Enterprise operation supervision multi-agent cooperation method and system based on large language model

Abstract

The invention discloses an enterprise operation supervision multi-agent cooperation method and system based on a large language model. The method comprises the steps that a data acquisition agent constructs an enterprise database; the risk identification agent trains a clustering model and a text abstract model according to historical data of each supervision object, and is used for identifying and summarizing risk-related events and topics; the regulation analysis agent retrieves the most relevant regulation file from the regulation knowledge base according to the received abstract and inputs the most relevant regulation file into the large language model, regulation opinions of the abstract are generated, and an event evolution regulation opinion tetrad is formed; the decision support agent evaluates the event evolution regulation opinion tetrad; if the evaluation result is that the regulation opinions are not clear enough or the event evolution information is insufficient, feedback opinions are sent to the regulation analysis intelligent agent or the risk identification intelligent agent, and related data are regenerated. According to the invention, the intelligentization, precision and real-time level of enterprise operation supervision are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of artificial intelligence, big data analysis, and enterprise operation supervision, and relates to a multi-agent collaborative method and system for enterprise operation supervision based on a large language model, which utilizes a multi-agent system driven by a large language model to perform real-time monitoring of enterprise operation behaviors, risk assessment, rule interpretation and compliance guidance, and an integrated solution for intelligent decision support. Background Art

[0002] With the rapid development of the digital economy, corporate business activities are becoming increasingly complex and the amount of data is growing explosively. Traditional means of corporate business supervision face problems such as inefficient information processing, untimely identification of violations, and inconsistent interpretation and implementation of regulatory policies.

[0003] As an artificial intelligence technology with powerful natural language understanding and generation capabilities, the large language model has demonstrated outstanding application value in many fields. At the same time, multi-agent systems have shown great potential in simulating complex interactive environments and achieving efficient collaborative decision-making. However, the deep integration of large language models and multi-agent systems and their application in corporate management supervision and governance scenarios is still in its infancy. Summary of the invention

[0004] The present invention aims to provide a multi-agent collaborative method and system for enterprise management supervision based on a large language model. By constructing a multi-agent ecosystem, each agent is responsible for data collection, rule analysis, risk identification, decision-making advice and other functions, and through the large language model, cross-agent information sharing, semantic understanding and decision coordination are realized to improve the intelligence, precision and real-time level of enterprise management supervision.

[0005] The technical solution of the present invention is: A multi-agent collaborative method for enterprise management supervision based on a large language model, the steps of which include: 1) The data collection agent obtains enterprise information from the authorized enterprise credit information disclosure system and builds an enterprise database; each enterprise in the enterprise database is regarded as a supervision object, and the dynamic data of each supervision object is obtained in real time from various public and authorized data sources and processed, and the structured data corresponding to each supervision object is obtained and sent to the risk identification agent; 2) The risk identification agent trains a clustering model and a text summary model based on the historical data of each regulatory object; wherein, the trained and updated clustering model is used to segment the currently received structured data, and the segmented words with a frequency exceeding a set threshold are extracted as high-frequency keywords, and then the structured data containing the same high-frequency keyword is used as the supporting text of the high-frequency keyword to obtain the supporting text set corresponding to the high-frequency keyword, and then the supporting text sets are clustered by a clustering algorithm to obtain a number of cluster clusters; and the training and updated text summary model is used to generate a summary of each of the supporting text sets and send it to the regulatory analysis agent; 3) The regulatory analysis agent retrieves the most relevant regulatory documents from the regulatory knowledge base based on the received summary and inputs it into the large language model, generates regulatory opinions for the summary, forms a quadruple of event evolution regulatory opinions (time, high-frequency keywords, event evolution, regulatory opinions) and sends it to the decision support agent; 4) The decision support agent evaluates the regulatory opinion quadruple of the event evolution; if the evaluation result is that the regulatory opinion is not clear enough or the event evolution information is insufficient, the feedback is sent to the regulatory analysis agent or the risk identification agent to regenerate the relevant data.

