Intelligent auditing method and system for contracts and projects in hospital

Abstract

The invention discloses a hospital internal contract and project intelligent auditing method and system, and the method comprises the steps: inputting to-be-examined files into a multi-agent cooperation network model of an auditing vertical domain large model, carrying out the document format conversion of a plurality of to-be-examined files through a perception agent, and obtaining a plurality of text files; performing intention recognition according to the to-be-examined file, the audit scene and the audit type to obtain an audit rule list, and performing extraction, document slicing and assembly on the to-be-examined content according to the audit rule list to obtain an atomic file corresponding to each audit rule; performing risk judgment on the atomic file corresponding to each auditing rule by using the analysis agent to obtain a corresponding atomic result, performing summarization by using the summarization agent to obtain a risk prompt list, and performing summarization based on the risk prompt list to obtain auditing suggestions. According to the method, the files in different formats are processed through stepped task decomposition, and the processing results of the intelligent agents are integrated, so that the efficiency of the audit results is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent recognition model applications, and particularly to an intelligent audit method, system, device and medium for hospital internal contracts. Background Art

[0002] The operation of hospitals is becoming increasingly complex and diversified. Internal audit, as an important means for hospitals to strengthen internal management and supervision, is becoming increasingly important, especially for the compliance audit of contracts and procurement. The existing in-hospital contract and procurement audits involve two scenarios: in-process audit, generally other functional departments of the hospital submit contracts for audit and submit contract attachments, including requirement documents, project approval documents, bidding documents, tender documents, winning bid notices, etc. Auditors conduct compliance audits based on the submitted materials; post-event special audit, auditing the compliance of the entire process from project approval, procurement to acceptance, involving the verification and comparison of up to 10 - 20 documents. Auditors often have to manually check one by one, and the mechanical and repetitive work takes up a large amount of time and energy of auditors. However, because the audit documents are all unstructured long text segments, and even have different formats (tables, pictures, pdf, etc.), and the writing styles are diverse, the previous technologies of RPA robots and deep learning-based methods are difficult to achieve good results. Summary of the Invention

[0003] To solve the above technical problems, embodiments of the present invention provide an intelligent audit method, system, device and medium for hospital internal contracts, so as to solve the technical problem that the existing technology cannot comprehensively analyze and understand hospital internal contracts in various formats, resulting in low audit efficiency.

[0004] The first aspect of the embodiments of the present invention provides an intelligent audit method for hospital internal contracts, and the method includes:

[0005] Obtain multiple files to be reviewed, where the files to be reviewed are contract files or project files;

[0006] Input the files to be reviewed into a multi-agent collaboration network model based on an audit vertical domain large model, and use a perception agent to perform document format conversion on the multiple files to be reviewed to obtain multiple text files, and perform intent recognition on the files to be reviewed to obtain an audit rule list. According to the audit rule list, determine the positions of the content to be reviewed corresponding to each audit rule, and based on the positions of each content to be reviewed, extract and slice the content to be reviewed to obtain multiple sliced documents, and assemble the multiple sliced documents to obtain an atomic file corresponding to each audit rule;

[0007] After using the analysis agent to judge the risks of the atomic files corresponding to each audit rule and obtaining the atomic results corresponding to each audit rule, the summary agent is used to summarize the individual atomic results to obtain a risk prompt list. Based on the risk prompt list, the atomic results with risk points are summarized to obtain the audit opinions.

[0008] In a possible implementation manner of the first aspect, the perception agent is used to perform document format conversion on multiple files to be reviewed, obtaining multiple text files, including:

[0009] Using the pdf parsing and conversion technology to perform image conversion on each page of multiple files to be reviewed, obtaining the corresponding converted files;

[0010] Based on each converted file, using multi-modal OCR technology to locate and extract computer text and handwritten text, obtaining the corresponding extraction results and text coordinates, and embedding the corresponding extraction results into the corresponding file to be reviewed, obtaining the text file corresponding to each file to be reviewed.

