Method and system for intelligently generating exhaustion report of financial unfavorable assets

Abstract

The invention provides a method and a system for intelligently generating a complete dispatch report for financial unfavorable assets, and relates to the technical field of data investigation and arrangement, comprising Sp1, complete dispatch demand acquisition, Sp2, intelligent reasoning planning, Sp3, data acquisition and processing, Sp4, public opinion retrieval and analysis, Sp5, legal information extraction and verification, Sp6, data integration and report generation, and Sp7, report modification and output, and the method comprises the steps of Sp1, complete dispatch demand acquisition, Sp2, intelligent reasoning planning, Sp3, data acquisition and processing, Sp4, public opinion retrieval and analysis, Sp5, legal information extraction and verification, Sp6, data integration and report generation, and Sp7, report modification and output. By integrating a large model, an online retrieval tool, a data table tool and a key point extraction tool, automation and intellectualization of a complete dispatching process are realized. The system can automatically acquire, process and integrate data, greatly improves the efficiency of data acquisition, analysis and report generation, and reduces the probability of manual intervention and errors.

Description

Technical Field

[0001] The present invention relates to the technical field of data investigation and collation, and specifically to a method and system for intelligently generating a due diligence report on financial non-performing assets. Background Art

[0002] According to the due diligence report generation method, device, electronic device, storage medium and product disclosed in the Chinese patent publication number "CN118897873A", the method describes the method of obtaining query request information of the enterprise to be due diligence; based on the query request information, using the information retrieval algorithm, obtaining the due diligence data of the enterprise to be due diligence from a pre-built enterprise knowledge base; based on the due diligence data of the enterprise to be due diligence and the pre-built due diligence report standard format, generating a due diligence report on the enterprise to be due diligence.

[0003] According to the due diligence method, apparatus and terminal device based on data collection disclosed in the Chinese patent publication number "CN110264222A", a due diligence method, apparatus, terminal device and computer-readable storage medium based on data collection are provided, including: assigning due diligence tasks to target clients; if the registration data corresponding to all the projects to be adjusted are the same as the field data uploaded by the target client, then screening out the comparison projects, obtaining the default data and default reasons corresponding to the comparison projects of at least two defaulting companies in the database; calculating the difference measurement value between the default data corresponding to the comparison projects and the field data; determining the defaulting company corresponding to the lowest difference measurement value as the target defaulting company, sending a reminder speech to the target client according to the default reason of the target defaulting company, and obtaining the company response information corresponding to the reminder speech uploaded by the target client; generating a due diligence report based on the field data, the default reason of the target defaulting company and the company response information.

[0004] The above patent documents and prior art have the following technical problems when used: The problems existing in the existing technologies in generating due diligence reports on financial non-performing assets are mainly reflected in data processing capabilities, intelligence, flexibility, and industry adaptability. The main problems include: limitations on data sources and insufficient flexibility; frequent manual intervention, resulting in low efficiency; insufficient ability to intelligently process unstructured data; single technical methods, lack of tool integration and automated processing; and deficiencies in generating reports that comply with laws and industry regulations. Summary of the invention

[0005] Technical issues solved In view of the deficiencies of the prior art, the present invention provides a method and system for intelligently generating due diligence reports on financial non-performing assets, which solves the problems of the prior art.

[0006] Technical Solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for intelligently generating a due diligence report on financial non-performing assets, the method comprising the following steps: Sp1. Obtaining due diligence requirements: a. Receive due diligence requirements input by the user through the human-computer interaction interface, including the name of the company to be investigated, the format and content requirements of the due diligence report; b. Use natural language processing technology to parse user input, identify key elements and target information in due diligence tasks, and dynamically generate report requirement templates based on laws, regulations and industry standards in the field of financial non-performing assets.

[0008] Sp2, Intelligent Reasoning Planning: a. Based on the multi-agent framework, combined with the preset legal and regulatory database and due diligence standards, design the due diligence report generation process; b. According to user needs and task characteristics, an execution plan consisting of task modules is constructed. The task modules include a data collection module, a public opinion retrieval module, and a key point extraction module, and the task objectives and output standards of each module are formulated.

[0009] Sp3, data collection and processing: a. Call the structured data query module through the data table tool to extract key fields from external data sources such as industry and commerce, taxation, justice, and electricity; b. Use NL2SQL technology to convert natural language query requirements into SQL statements that are suitable for different databases; c. Automatically generate table content that meets reporting requirements, including basic information of the company, tax records, judicial litigation data and asset and liability details.

[0010] Sp4, public opinion retrieval and analysis: a. Use online search tools to search for public opinion information related to the due diligence target on news platforms, social media and financial websites; b. Apply sentiment analysis algorithms to identify the sentiment tendency of public opinion content (positive, negative, neutral), and perform risk scoring based on click and comment information; c. Output public opinion summaries and risk assessment reports to clarify risk information related to corporate reputation and market behavior.

[0011] Sp5, Legal Information Extraction and Verification: a. Extract basic case information from judgment documents and legal texts through a large-scale language model (AI model fine-tuned by LoRA), including plaintiff, defendant, cause of action, property clues, and cited legal provisions; b. Compare the extracted legal provisions with the current legal and regulatory database to verify the applicability and accuracy of the cited legal provisions; c. Comprehensively analyze case information and generate a comprehensive assessment of the company’s legal risks.

[0012] Sp6, data integration and report generation: a. Integrate the collected structured data, public opinion analysis results and legal information according to the preset report template; b. Automatically fill in various modules of the due diligence report, including corporate information overview, financial status analysis, public opinion and market evaluation, legal risk assessment and comprehensive conclusion; c. Ensure that the content of the generated report complies with legal and industry standards.

[0013] Sp7, report modification and output: a. The user adjusts the generated due diligence report through the interactive interface and proposes specific modification requirements; b. The system calls relevant modules to update the report content in real time according to user needs and regenerates the report; c. The final report is output as a Word document that complies with industry standards and supports export, sharing and printing.

[0014] Preferably, the natural language processing technology comprises the following steps: Convert the due diligence requirements entered by the user into a task instruction tree, and parse out the key information of the company name, report type, and target module; Based on the task instruction tree, semantic parsing technology is used to identify the logical structure of task requirements and dynamically generate dependencies between task modules; Dynamically call the semantic library and domain knowledge base to optimize input parsing results and ensure accurate understanding of industry-specific terminology.