[0006] A multi-agent collaborative system for enterprise management supervision based on a large language model, characterized by comprising a data collection agent, a risk identification agent, a regulation analysis agent and a decision support agent; The data collection agent is used to obtain enterprise information from the authorized enterprise credit information disclosure system in the initial stage and build an enterprise database; and in the operation stage, each enterprise in the enterprise database is regarded as a supervision object, and the dynamic data of each supervision object is obtained in real time from various public and authorized data sources and processed, and the structured data corresponding to each supervision object is obtained and sent to the risk identification agent; The risk identification agent is used to train a clustering model and a text summary model based on the historical data of each regulatory object; wherein, the currently received structured data is segmented using the clustering model after training and updating, and the segmented words whose word frequency exceeds the set threshold are extracted as high-frequency keywords, and then the structured data containing the same high-frequency keyword is used as the supporting text of the high-frequency keyword to obtain the supporting text set corresponding to the high-frequency keyword, and then the supporting text sets are clustered by a clustering algorithm to obtain a number of cluster clusters; and the training and updated text summary model is used to generate a summary of each of the supporting text sets and send it to the regulatory analysis agent; The regulatory analysis agent is used to retrieve the most relevant regulatory documents from the regulatory knowledge base according to the received summary and input it into the large language model, generate regulatory opinions of the summary, form a quadruple of event evolution regulatory opinions (time, high-frequency keywords, event evolution, regulatory opinions) and send it to the decision support agent; The decision support agent is used to evaluate the four-tuple of regulatory opinions on the event evolution; if the evaluation result is that the regulatory opinions are not clear enough or the event evolution information is insufficient, feedback is sent to the regulatory analysis agent or the risk identification agent to regenerate relevant data.

[0007] Furthermore, the summary is an event evolution triple (time, high-frequency keywords, event evolution).

[0008] Furthermore, the regulatory analysis agent constructs the regulatory knowledge base based on laws, regulations, regulatory policies, and industry standard documents.

[0009] Furthermore, the enterprise information includes the enterprise's name, place of registration, registration time, main business, and administrative penalties.

[0010] The present invention provides four types of intelligent agent role definitions and division of labor for enterprise management supervision, including: 1) Data collection agent: Initialization stage: Build an enterprise database from the authorized enterprise credit information disclosure system, including the enterprise's name, registration place, registration time, main business, administrative penalties and other information.

[0011] Operation phase: Obtain relevant dynamic data of each enterprise in the enterprise database in real time from various public and authorized data sources, clean and standardize unstructured data through natural language processing technology, and output structured data related to the regulatory objects to ensure that the regulatory authorities can fully understand the dynamics of the enterprise.

[0012] 2) Risk Identification Agent: Initialization phase: Clustering models such as K-Means, DBSCAN and text summarization models such as PEGASUS or T5 are trained based on historical data to identify and summarize risk-related events and topics.

[0013] Operation phase: retrieve enterprise-related data of the specified time interval from the data collection agent and perform data preprocessing, such as removing HTML tags, special characters and duplicate data. Perform keyword extraction, use TF-IDF, TextRank and other models to extract high-frequency keywords that exceed the threshold, and collect the supporting text set corresponding to each high-frequency keyword, then cluster the supporting text set through the clustering algorithm, and use the text summary model to summarize each supporting text group to form an event evolution triple (time, high-frequency keyword, event evolution).

[0014] 3) Regulatory Analysis Agent: Initialization stage: Summarize laws, regulations, regulatory policies, industry standards and other documents to build a three-level hierarchical regulatory knowledge base of "document-chapter-entry". The knowledge base has a clear structure and is convenient for quick retrieval and matching of relevant regulations.

[0015] Operation phase: retrieve the event evolution triples from the risk identification agent, use the embedding model (such as BGE-M3) to retrieve the most relevant regulatory documents, and try to generate regulatory opinions to form an event evolution regulatory opinion quadruple (time, high-frequency keywords, event evolution, regulatory opinions).