[0011] In a possible implementation manner of the first aspect, after obtaining multiple text files, it further includes:

[0012] Based on punctuation marks and grammar rules, perform intelligent sentence splitting on each text file to obtain multiple segmented text files;

[0013] Correct the incorrect and missing characters in each sentence of each segmented text file and assign a positioning identifier.

[0014] In a possible implementation manner of the first aspect, assembling multiple sliced documents to obtain the atomic file corresponding to each audit rule, including:

[0015] Assemble the multiple sliced documents corresponding to each audit rule to obtain the assembled file corresponding to each audit rule;

[0016] Obtain the atomic file corresponding to each audit rule according to the audit rule number, text file number, positioning identifier, assembled file, and the content of the next task.

[0017] In a possible implementation manner of the first aspect, after obtaining the audit opinions, it further includes:

[0018] Use the execution agent to encapsulate the risk prompt list into the first document and the audit opinions into the second document;

[0019] Based on the first document and the second document, use the numbers and positioning identifiers of each file to be reviewed to annotate the files to be reviewed, obtaining the annotation results.

[0020] In a possible implementation of the first aspect, the audit vertical domain large model is obtained by pre-training a general large model using audit domain text data, including:

[0021] Obtain audit domain text data, where the audit domain text data includes audit domain laws and regulations text data, audit cases, and audit rules;

[0022] Use NLP technology to tokenize and perform entity recognition on the audit domain text data to obtain key information, and construct a knowledge graph based on the entities and the relationships between the entities in the key information;

[0023] Select a general large model as the base model, insert the entity vectors in the knowledge graph as key values into the base model, and use an adapter structure to isolate the knowledge parameters, dynamically adjust the learning rate and mask ratio to obtain an initial audit vertical domain large model;

[0024] According to the business requirements and model optimization objectives of the audit vertical domain, divide them into multiple tasks, construct the supervision data and instruction sets for each task, and based on the supervision data and instruction sets of each task, use reinforcement learning to design a reward model and optimize the strategy based on the PPO algorithm to optimize the initial audit vertical domain large model to obtain the audit vertical domain large model.

[0025] In a possible implementation of the first aspect, each agent in the multi-agent collaboration network model is obtained by fine-tuning the audit vertical domain large model, including:

[0026] Based on the audit vertical domain large model, a multi-agent collaboration network model is constructed by combining prompt engineering and engineering functions based on Python. The multi-agent collaboration network model includes a perception agent, an analysis agent, a summary agent, and an execution agent.

[0027] The second aspect of the embodiments of the present invention provides a hospital internal contract and project intelligent audit system, and the system includes:

[0028] An acquisition module, configured to, in response to a review operation, determine an audit scenario and an audit type, and obtain multiple documents to be reviewed according to the audit scenario and the audit type, where the documents to be reviewed are contract documents or project documents;

[0029] A conversion module, configured to input a document to be reviewed into a multi-agent collaboration network model based on an audit vertical domain large model, and apply a perception agent to perform document format conversion on multiple documents to be reviewed to obtain multiple text files, and perform intention recognition on the documents to be reviewed to obtain an audit rule list. According to the audit rule list, determine the positions of the content to be reviewed corresponding to each audit rule, extract and slice the content to be reviewed based on the positions of each content to be reviewed to obtain multiple sliced documents, and assemble the multiple sliced documents to obtain an atomic file corresponding to each audit rule;

[0030] A summarization module, configured to use an analysis agent to perform risk judgment on the atomic file corresponding to each audit rule to obtain an atomic result corresponding to each audit rule, and then use a summary agent to summarize each atomic result to obtain a risk prompt list, and summarize the atomic results with risk points based on the risk prompt list to obtain an audit opinion.

[0031] A third aspect of the embodiments of the present invention provides a computer device, including:

[0032] A memory, configured to store a computer program;

[0033] A processor, configured to implement the steps of the hospital internal contract and project intelligent audit method as in the first aspect when executing the computer program.

[0034] A fourth aspect of the embodiments of the present invention provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the hospital internal contract and project intelligent audit method as in the first aspect are implemented.