[0015] Preferably, the NL2SQL technology achieves dynamic adaptation in the following ways: Provide standard SQL query templates, covering data fields and query requirements of different industrial and commercial, tax, and judicial databases; Use prompt word technology to convert natural language queries into SQL query statements, where prompt words include target fields, query scope, and filter conditions; The generated SQL statements are checked for legality through a syntax checker, and the statement structure, paging, sorting, and grouping operations are adjusted according to database characteristics.

[0016] Preferably, the public opinion retrieval and analysis module includes the following functional modules: Data source selection module: prioritize retrieval sources from news, social media, and industry websites; Deduplication algorithm module: removes duplicate public opinion information through hash comparison and semantic similarity calculation; Time series analysis module: sort the public opinion content by time, giving priority to analyzing the latest released information; Risk scoring module: Generates a comprehensive risk score based on the sentiment tendency of public opinion, the popularity of dissemination and the authority of the publishing platform.

[0017] Preferably, the key point extraction module specifically includes: Case information extraction: extract basic case fields from judgment documents, including case number, cause of action, and amount involved; Property clue analysis: Use regular expressions and language models to identify property clues from unstructured text, such as property location, vehicle information, and bank accounts; Legal article comparison and verification: Call the local legal and regulatory database and check the cited legal articles in the judgment documents one by one to ensure the accuracy and legality of the cited clauses; Multi-stage case linkage: Establish a case linkage network at different trial stages through the time and logical structure of the judgment, and evaluate the availability of case clues.

[0018] Preferably, the report modification and output is automatically formatted through the following steps: Automatically identify the chapter and module structure in the report content, and adjust the font and paragraph alignment according to the preset template; Automatically adjust table contents according to column width and row height to ensure clarity of data display; Apply paragraph numbering rules and automatically generate a table of contents and cross-reference links to improve the readability and structure of your report.

[0019] Preferably, the system for intelligently generating due diligence reports on financial non-performing assets comprises the following modules: User interaction module: used to receive due diligence requirements and subsequent modification opinions; Intelligent reasoning planning module: used to formulate an execution plan for the due diligence process; Data collection module: extract data from structured databases through NL2SQL tools; Public opinion retrieval module: used for obtaining public opinion information and analyzing sentiment tendency; Key point extraction module: extract key case information from unstructured text; Information integration module: integrates data and fills it into report templates; Report generation module: used to generate standardized due diligence reports and supports dynamic modification.

[0020] Preferably, the user interaction module is based on natural language processing technology, understands the user's customized modification requirements through semantic analysis, and dynamically calls other modules to update the report content.

[0021] Preferably, the information integration module dynamically calls a multi-task processing framework to schedule data spreadsheet tools and public opinion retrieval and analysis modules in real time to ensure the consistency and real-time nature of report content and data.

[0022] Preferably, the public opinion retrieval module generates a public opinion risk index for the due diligence object based on a sentiment analysis model and a risk factor weighting algorithm.

[0023] Beneficial Effects

[0024] The present invention provides a method and system for intelligently generating due diligence reports on financial non-performing assets. It has the following beneficial effects: The present invention realizes the automation and intelligence of the due diligence process by integrating large models, online search tools, data table tools and key point extraction tools. The system can automatically acquire, process and integrate data, greatly improving the efficiency of data collection, analysis and report generation, and reducing the probability of manual intervention and errors.

[0025] This invention combines the local legal and regulatory database and judicial document database, and improves the reasoning ability and information extraction accuracy in the legal field by fine-tuning the large model. Compared with traditional methods, it can accurately extract key information in legal texts (such as case subject matter, court judgments, property clues, etc.), and compare it with the latest legal provisions to ensure the legal professionalism and accuracy of the due diligence report.

[0026] The present invention integrates online search tools to retrieve public opinion information related to the company to be investigated from multiple data sources such as social media and news platforms in real time. The system can automatically analyze the sentiment trend of public opinion and judge the public opinion risk of the company. This function is rare in the prior art and greatly enhances the risk assessment ability of the due diligence report.

[0027] The present invention provides an interactive modification function between the user and the system. After the report is generated, the user can input the modification requirements through natural language, and the system can respond and adjust the report content in real time. This flexibility and customization capability enables the report to better meet the specific needs of the user and improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a system flow chart of the present invention; Figure 2 is a flow chart of the data table tool of the present invention; Figure 3 is a flow chart of the online search tool of the present invention; Figure 4 A flow chart of the key points extraction tool of the present invention; Figure 5 It is the overall architecture diagram of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Specific embodiment 1

[0031] like Figure 1-5 As shown, a method for intelligently generating a due diligence report on financial non-performing assets includes the following steps: Sp1. Obtaining due diligence requirements: a. Receive due diligence requirements input by users through the human-computer interaction interface, including the name of the company to be investigated, the format and content requirements of the due diligence report; b. Use natural language processing technology to parse user input, identify key elements and target information in due diligence tasks, and dynamically generate report requirement templates based on laws, regulations and industry standards in the field of financial non-performing assets.

[0032] In the process of obtaining due diligence requirements, natural language processing (NLP) technology is involved. We convert the input due diligence requirements into a task instruction tree and extract key elements through semantic parsing technology. The following is a detailed description of the algorithm: 1.1 Task instruction tree construction, convert the due diligence requirements input in natural language (such as "Company A's due diligence report") into a task instruction tree (TaskInstructionTree). First, the input text is segmented by sentence, and the sentences are segmented into words or phrases. Each word is annotated using part-of-speech tagging to indicate its grammatical role (such as noun, verb, etc.), and entities in the text are identified, such as company name, report type, target module, etc. You can use the named entity recognition (NER) algorithm, use the CRF (conditional random field) or BERT model, analyze the dependencies between words, build a dependency tree, extract the logical structure of the task from the dependency tree, and generate a task instruction tree.

[0033] Suppose the input sentence is , the task instruction tree is , the dependency is , then the task tree can be generated through the following steps:

[0034] in, It is a function that converts the input sentence into a dependency syntactic tree and then further converts it into a task tree structure.

[0035] Identify the task requirements (such as target module, field requirements, etc.) in the sentence through semantic parsing technology. First, eliminate synonym ambiguity and determine the specific meaning of each word. "Asset" and "property" may be synonyms. Annotate the semantic roles in the sentence, such as subject, predicate, and object, further understand the dependency relationship between words, and generate a semantic dependency graph. Extract task requirements and target module information based on the semantic dependency graph.

