Multi-model cooperative enterprise credit risk analysis method and system

Abstract

The application discloses a multi-model cooperative enterprise credit risk analysis method and system, relates to the technical field of risk analysis, and comprises the following steps: configuring a standard enterprise name mapping of a target enterprise; establishing an enterprise basic knowledge base; initializing a multi-model cooperative identification framework by using the standard enterprise name mapping, wherein the identification behavior of the multi-model cooperative identification framework comprises lightweight identification, deep identification, behavior evolution identification and graph prediction; after data preprocessing of the enterprise basic knowledge base, the data are sent to the initialized multi-model cooperative identification framework; multi-source data are cooperatively authenticated based on the multi-model cooperative identification framework, and a cooperative authentication analysis result is generated; and enterprise credit risk is reported according to the cooperative authentication analysis result. The application solves the technical problems that the existing enterprise credit risk analysis technology depends on a single data source and model and cannot fully integrate various information sources and identify the complexity of enterprise behavior, and achieves the technical effect of improving the accuracy of enterprise credit risk analysis results.

Description

Technical Field

[0001] The present invention relates to the technical field of risk analysis, and in particular to a method and system for analyzing corporate credit risk in a multi-model collaborative manner. Background Art

[0002] Existing corporate credit risk analysis typically relies on a single data source and model. This approach fails to effectively integrate disparate data from multiple channels and fails to fully recognize the complexity of corporate behavior. For example, traditional credit scoring models typically rely on fixed, static information such as financial data and credit records, ignoring the impact of changing corporate behavior, external environmental factors, and unstructured data. This often leads to biased credit risk assessment results, making it difficult to adapt to rapidly changing market environments and internal corporate dynamics, and reducing the accuracy of risk analysis. Summary of the Invention

[0003] This application provides a multi-model collaborative enterprise credit risk analysis method and system, which is used to solve the technical problems of existing enterprise credit risk analysis relying on a single data source and model, failing to fully integrate multiple information sources and identifying the complexity of enterprise behavior.

[0004] In view of the above problems, this application provides a multi-model collaborative enterprise credit risk analysis method and system.

[0005] In a first aspect, the present application provides a multi-model collaborative enterprise credit risk analysis method, the method comprising:

[0006] Configure a standard enterprise name mapping for the target enterprise, wherein the standard enterprise name mapping is set with a mapping trust; establish an enterprise basic knowledge base, wherein the enterprise basic knowledge base includes network storage data and network question and answer data; use the standard enterprise name mapping to initialize a multi-model collaborative recognition framework, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction; perform data preprocessing on the enterprise basic knowledge base and send it to the initialized multi-model collaborative recognition framework; perform multi-source data collaborative authentication based on the multi-model collaborative recognition framework to generate collaborative authentication analysis results; and report enterprise credit risk based on the collaborative authentication analysis results.

[0007] A second aspect of the present application provides a multi-model collaborative enterprise credit risk analysis system, the system comprising:

[0008] A configuration module is used to configure the standard enterprise name mapping of the target enterprise, and the standard enterprise name mapping is set with a mapping trust; a knowledge base establishment module is used to establish an enterprise basic knowledge base, and the enterprise basic knowledge base includes network storage data and network question and answer data; an initialization module is used to initialize the multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behavior of the multi-model collaborative recognition framework includes lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction; a preprocessing module is used to preprocess the data of the enterprise basic knowledge base and send it to the initialized multi-model collaborative recognition framework; a collaborative authentication module is used to perform multi-source data collaborative authentication based on the multi-model collaborative recognition framework and generate collaborative authentication analysis results; a credit risk reporting module is used to report the enterprise credit risk according to the collaborative authentication analysis results.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application configures a standard enterprise name mapping for the target enterprise, and the standard enterprise name mapping is provided with a mapping trust; establishes an enterprise basic knowledge base, and the enterprise basic knowledge base includes network storage data and network question and answer data; uses the standard enterprise name mapping to initialize a multi-model collaborative recognition framework, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction; after data preprocessing of the enterprise basic knowledge base, it is sent to the initialized multi-model collaborative recognition framework; performs multi-source data collaborative authentication based on the multi-model collaborative recognition framework, and generates collaborative authentication analysis results; reports enterprise credit risk based on the collaborative authentication analysis results. The present invention solves the technical problems of the prior art that enterprise credit risk analysis relies on a single data source and model, cannot fully integrate multiple information sources, and cannot identify the complexity of enterprise behavior. By adopting a multi-model collaborative framework such as lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction, the technical effect of improving the accuracy of enterprise credit risk analysis results is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A flowchart of a multi-model collaborative enterprise credit risk analysis method provided in an embodiment of the present application;

[0013] Figure 2Schematic diagram of the structure of the multi-model collaborative enterprise credit risk analysis system provided in an embodiment of the present application.