[0035] The technical solution of the present invention has the following advantages:

[0036] The intelligent audit method for hospital internal contracts provided by the embodiments of the present invention determines the audit scenario and audit type in response to a review operation, obtains multiple files to be reviewed according to the audit scenario and audit type, inputs the files to be reviewed into a multi-agent collaboration network model based on an audit vertical domain large model, uses a perception agent to perform document format conversion on the multiple files to be reviewed to obtain multiple text files, performs intention recognition based on the files to be reviewed, audit scenario and audit type to obtain an audit rule list, determines the positions of the content to be reviewed corresponding to each audit rule according to the audit rule list, extracts and slices the content to be reviewed based on the positions of each content to be reviewed to obtain multiple sliced documents, assembles the multiple sliced documents corresponding to each audit rule to obtain atomic files corresponding to each audit rule, uses an analysis agent to perform risk judgment on the atomic files corresponding to each audit rule, obtains atomic results corresponding to each audit rule, then uses a summary agent to summarize the atomic results to obtain a risk prompt list, and based on the risk prompt list, summarizes the atomic results with risk points to obtain an audit opinion. The above method processes files in different formats through step-by-step task decomposition and synthesizes the processing results of each agent, realizing intelligent auditing and improving the efficiency of audit results. Description of the Drawings

[0037] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0038] Figure 1 It is a flowchart of the audit method for the intelligent audit method of hospital internal contracts and projects in the embodiments of the present invention;

[0039] Figure 2 It is a function introduction diagram of the intelligent audit system for the intelligent audit method of hospital internal contracts and projects in the embodiments of the present invention;

[0040] Figure 3 It is a working flowchart of the audit intelligent agent collaboration network for the intelligent audit method of hospital internal contracts and projects in the embodiments of the present invention;

[0041] Figure 4 It is a system block diagram of the intelligent audit system for hospital internal contracts and projects in the embodiments of the present invention. Detailed Embodiments

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Please refer to Figure 1 , which is a schematic flowchart of an embodiment of the intelligent audit method for hospital internal contracts and projects provided by the embodiments of the present invention, including steps S101 to S103.

[0044] S101. In response to a review operation, determine the audit scenario and audit type, and obtain multiple documents to be reviewed according to the audit scenario and audit type, where the documents to be reviewed are contract documents or project documents.

[0045] In this embodiment, by designing an intelligent audit system for hospital internal contracts and projects based on an agent collaboration network, it is deeply integrated with the in-hospital operation system, project management contracts, etc., to realize the intelligent audit of hospital internal contracts and project special items. As Figure 2 shown, the audit system includes an AI consultation module, an AI contract audit module, an AI project special audit module, an AI knowledge base, and a rule base management module. AI consultation retrieves relevant cases through natural language queries. The AI is based on laws, regulations, cases, and FAQ knowledge bases related to audits, and summarizes materials to answer. Specifically, the intelligent agent receives a natural language query, identifies the intent through a model, and determines the query type. According to the query type, the intelligent agent retrieves relevant cases, laws, regulations, and FAQ and other materials from the audit knowledge base, integrates the retrieval results, generates an answer, and returns it to the user. Through natural language queries, relevant cases are retrieved, and the AI is based on laws, regulations, cases, and FAQ knowledge bases related to audits, and summarizes materials to answer.

[0046] The AI contract audit module is used for contract submission for review both during off-campus bidding and contract signing. In this link, the auditor clicks the intelligent audit button, selects the document to be submitted for review and the document to be compared. After being reviewed by the audit intelligent agent collaboration network model, it outputs a review opinion, an Excel version of the risk review situation table, inconsistent comparison results, and an annotated version of the contract review document.

[0047] The AI project special audit module is used for auditing after the project is completed. The auditor clicks intelligent audit, and it is reviewed by the audit intelligent agent collaboration network model, and outputs an Excel version of the risk review situation table and a Word version of the review opinion. The AI knowledge base and rule base management module is used to support the auditor to set the knowledge base and rule base by himself. The auditor can set the knowledge base and rule base by himself, including information such as regulations, cases, and audit rules.

[0048] AI Knowledge Base and Rule Base Management: Support auditors to set up the knowledge base and rule base by themselves. Auditors can set up the knowledge base and rule base by themselves, including information such as regulations, cases, audit rules, etc.