[0036] set up is the vocabulary set, is a task instruction tree, the semantic analysis is expressed as:

[0037] Among them, SemanticParse is a semantic parsing function that takes the input vocabulary set Convert to task command tree .

[0038] Sp2, Intelligent Reasoning Planning: a. Based on the multi-agent framework, combined with the preset legal and regulatory database and due diligence standards, design the due diligence report generation process; b. Based on user needs and task characteristics, construct an execution plan consisting of task modules (data collection module, public opinion retrieval module and key point extraction module), and formulate the task objectives and output standards for each module.

[0039] In the intelligent reasoning planning step, the scheduling of task modules and the optimization of execution strategies are involved. The multi-agent framework is designed to dynamically generate the execution plan for each task.

[0040] Based on the task dependencies and priorities, the task execution order is arranged reasonably. The dependencies between tasks are modeled as a directed acyclic graph (DAG). Among them, nodes represent tasks and edges represent dependencies between tasks. Topological sorting algorithm is used to schedule tasks. Let the task set be , the dependency is , the scheduling algorithm is the topological sorting algorithm TopSort , then the scheduling result is:

[0041] in, is the order of tasks after scheduling.

[0042] Sp3, data collection and processing: a. Call the structured data query module through the data table tool to extract key fields from external data sources such as industry and commerce, taxation, justice, and electricity; b. Use NL2SQL technology to convert natural language query requirements into SQL statements that are suitable for different databases; c. Automatically generate table content that meets reporting requirements, including but not limited to basic information of the company, tax records, judicial litigation data and asset and liability details.

[0043] In the process of data collection and processing, NL2SQL technology is mainly used to convert natural language into SQL query statements.

[0044] Convert natural language query requirements (such as "query the business registration information of the enterprise") into SQL query statements. ·Algorithm flow: 1. Question understanding: Understand the query target through semantic analysis, and extract information such as target fields and screening conditions. 2. SQL template matching: Select the appropriate SQL template according to the natural language question 3. Query generation: Fill in the placeholders in the SQL template to generate a complete SQL query statement. ·Mathematical formula: Let the natural language question be , the SQL query template is , then the NL2SQL conversion can be expressed as:

[0045] in, The generated SQL query statement.

[0046] Optimize the generated SQL query statements according to the characteristics of the database to improve query efficiency. Check the legality of the generated SQL statements to ensure that the syntax is correct. Select the optimal query plan (such as index scan or full table scan) according to the database execution plan generation strategy. Suppose the query statement is , the execution plan is , then the query optimization process can be expressed as:

[0047] Sp4, public opinion retrieval and analysis: a. Use online search tools to search for public opinion information related to the due diligence target on news platforms, social media and financial websites; b. Apply sentiment analysis algorithms to identify the sentiment tendency of public opinion content (positive, negative, neutral), and perform risk scoring based on click and comment information; c. Output public opinion summaries and risk assessment reports to clarify risk information related to corporate reputation and market behavior.

[0048] Sentiment analysis and public opinion scoring models play a key role in public opinion retrieval and analysis. Perform sentiment analysis on public opinion content to identify its positive, negative or neutral sentiment. Build a sentiment dictionary suitable for the financial industry, including positive, negative and neutral words. Use a deep learning model (LSTM or BERT) to classify public opinion content. Suppose the public opinion text is , the emotion category is , the sentiment analysis model is , then the result of sentiment analysis is:

[0049] in, positive, negative, neutral

[0050] Generate a comprehensive risk score based on sentiment analysis results, communication heat, platform authority and other information. Assign weights to each public opinion information, and calculate the comprehensive risk score of public opinion by combining sentiment tendency, heat and platform authority through weighted average method. Suppose the public opinion information is , the weight is , the risk score is , then the comprehensive risk score is:

[0051] in, For the The sentiment score of the public opinion. For its weight.

[0052] Legal information extraction and verification: a. Extract basic case information from judgment documents and legal texts through a large-scale language model (AI model fine-tuned by LoRA), including plaintiff, defendant, cause of action, property clues, and cited legal provisions; b. Compare the extracted legal provisions with the current legal and regulatory database to verify the applicability and accuracy of the cited legal provisions; c. Comprehensively analyze case information and generate a comprehensive assessment of the company’s legal risks.

[0053] In legal information extraction and verification, regular expressions and deep learning models are used to process legal documents and extract property clues from legal documents, such as real estate, vehicles, bank accounts, and other information. Use regular expressions or named entity recognition (NER) models to identify property information. Associate the identified property information with the basic information of the case to ensure the integrity of the information. Suppose the legal document is , the extracted property information is , then the property clue extraction can be expressed as:

[0054] Sp6, data integration and report generation: a. Integrate the collected structured data, public opinion analysis results and legal information according to the preset report template; b. Automatically fill in various modules of the due diligence report, including corporate information overview, financial status analysis, public opinion and market evaluation, legal risk assessment and comprehensive conclusion; c. Ensure that the content of the generated report complies with legal and industry standards.

[0055] In data integration and report generation, use automatic typesetting algorithms to ensure that the report meets format requirements and automatically adjust the format of the report content, including chapters, tables, paragraphs, etc. Identify the chapter structure of the report and determine the layout of each part. Automatically adjust the font, line spacing, table format, etc. according to the report template. Suppose the report content is , the template is , then the automatic typesetting process is:

[0056] Sp7, report modification and output: a. The user adjusts the generated due diligence report through the interactive interface and proposes specific modification requirements; b. The system calls relevant modules to update the report content in real time according to user needs and regenerates the report; c. The final report is output as a Word document that complies with industry standards and supports export, sharing and printing.

[0057] Natural language processing techniques include the following steps: Convert the due diligence requirements entered by the user into a task instruction tree, and parse out the key information of the company name, report type, and target module; Based on the task instruction tree, semantic parsing technology is used to identify the logical structure of task requirements and dynamically generate dependencies between task modules; Dynamically call the semantic library and domain knowledge base to optimize input parsing results and ensure accurate understanding of industry-specific terminology.

[0058] NL2SQL technology achieves dynamic adaptation in the following ways: Provide standard SQL query templates, covering data fields and query requirements of different industrial and commercial, tax, and judicial databases; Use prompt word technology to convert natural language queries into SQL query statements, where prompt words include target fields, query scope, and filter conditions; The generated SQL statements are checked for legality through a syntax checker, and the statement structure, paging, sorting, and grouping operations are adjusted according to database characteristics.