[0014] Explanation of the accompanying drawings: configuration module 11, knowledge base establishment module 12, initialization module 13, pre-processing module 14, collaborative authentication module 15, credit risk reporting module 16. DETAILED DESCRIPTION

[0015] This application provides a multi-model collaborative enterprise credit risk analysis method and system to solve the technical problems of existing enterprise credit risk analysis relying on a single data source and model, failing to fully integrate multiple information sources and identify the complexity of enterprise behavior. By adopting a multi-model collaborative framework such as lightweight recognition, deep recognition, behavior evolution recognition and graph prediction, the application achieves the technical effect of improving the accuracy of enterprise credit risk analysis results.

[0016] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] It should be noted that any variations of the terms "include" and "have" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0018] Example 1, as Figure 1 As shown, the present application provides a multi-model collaborative enterprise credit risk analysis method, the method comprising:

[0019] Step S100: configuring a standard enterprise name mapping of a target enterprise, wherein the standard enterprise name mapping is set with a mapping trust level.

[0020] In this embodiment of the present application, when configuring a standard enterprise name mapping for a target enterprise, a name collection database containing multiple target enterprise name data is first obtained. The names in the name collection database are then matched against the target enterprise, performing enterprise identification matching and generating an identification matching result. A risk analysis of the target enterprise is then performed based on the identification matching result, and a mapping confidence level is established based on this risk analysis. Finally, the standard enterprise name mapping is completed using the name collection database and the mapping confidence level.

[0021] Furthermore, in the method provided in the embodiment of the application, configuring the standard enterprise name mapping of the target enterprise further includes:

[0022] Obtain a name collection database of the target enterprise, wherein the name collection database integrates name data of multiple target enterprises; perform enterprise identification matching using the names in the name collection database as matching data, and establish an identification matching result; perform identification risk analysis based on the identification matching result and the target enterprise, and establish the mapping trust; complete standard enterprise name mapping based on the name collection database and the mapping trust.

[0023] In an embodiment of the present application, a pre-prepared name collection database of target enterprises is first obtained. The name collection database contains name data of multiple target enterprises, which is composed of enterprise names from different channels such as industrial and commercial registration information, industry databases, and public network resources.

[0024] Next, using the names in the name collection database as matching data, we perform company identification matching. This process searches and extracts names similar to the target company name from the internet. Using string similarity algorithms (such as Jaccard similarity), we compare the target company name with other names in the database to identify other names similar to the target company name. This process generates identification matching results, which represent company names similar to the target company name.

[0025] Identification risk analysis is then performed based on the identification and matching results and the target company name. In this step, the target company name's proportion of all identification and matching results is calculated as the mapping confidence level. Mapping confidence is a quantitative measure that represents the ratio of the target company name to similar names in the identification and matching results. If the target company name accounts for a large proportion of the identification and matching results, it indicates a high degree of match with similar names and a high mapping confidence level. If the target company name accounts for a small proportion of the identification and matching results, it indicates a low degree of match, a low mapping confidence level, and a high identification risk.

[0026] Finally, the data in the name collection database and the calculated mapping trust are used to complete the standard enterprise name mapping.

[0027] Step S200: Establishing an enterprise basic knowledge base, wherein the enterprise basic knowledge base includes network storage data and network question and answer data.

[0028] In the embodiment of the present application, when establishing an enterprise basic knowledge base, the first step is to obtain access rights to access the target enterprise's associated network data and download the data. The downloaded data includes network storage data and network question and answer data, where each data source is clearly identified by the data source. In the downloaded results, the network storage data mainly includes structured data such as the enterprise's registration information, financial statements, and industry data, while the network question and answer data includes unstructured data such as comments on social media, forum discussions, and feedback on online question and answer platforms. Based on this data, a complete enterprise basic knowledge base is constructed.

[0029] Furthermore, in the method provided in the embodiment of the application, the establishment of the enterprise basic knowledge base further includes:

[0030] After obtaining access rights, the associated network data of the target enterprise is accessed and data download is performed, wherein the associated network data is provided with a data source identifier; the data download result includes network storage data and network question and answer data, and the enterprise basic knowledge base is established according to the data download result.

