Breakdown data analysis method and device based on large model and medium

Abstract

The invention discloses a large model-based production breaking data analysis method, equipment and a medium. The method comprises the following steps: acquiring multi-channel production breaking data through a distributed data acquisition system, and realizing automatic classified storage of the data by adopting OCR and text extraction technologies; a deep learning model is utilized to extract a production breaking field entity, and an enterprise production breaking knowledge graph is constructed through a mutual attention mechanism; an oversampling algorithm is applied to process a data imbalance problem, and a multi-dimensional feature system is constructed in combination with a knowledge graph; a Qwen-plus large language model and an improved V-detector algorithm are adopted to realize dynamic risk early warning; similar historical cases are retrieved based on the knowledge graph, and a personalized disposal scheme is generated in combination with rule reasoning; according to the method, the knowledge graph and the deep learning technology are creatively fused, the dynamic monitoring, accurate early warning and intelligent decision support of the production breaking risk are realized, and the accuracy and timeliness of the production breaking data analysis are remarkably improved. The method can be widely applied to the field of financial institution, court and enterprise risk management.

Description

Technical Field

[0001] This document relates to the field of data analysis technology, and in particular to a bankruptcy data analysis method, device and medium based on a large model. Background Technology

[0002] In the booming wave of the digital economy, the amount of data generated during enterprise operations is growing exponentially, and the diversity and complexity of this data are also constantly increasing. This massive amount of data contains rich information about enterprise operations and the market environment, becoming an important basis for enterprises to understand their own development trends and prevent operational risks. Bankruptcy data analysis, as a key component of enterprise risk management, aims to accurately identify potential bankruptcy risks through in-depth analysis of various types of data, providing strong support for enterprises to formulate risk response strategies and for regulatory agencies to implement effective supervision.

[0003] However, current traditional bankruptcy data analysis techniques have revealed the following drawbacks in the context of big data:

[0004] 1. Insufficient processing capability for multi-source heterogeneous data: Existing technologies are unable to efficiently analyze implicit bankruptcy risk signals in unstructured text, resulting in a high false negative rate;

[0005] 2. Weak ability to capture dynamic risk factors: Traditional methods rely on static rule bases and cannot adapt in real time to the impact of dynamic factors such as economic environment and policy changes on bankruptcy risk;

[0006] 3. Insufficient contextual understanding and semantic understanding: When extracting key information from long texts, existing models lack in-depth analysis of industry background and upstream and downstream relationships of enterprises, leading to misjudgments, such as mislabeling short-term cash flow problems as bankruptcy risks.

[0007] Therefore, there is an urgent need to integrate the diverse processing capabilities of big data technologies to construct an intelligent analysis system covering the entire data lifecycle. By introducing Natural Language Processing (NLP) technology to parse unstructured text, utilizing knowledge graphs to model industry ecosystem relationships, and combining dynamic machine learning algorithms to capture economic environmental variables in real time, this approach overcomes the bottlenecks of traditional methods in terms of data dimensions, timeliness, and semantic understanding. This will improve the accuracy and foresight of bankruptcy risk identification, providing more timely technical support for enterprise risk prevention and regulatory decision-making. Summary of the Invention

[0008] This invention provides a bankruptcy data analysis method, device, and medium based on a large model, aiming to solve the above-mentioned problems.

[0009] According to an embodiment of the present invention, a bankruptcy data analysis method based on a large model is provided, comprising:

[0010] S1. Obtain bankruptcy-related data from multiple channels through a pre-built distributed data acquisition system, and automatically classify and process the bankruptcy-related data according to its editability before storing it in a dedicated database; wherein, the automatic classification and processing specifically includes: directly extracting text from editable documents, extracting text from non-editable documents after OCR recognition, and storing the processed structured data in a dedicated database, where OCR is optical character recognition technology;

[0011] S2. A deep learning-based entity recognition model is used to extract bankruptcy-related entities from the dedicated database. An entity relationship extraction model is constructed through a mutual attention mechanism. Based on the extracted entities and relationships, an enterprise bankruptcy knowledge graph is constructed and stored in the Neo4j graph database.