[0049] When conducting audits using the AI Contract Audit Module and the AI Project Audit Module, first obtain multiple documents to be reviewed and input them into the audit intelligent agent collaboration network model for review. Information acquisition steps: Through synchronous methods such as API interfaces and message subscriptions, obtain the review scenario (contract audit / special audit), audit type (service, engineering, goods, information system), project / contract basic information (name, ID number), and submitted review attachments (original documents). For example: When the review scenario is contract audit, the audit type is information system, the project / contract basic information is the H001 OA system upgrade project contract, and the submitted review documents include the contract document in docx format and the tender document in pdf format.

[0050] When the review scenario is special audit, the audit type is engineering Z002, and the submitted review attachments include the project approval notice in pdf format, the budget allocation notice in jpg format, the tender document in pdf format, the winning bid notice in doc format, the contract in pdf format, the advance payment approval form in pdf format, the starting order in jpg format, and the acceptance report in pdf format.

[0051] It should be noted that the documents to be reviewed refer to contract documents or project documents.

[0052] In one embodiment, use the perception intelligent agent to perform document format conversion on multiple documents to be reviewed to obtain multiple text files, including:

[0053] Use pdf parsing and conversion technology to perform image conversion on each page of multiple documents to be reviewed to obtain the corresponding converted documents;

[0054] Based on each of the converted documents, use multi-modal OCR technology to locate and extract computer text and handwritten text, obtain the corresponding extraction results and text coordinates, and embed the corresponding extraction results into the corresponding documents to be reviewed to obtain the text files corresponding to each document to be reviewed.

[0055] In this embodiment, first, the PDF parsing and conversion technology is used to convert each page of multiple documents to be reviewed into pictures, and then the multi-modal OCR technology is used to locate and extract computer text and handwritten text from the converted picture files. Then the text is embedded into the original PDF review document to obtain the converted text document, and the text coordinates are used as the sorting basis for the positioning identifiers of all independent text segments. Through the above method, for any pdf or word document, the risk positions of the original text can be located and traced back, the hallucination of the large model can be reduced, and it is more convenient for auditors to view and use.

[0056] In one embodiment, after obtaining multiple text documents, it further includes:

[0057] Based on punctuation marks and grammar rules, perform intelligent sentence segmentation processing on each text document to obtain multiple segmented text documents;

[0058] Correct the incorrect characters and missing characters in each sentence of each segmented text document, and assign a positioning identifier.

[0059] In this embodiment, the converted text document will be subjected to intelligent sentence segmentation processing. Combining natural language processing (NLP) technology, the text in the text document is segmented into independent sentences according to punctuation marks and grammar rules, and automatic correction of possible incorrect and missing characters is performed. In addition, to ensure the traceability of the text, each sentence will be assigned a unique positioning identifier, that is, the line number ID, which can be understood as the line number, for subsequent retrieval, comparison and analysis.

[0060] S102. Input the document to be reviewed into the multi-agent collaborative network model based on the audit vertical domain large model. Apply the perception agent to perform document format conversion on multiple documents to be reviewed to obtain multiple text documents, and perform intention recognition on the documents to be reviewed to obtain an audit rule list. According to the audit rule list, determine the positions of the content to be reviewed corresponding to each audit rule. Based on the positions of each content to be reviewed, extract and slice the content to be reviewed to obtain multiple sliced documents, and assemble the multiple sliced documents to obtain the atomic document corresponding to each audit rule.

[0061] In this embodiment, as Figure 3As shown in the figure, the document to be reviewed is input into the audit vertical domain large model. The perception agent is used to convert the format of the document to be reviewed, and all documents are converted into text files. Then, intention recognition is performed based on the document to be reviewed, the audit scenario, and the audit type to generate an audit rule list. According to each rule list, the document where the content to be reviewed is located and its location in the document are located. Data extraction and document slicing are performed based on the location, and file assembly is carried out to output the atomic file of this rule {rule ID, document ID, line number ID, assembled file, next task content}. By processing multi-modal data more flexibly, through the combination of multi-stage and specific-function agents, complex tasks are disassembled and processed to improve the information extraction effect, and a comprehensive understanding and analysis of the content of different format files are achieved.