[0059] The public opinion retrieval and analysis module includes the following functional modules: Data source selection module: prioritize retrieval sources from news, social media, and industry websites; Deduplication algorithm module: removes duplicate public opinion information through hash comparison and semantic similarity calculation; Time series analysis module: sort the public opinion content by time, giving priority to analyzing the latest released information; Risk scoring module: Generates a comprehensive risk score based on the sentiment tendency of public opinion, the popularity of dissemination and the authority of the publishing platform.

[0060] The key point extraction module specifically includes: Case information extraction: extract basic case fields from judgment documents, including case number, cause of action, and amount involved; Property clue analysis: Use regular expressions and language models to identify property clues from unstructured text, such as property location, vehicle information, and bank accounts; Legal article comparison and verification: Call the local legal and regulatory database and check the cited legal articles in the judgment documents one by one to ensure the accuracy and legality of the cited clauses; Multi-stage case linkage: Establish a case linkage network at different trial stages through the time and logical structure of the judgment, and evaluate the availability of case clues.

[0061] Report modification and output are automatically formatted through the following steps: Automatically identify the chapter and module structure in the report content, and adjust the font and paragraph alignment according to the preset template; Automatically adjust table contents according to column width and row height to ensure clarity of data display; Apply paragraph numbering rules and automatically generate a table of contents and cross-reference links to improve the readability and structure of your report. Specific embodiment 2: like Figure 1-5 As shown, a system for intelligently generating due diligence reports on financial non-performing assets includes the following modules: User interaction module: used to receive due diligence requirements and subsequent modification opinions; runs on a computer or mobile device, relying on input devices such as a touch screen, keyboard, and mouse to interact with the user. The hardware configuration requirements are relatively basic, usually devices that support a graphical user interface. The system must also have strong network transmission capabilities to ensure real-time interaction and feedback of data. The user interaction module is the front-end interface of the entire system, responsible for receiving due diligence requirements input by the user and displaying the report generated by the system. The module first receives the user's due diligence requirements through a user-friendly graphical interface or text box, such as the name of the company to be due diligenced, the content of the report, and the format requirements. After the input content is processed by natural language processing (NLP), the user's intention is parsed and converted into structured data that can be understood by the machine. Users can also modify the generated report through this module and put forward specific modification requirements. The system will respond and update the report in real time. This module needs to exchange data with the system's back-end modules (such as the intelligent reasoning planning module, data acquisition module, etc.) to ensure that the data and modifications entered by the user can be correctly transmitted and fed back in real time.

[0063] Intelligent reasoning planning module: used to formulate the execution plan of the due diligence process; the intelligent reasoning planning module usually relies on high-performance computing resources on the server side, especially processors (CPU / GPU) and large-scale memory to support efficient computing and data storage. This module usually works with deep learning servers, cloud computing platforms or clusters, and uses parallel computing to improve the efficiency of task planning. The main task of the intelligent reasoning planning module is to automatically formulate a task execution plan based on the due diligence requirements input by the user. The module first receives the due diligence requirements from the user interaction module, and then designs the execution process of the task according to the preset laws, industry standards and due diligence requirements. The execution plan includes multiple task modules such as data collection, public opinion analysis, and case information extraction, and determines the specific task objectives and execution order of each module. The intelligent reasoning planning module also needs to determine the dependencies between modules to ensure that the tasks are executed in a reasonable order. In addition, the module adjusts the task execution strategy through the intelligent agent framework and reasoning algorithm, and dynamically optimizes the steps in the report generation process to achieve the final goal more efficiently.

[0064] Data collection module: extracts data from structured databases through the NL2SQL tool; the data collection module usually needs to be connected to a database system (such as a relational database, NoSQL database, etc.) and has a certain amount of storage space for caching intermediate data. This module requires strong computing power and network connection, especially when obtaining a large amount of data from an external API or database, which requires high I / O performance. The core task of the data collection module is to extract the required due diligence data from the structured database through the NL2SQL technology. The module first receives the query requirements from the intelligent reasoning planning module and converts them into a data query format (SQL) suitable for a specific database. Using natural language to SQL (NL2SQL) technology, the module can automatically generate SQL query statements based on user input to query the company's basic information, financial data, tax records and other related data stored in the database. When executing SQL queries, the data collection module needs to ensure the accuracy of the query statements and optimize the syntax according to the characteristics of different databases to ensure query efficiency and data accuracy. The module should also be able to obtain data from different types of databases (such as relational databases, document databases) and handle format conversion between different data sources.

[0065] Public opinion retrieval module: used for public opinion information acquisition and sentiment analysis; the public opinion retrieval module needs to connect to multiple external platforms and information sources, relies on Internet connection and requires stable network bandwidth. In addition, in order to ensure the real-time nature of public opinion information, the public opinion retrieval module is often deployed in a high-performance computing environment, especially when processing large-scale data, which has high requirements for network and storage. The main function of the public opinion retrieval module is to obtain public opinion information related to the due diligence object from sources such as news platforms, social media, and industry websites, and perform sentiment analysis and risk assessment on it. The module first obtains real-time public opinion data from multiple data sources (such as Sina, Weibo, Zhihu, etc.) through API or crawler technology. Then, the obtained public opinion text is classified through sentiment analysis algorithms (such as LSTM, BERT, etc.) to identify the sentiment tendency (positive, negative, neutral) in the text. The public opinion retrieval module also assigns weights to public opinion information based on the spread heat (such as clicks, forwarding, comments, etc.) and platform authority, calculates a comprehensive public opinion risk score, and then outputs a public opinion summary and risk assessment report. This module supports batch processing and real-time updates to ensure comprehensive monitoring of public opinion on due diligence targets.

[0066] Key point extraction module: extracts key case information from unstructured text; key point extraction modules usually require powerful computing power, especially GPU-based accelerated computing, to process large amounts of unstructured text data. The module is usually integrated with a deep learning platform and can perform efficient batch processing on a cloud server or local server. The task of the key point extraction module is to extract key information from unstructured text (such as judgment documents, legal provisions, etc.) to ensure the accuracy of legal analysis and asset information in the due diligence report. The module first uses named entity recognition (NER) technology to extract basic case-related information (such as case number, cause of action, parties, etc.) from the text, and then uses regular expressions and deep learning models (such as BERT) to extract property clues (such as real estate, vehicles, bank accounts, etc.). For complex legal texts, the module can also understand the logical relationship between clauses through semantic reasoning technology, and associate different trial stages (such as first instance and second instance) in the same case to generate a complete case analysis. The key point extraction module is also responsible for comparing and verifying the extracted legal provisions to ensure that the cited legal provisions are consistent with the current law.