[0012] S3. Apply oversampling algorithm to address the imbalance problem of bankruptcy-related data, screen core feature variables through chi-square test, and construct a multi-dimensional feature system by combining entity attributes of enterprise bankruptcy knowledge graph;

[0013] S4. Obtain incremental update data from the dedicated database, input the incremental update data into the Qwen-plus big language model to extract bankruptcy progress features, construct a risk identifier based on the improved V-detector algorithm, and dynamically generate risk warnings by combining the risk transmission path of the enterprise bankruptcy knowledge graph.

[0014] S5. Extract historical case data from the enterprise bankruptcy knowledge graph and represent it as a triple of <entity, relationship, disposal result>. Search for historical cases similar to the current case using cosine similarity and generate a personalized disposal plan by combining rule-based reasoning.

[0015] S6. Receive real-time dynamic data from enterprises, continuously update the knowledge graph through meta-learning algorithms, and optimize model parameters and early warning thresholds based on few-shot learning.

[0016] According to an embodiment of the present invention, an electronic device is provided, comprising:

[0017] Processor; and,

[0018] A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the bankruptcy data analysis method based on the large model described above.

[0019] According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the steps of the bankruptcy data analysis method based on the large model described above.

[0020] This invention addresses the problem of low efficiency in processing multi-source heterogeneous bankruptcy data by employing a distributed data acquisition system and intelligent classification technology, achieving efficient collection and structured storage of bankruptcy-related data. Through deep learning-based entity recognition models and knowledge graph construction technology, it overcomes the shortcomings of traditional methods in analyzing the correlation of bankruptcy risk factors, enabling comprehensive modeling and visual analysis of corporate bankruptcy risks. An improved V-detector algorithm and dynamic feature evolution model address the issue of poor timeliness in static risk assessment models, enabling real-time monitoring and accurate early warning of bankruptcy risks. Historical case retrieval and Monte Carlo simulation technology address the lack of personalized references for bankruptcy resolution plans, achieving data-driven intelligent decision support. This application integrates cutting-edge technologies such as knowledge graphs and large language models to construct an intelligent analysis system covering the entire process of data collection, risk identification, and early warning decision-making, significantly improving the accuracy, timeliness, and practicality of bankruptcy risk analysis, providing reliable risk management tools for financial institutions, judicial organs, and enterprises. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a bankruptcy data analysis method based on a large model, according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the automatic classification processing based on the editability of bankruptcy-related data in an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0025] Method Implementation Examples

[0026] According to embodiments of the present invention, a method, apparatus, and system for bankruptcy data analysis based on a large model are provided. Figure 1 This is a flowchart of a bankruptcy data analysis method based on a large model, according to an embodiment of the present invention. Figure 1 As shown, the bankruptcy data analysis method based on a large model according to an embodiment of the present invention includes:

[0027] S1. Obtain bankruptcy-related data from multiple channels through a pre-built distributed data acquisition system, and automatically classify and process the bankruptcy-related data according to its editability before storing it in a dedicated database; wherein, the automatic classification and processing specifically includes: directly extracting text from editable documents, extracting text from non-editable documents after OCR recognition, and storing the processed structured data in a dedicated database. Figure 2 This is a schematic diagram illustrating the automatic classification processing based on the editability of bankruptcy-related data according to an embodiment of the present invention;

[0028] First, relevant data on corporate bankruptcy is collected from websites such as government announcements and court notices. This data is then cleaned and preprocessed to remove invalid information and standardize its format. Next, the processed structured data is stored in a dedicated database for efficient querying and analysis. Finally, the progress of bankruptcy cases is assessed based on specific corporate bankruptcy data in the database. This process not only improves decision-making efficiency and reduces labor costs, but also enhances the ability to predict future trends through comprehensive analysis of historical and current data. This helps to rationally allocate resources, optimize the allocation of judicial and social resources, increase the transparency and fairness of bankruptcy case handling, provide important references for decision-makers, and promote stable economic and social development.

[0029] Furthermore, in a specific implementation of this embodiment of the invention, S1 includes:

[0030] S11: Construct an RPA (Robotic Process Automation) distributed data acquisition system based on the enterprise bankruptcy reorganization network;

[0031] S12: Collect the list pages of corporate bankruptcy reorganization websites, obtain the details page links, and obtain full and incremental data;

[0032] S13: Download relevant bankruptcy documents via RPA access to the details page link;

[0033] S14: Store the title, publication date, and bankruptcy document in OSS (Object Storage Service).