[0062] As an example of this embodiment, for instance: According to the input content, the audit intention is identified as a special audit of an engineering project, and an engineering project special audit rule list is generated, with a total of 11 rules. Rule 1, identify risk points: whether there is procurement and expenditure without a budget. The content to be reviewed is: the notice time in the budget release notice, the contract signing time in the contract, and the approval time in the advance payment approval form. The content to be reviewed in the three documents is extracted to form assembled file 1, and the next task content is generated: analyze the chronological relationship of the three times, requiring that the contract signing time is later than the budget release notice time, and the approval time of the hospital leadership's approval opinion in the advance payment approval form is later than the contract signing time; otherwise, it violates the regulations.

[0063] In one embodiment, the audit vertical domain large model is obtained by pre-training a general large model using audit domain text data, including:

[0064] Obtain audit domain text data, where the audit domain text data includes audit domain laws and regulations text data, audit cases, and audit rules;

[0065] Use NLP technology to perform word segmentation and entity recognition on the audit domain text data to obtain key information, and construct a knowledge graph based on the entities and the relationships between the entities in the key information;

[0066] Select a general large model as the base model, insert the entity vectors in the knowledge graph as key values into the base model, and use an adapter structure to isolate the knowledge parameters, dynamically adjust the learning rate and mask ratio to obtain the initial audit vertical domain large model;

[0067] According to the business requirements of the audit vertical domain and the model optimization goal, divide it into multiple tasks, construct the supervision data and instruction sets for each task, and based on the supervision data and instruction sets of each task, use reinforcement learning to design a reward model and optimize the initial audit vertical domain large model using the PPO algorithm to obtain the audit vertical domain large model.

[0068] In this embodiment, the audit vertical domain large model is obtained by pre-training a general large model using audit domain text data. The specific steps of pre-training are as follows:

[0069] S21: Obtain audit domain text data, which includes audit laws and regulations knowledge, audit cases, and audit rules. Use NLP technology to tokenize and identify entities in the audit domain text data to obtain key information, which includes audit issues, violation types, audit procedures, etc. Construct a knowledge graph based on the entities in the key information and the relationships between the entities;

[0070] S22: Select a general large model as the base model, such as llama3, Claude, Wenxin Big Model 3.5, Alibaba Tongyi Big Model, deepseek, doubao pro, etc., for continued pre-training (Continual Pre-training). Insert the entity vectors of the knowledge graph established in S21 as key-value into the base model's transfomer layer. Use the adapter structure to isolate the knowledge parameters, and dynamically adjust the learning rate (1e-5) and mask ratio (0.3) to obtain the initial audit vertical domain large model;

[0071] S23: According to the business requirements and model optimization goals of the audit vertical domain, divide it into 3 tasks, and construct the supervision data and instruction sets for each task.

[0072] 1) Question-answering task: Question-answer pairs, such as "instruction: How much amount is above for government procurement; output: For goods and services projects, it is above 1 million yuan, and for engineering projects, it is above 1.2 million yuan."

[0073] 2) Perception classification task: Combinations of the document to be reviewed, audit scenario, and audit type - audit intention - audit rule list, such as "instruction: According to the submitted document and scenario, identify the intention and output the rule list; input: Contract audit, information system, H001 OA system upgrade project contract; output: Special audit for information contracts, with a total of 11 rule lists."

[0074] 3) Analysis task: Apply each rule to the Whitelist-Guided Question Chain Generation (WGQG) technique to transform it into a step-by-step reasoning chain of thought to guide the large model to think. "Input: Sliced original contract text, Rule: Whether the payment terms are compliant; Output1: Chain of thought process: What is the amount of this contract - Whether it belongs to a government procurement project (projects with an amount greater than 1 million yuan are determined as government procurement projects) - Whether the time for all payments exceeds 10 working days (government procurement projects require less than 10 working days); Output2: There is a risk and it needs to be modified."

