORA algorithm is used to fine-tune the AI ​​large model. That is, the inherent low-rank characteristics of the large model are used to simulate the effect of full parameter fine-tuning by adding low-rank matrices to achieve efficient model adaptive adjustment, ensuring that the professionalism in the field of engineering auditing is enhanced while retaining general intelligence. The specific fine-tuning steps are as follows: Step S2-1, keep the parameter weight matrix W of the pre-trained AI large model unchanged; Step S2-2: Introduce low-rank weight matrix A m,p , B p,n , where m, p, and n represent the rows of matrix A, the columns of matrix A, and the columns of matrix B, respectively. The product of these two matrices is C m,n It will be used to simulate the parameter weight changes of full parameter fine-tuning; in particular, p is much smaller than m, which means that most of the information in the matrix can be represented by a few main components, which helps to reduce the complexity of model parameter learning.

[0023] Step S2-3, based on the training data set T, the low-rank weight matrix A is conditioned by the gradient descent method. m,p , B p,n After training, after the number of iterations (the number of iterations epoch is set to 200~500), the trained weight matrix A is obtained * m,p , B * p,n ; Step S2-4, merge weights; * m,p , B * p,n Multiply them together to get the weight update matrix C of the AI ​​large model parameters * m,n , and update the matrix C * m,n Merge into the parameters of the pre-trained model to obtain the fine-tuned parameter weight matrix W * m,n ; The specific formula for merging weights is: W * m,n =W m,n +C * m,n In the above steps, through the low-rank matrix A m,p , B p,n , only a small amount of computing resources are needed to fine-tune the weight matrix of the pre-trained AI large model, so that it not only retains the knowledge of the pre-trained model, but also can adapt to the needs of the specific tasks faced by the present invention.

[0024] Step S3: Input the text data of the audit policy of the target project as the AI ​​big model, and use the prompt word technology to extract the audit entities and audit relationships to form a candidate audit triple set AT_CAND; obtain the specific audit policy text data of the target project, and perform necessary formatting and preprocessing including: noise removal, text standardization, word segmentation, and stop word removal. Input the preprocessed audit policy text data into the AI ​​big model; then, use the prompt word technology to let the AI ​​big model automatically extract a series of audit entities and attached audit relationships; The audit relationship prompt word library P = {prompt1, prompt2, ...} is designed to prompt the AI ​​big model to automatically identify and extract relevant audit entities and their associated audit relationships to generate structured audit triples. The prompt words in the audit relationship prompt word library P are created through multiple practices of auditors, with the aim of fully extracting audit relationships containing various audit elements from policy documents. A possible example based on an engineering project is shown in the table below.

[0025] Step S4: Send the obtained triples in the candidate audit triple set AT_CAND to the AI ​​big model for reverse verification, remove inaccurate audit triples, and retain the audit triples that have passed the verification to form a set AT_VALID; Research has shown that large AI models can perform explainable self-verification of their own reasoning conclusions, solving the problem of deviation from the correct thinking process in large model thinking chain technology, thereby improving reasoning performance.

[0026] The element AT_CAND in the candidate audit triple set AT_CAND formed in step S3 is i , and then send them to the AI ​​big model for reverse verification through the prompt word technology, asking whether the triple accurately reflects the rule facts in the policy document, and let the big model return a result from the set {accurate, inaccurate}; if the result is "accurate", AT_CAND i Add the verified triples to the set AT_VALID; if it returns "inaccurate", let it return a similar revised audit triple and verify it again. If the above steps are repeated more than 3 times, discard the triple; otherwise, add the verified triples to the set AT_VALID.

[0027] Through the reverse verification process of this step, a relatively accurate audit triple set AT_VALID is obtained. The triples in this set provide a reasoning basis for guiding the AI ​​big model to generate a specific audit model.

[0028] Step S5: Based on the information of each audit unit in the audit unit set S, the audit triples of the set AT_VALID are used to prompt the AI ​​large model to obtain the candidate audit model set AM_CAND; In order to make the AI ​​big model understand the purpose and scheduling method of the audit unit and use it to extract specific audit data, the information of each audit unit of the audit unit set S is sent to the AI ​​big model for learning; then, according to the audit triple AT_VALID of the set AT_VALID i Prompt the AI ​​model so that it can reasonably call the audit unit in the set S and verify the result according to the audit triple AT_VALID i Design executable audit code to form a candidate audit model AM_CAND that can detect audit doubts i , and added to the set AM_CAND.