[0034] By constructing an RPA-based distributed data acquisition system for corporate bankruptcy reorganization websites, human operations can be simulated to automatically access target websites and retrieve required information. This highly automated process not only covers the accurate identification and retrieval of links to corporate bankruptcy reorganization website listing pages and detail pages, but also includes the efficient acquisition of full and incremental data. In the specific implementation process, RPA tools are programmed to perform a series of complex tasks, including but not limited to navigating to specific web pages, parsing page content, clicking detail page links, and downloading relevant bankruptcy documents. These operations ensure the integrity and accuracy of the collected data, while significantly reducing the time and resources required for traditional manual processing. More importantly, to achieve long-term data preservation and efficient management, the system uploads and stores the retrieved titles, publication dates, and bankruptcy documents to OSS, a highly available and scalable cloud storage solution. This not only ensures the data's longevity but also... The system provides security and accessibility, as well as flexible data access capabilities for enterprises, supporting subsequent data analysis and application needs. The highly automated nature of the entire process significantly improves the efficiency and quality of data collection, effectively avoiding errors and omissions that may be caused by human operation. In addition, the application of this technical solution has profound significance for enterprises. It can not only help enterprises keep abreast of the latest bankruptcy reorganization information and provide decision support for management, but also promote resource sharing and optimal allocation among enterprises to a certain extent, driving the overall development of the industry. Through rapid response to market dynamics, enterprises can more calmly face fierce market competition, formulate more scientific and reasonable strategic plans, and thus maintain a competitive advantage in the ever-changing business environment. In conclusion, the RPA-based enterprise bankruptcy reorganization network data collection system is not only a manifestation of technological innovation, but also an effective means to improve enterprise management level and market competitiveness.

[0035] The distributed data acquisition system of this invention includes:

[0036] The dynamic load balancing module dynamically adjusts the crawler resource allocation strategy based on reinforcement learning algorithms, monitors the response latency, CAPTCHA trigger frequency and blocking probability of each data source in real time, and optimizes resource allocation weights through Q-learning algorithm;

[0037] The anti-crawler strategy module includes a request fingerprint randomization component to dynamically generate browser fingerprints, a hybrid proxy pool component to integrate a three-layer proxy architecture of data center IP, residential IP and TOR network, and a behavior simulation engine component to simulate human operation trajectories based on generative adversarial networks.

[0038] The data quality verifier uses an anomaly detection model based on the Transformer architecture to identify dirty data and implements field-level integrity verification rules, including non-empty verification, logical conflict detection, and statistical distribution verification. It also triggers an automatic completion process for incomplete data.

[0039] S2. A deep learning-based entity recognition model is used to extract bankruptcy-related entities from the dedicated database. An entity relationship extraction model is constructed through a mutual attention mechanism. Based on the extracted entities and relationships, an enterprise bankruptcy knowledge graph is constructed and stored in the Neo4j graph database.

[0040] The entity recognition model based on deep learning in this embodiment of the invention adopts an ELMo-BiLSTM-CRF hybrid architecture, and the specific construction steps include:

[0041] Based on the pre-training of general corpus, a second-stage pre-training is carried out using bankruptcy legal documents and financial reports. A gradient inversion layer is introduced to eliminate domain distribution differences and jointly optimize the adversarial loss of the domain classifier and entity recognizer.

[0042] A conceptual system for the bankruptcy domain is constructed, with a professional terminology dictionary embedded in the model as an external memory module. Relevant terms are dynamically retrieved through a key-value attention mechanism, and a knowledge graph-based transition constraint matrix is ​​added to the CRF layer to restrict the generation of illegal entity sequences.

[0043] The bottom encoder uses cosine annealing for learning rate scheduling, with the initial learning rate set to the first preset value. The middle transition layer uses a linear warming strategy to gradually increase the learning rate to the second preset value. The top classifier uses a linear decay strategy with warmup, applies L2 regularization to the CRF layer parameters separately, and introduces an adaptive learning rate adjustment factor. The top learning rate step size is dynamically adjusted according to the loss difference between the domain classifier and the entity recognizer, where the second preset value is a preset multiple of the first preset value.