[0016] 4) Decision support agent: Operation phase: Evaluate each event evolution regulatory opinion quaternary to decide whether to submit it to decision makers. The evaluation process is as follows: (1) Data integrity evaluation: Input the event evolution regulatory opinion quaternary (time, high-frequency keywords, event evolution, regulatory opinion), check whether each field of the quaternary is complete, and if it is found that the data is insufficient or key information is missing, send feedback to the data collection agent to adjust the collection strategy or improve the data cleaning process. (2) Risk identification evaluation: If the regulatory analysis agent finds that the required data is not specific or the keyword clustering results are inaccurate in the subsequent regulatory matching stage, it can provide feedback and require the risk identification agent to review the keyword extraction algorithm or clustering logic. (3) Regulatory opinion quality evaluation: Input the complete event evolution regulatory opinion quaternary to determine whether the regulatory opinion can actually be applied to the current event. If it cannot be applied, feedback is given to the regulatory analysis agent to request more detailed analysis or other regulatory options.

[0017] The present invention provides a multi-agent collaborative enterprise management supervision operation process driven by a large language model, including: The data collection agent is responsible for collecting enterprise-related data in real time from a variety of authorized public data sources, such as social media, complaint platforms, etc. The data is initially cleaned and standardized through natural language processing technology to generate structured data related to the enterprise.

[0018] The risk identification agent receives structured data from the data collection agent, applies advanced text analysis techniques such as TF-IDF and TextRank to extract keywords, and performs text clustering.

[0019] The regulatory analysis agent searches the internal regulatory knowledge base based on the event summary provided by the risk identification agent, finds relevant regulatory provisions, and generates specific regulatory recommendations.

[0020] The decision support agent evaluates the regulatory recommendations proposed by the regulatory analysis agent and forms a decision opinion. According to the urgency of the event and the importance of the regulatory recommendations, an appropriate execution strategy is selected. The present invention provides a multi-agent collaborative enterprise management supervision system architecture driven by a large language model, including:

[0021] The data access layer is responsible for data collection and preprocessing; the intelligent agent cluster calls machine learning models, knowledge bases, etc. to perform their respective tasks and collaborate based on the large language model; the large language model core provides knowledge processing and cross-agent communication services; the decision support interface provides decision-making basis to regulators; the user interface displays the regulatory situation, early warning information and decision recommendations.

[0022] Compared with the prior art, the present invention has the following positive effects: The present invention improves the intelligence, precision and real-time level of business operation supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the system architecture in an embodiment of the present invention.

[0024] Figure 2 It is a schematic diagram of the operation flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings.

[0026] This manual provides the definition and division of roles of intelligent agents, including: Specific implementation of data collection agent: Initialization stage: Obtain basic enterprise information from the National Enterprise Credit Information Publicity System, such as name, place of registration, registration time, etc., as well as other important information such as main business and historical administrative penalties. This information is used to establish an initial enterprise database, which will provide basic data support for subsequent regulatory activities.

[0027] Operational stage: Real-time access to a variety of public and authorized data sources, such as Weibo, Douyin, Black Cat Complaints and other social platforms and complaint platforms, to collect the latest information about registered companies. The collected data is first processed by noise removal, such as filtering irrelevant content and removing duplicate information. Then, natural language processing techniques such as word segmentation, part-of-speech tagging, named entity recognition, etc. are used to structure the unstructured data and extract key information such as corporate activities and consumer feedback. The processed data is converted into a structured format to facilitate query and further analysis by regulatory authorities, ensuring that regulatory authorities can fully understand the latest developments and business behaviors of enterprises.

[0028] Specific implementation of risk identification agent: Initialization phase: Use historical data sets to train clustering models such as K-Means, DBSCAN and text summarization models such as PEGASUS or T5 to identify and summarize risk-related events and topics.