[0031] In this embodiment of the application, we first apply for permission from the relevant data provider to ensure that we can access the target enterprise's associated network data. This process involves issuing an access request to the enterprise or data platform and obtaining API keys, account permissions, or other forms of access rights. Through this process, we obtain access to a specific data source.

[0032] Once access is granted, the attacker begins accessing the target company's associated web data and downloading it. This process involves ingesting data related to the target company from various open and private sources, such as APIs, web crawlers, and database queries. This data includes business registration information, financial statements, industry databases, social media, and online reviews, ensuring comprehensiveness and diversity. To ensure data credibility, each piece of downloaded data is marked with a source identifier, facilitating tracing its origin and ensuring its legitimacy.

[0033] Data downloads include online storage data and online Q&A data. Online storage data is typically structured, such as company registration information, financial reports, and industry analysis reports. This data provides objective, quantitative support for a company's operations. Online Q&A data is unstructured data, sourced from social media, forums, and Q&A platforms. It contains discussions, comments, and feedback from the public, customers, and industry experts regarding a company.

[0034] Finally, an enterprise basic knowledge base is established based on the downloaded network storage data and network question and answer data.

[0035] Step S300: Initialize a multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction.

[0036] In an embodiment of the present application, when initializing the multi-model collaborative recognition framework using the standard enterprise name mapping, we first start by loading the standard enterprise name mapping, and eliminate the name differences in different data sources by ensuring the consistency between the target enterprise name and the standard enterprise name. On this basis, the multi-model collaborative recognition framework is initialized, which contains multiple recognition modules, including lightweight recognition, deep recognition, behavioral evolution recognition and graph prediction. During the initialization process, the lightweight recognition module is loaded first, which is used to quickly identify the basic information of the enterprise, such as the enterprise name, registration information, etc., so as to efficiently complete the basic data processing. Then, the deep recognition module analyzes more complex enterprise data, such as financial status and market performance, through a deep learning model, to provide a more accurate enterprise portrait. The behavioral evolution recognition module predicts the future development direction or potential risks of the enterprise by analyzing the historical behavior and data change trends of the enterprise, while the graph prediction module constructs a relationship map between enterprises, and helps identify potential connections and risks between enterprises through association analysis.

[0037] Step S400: After data preprocessing is performed on the enterprise basic knowledge base, the data is sent to the initialized multi-model collaborative recognition framework.

[0038] In the embodiments of the present application, data preprocessing is first performed on the data in the enterprise basic knowledge base. This step includes cleaning, formatting, and normalizing the network storage data and online question and answer data. Specifically, invalid or duplicate data is removed and the data format is unified. For example, fields such as dates, numbers, and addresses from different sources are standardized. In addition, unstructured data, such as online question and answer data, is processed using natural language processing (NLP) technology to extract valuable information, such as keywords and entity recognition, and thus converted into structured data.

[0039] After completing data preprocessing, the processed data is sent to the initialized multi-model collaborative recognition framework.

[0040] Step S500: Perform multi-source data collaborative authentication based on the multi-model collaborative recognition framework to generate collaborative authentication analysis results.

[0041] In an embodiment of the present application, in the process of collaborative authentication of multi-source data based on a multi-model collaborative recognition framework, the lightweight model in the framework is first activated, and the model is used to classify and identify the complexity of the sentences in the data preprocessing results, and a classification identification result is established. The lightweight model then uses the standard enterprise name mapping to locate the main entity of the data, generate a positioning result, and set a mapping trust identifier for these results. Then, for simple sentences in the classification identification result, a rapid recognition is performed based on the positioning result to generate a first recognition result, which is then synchronized with the classification identification result and the positioning result to the multi-model collaborative recognition framework.

[0042] After synchronizing the data, the classification identification results are used to locate complex sentences, and the deep recognition model is activated for further analysis. The model first extracts behavioral trigger words and limiting conditions from complex sentences to generate extraction results, and then performs behavioral driving analysis through the semantic motivation analyzer to form the first temporary result. Next, the dependency analyzer splits the multi-layer nested modification chain of the complex sentence, reconstructs the semantic link based on the splitting results, and generates the second temporary result. Then, the feature retrieval engine performs reverse query reasoning on the enterprise graph based on behavioral characteristics, industry characteristics, and time characteristics to obtain the third temporary result. Finally, these three parts of the temporary results are combined to generate the second recognition result and synchronized to the framework.