[0044] S3. Apply oversampling algorithm to address the imbalance problem of bankruptcy-related data, screen core feature variables through chi-square test, and construct a multi-dimensional feature system by combining entity attributes of enterprise bankruptcy knowledge graph;

[0045] The oversampling algorithm is the SMOTE-Borderline1-GAN algorithm, specifically including:

[0046] K-nearest neighbors are used to identify bankrupt enterprise samples at the category decision boundary, and the boundary coefficient of the samples is calculated as the basis for hierarchical processing.

[0047] For dangerous boundary samples, a conditional generative adversarial network is used to generate synthetic samples, while for ordinary boundary samples, an improved SMOTE algorithm is applied and a dynamic weight matrix based on feature importance is introduced.

[0048] Outliers in the generated samples are removed using the Isolation Forest algorithm, and graph consistency checks are used to ensure that the related entities in the synthesized samples have reasonable paths in the knowledge graph.

[0049] The construction of the multi-dimensional feature system in S3 includes:

[0050] The financial dimension features are constructed, including: selecting 12 core financial indicators such as debt-to-equity ratio, current ratio, quick ratio, interest coverage ratio, accounts receivable turnover, inventory turnover, total asset turnover, operating profit margin, return on equity, cash flow ratio, capital adequacy ratio, and contingent liability ratio, and calculating the quarterly volatility and annual trend slope of each indicator through time series analysis.

[0051] Judicial dimension feature extraction includes: integrating eight types of judicial features, such as frequency of litigation, amount of execution target, number of times of execution, record of dishonesty, number of times of property preservation, number of times of bankruptcy application, amount of debt claim, and number of times of guarantee recovery; and using natural language processing technology to extract six types of structured information from judgment documents, such as amount of case involved, type of litigation, judgment result, execution progress, property seizure status, and debt restructuring plan.

[0052] The industry-related characteristic calculation, based on a corporate bankruptcy knowledge graph, quantifies the industry risk transmission coefficient through a random walk algorithm. Specifically, it includes: an upstream supplier risk transmission weight calculation module, determined based on three parameters: the proportion of purchase amount, payment period length, and the number of overdue days for accounts payable. The overdue days for accounts payable parameter is calibrated by analyzing the amount of the execution target and the number of executions. A downstream customer risk transmission weight calculation module is also included, determined based on three parameters: sales concentration, collection cycle, and the proportion of advance payments. The collection cycle parameter is dynamically adjusted based on the amount involved and the status of asset seizure.

[0053] Dynamic feature evolution modeling includes: constructing a corporate feature evolution trajectory model using a bidirectional LSTM network; the input layer receives multi-dimensional feature data for 12 consecutive quarters; the output layer predicts the risk level change trend for the next 3 quarters; and the intermediate layer introduces an attention mechanism to highlight feature mutations at key time nodes, where the attention weight is jointly determined by the fluctuation range of financial indicators and the severity of legal events.

[0054] S4. Obtain incremental update data from the dedicated database, input the incremental update data into the Qwen-plus big language model to extract bankruptcy progress features, construct a risk identifier based on the improved V-detector algorithm, and dynamically generate risk warnings by combining the risk transmission path of the enterprise bankruptcy knowledge graph.

[0055] By extracting updated data from the database and inputting it into the Qwen-plus large language model, the system can obtain the latest corporate bankruptcy reorganization information in real time and perform in-depth analysis using the model's powerful natural language processing capabilities. Given fixed labels, the large language model can accurately filter information matching the labels from massive amounts of data, thus achieving efficient and accurate information output. This process not only significantly improves the automation and efficiency of information processing but also enhances the readability and usability of the data, providing enterprises with timely and comprehensive decision support to help them better cope with market changes and risk challenges.

[0056] The risk detector in this embodiment of the invention employs an improved V-detector algorithm, the improvements of which include:

[0057] The detection radius is dynamically adjusted based on the spatial distribution of corporate bankruptcy risk indicators. When concentrated financial anomalies occur in a certain industry or region, the detection radius is automatically reduced to improve the accuracy of risk identification. When isolated financial fluctuations are detected, the detection radius is expanded to cover the risks of related enterprises.