[0029] Running phase: Receive data from the data collection agent and perform further cleaning, including removing HTML tags, special characters, and cleaning up duplicate data. Apply text analysis techniques such as TF-IDF and TextRank to extract high-frequency keywords in the text, such as "Shanghai taxi refused to pick up passengers". Cluster the text associated with the extracted keywords using the equal clustering algorithm. Use the text summary model to summarize the text of each cluster, generate a refined description of each event, and form an event evolution triple (time, high-frequency keyword, event evolution).

[0030] Specific implementation of regulatory analysis agent: Initialization phase: Summarize and organize text materials such as laws, regulations, regulatory policies, industry standards, etc., and establish a structured regulatory knowledge base to facilitate rapid retrieval and accurate matching of relevant regulations.

[0031] Operational stage: Receive event evolution triples from risk identification agents, and use embedding models such as BGE-M3 to retrieve regulatory documents related to the event. Try to generate specific regulatory suggestions based on the retrieved regulatory content using a large language model, with the prompt template being "Please generate reasonable regulatory opinions based on the provided event evolution and retrieved regulatory content.", forming a quadruple of event evolution regulatory opinions (time, high-frequency keywords, event evolution, regulatory opinions).

[0032] Specific implementation of decision support agent: Each event evolution regulatory opinion quadruple is evaluated to decide whether to submit it directly to decision-makers for review or action. The evaluation process is as follows: (1) Data integrity evaluation: Input the event evolution regulatory opinion quadruple (time, high-frequency keywords, event evolution, regulatory opinion), check whether each field of the quadruple is complete, and if insufficient data or key information is missing, send feedback to the data collection agent to adjust the collection strategy or improve the data cleaning process. (2) Risk identification evaluation: If the regulatory analysis agent finds that the required data is not specific or the keyword clustering results are inaccurate in the subsequent regulatory matching stage, it can provide feedback and require the risk identification agent to review the keyword extraction algorithm or clustering logic. (3) Regulatory opinion quality evaluation: Input the complete event evolution regulatory opinion quadruple to determine whether the regulatory opinion can actually be applied to the current event. If not, feedback is given to the regulatory analysis agent to request more detailed analysis or other regulatory options.

[0033] This manual provides the operation process, such as Figure 2 As shown, including: S101: The data collection agent collects enterprise-related data in real time from a variety of authorized public data sources, such as social media and complaint platforms. It uses natural language processing technology to perform preliminary data cleaning and standardization to generate structured data.

[0034] S102: Receive structured data from the data collection agent and perform further data preprocessing, extract keywords using TF-IDF, TextRank and other technologies, and perform text clustering.

[0035] S103: Based on the event summary provided by the risk identification agent, the internal regulatory knowledge base is searched to find relevant regulations, and then specific regulatory recommendations are generated for decision support.

[0036] S104: Evaluate the recommendations of the regulatory analysis agent and select an appropriate implementation strategy based on the urgency of the event and the importance of the recommendations. Implement the decided strategy and monitor its effectiveness. Based on the feedback of the implementation results, the decision support agent can adjust its decision model or request the aforementioned agent to provide more accurate or updated data and analysis.

[0037] This manual provides the system architecture, such as Figure 1 As shown, including: Data access layer: deploys data interfaces and crawler tools, responsible for accessing and initially processing external data sources.

[0038] Agent cluster: includes data collection agents, regulatory analysis agents, risk identification agents and decision support agents, and calls external machine learning models, knowledge bases, etc., each performing its own duties and collaborating with each other.

[0039] Large language model core: Integrates a large language model, responsible for knowledge processing, cross-agent communication support, and generation of high-quality text.

[0040] Decision support interface: Provides a user-friendly interface that displays regulatory insights, early warning information and decision recommendations for regulators to review and interact with.

[0041] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.