[0043] Afterwards, the first, second, and third temporarily stored results are temporarily authenticated for the same sentence. Cross-sentence correlation analysis is used to generate a second recognition result, which is then synchronized to the framework. Based on this, the framework activates the time evolution model, identifies time nodes in the preprocessed results, establishes a data time series, and further optimizes the time series using the first and second recognition results. The identified data is then fed into the time evolution model, which generates a third recognition result based on temporal trends.

[0044] Finally, the first identification result, the second identification result and the third identification result together constitute the collaborative authentication analysis result.

[0045] Furthermore, in the method provided in the embodiment of the application, the multi-source data collaborative authentication based on the multi-model collaborative recognition framework further includes:

[0046] Activate the lightweight model in the multi-model collaborative recognition framework, use the lightweight model to classify and identify the sentence complexity of the data preprocessing results, and establish a classification identification result; use the lightweight model to call the standard enterprise name mapping to locate the main entity of the data preprocessing results, and establish a positioning result, and the positioning result is provided with a trust identifier based on the mapping trust; quickly identify simple sentences in the classification identification results based on the positioning result to generate a first recognition result; synchronize the first recognition result, the classification identification result, and the positioning result to the multi-model collaborative recognition framework.

[0047] In an embodiment of the present application, the lightweight model in the multi-model collaborative recognition framework is first activated. The lightweight model is intended to perform preliminary screening and classification of data through simple calculations. At this step, the lightweight model will classify and identify the complexity of the sentences for the data preprocessing results. Specifically, a grammatical analysis tool or a rule-based method (such as syntax tree analysis) is used to analyze the sentences in the text to determine the complexity of the sentences. Then, complex sentences and simple sentences are distinguished, and a classification identification result is generated to indicate whether each sentence belongs to a simple sentence or a complex sentence. For example, the sentence "The company has completed the audit of financial statements and plans to proceed to the next stage of financial restructuring." is analyzed as a complex sentence because it contains multiple actions and different events. Simple sentences such as "The company has completed the audit" are marked as simple sentences.

[0048] Next, a lightweight model is used to map standard company names to the data preprocessing results to locate the primary entity. Specifically, the target company name is compared with names in the standard database. The best matching standard company name is found by calculating string similarity or using a fuzzy matching algorithm (such as Jaccard similarity). This company name is then designated as the primary entity. This process generates a location result, in which the matching relationship between each company name and the standard name is determined. These location results are then assigned a trust indicator based on the mapping confidence level.

[0049] The simple sentences in the classification and identification results are then quickly identified based on the generated positioning results. Using a rule-matching algorithm (such as a regular expression-based matching method), the company name and related information in the simple sentence are quickly identified, for example, "Huawei Technologies Co., Ltd." is used as the primary entity, and its related action information is extracted. This process generates the first recognition result, which identifies the key information in the simple sentence.

[0050] Finally, the first recognition result, classification identification result and positioning result are synchronized to the multi-model collaborative recognition framework for further processing by other models in the framework.

[0051] Furthermore, in the method provided in the embodiment of the application, after synchronizing the first recognition result, the classification identification result, and the positioning result to the multi-model collaborative recognition framework, the method further includes:

[0052] The complex sentence is located using the classification identification result, and the deep recognition model in the multi-model collaborative recognition framework is activated to update the location result to the deep recognition model; the behavior trigger words and limiting conditions of the complex sentence are extracted to establish the extraction result; the semantic motivation analyzer of the deep recognition model is called to perform behavior-driven semantic analysis of the extraction result to establish a first temporary result; the dependency analyzer of the deep recognition model is called to perform multi-layer nested modification chain splitting of the complex sentence, reconstruct the semantic link according to the splitting result, and recycle the backbone structure, and use the recycled backbone structure to perform semantic recognition to establish a second temporary result; the feature retrieval engine of the deep recognition model is called, and the feature retrieval engine is used to perform enterprise graph reverse query reasoning based on behavior characteristics, industry characteristics, and time characteristics to establish a third temporary result; a second recognition result is generated based on the first temporary result, the second temporary result, and the third temporary result; the second recognition result is synchronized to the multi-model collaborative recognition framework.