[0058] By analyzing the risk transmission path in the corporate bankruptcy knowledge graph through graph attention network analysis, we focus on monitoring key nodes in high-risk relationship networks such as guarantee chains and related-party transactions, and use the risk correlation between upstream and downstream of the industrial chain as the basis for adjusting the detection weight.

[0059] Establish a bankruptcy risk pattern memory bank to store characteristic patterns from historical bankruptcy cases. When a newly emerging combination of financial indicators matches a high-risk pattern in the memory bank with a degree exceeding a threshold, a priority warning is triggered.

[0060] S5. Extract historical case data from the enterprise bankruptcy knowledge graph and represent it as a triple of <entity, relationship, disposal result>. Search for historical cases similar to the current case using cosine similarity and generate a personalized disposal plan by combining rule-based reasoning.

[0061] The generation of personalized treatment schemes in this embodiment of the invention includes:

[0062] After initial screening based on corporate fundamentals and industry characteristics, bankruptcy cases with similar debt structures, asset status and litigation history were selected as a candidate set by combining knowledge graph analysis.

[0063] A multi-dimensional similarity assessment method was adopted, which comprehensively considered factors such as the similarity of financial indicators, the similarity of creditors, and the matching degree of judicial procedures to rank and select the best candidate cases.

[0064] The disposal plan of the selected cases is verified against the compliance requirements of the current bankruptcy laws and regulations. The asset recovery rate and creditor recovery ratio under different plans are predicted through Monte Carlo simulation, and a feasibility assessment report is output.

[0065] S6. Receive real-time dynamic data from enterprises, continuously update the knowledge graph through a meta-learning algorithm, and optimize model parameters and warning thresholds based on few-sample learning; the knowledge graph update mechanism of S6 specifically includes:

[0066] For newly generated bankruptcy rulings, financial reports, and other incremental data, priority should be given to updating the subgraph structure involving major asset disposals and key creditors;

[0067] When new data contradicts existing knowledge, a dual verification mechanism is activated, involving both the verification of the authority of judicial documents and the assessment of the credibility of financial data.

[0068] Preserve the local characteristics of bankruptcy case handling in each jurisdiction, establish a regional judicial practice difference labeling system in the knowledge graph, and support the retrieval of reference cases by regional characteristics.

[0069] By employing the embodiments of the present invention, the following beneficial effects are achieved:

[0070] This invention addresses the problem of low efficiency in processing multi-source heterogeneous bankruptcy data by employing a distributed data acquisition system and intelligent classification technology, achieving efficient collection and structured storage of bankruptcy-related data. Through deep learning-based entity recognition models and knowledge graph construction technology, it overcomes the shortcomings of traditional methods in analyzing the correlation of bankruptcy risk factors, enabling comprehensive modeling and visual analysis of corporate bankruptcy risks. An improved V-detector algorithm and dynamic feature evolution model address the issue of poor timeliness in static risk assessment models, enabling real-time monitoring and accurate early warning of bankruptcy risks. Historical case retrieval and Monte Carlo simulation technology address the lack of personalized references for bankruptcy resolution plans, achieving data-driven intelligent decision support. This application integrates cutting-edge technologies such as knowledge graphs and large language models to construct an intelligent analysis system covering the entire process of data collection, risk identification, and early warning decision-making, significantly improving the accuracy, timeliness, and practicality of bankruptcy risk analysis, providing reliable risk management tools for financial institutions, judicial organs, and enterprises.

[0071] Device Example 1

[0072] According to an embodiment of the present invention, an electronic device is provided, comprising:

[0073] Processor; and,

[0074] A memory is configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the method embodiments described above.