[0075] S24: Design a reward model using Reinforcement Learning from Human Feedback (RLHF), fine-tune the optimization strategy based on the Proximal Policy Optimization (PPO) algorithm, optimize the large model for the audit vertical domain, and improve the task effect.

[0076] In one embodiment, each agent in the multi-agent collaboration network model is obtained by fine-tuning the large model for the audit vertical domain, including:

[0077] Based on the large model for the audit vertical domain, a multi-agent collaboration network model is constructed by combining prompt engineering and Python-based engineering functions. The multi-agent collaboration network model includes a perception agent, an analysis agent, a summary agent, and an execution agent.

[0078] In this embodiment, the process of constructing the agents is as follows: Based on S24, by combining prompt engineering and Python-based engineering functions, a perception agent, an analysis agent, a summary agent, and an execution agent are constructed, thus building an agent collaboration network model. By adopting the method of combining multiple agents, each agent focuses on a specific function, improving the execution effect of each stage. Through step-by-step task decomposition and integrating the results of each agent, the overall task is finally completed efficiently, solving the problem of poor performance in processing multi-format documents in the prior art.

[0079] In one embodiment, assembling multiple sliced documents to obtain the atomic file corresponding to each audit rule includes:

[0080] Assembling the multiple sliced documents corresponding to each audit rule to obtain the assembled file corresponding to each audit rule;

[0081] According to the audit rule number, text file number, location identifier, assembled file, and next task content, the atomic file corresponding to each audit rule is obtained.

[0082] In this embodiment, after data extraction and document slicing according to the location, multiple sliced documents are obtained. The multiple sliced documents are assembled into files, and the atomic file of this rule is output. The content of the atomic file includes the rule ID, document ID, line number ID, assembled file, and the content of the next task.

[0083] S103. Use the analysis agent to perform risk judgment on the atomic file corresponding to each audit rule. After obtaining the atomic result corresponding to each audit rule, use the summary agent to summarize each atomic result to obtain a risk prompt list, and based on the risk prompt list, summarize the atomic results with risk points to obtain an audit opinion.

[0084] In this embodiment, the atomic file of each audit rule is obtained. The analysis agent is used to judge the content of each atomic file according to the task requirements and the rule knowledge base to determine whether there are risk points. If there are risk points, the risk point content, violation regulations, and rectification suggestions are output to form the atomic result of each rule. The content of the atomic result includes the rule ID, document ID, line number ID, and output result. Then, the atomic results of each of the above rules are obtained, and all the results are summarized to form a risk prompt list. The risk prompt list is shown in Table 1.

[0085] Table 1 Risk Prompt List

[0086]

[0087]

[0088] After obtaining the atomic results of each of the above rules, use the summary agent to summarize all the results to form a risk prompt list, and further summarize the entries with risks, and combine the basic situation of the project to form an audit opinion draft, that is, an audit opinion.

[0089] In one embodiment, after obtaining the audit opinion, it further includes:

[0090] Enable the audit vertical domain large model to use the execution agent to encapsulate the risk prompt list into a first document and encapsulate the audit opinion into a second document;

[0091] Based on the first document and the second document, use the numbers and positioning identifiers of each file to be reviewed to annotate the file to be reviewed to obtain an annotation result.

[0092] In this embodiment, an execution agent is used to encapsulate the above-generated risk prompt list into a separate file. The list is encapsulated in the excel format, and the audit opinion working papers are encapsulated in word. According to the document ID, line number ID, and output result, the output result is annotated at the corresponding position of the original text for easy viewing. The files are packaged and the interface is called to send them back to the native system. Through this method, the result can be accurately annotated at the positioning position, which is convenient for further viewing of the problem.

[0093] It should be noted that the first document refers to a document in the excel format, and the second document refers to a document in the word format.

[0094] The intelligent audit method for hospital internal contracts proposed by the present invention fits specific business scenarios, organically integrates agents with the audit system, and realizes an end-to-end usage mode. In the close combination with the audit system, not only the actual application effect of the agent is improved, but also the overall performance of the audit system is optimized. By deeply integrating the agent and the audit system, the deficiencies in system integration and optimization in the prior art are made up, and the intelligent level and work efficiency of the audit work are improved.