Claims

1. A multi-agent collaborative method for enterprise management supervision based on a large language model, the steps of which are include: 1) The data collection agent obtains enterprise information from the authorized enterprise credit information disclosure system and builds an enterprise database; Taking each enterprise in the enterprise database as a supervision object, obtaining dynamic data of each supervision object from various public and authorized data sources in real time and processing it, obtaining structured data corresponding to each supervision object and sending it to the risk identification intelligent agent; 2) The risk identification agent trains a clustering model and a text summary model based on the historical data of each regulatory object; wherein, the trained and updated clustering model is used to segment the currently received structured data, and the segmented words with a frequency exceeding a set threshold are extracted as high-frequency keywords, and then the structured data containing the same high-frequency keyword is used as the supporting text of the high-frequency keyword to obtain the supporting text set corresponding to the high-frequency keyword, and then the supporting text sets are clustered by a clustering algorithm to obtain a number of cluster clusters; and the training and updated text summary model is used to generate a summary of each of the supporting text sets and send it to the regulatory analysis agent; 3) The regulatory analysis agent retrieves the most relevant regulatory documents from the regulatory knowledge base based on the received summary and inputs it into the large language model, generates regulatory opinions for the summary, forms a quadruple of event evolution regulatory opinions (time, high-frequency keywords, event evolution, regulatory opinions) and sends it to the decision support agent; 4) The decision support agent evaluates the regulatory opinion quadruple of the event evolution; if the evaluation result is that the regulatory opinion is not clear enough or the event evolution information is insufficient, the feedback is sent to the regulatory analysis agent or the risk identification agent to regenerate the relevant data.

2. The method according to claim 1, It is characterized in that The summary is an event evolution triple (time, high-frequency keywords, event evolution).

3. The method according to claim 1, It is characterized in that The regulatory analysis agent constructs the regulatory knowledge base based on laws, regulations, regulatory policies, and industry standard documents.

4. The method according to claim 1, 2 or 3, It is characterized in that The enterprise information includes the enterprise's name, place of registration, registration time, main business, and administrative penalties.

5. A multi-agent collaborative system for enterprise management supervision based on a large language model. It is characterized in that Including data collection agents, risk identification agents, regulatory analysis agents and decision support agents; The data collection agent is used to obtain enterprise information from the authorized enterprise credit information disclosure system in the initial stage and build an enterprise database; And in the operation stage, each enterprise in the enterprise database is regarded as a supervision object, and the dynamic data of each supervision object is obtained in real time from various public and authorized data sources and processed, so as to obtain the structured data corresponding to each supervision object and send it to the risk identification intelligent body; The risk identification agent is used to train a clustering model and a text summary model based on the historical data of each regulatory object; wherein, the currently received structured data is segmented using the clustering model after training and updating, and the segmented words whose word frequency exceeds the set threshold are extracted as high-frequency keywords, and then the structured data containing the same high-frequency keyword is used as the supporting text of the high-frequency keyword to obtain the supporting text set corresponding to the high-frequency keyword, and then the supporting text sets are clustered by a clustering algorithm to obtain a number of cluster clusters; and the training and updated text summary model is used to generate a summary of each of the supporting text sets and send it to the regulatory analysis agent; The regulatory analysis agent is used to retrieve the most relevant regulatory documents from the regulatory knowledge base according to the received summary and input it into the large language model, generate regulatory opinions of the summary, form a quadruple of event evolution regulatory opinions (time, high-frequency keywords, event evolution, regulatory opinions) and send it to the decision support agent; The decision support agent is used to evaluate the four-tuple of regulatory opinions on the event evolution; if the evaluation result is that the regulatory opinions are not clear enough or the event evolution information is insufficient, feedback is sent to the regulatory analysis agent or the risk identification agent to regenerate relevant data.

6. The system according to claim 5, It is characterized in that The summary is an event evolution triple (time, high-frequency keywords, event evolution).

7. The system according to claim 5, It is characterized in that The regulatory analysis agent constructs the regulatory knowledge base based on laws, regulations, regulatory policies, and industry standard documents.

8. A system according to claim 5, 6 or 7, It is characterized in that The enterprise information includes the enterprise's name, place of registration, registration time, main business, and administrative penalties.

Images