[0075] Device Example 2

[0076] According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the steps of the method embodiments described above.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bankruptcy data analysis method based on a large model, characterized in that... include: S1. Obtain multi-channel bankruptcy-related data through a pre-built distributed data acquisition system, preprocess the multi-channel bankruptcy-related data, automatically classify and process it according to the editability of the bankruptcy-related data, and then store it in a dedicated database. S2. A deep learning-based entity recognition model is used to extract bankruptcy-related entities from the dedicated database. An entity relationship extraction model is constructed through a mutual attention mechanism. Based on the extracted entities and relationships, an enterprise bankruptcy knowledge graph is constructed and stored in the Neo4j graph database. S3. Apply oversampling algorithm to address the imbalance problem of bankruptcy-related data, screen core feature variables through chi-square test, and construct a multi-dimensional feature system by combining entity attributes of enterprise bankruptcy knowledge graph; S4. Obtain incremental update data from the dedicated database, input the incremental update data into the Qwen-plus big language model to extract bankruptcy progress features, construct a risk identifier based on the improved V-detector algorithm, and dynamically generate risk warnings by combining the risk transmission path of the enterprise bankruptcy knowledge graph. S5. Extract historical case data from the enterprise bankruptcy knowledge graph and represent it as a triple of <entity, relationship, disposal result>. Search for historical cases similar to the current case using cosine similarity and generate a personalized disposal plan by combining rule-based reasoning. S6. Receive real-time dynamic data from enterprises, continuously update the knowledge graph through meta-learning algorithms, and optimize model parameters and early warning thresholds based on few-sample learning. Specifically, the automatic classification process includes: directly extracting text from editable documents, extracting text from non-editable documents after OCR recognition, and storing the processed structured data in a dedicated database.

2. The method according to claim 1, characterized in that, The pre-built distributed data acquisition system includes: The dynamic load balancing module dynamically adjusts the crawler resource allocation strategy based on reinforcement learning algorithms, monitors the response latency, CAPTCHA trigger frequency and blocking probability of each data source in real time, and optimizes resource allocation weights through Q-learning algorithm; The anti-crawler strategy module includes a request fingerprint randomization component to dynamically generate browser fingerprints, a hybrid proxy pool component to integrate a three-layer proxy architecture of data center IP, residential IP and TOR network, and a behavior simulation engine component to simulate human operation trajectories based on generative adversarial networks. The data quality verifier uses an anomaly detection model based on the Transformer architecture to identify dirty data and implements field-level integrity verification rules, including non-empty verification, logical conflict detection, and statistical distribution verification. It also triggers an automatic completion process for incomplete data.

3. The method according to claim 1, characterized in that, The deep learning-based entity recognition model adopts an ELMo-BiLSTM-CRF hybrid architecture, and the specific construction steps include: Based on the pre-training of general corpus, a second-stage pre-training is carried out using bankruptcy legal documents and financial reports. A gradient inversion layer is introduced to eliminate domain distribution differences and jointly optimize the adversarial loss of the domain classifier and entity recognizer. A conceptual system for the bankruptcy domain is constructed, with a professional terminology dictionary embedded in the model as an external memory module. Relevant terms are dynamically retrieved through a key-value attention mechanism, and a knowledge graph-based transition constraint matrix is ​​added to the CRF layer to restrict the generation of illegal entity sequences. The bottom encoder uses cosine annealing for learning rate scheduling, with the initial learning rate set to the first preset value. The middle transition layer uses a linear warming strategy to gradually increase the learning rate to the second preset value. The top classifier uses a linear decay strategy with warmup, applies L2 regularization to the CRF layer parameters separately, and introduces an adaptive learning rate adjustment factor. The top learning rate step size is dynamically adjusted according to the loss difference between the domain classifier and the entity recognizer, where the second preset value is a preset multiple of the first preset value.

4. The method according to claim 1, characterized in that, The oversampling algorithm is the SMOTE-Borderline1-GAN algorithm, specifically including: K-nearest neighbors are used to identify bankrupt enterprise samples at the category decision boundary, and the boundary coefficient of the samples is calculated as the basis for hierarchical processing. For dangerous boundary samples, a conditional generative adversarial network is used to generate synthetic samples, while for ordinary boundary samples, an improved SMOTE algorithm is applied and a dynamic weight matrix based on feature importance is introduced. Outliers in the generated samples are removed using the Isolation Forest algorithm, and graph consistency checks are used to ensure that the related entities in the synthesized samples have reasonable paths in the knowledge graph.