[0053] In an embodiment of the present application, complex sentences in the classification identification results are first located. By distinguishing simple sentences from complex sentences, it is identified which sentences require deeper processing, and the deep recognition model in the multi-model collaborative recognition framework is activated. The deep recognition model is built based on deep learning methods (such as the Transformer architecture or the BERT model). These models are capable of processing complex text inputs and extracting deep semantic information from the text. By loading pre-trained deep learning models and fine-tuning them for target tasks, the ability to specifically process enterprise data and behavior patterns is obtained.

[0054] Next, the positioning results (i.e., company names and related information) obtained from the lightweight model are updated to the deep recognition model to ensure that the model can utilize accurate company entity data for further analysis. Based on this foundation, the deep recognition model begins extracting action triggers and qualifiers from complex sentences. Using a trained named entity recognition (NER) module and relation extraction technology, the model automatically extracts action-related verbs (such as "plan") and qualifiers (such as "before the end of 2023") from complex sentences. For example, in the sentence "Huawei plans to complete its financial audit by the end of 2023," "plan" is extracted as the action trigger and "before the end of 2023" as the qualifier, generating the extraction results.

[0055] Subsequently, the semantic cause analyzer of the deep recognition model is called to perform behavior-driven semantic analysis on the extraction results. The goal of this step is to analyze the underlying reasons behind the behavior trigger words, such as why Huawei plans to conduct a financial audit, analyze the driving factors of the behavior (such as annual audit, financial compliance, etc.), and generate the first temporary result, which is the semantic analysis result of the behavior and its causes.

[0056] Next, the dependency analyzer of the deep recognition model is used to split the multi-layer nested modification chain of complex sentences. The dependency analyzer analyzes the dependency relationships between words in the sentence and decomposes complex sentence structures. For example, the sentence "Huawei plans to complete financial audit by the end of 2023" is analyzed as the relationship between the subject "Huawei" and the behavior "plan", while separating the modification parts such as "by the end of 2023" and "financial audit". Based on these analysis results, the semantic link is reconstructed and the main structure is recovered, so that the core information of the sentence is accurately extracted, and finally the second temporary result, i.e. the complete semantic structure, is generated.

[0057] Next, the feature retrieval engine of the deep recognition model is activated to use enterprise graph reverse reasoning to analyze behavior features, industry features, and time features to generate the third temporary result. This step uses the enterprise association graph to combine historical behavior and industry background for reasoning to analyze the behavior patterns and potential risks of the enterprise. For example, the deep recognition model predicts whether a certain enterprise may change at a certain time node based on historical audit data and industry trends, thereby identifying potential risk points.

[0058] Finally, the first temporary result, the second temporary result, and the third temporary result are combined to generate the second recognition result, which integrates the information of behavior analysis, semantic analysis, and graph reasoning to provide more accurate support for enterprise credit risk assessment. Finally, the second recognition result is synchronized to the multi-model collaborative recognition framework.

[0059] Further, the method provided by the application embodiment further comprises:

[0060] According to the first temporary result, the second temporary result, and the third temporary result, the temporary authentication of the same sentence is performed, and the sentence temporary result is updated; the updated sentence temporary result is used for cross-sentence association analysis, and the cross-sentence authentication is performed using the association analysis result to establish the second recognition result.

[0061] In an embodiment of the present application, a temporary authentication is first performed on the same statement based on the first, second, and third temporary results. Specifically, the portions of the three data sets involving the same statement are compared, and string comparison methods (e.g., Jaccard similarity) are used to evaluate the consistency between these results. If inconsistencies are found in parts of the same statement (e.g., different time points described or inconsistent information), information merging rules are used, such as taking the most frequently occurring time or the most accurate entity information, to correct the inconsistencies and update the temporary statement results.

[0062] Next, cross-sentence association analysis is performed using the updated temporary statement results. This process uses keyword matching and contextual association analysis to identify commonalities across different statements. For example, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm analyzes high-frequency keywords (such as "audit" and "financial statement") across different statements. These keywords are then used to determine the associations between different statements. By analyzing shared themes or related behaviors, such as "Huawei" and "completing a financial audit," it is determined whether they describe the same event or belong to the same business process. This process identifies intrinsic connections across statements. For example, the association between "Huawei Technologies Co., Ltd. plans to complete its audit by the end of 2023" and "Huawei will complete its financial audit by the end of the year" indicates that they likely describe the same event. This process yields cross-sentence association analysis results.