[0095] It should be understood that although Figure 1 the steps in the flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in

[0096] can include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of the steps or stages in other steps or other steps.

[0096] In one embodiment, as Figure 4 shown, it shows a block diagram of a hospital internal contract and project intelligent audit system 400 provided by an embodiment of the present application, including an acquisition module 401, a conversion module 402, and a summary module 403, where:

[0097] The acquisition module 401 is configured to determine an audit scenario and an audit type in response to a review operation, and obtain multiple files to be reviewed according to the audit scenario and the audit type, where the files to be reviewed are contract files or project files;

[0098] The conversion module 402 is configured to input the document to be reviewed into the multi-agent collaboration network model based on the audit vertical domain large model. The application perception agent is used to perform document format conversion on multiple documents to be reviewed to obtain multiple text files, and perform intention recognition on the documents to be reviewed to obtain an audit rule list. According to the audit rule list, the positions of the content to be reviewed corresponding to each audit rule are determined. Based on the positions of the content to be reviewed, the content to be reviewed is extracted and sliced into documents to obtain multiple sliced documents, and the multiple sliced documents are assembled to obtain the atomic file corresponding to each audit rule.

[0099] The summarization module 403 is configured to use the analysis agent to perform risk judgment on the atomic file corresponding to each audit rule. After obtaining the atomic result corresponding to each audit rule, the summary agent is used to summarize the atomic results to obtain a risk prompt list, and based on the risk prompt list, the atomic results with risk points are summarized to obtain an audit opinion.

[0100] The specific implementation manner of the intelligent audit system for hospital internal contracts and projects is basically the same as the specific embodiments of the above-mentioned intelligent audit method for hospital internal contracts and projects, and will not be elaborated here.

[0101] In an embodiment of the present application, a computer device is provided. The computer device includes a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the above steps are implemented. The computer device provided in this embodiment has the same implementation principle and technical effects as the above method embodiment, and will not be elaborated here.

[0102] In an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the above steps are implemented. The computer-readable storage medium provided in this embodiment has the same implementation principle and technical effects as the above method embodiment, and will not be elaborated here.

[0103] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, it should be considered that the scope described in this specification.

[0104] The above specific embodiments have further elaborated the purpose, technical solution and beneficial effects of the present invention. It should be understood that the above description is only the specific embodiments of the present invention and is not used to limit the protection scope of the present invention. In particular, for those skilled in the art, any modifications, equivalent replacements, improvements, 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 intelligent auditing method for hospital internal contracts and projects, characterized in that, Including: In response to a review operation, determine an audit scenario and an audit type, and obtain multiple documents to be reviewed according to the audit scenario and the audit type, where the documents to be reviewed are contract documents or project documents; Input the documents to be reviewed into a multi-agent collaboration network model based on an audit vertical domain large model. Apply a perception agent to perform document format conversion on the multiple documents to be reviewed to obtain multiple text files, and perform intention recognition according to the documents to be reviewed, the audit scenario, and the audit type to obtain an audit rule list. According to the audit rule list, determine the positions of the content to be reviewed corresponding to each audit rule. Based on the positions of each piece of content to be reviewed, extract and slice the content to be reviewed to obtain multiple sliced documents, and assemble the multiple sliced documents to obtain an atomic file corresponding to each audit rule; Use an analysis agent to perform risk judgment on the atomic file corresponding to each audit rule. After obtaining the atomic result corresponding to each audit rule, use a summary agent to summarize the atomic results to obtain a risk prompt list, and based on the risk prompt list, summarize the atomic results with risk points to obtain an audit opinion.

2. The intelligent auditing method for hospital internal contracts and projects according to claim 1, characterized in that The step of using a perception agent to perform document format conversion on the multiple documents to be reviewed to obtain multiple text files includes: Use pdf parsing and conversion technology to perform image conversion on each page of the multiple documents to be reviewed to obtain corresponding converted files; Based on each of the converted files, use multi-modal OCR technology to locate and extract computer text and handwritten text to obtain corresponding extraction results and text coordinates, and embed the corresponding extraction results into the corresponding documents to be reviewed to obtain text files corresponding to each document to be reviewed.