5. The method according to claim 1, characterized in that, The construction of the multi-dimensional feature system in S3 includes: The financial dimension features are constructed, including: selecting 12 core financial indicators such as debt-to-equity ratio, current ratio, quick ratio, interest coverage ratio, accounts receivable turnover, inventory turnover, total asset turnover, operating profit margin, return on equity, cash flow ratio, capital adequacy ratio, and contingent liability ratio, and calculating the quarterly volatility and annual trend slope of each indicator through time series analysis. Judicial dimension feature extraction includes: integrating eight types of judicial features, such as frequency of litigation, amount of execution target, number of times of execution, record of dishonesty, number of times of property preservation, number of times of bankruptcy application, amount of debt claim, and number of times of guarantee recovery; and using natural language processing technology to extract six types of structured information from judgment documents, such as amount of case involved, type of litigation, judgment result, execution progress, property seizure status, and debt restructuring plan. The industry-related characteristic calculation, based on a corporate bankruptcy knowledge graph, quantifies the industry risk transmission coefficient through a random walk algorithm. Specifically, it includes: an upstream supplier risk transmission weight calculation module, determined based on three parameters: the proportion of purchase amount, payment period length, and the number of overdue days for accounts payable. The overdue days for accounts payable parameter is calibrated by analyzing the amount of the execution target and the number of executions. A downstream customer risk transmission weight calculation module is also included, determined based on three parameters: sales concentration, collection cycle, and the proportion of advance payments. The collection cycle parameter is dynamically adjusted based on the amount involved and the status of asset seizure. Dynamic feature evolution modeling includes: constructing a corporate feature evolution trajectory model using a bidirectional LSTM network; the input layer receives multi-dimensional feature data for 12 consecutive quarters; the output layer predicts the risk level change trend for the next 3 quarters; and the intermediate layer introduces an attention mechanism to highlight feature mutations at key time nodes, where the attention weight is jointly determined by the fluctuation range of financial indicators and the severity of legal events.

6. The method according to claim 1, characterized in that, The risk detector in S4 employs an improved V-detector algorithm, with improvements including: The detection radius is dynamically adjusted based on the spatial distribution of corporate bankruptcy risk indicators. When concentrated financial anomalies occur in a certain industry or region, the detection radius is automatically reduced to improve the accuracy of risk identification. When isolated financial fluctuations are detected, the detection radius is expanded to cover the risks of related enterprises. By analyzing the risk transmission path in the corporate bankruptcy knowledge graph through graph attention network analysis, we focus on monitoring key nodes in high-risk relationship networks such as guarantee chains and related-party transactions, and use the risk correlation between upstream and downstream of the industrial chain as the basis for adjusting the detection weight. Establish a bankruptcy risk pattern memory bank to store characteristic patterns from historical bankruptcy cases. When a newly emerging combination of financial indicators matches a high-risk pattern in the memory bank with a degree exceeding a threshold, a priority warning is triggered.

7. The method according to claim 1, characterized in that, The generation of personalized handling solutions in S5 includes: After initial screening based on corporate fundamentals and industry characteristics, bankruptcy cases with similar debt structures, asset status and litigation history were selected as a candidate set by combining knowledge graph analysis. A multi-dimensional similarity assessment method was adopted, which comprehensively considered factors such as the similarity of financial indicators, the similarity of creditors, and the matching degree of judicial procedures to rank and select the best candidate cases. The disposal plan of the selected cases is verified against the compliance requirements of the current bankruptcy laws and regulations. The asset recovery rate and creditor recovery ratio under different plans are predicted through Monte Carlo simulation, and a feasibility assessment report is output.

8. The method according to claim 1, characterized in that... The knowledge graph update mechanism of S6 specifically includes: For newly generated bankruptcy rulings, financial reports, and other incremental data, priority should be given to updating the subgraph structure involving major asset disposals and key creditors; When new data contradicts existing knowledge, a dual verification mechanism is activated, involving both the verification of the authority of judicial documents and the assessment of the credibility of financial data. Preserve the local characteristics of bankruptcy case handling in each jurisdiction, establish a regional judicial practice difference labeling system in the knowledge graph, and support the retrieval of reference cases by regional characteristics.

9. An electronic device, comprising: processor; as well as, A memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the bankruptcy data analysis method based on a large model as described in any one of claims 1-8.

10. A storage medium for storing computer-executable instructions, which, when executed, implement the steps of the bankruptcy data analysis method based on a large model as described in any one of claims 1-8.

Images