[0029] In this step, the audit triple AT_VALID i ={indirect expenses, not more than, 30%} is used as an example, the AI ​​big model uses the preset prompt words "according to the audit triples of the project, use the functions of each audit unit in the audit unit set to generate an executable program", and the generated program logic is as follows: a Calculate the total cost of the project using AU a Calculate the indirect expenses of the project. If the ratio of indirect expenses to total expenses is greater than 30%, an audit doubt is returned. a Represents the audit unit that calculates the total cost of the target project, which returns the total cost of the specified project from the database; AU b The audit unit that represents the indirect expenses of the target project returns the indirect expenses of the specified project from the database. By running the above audit model program on the digital platform, the audit doubts of the engineering project can be automatically discovered.

[0030] Step S6: Send the audit models in the candidate audit model set AM_CAND to the AI ​​big model for reverse verification, remove inaccurate audit models, and retain the audit models that have passed the verification to form a set AM_VALID; The audit model AM_CAND in the candidate audit model set AM_CAND formed in step S5 is i, through the prompt word technology, it is sent to the AI ​​big model for reverse verification, asking whether the audit model accurately reflects the relationship described by the corresponding audit triple, and asking the big model to return a result from the set {accurate, inaccurate}; if it returns "accurate", AM_CAND i Add the verified audit model set AM_VALID; if it returns "inaccurate", let it return the revised audit model and verify it again; if the above steps are repeated more than 3 times, discard the audit model; otherwise, add the verified audit model to the set AM_VALID.

[0031] Through the reverse verification process of this step, an accurate audit model set AM_VALID is obtained. The audit model codes in this set can help auditors find audit doubts in engineering projects.

[0032] Step S7, run the audit model code in the audit model set AM_VALID to find audit doubts about the engineering project; for the audit model set AM_VALID formed in step S6, run the audit model codes it contains one by one on the internal digital platform of the enterprise, determine the specific audit doubts about the engineering project according to the returned results, guide the auditors to pay attention to these specific doubts, conduct in-depth investigation, verification and evidence collection, and realize the "end-to-end" automatic audit method for engineering projects from the "policy document end" to the "audit doubt end".

[0033] In the description of the present invention, the term "plurality" refers to two or more than two. Unless otherwise clearly defined, the orientation or positional relationship indicated by the terms "upper" and "lower" is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention; the terms "connection", "installation", "fixation", etc. should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0034] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "specific embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0035] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An automatic audit method for engineering projects based on a large model, characterized in that: The specific steps include: Step S1: Based on the internal digital platform of the enterprise, create an audit unit set S for audit work to extract specific audit data respectively; Step S2: Select the document policies used in the project audit as the audit knowledge data, perform preprocessing, and use the obtained data set to fine-tune the AI ​​model; The document policies used in project audits are selected as audit knowledge data, including four categories: audit policies, laws and regulations, internal enterprise regulations, and project audit cases. The collected audit knowledge data is processed through standard preprocessing steps including textualization, cleaning, deduplication, and denoising. The preprocessed text data is converted into a text format suitable for AI large model training to obtain the fine-tuning training data set T. Next, the LORA algorithm is used to fine-tune the AI ​​large model. That is, the inherent low-rank characteristics of the large model are used to simulate the effect of full parameter fine-tuning by adding low-rank matrices to achieve efficient model adaptive adjustment, ensuring that the professionalism in the field of engineering auditing is enhanced while retaining general intelligence. Step S3: Use the text data of the audit policy of the target project as the input of the AI ​​big model, and use the prompt word technology to extract the audit entities and audit relations to form a candidate audit triple set AT_CAND; Obtain the specific audit policy text data of the target project, and format and preprocess it, including: noise removal, text standardization, word segmentation, and stop word removal; and input the preprocessed audit policy text data into the AI ​​big model; then, through the prompt word technology, let the AI ​​big model automatically extract a series of audit entities and attached audit relationships; Step S4: Send the obtained triples in the candidate audit triple set AT_CAND to the AI ​​big model for reverse verification, remove inaccurate audit triples, and retain the audit triples that have passed the verification to form a set AT_VALID; Step S5: Based on the information of each audit unit in the audit unit set S, the audit triples of the set AT_VALID are used to prompt the AI ​​large model to obtain the candidate audit model set AM_CAND; Step S6: Send the audit models in the candidate audit model set AM_CAND to the AI ​​big model for reverse verification, remove inaccurate audit models, and retain the audit models that have passed the verification to form a set AM_VALID; Step S7, run the audit model code in the audit model set AM_VALID to find the audit doubts of the engineering project; for the audit model set AM_VALID formed in step S6, run the audit model codes it contains one by one on the internal digital platform of the enterprise, and determine the specific audit doubts of the engineering project based on the returned results.