[0067] Information integration module: Integrate data and fill it into the report template; the information integration module usually requires large-capacity memory and efficient data processing capabilities, especially when integrating data from multiple data sources (such as structured data, public opinion analysis data, legal analysis data, etc.). This module is usually deployed on high-performance computers or servers to support real-time information processing. The function of the information integration module is to integrate the data provided by each module and fill it into the preset report template. The module first obtains the structured information of the enterprise from the data collection module, obtains the public opinion risk data from the public opinion retrieval module, and obtains the legal analysis data from the key point extraction module. The integration module needs to associate data from different sources (industrial and commercial information, financial data, public opinion analysis results, and legal document analysis results) to ensure information consistency and integrity. Then, the module fills the integrated information into different chapters of the report according to the preset report template. The report structure includes company profile, financial status, public opinion analysis, legal risk and other parts. The information integration module needs to dynamically generate the report content according to the template and adjust the presentation of the content according to the requirements of the report chapter.

[0068] Report generation module: used to generate standardized due diligence reports and support dynamic modification. Report generation modules usually rely on powerful computing and storage devices, especially when generating reports, they have high requirements for format processing and content filling. This module also needs to support file export and printing functions, so it needs to have the ability to connect to a printer or document management system. The core task of the report generation module is to integrate the final information output by the module and generate a compliant due diligence report in a preset format. The module first typeset the layout, chapter structure and font style of the report according to the report template, and automatically fills in the integrated content. The specific content includes the company's basic information, financial data, public opinion analysis results, legal risk assessment, etc. The report generation module supports dynamic modification. Users can make modification suggestions through the user interaction module after the report is generated, and the module will automatically update the report content. In addition, the module will automatically adjust the table format, paragraph numbering, title and directory generation according to the report template, and finally output the report in Word or PDF format that meets industry standards for users to download, share or print.

[0069] The user interaction module is based on natural language processing technology. It understands the user's customized modification requirements through semantic analysis and dynamically calls other modules to update the report content.

[0070] The information integration module dynamically calls the multi-task processing framework and schedules the data spreadsheet tool and public opinion retrieval and analysis module in real time to ensure the consistency and real-time nature of the report content and data.

[0071] The public opinion retrieval module generates the public opinion risk index of the due diligence object based on the sentiment analysis model and risk factor weighting algorithm. Specific embodiment three

[0073] Based on the technical solutions of the first and second specific embodiments, further description is given as follows: Working steps see Figure 1 : 1. Obtain due diligence requirements: In the first step of the present invention, the system obtains due diligence requirements through human-computer interaction dialogue. The user (such as due diligence personnel or financial institutions) interacts with the system and inputs the name of the debtor company to be due diligenced. The system understands the user input through natural language processing technology, and automatically analyzes and determines the required due diligence information types and report structures according to the legal and industry requirements in the financial non-performing asset due diligence scenario. Specifically, the system identifies and extracts key data types that require due diligence, such as basic corporate information, financial status, debt performance, legal dispute records, etc., according to different legal and regulatory requirements. At the same time, the system will automatically generate an adaptive due diligence report structure to clarify the various information modules and format requirements that should be included in the report. This process ensures that the content of the report complies with relevant legal regulations and can meet the specific needs of different due diligence scenarios, laying the foundation for subsequent data collection and analysis.

[0074] 2. Intelligent reasoning planning: In the second step of the present invention, the system automatically formulates a detailed work plan based on the due diligence scenario requirements input by the user through the reasoning planning tool in the intelligent agent framework. First, the system analyzes the legal requirements and business objectives of the due diligence scenario, and identifies the core data and information categories required for report generation, such as corporate financial data, legal risks, debt fulfillment, etc. Then, based on these requirements, the system intelligently infers the operating steps and execution strategies for each task link, including data collection, analysis methods and required tools (such as data tables, online retrieval or key point extraction tools). At the same time, the system will also generate corresponding prompt word contexts based on the report structure to guide the large model to generate accurate report content. The system will automatically derive the required legal terms, financial analysis methods and industry standards, providing clear guidance for subsequent data queries and report content filling. This reasoning planning process ensures the efficiency, intelligence and accuracy of due diligence report generation, while meeting the specific requirements of laws and industries.

[0075] 3. Data table: In the third step of the present invention, the system automatically performs data query for task links involving structured data such as judicial, industrial and commercial, taxation, and electricity through the data table tool in the intelligent body framework, and generates specific query results that meet the due diligence requirements based on the set report templates and screening conditions. Figure 2 .

[0076] The specific process is as follows: 3.1 Query condition setting: First, based on the results of the previous step of reasoning and planning, the system combines the report template to set up detailed task links and clarify the information types required for each query link. At this point, the system automatically extracts relevant data fields, such as business information, tax records, environmental penalties, judicial litigation, restrictions on high consumption, etc. The system will also set screening conditions, output fields, sorting methods, etc. according to the report structure to ensure the accuracy and completeness of the query.

[0077] 3.2 Query template construction: The system automatically builds a query template based on the user's due diligence requirements and set screening conditions. The template will contain specific database tables, field names, screening conditions, sorting methods, etc. to ensure that relevant data is extracted from different data sources. In order to adapt to the differences between different databases, the system can generate customized query templates for different data sources.

[0078] 3.3 Prompt word auxiliary tools: During this process, the system also uses prompt word auxiliary tools to convert the constructed query template into a prompt word context that adapts to the big model. The system maps the various components of the query template (such as database table names, field names, filter conditions, etc.) into a prompt word structure that is easy for the big model to understand and process. These prompt word contexts can help the big model better understand the intent and structure of the query, thereby improving the efficiency and accuracy of the query.

[0079] 3.4 Generate query requests using NL2SQL technology: The system uses the NL2SQL (natural language to SQL) technology in the big model to convert the constructed prompt word context into an actual SQL query request. The big model will understand the user's query needs based on the context information, and automatically execute SQL queries to extract structured data related to the due diligence task. For example, query whether the debtor company is involved in judicial cases from the judicial database, extract the company's basic industrial and commercial registration information from the industrial and commercial database, and retrieve major tax violation records from the tax database.