3. The hospital internal contract and project intelligent auditing method according to claim 2, wherein After obtaining the multiple text files, it further includes: Based on punctuation and grammar rules, perform intelligent sentence splitting on each of the text files to obtain multiple segmented text files; Correct the incorrect and missing characters in each sentence of each segmented text file, and assign a positioning identifier.

4. The intelligent auditing method for hospital internal contracts and projects according to claim 1, characterized in that, The step of assembling the multiple sliced documents to obtain an atomic file corresponding to each audit rule includes: Assemble the multiple sliced documents corresponding to each audit rule to obtain an assembled file corresponding to each audit rule; Obtain an atomic file corresponding to each audit rule according to the audit rule number, text file number, positioning identifier, assembled file, and next task content.

5. The intelligent audit method for hospital internal contracts and projects as claimed in claim 1, wherein After obtaining the audit opinion, it further includes: Use an execution agent to encapsulate the risk prompt list into a first document and encapsulate the audit opinion into a second document; Based on the first document and the second document, annotate the documents to be reviewed using the numbers of the documents to be reviewed and the positioning identifier to obtain an annotation result.

6. The intelligent auditing method for hospital internal contracts and projects according to claim 1, wherein The audit vertical domain large model is obtained by pre-training a general large model using audit domain text data, including: Obtain text data in the audit field, where the text data in the audit field includes text data of laws and regulations in the audit field, audit cases, and audit rules; Use NLP technology to perform word segmentation and entity recognition on the text data in the audit field to obtain key information, and construct a knowledge graph based on the entities and the relationships between the entities in the key information; Select a general large model as the base model, insert the entity vectors in the knowledge graph as key values into the base model, and use an adapter structure to isolate the knowledge parameters, dynamically adjust the learning rate and mask ratio to obtain an initial large model for the vertical audit domain; According to the business requirements and model optimization objectives of the vertical audit domain, divide them into multiple tasks, construct the supervision data and instruction sets for each task, and based on the supervision data and instruction sets for each task, use reinforcement learning to design a reward model and optimize the strategy based on the PPO algorithm to optimize the initial large model for the vertical audit domain to obtain a large model for the vertical audit domain.

7. The intelligent auditing method for hospital internal contracts and projects according to claim 1, wherein Each agent in the multi-agent collaboration network model is obtained by fine-tuning the large model for the vertical audit domain, including: Based on the large model for the vertical audit domain, a multi-agent collaboration network model is constructed by combining prompt engineering and engineering functions based on Python, where the multi-agent collaboration network model includes a perception agent, an analysis agent, a summary agent, and an execution agent.

8. An intelligent audit system for hospital internal contracts and projects, characterized in that, Including: An acquisition module, configured to, in response to a review operation, determine an audit scenario and an audit type, and obtain multiple files to be reviewed according to the audit scenario and the audit type, where the files to be reviewed are contract files or project files; A conversion module, configured to input the files to be reviewed into a multi-agent collaboration network model based on the large model for the vertical audit domain, and use the perception agent to perform document format conversion on the multiple files to be reviewed to obtain multiple text files, and perform intent recognition on the files to be reviewed to obtain an audit rule list, and according to the audit rule list, determine the positions of the content to be reviewed corresponding to each audit rule, and based on the positions of the content to be reviewed, extract and slice the content to be reviewed to obtain multiple sliced documents, and assemble the multiple sliced documents to obtain an atomic file corresponding to each audit rule; A summarization module, configured to use the analysis agent to perform risk judgment on the atomic file corresponding to each audit rule, and after obtaining the atomic result corresponding to each audit rule, use the summary agent to summarize the atomic results to obtain a risk prompt list, and based on the risk prompt list, summarize the atomic results with risk points to obtain an audit opinion.

9. A computer device, characterized in that, Including: A memory, configured to store a computer program; A processor, configured to implement the intelligent audit method for hospital internal contracts and projects as described in any one of claims 1 to 7 when executing the computer program.

10. A storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by the processor, the steps of the intelligent audit method for hospital internal contracts and projects as described in any one of claims 1 to 7 are implemented.

Images