2. According to the automatic audit method of engineering projects based on large models as described in claim 1, it is characterized in that: The step S1 specifically includes the following steps: for the project audit, an audit unit set S is created, which includes multiple audit units with parameter lists, each audit unit is used to automatically extract eight types of original audit data from the database of the digital platform, including funding, progress, contract, quality, personnel information, project information, equipment information, and material. The created audit unit AU i It can be expressed as AU i ={Unit DescriptionAU info , parameter list AU param , executable operation behavior AU opt }, where the audit unit describes the AU info Includes description of audit unit functions, parameters, and return types, which are used to be recognized, understood, and used by the AI ​​big model; parameter list AU param Used to specify the specific parameters of the operation performed by the audit unit; executable operation behavior AU opt It is a sequence of operations that can be performed by the audit digital platform to extract specific audit data based on a list of parameters.

3. According to the automatic audit method of engineering projects based on large models as described in claim 1, it is characterized in that: In step S2, fine-tuning the AI ​​large model using the LORA algorithm specifically includes the following steps: Step S2-1, keep the parameter weight matrix W of the pre-trained AI large model unchanged; Step S2-2: Introduce low-rank weight matrix A m,p , B p,n , where m, p, and n represent the rows of matrix A, the columns of matrix A, and the columns of matrix B, respectively. The product of these two matrices is C m,n It will be used to simulate the parameter weight changes of full parameter fine-tuning; in particular, p is much smaller than m, which means that most of the information in the matrix can be represented by a few main components, which helps to reduce the complexity of model parameter learning; Step S2-3, based on the training data set T, the low-rank weight matrix A is conditioned by the gradient descent method. m,p , B p,n After training, the trained weight matrix A is obtained after the number of iterations. * m,p , B * p,n ; Step S2-4, merge weights; * m,p , B * p,n Multiply them together to get the weight update matrix C of the AI ​​large model parameters * m,n , and update the matrix C * m,n Merge into the parameters of the pre-trained model to obtain the fine-tuned parameter weight matrix W * m,n ; The specific formula for merging weights is: W * m,n =W m,n +C * m,n 。 4. According to the automatic audit method of a large model-based engineering project according to claim 2, it is characterized in that: The number of iterations epoch in step S2-4 is set to 200-500.

5. The automatic audit method for engineering projects based on a large model according to claim 1 is characterized in that: The step S3 specifically includes the following steps: designing an audit relationship prompt vocabulary P={prompt1, prompt2,...} to prompt the AI ​​big model to automatically identify and extract relevant audit entities and their associated audit relationships to generate structured audit triples.

6. The automatic audit method for engineering projects based on a large model according to claim 1 is characterized in that: The step S4 specifically includes the following steps: adding the element AT_CAND in the candidate audit triple set AT_CAND formed in step S3 i , and then send them to the AI ​​big model for reverse verification through the prompt word technology, asking whether the triple accurately reflects the rule facts in the policy document, and let the big model return a result from the set {accurate, inaccurate}; if it returns "accurate", AT_CAND i Add the verified triples to the set AT_VALID; if it returns "inaccurate", let it return a similar revised audit triple and verify it again; if the above steps are repeated more than 3 times, discard the triple; otherwise, add the verified triples to the set AT_VALID.

7. The automatic audit method for engineering projects based on a large model according to claim 1 is characterized in that: The step S5 specifically includes the following steps: sending the information of each audit unit of the accounting unit set S to the AI ​​big model for learning; then, according to the audit triple AT_VALID of the set AT_VALID i Prompt the AI ​​model so that it can reasonably call the audit unit in the set S and verify the result according to the audit triple AT_VALID i Design executable audit code to form a candidate audit model AM_CAND that can detect audit doubts i , and added to the set AM_CAND.

8. The automatic audit method for engineering projects based on a large model according to claim 1 is characterized in that: The step S6 specifically includes the following steps: the audit model AM_CAND in the candidate audit model set AM_CAND formed in step S5 is i , through the prompt word technology, it is sent to the AI ​​big model for reverse verification, asking whether the audit model accurately reflects the relationship described by the corresponding audit triple, and asking the big model to return a result from the set {accurate, inaccurate}; if it returns "accurate", AM_CAND i Add the verified audit model set AM_VALID; if it returns "inaccurate", let it return the revised audit model and verify it again; if the above steps are repeated more than 3 times, discard the audit model; otherwise, add the verified audit model to the set AM_VALID.

Images