[0080] 3.5 Data return: After the query results are returned, the system will organize the data obtained from the database into a Markdown table format that conforms to the report format. These tables are arranged according to the preset output fields and sorting methods to ensure that each data is clearly presented to facilitate subsequent analysis and decision-making. The table content includes but is not limited to information such as company name, registration number, tax violation record, environmental protection penalty record, and judicial case.

[0081] The system may return query results in the following format: |Company name|Registration number|Legal litigation|Environmental penalties|Tax violations|Restrictions on high consumption| |------------|-----------|------------|----------|---------- |------------| |XX Company|12345678|Under litigation|No|Major tax violation|Yes| |YY Company|87654321|Case closed|Minor fine|None|No| 3.6 Content Generation: The acquired structured data will be automatically filled into the corresponding position of the report template. The system will automatically integrate this information with other data sources (such as online search, key point extraction) according to the template settings to form a complete due diligence report content. Users can view and conduct a final review of the generated report in the system.

[0082] Through this process, the system can efficiently and accurately query and organize structured data in the fields of justice, industry and commerce, taxation, etc., provide comprehensive and authoritative support for due diligence reports on financial non-performing assets, and greatly improve work efficiency and accuracy.

[0083] 4. Online search: In the fourth step of the present invention, the system uses the online search tool in the intelligent agent framework to comprehensively search, analyze and evaluate the public opinion information related to the debtor enterprise to be investigated, and finally forms a public opinion risk judgment and presents the relevant information in the due diligence report. Figure 3 The specific steps are as follows: 4.1 Public opinion information retrieval: The system first uses the online search tool in the intelligent agent framework to automatically search for public opinion articles related to the debtor company to be investigated from multiple authoritative and widely covered online platforms. These platforms include but are not limited to: News platforms: such as Sohu News, Toutiao, Sina News, NetEase News, etc.; Social media platforms: such as WeChat, Douyin, Kuaishou, Guba, Xueqiu, etc.; Financial platforms: such as China Finance Network, Securities Star, and Flush, etc.

[0084] Through extensive coverage of multiple data sources, the system can ensure that various public opinion information related to the enterprise is retrieved, especially negative news or comments that may affect the enterprise's reputation, market performance or financial situation.

[0085] 4.2 Data deduplication and cleaning: In order to avoid retrieving duplicate public opinion articles, the system uses a deduplication algorithm to clean and deduplicate search results. The elimination of duplicate information can ensure that the public opinion data for the final analysis is not interfered with by redundant content, improving data quality and processing efficiency. This process not only eliminates multiple occurrences of the same news source, but also removes some invalid information that is not related to the debtor's company.

[0086] 4.3 Hotness Statistics Sorting: The system counts the popularity of public opinion articles after deduplication, and evaluates the social attention and spread of each article based on the number of clicks, forwarding, comments and other data of the article. At the same time, the system sorts the articles according to the time of publication to ensure that the latest and most timely public opinion information is processed first. This ensures that the public opinion analysis results reflect the current public opinion atmosphere of the enterprise, rather than outdated information.

[0087] 4.4 Public Opinion Summary Generation: After information screening and sorting, the system will use the big model to summarize the retrieved public opinion articles. The big model uses natural language processing (NLP) technology to extract key information from each article and generate a concise public opinion summary. The system automatically identifies key information in the article, such as corporate disputes, negative events, legal proceedings, etc., and converts these contents into concise summaries for subsequent analysis and decision-making.

[0088] 4.5 Sentiment Analysis: The system performs sentiment analysis on the public opinion summary. Through the sentiment analysis model, the system can identify the sentiment of each public opinion article and determine whether it is positive, negative or neutral. The specific analysis steps include: Sentiment dictionary training: Build sentiment vocabulary that is suitable for financial non-performing asset scenarios. These vocabulary usually include positive words, negative words, and neutral words. Through manual annotation, dictionary expansion, and automated learning, a sentiment vocabulary library covering this field is established.

[0089] Sentiment category prediction: The large model predicts the sentiment category of the input text based on the contextual information of the input text. The article uses a large number of positive sentiment words, such as "success", "leadership", "innovation", etc., and is predicted to be positive public opinion. The article uses a large number of negative sentiment words, such as "pollution", "violation", "prosecution", etc., and is predicted to be negative public opinion.

[0090] Comprehensive tendency judgment: Through semantic understanding of the context, the large model can determine whether the sentiment in the article is consistent with the overall context of the article. Even if the sentiment words are relatively neutral, their sentiment tendency can be judged based on the context. The model can identify the different sentiment meanings of a word in different contexts. The word "innovation" is usually positive in the field of technology, but it may have implicit negative sentiment in some business fields.

[0091] The core goal of sentiment analysis is to objectively and accurately evaluate the emotional bias of public opinion content through machine learning algorithms, thereby providing a strong basis for subsequent risk assessment.

[0092] 4.6 Public Opinion Risk Assessment: After completing the sentiment analysis, the system comprehensively considers the public opinion sentiment, article popularity and time sensitivity of each article, and finally comes up with a public opinion risk score and judgment for the enterprise. The specific judgment steps include: Public opinion sentiment weighting: The system gives a corresponding weight to each public opinion article based on its sentiment and popularity. Negative public opinion articles may be weighted higher than positive ones, especially when they involve major legal risks or social responsibility incidents of the enterprise.

[0093] Risk assessment model: Combining the sentiment of public opinion articles and historical public opinion data, the system conducts a comprehensive assessment of corporate public opinion risks through machine learning models (such as support vector machines, random forests, etc.). The model analyzes the relationship between corporate public opinion fluctuations and other business risks (such as financial status, legal disputes, etc.) to form a comprehensive public opinion risk index.

[0094] Public opinion risk assessment: Finally, the system generates a public opinion risk level for the company, such as "low risk", "medium risk" or "high risk". High-risk companies may face serious damage to their public image or potential legal disputes. This information will have an important impact on the preparation of due diligence reports and risk management decisions.

[0095] 4.7 Report content generation: After judging the public opinion risk, the system will integrate the public opinion analysis results into the due diligence report. The report will present the public opinion summary, sentiment analysis results, popularity statistics and risk judgment in detail to ensure that due diligence personnel can fully understand the public opinion status of the company. In addition, public opinion risk assessment will be used as one of the decision-making bases to help due diligence personnel judge the overall risk level of the company.

[0096] Through this series of intelligent public opinion analysis and risk assessment, the system can provide more accurate corporate public opinion information for due diligence reports on financial non-performing assets, help financial institutions fully and deeply understand the social reputation and potential risks of enterprises, and further support asset evaluation and decision-making.

[0097] 5. Key points extraction: In the fifth step of the present invention, the system uses the key point extraction tool in the intelligent agent framework, utilizes the large model base with tens of billions of parameters, and combines the local legal and regulatory database and judicial document database to accurately extract the key information required for the due diligence of non-performing assets, further improving the legal professionalism and accuracy of the report. Figure 4 The specific process is as follows: 5.1 Large Model Fine-tuning: The base model uses the Qwen2.5-14B large model developed by Alibaba Cloud, with a parameter scale of 14 billion. It uses data of more than 3 trillion tokens for pre-training and has strong natural language understanding and reasoning capabilities. Due diligence of financial non-performing assets involves a large amount of legal texts and industry-specific knowledge. The system uses LoRA fine-tuning technology to train proprietary corpus in the legal field (such as laws, regulations, and judicial documents) to improve the model's vertical field reasoning ability in legal text processing. This fine-tuning process enables the model to more accurately understand and extract complex information related to the law, ensuring that various complex cases and legal clauses can be handled in legal due diligence.

[0098] 5.2 Accurately extract information: The system uses a fine-tuned large model to accurately extract a series of key fields from judicial documents. These fields are the basic data for legal due diligence and non-performing asset evaluation. For example: Basic case information: plaintiff, defendant, court, case number, cause of action, subject matter, etc.

[0099] Property clues: property location, property valuation, vehicle model, bank account number, frozen amount, actual controlled amount, etc.

[0100] The pattern of basic case information is relatively fixed, and it can be accurately extracted through large models and pattern matching.

[0101] The real estate location information is a combination of multiple elements such as province, city, district, street, building, unit, house number, type, etc., which is extracted through a large model plus prompt word engineering and regular expressions.

[0102] Vehicle model information requires the collection of automobile brand information from various car manufacturers and is extracted through large models, prompt word engineering, and semantic disambiguation.

[0103] The bank account number, frozen amount, and actual controlled amount need to refer to the business logic of non-performing asset preservation, and extract and associate key points through large models, bank entity recognition, sentence / inter-sentence position information, and chapter information.

[0104] 5.3 Case chain analysis: The system can connect multiple trial records of the same case, such as first instance, second instance, judgment and execution cases, and establish the time and legal logic relationship between cases. Through these connections, the system can determine whether the property clues of the case are timely and effective, and whether there are potential legal risks. A property clue was not effectively frozen in the first instance, but it may become crucial in the execution case. The system comprehensively analyzes the judgment information of each trial stage, evaluates the effectiveness, timeliness and potential value of property clues, and helps due diligence personnel more accurately judge the feasibility of disposing of non-performing assets.

[0105] 5.4 Legal provisions verification: In the process of legal due diligence, the citation of legal provisions in judicial documents is the core of analyzing cases and judging their legal risks. The system compares the legal provisions in judicial documents with the original legal texts in the local legal database to ensure that each cited legal provision matches the current laws and regulations and verify whether it is properly applied in the case. Through this comparison, the system can automatically identify whether the application of laws in judicial documents complies with current legal norms, and improve the legal professionalism of due diligence reports.

[0106] When a judgment document cites a legal provision related to debt performance, the system compares it with the relevant provisions in the latest Contract Law or Civil Code to ensure that the cited legal provision is accurate. If there is any inconsistency, the system will automatically mark it and remind the due diligence personnel to further verify it to ensure that the legal basis and analysis results of the report are accurate.

[0107] 5.5 Report content generation: Through accurate extraction of key points and comparison of legal provisions, the system can not only provide detailed information on cases and properties, but also enhance the legal professionalism in due diligence reports. The report will comprehensively display the legal risks, property status and potential asset value of the enterprise based on the legal provisions and court decisions in the judgment documents, combined with the analysis results of property clues. This information will help financial institutions more accurately assess the legal status and market prospects of non-performing assets, so as to make more scientific investment decisions.

[0108] Through this process, the system can effectively extract, correlate and analyze key information in legal cases, and accurately compare them with the latest laws and regulations, greatly improving the quality and professionalism of due diligence reports on financial non-performing assets. This function is of great significance for asset evaluation, risk control and decision support.

[0109] 6. Information Integration: In the sixth step, the system effectively integrates the data table results, online search summaries, and key point extraction results obtained in the previous step under the guidance of the reasoning planning tool. Specifically, the system combines the corporate financial, judicial, industrial and commercial information obtained from the data table with the public opinion summary retrieved online to form a comprehensive view of the company. At the same time, the system embeds key information extracted from judicial documents and legal clauses into the report to ensure the legal professionalism of the report. The integrated information will be automatically filled in according to the report template to generate a detailed, clearly structured due diligence report that complies with the requirements of laws and regulations, covering key information such as the company's risk assessment, asset analysis, and legal background, to help financial institutions make scientific decisions.

[0110] 7. Interactive modification: In the seventh step, the system allows users to make further modifications and adjustments through human-computer interactive dialogue after the main body of the report is generated. Users can request modifications to specific chapters, table contents, or data parts of the report. The system analyzes the user's modification requirements through natural language understanding technology and accurately identifies the target content of the modification and the additional information required.

[0111] According to the user's needs, the agent will automatically retrieve the relevant chapter context and call the corresponding tools according to the modification requirements: Data spreadsheet tools: If a user requests an update to a company's financial data or legal records, the system will requery the database to obtain the latest information.

[0112] Online search tools: If the user requests to add the latest public opinion data or industry trends, the system will search for relevant news and social media content.

[0113] Key points extraction tool: If the user requests to modify the specific content of the legal document analysis or the judgment document, the system will re-extract the latest legal provisions or judgment points.

[0114] After acquiring new information, the system will automatically integrate it into the report and update the corresponding chapters or tables to ensure the accuracy and timeliness of the report content. Finally, the revised report will be re-presented to the user to ensure that the report meets actual needs and user expectations.

[0115] 8. Report output: In the eighth step, the system automatically fills in the integrated information in the corresponding position of the report according to the preset report template. For information such as query results, public opinion summaries and legal points, the system will ensure that they are accurately filled in the various chapters of the report. For tables presented in Markdown format, the system will automatically convert them into standard Word tables and adjust their formats according to the requirements of the report template. The system will accurately set the font, font size, indentation, line spacing, paragraph spacing and other layout details to ensure that the report style is consistent and professional. At the same time, the column width, row height, cell alignment, etc. of the table will also be automatically adjusted according to the template, so that the final report is both compliant with the specifications and has good readability. Finally, the system will generate a Word document that conforms to the template format for users to download and further use.

[0116] It should be made clear that the mentor intelligent agent in this case is a packaged AI system.

[0117] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0118] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently generating a due diligence report on financial non-performing assets, characterized in that: The method comprises the following steps: Sp1. Obtaining due diligence requirements: a. Receive due diligence requirements input by the user through the human-computer interaction interface, including the name of the company to be investigated, the format and content requirements of the due diligence report; b. Use natural language processing technology to parse user input, identify key elements and target information in due diligence tasks, and dynamically generate report requirement templates based on laws, regulations and industry standards in the field of financial non-performing assets; Sp2, Intelligent Reasoning Planning: a. Based on the multi-agent framework, combined with the preset legal and regulatory database and due diligence standards, design the due diligence report generation process; b. Construct an execution plan consisting of task modules according to user needs and task characteristics. The task modules include data collection module, public opinion retrieval module and key point extraction module, and formulate task objectives and output standards for each module; Sp3, data collection and processing: a. Call the structured data query module through the data table tool to extract key fields from external data sources such as industry and commerce, taxation, justice, and electricity; b. Use NL2SQL technology to convert natural language query requirements into SQL statements that are suitable for different databases; c. Automatically generate table contents that meet the reporting requirements, including basic information of the enterprise, tax records, judicial litigation data and asset and liability details; Sp4, public opinion retrieval and analysis module: a. Use online search tools to search for public opinion information related to the due diligence target on news platforms, social media and financial websites; b. Apply sentiment analysis algorithms to identify the sentiment tendency of public opinion content and perform risk scoring based on click and comment information; c. Output public opinion summaries and risk assessment reports to clarify risk information related to corporate reputation and market behavior; Sp5, Legal Information Extraction and Verification: a. Extract basic case information from judicial documents and legal texts through large-scale language models, including plaintiff, defendant, cause of action, property clues, and cited laws; b. Compare the extracted legal provisions with the current legal and regulatory database to verify the applicability and accuracy of the cited legal provisions; c. Comprehensively analyze case information and generate comprehensive assessment results of corporate legal risks; Sp6, data integration and report generation: a. Integrate the collected structured data, public opinion analysis results and legal information according to the preset report template; b. Automatically fill in various modules of the due diligence report, including corporate information overview, financial status analysis, public opinion and market evaluation, legal risk assessment and comprehensive conclusion; c. Ensure that the content of the generated report complies with laws and industry standards; Sp7, report modification and output: a. The user adjusts the generated due diligence report through the interactive interface and proposes specific modification requirements; b. The system calls relevant modules to update the report content in real time according to user needs and regenerates the report; c. The final report is output as a Word document that complies with industry standards and supports export, sharing and printing.

2. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The natural language processing technique comprises the following steps: Convert the due diligence requirements entered by the user into a task instruction tree, and parse out the key information of the company name, report type, and target module; Based on the task instruction tree, semantic parsing technology is used to identify the logical structure of task requirements and dynamically generate dependencies between task modules; Dynamically call the semantic library and domain knowledge base to optimize input parsing results and ensure accurate understanding of industry-specific terminology.

3. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The NL2SQL technology achieves dynamic adaptation in the following ways: Provide standard SQL query templates, covering data fields and query requirements of different industrial and commercial, tax, and judicial databases; Use prompt word technology to convert natural language queries into SQL query statements, where prompt words include target fields, query scope, and filter conditions; The generated SQL statements are checked for legality through a syntax checker, and the statement structure, paging, sorting, and grouping operations are adjusted according to database characteristics.

4. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The public opinion retrieval and analysis module includes the following functional modules: Data source selection module: prioritize retrieval sources from news, social media, and industry websites; Deduplication algorithm module: removes duplicate public opinion information through hash comparison and semantic similarity calculation; Time series analysis module: sort the public opinion content by time, giving priority to analyzing the latest released information; Risk scoring module: Generates a comprehensive risk score based on the sentiment tendency of public opinion, the popularity of dissemination and the authority of the publishing platform.

5. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The key point extraction module specifically includes: Case information extraction: extract basic case fields from judgment documents, including case number, cause of action, and amount involved; Property clue analysis: Use regular expressions and language models to identify property clues from unstructured text, such as property location, vehicle information, and bank accounts; Legal article comparison and verification: Call the local legal and regulatory database and check the cited legal articles in the judgment documents one by one to ensure the accuracy and legality of the cited clauses; Multi-stage case linkage: Establish a case linkage network at different trial stages through the time and logical structure of the judgment, and evaluate the availability of case clues.

6. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The report modification and output are automatically formatted through the following steps: Automatically identify the chapter and module structure in the report content, and adjust the font and paragraph alignment according to the preset template; Automatically adjust table contents according to column width and row height to ensure clarity of data display; Apply paragraph numbering rules and automatically generate a table of contents and cross-reference links to improve the readability and structure of your report.

7. A system for intelligently generating a due diligence report on financial non-performing assets according to any one of claims 1 to 6, characterized in that: The system includes the following modules: User interaction module: used to receive due diligence requirements and subsequent modification opinions; Intelligent reasoning planning module: used to formulate an execution plan for the due diligence process; Data collection module: extract data from structured databases through NL2SQL tools; Public opinion retrieval module: used for obtaining public opinion information and analyzing sentiment tendency; Key point extraction module: extract key case information from unstructured text; Information integration module: integrates data and fills it into report templates; Report generation module: used to generate standardized due diligence reports and supports dynamic modification.

8. The system for intelligently generating a due diligence report on financial non-performing assets according to claim 7, characterized in that: The user interaction module is based on natural language processing technology, understands the user's customized modification requirements through semantic analysis, and dynamically calls other modules to update the report content.

9. The system for intelligently generating a due diligence report on financial non-performing assets according to claim 7, characterized in that: The information integration module dynamically calls the multi-task processing framework, and dispatches the data spreadsheet tool and the public opinion retrieval and analysis module in real time to ensure the consistency and real-time nature of the report content and data.

10. The method for intelligently generating a due diligence report on financial non-performing assets according to claim 1, characterized in that: The public opinion retrieval module generates a public opinion risk index for the due diligence object based on a sentiment analysis model and a risk factor weighting algorithm.

Images