[0063] Finally, based on the results of cross-statement correlation analysis, cross-statement authentication is performed. During this process, logical reasoning algorithms (such as rule-based reasoning systems) are used to verify the consistency of information across statements. Specifically, the process checks whether the times, actions, or conditions mentioned in two or more statements match each other, ensuring that there are no conflicts in the information flow across multiple statements. For example, the process might verify whether "the end of 2023" and "the next quarter" fall within the same timeframe. Based on this logic, the final second recognition result, or the final recognition result after cross-statement authentication, is generated.

[0064] Furthermore, in the method provided in the embodiment of the application, after synchronizing the second recognition result to the multi-model collaborative recognition framework, the method further includes:

[0065] Activate the time evolution model of the multi-model collaborative recognition framework; after marking the positioning results and time nodes of the preprocessing results, establish a data time series; use the first recognition result and the second recognition result to mark the data time series, and input it into the time evolution model; establish a third recognition result based on the time evolution result, and upload the third recognition result to the multi-model collaborative recognition framework.

[0066] In this embodiment of the present application, the time evolution model of the multi-model collaborative recognition framework is first activated. The time evolution model is trained using an LSTM (Long Short-Term Memory) network and uses historical data to capture the changing trends of corporate behavior over time. Specifically, time series data is extracted from historical data, including information such as corporate-related behaviors, events, and time nodes. This data is used to train the time evolution model, ensuring that the model can identify long-term patterns and time-related changing trends in corporate behavior.

[0067] Next, the preprocessing results are mapped to location and time nodes. During this process, time information is extracted from the preprocessed data using time-tagging methods (such as regular expressions or time extraction tools), and the time node is annotated on each piece of data. At the same time, the company information in the data is consistent with the corresponding time node, so that each event or behavior can be associated with a specific time.

[0068] Once the time nodes and positioning results are identified, the data time series is established. In this step, different behavioral data are arranged in chronological order according to the time nodes to generate a continuous time series. For example, corporate audit reports, financial decision-making, and other behaviors at different time points can be arranged in chronological order to ensure that the data shows the time evolution of corporate behavior.

[0069] The first recognition result (basic enterprise information from the lightweight model) and the second recognition result (enterprise behavior analysis from the deep recognition model) are then labeled with the data time series. During this process, the first and second recognition results are matched with the corresponding time nodes to ensure that all behaviors and events can be accurately attributed to their occurrence time.

[0070] Next, the labeled data time series is fed into the time evolution model. In this step, the time series data is fed into the LSTM model for analysis. For example, by feeding in past financial statements and behavioral data, it can predict a company's likely financial status or behavioral changes at a certain point in the future.

[0071] Based on the output of the time evolution model, a third-party recognition result is generated. The time evolution result reflects the prediction of corporate behavior over time, for example, predicting that "Huawei will complete the audit by the end of 2023." These predictions are used to generate a third-party recognition result based on the company's historical behavior and the trends output by the model.

[0072] Finally, the third recognition result is uploaded to the multi-model collaborative recognition framework.

[0073] Furthermore, in the method provided in the embodiment of the application, uploading the third recognition result to the multi-model collaborative recognition framework further includes:

[0074] Perform identification conflict analysis on the first identification result, the second identification result, and the third identification result; if the identification conflict analysis result triggers a preset abnormality threshold, activate the review channel; perform review authentication according to the review channel to update the collaborative authentication analysis result.

[0075] In an embodiment of the present application, first, an identification conflict analysis is performed on the first identification result, the second identification result, and the third identification result to generate an identification conflict analysis result. Specifically, the key data in the three identification results are compared one by one by using rule matching and data comparison methods. At this time, the core information in each identification result is compared, such as the company name, time node, behavior description, etc. For example, check whether the time in the first identification result is consistent with the time in the second identification result, or whether the corporate behavior in the two identification results is consistent. These data items are compared through string comparison (such as Jaccard similarity) or grammatical analysis, and conflicts are marked. If the "financial audit" time mentioned in the first identification result is "December 2023", and the second identification result mentions "the audit is postponed to January 2024", the time conflict between the two results is identified, and an identification conflict analysis result is generated, recording the specific type of conflict.

[0076] Next, the conflict analysis results are compared against pre-set anomaly thresholds. These thresholds can be quantitative data (e.g., time discrepancies greater than 30 days, or inconsistencies in described behaviors exceeding 20%) or rules (e.g., behavior descriptions must fall within the same timeframe, or multiple outcomes for the same event must be completed within a month). For example, if the time discrepancy found in the conflict analysis results exceeds the pre-set 30-day threshold, or if the differences between multiple described enterprise behaviors are significant, exceeding the set tolerance, an anomaly conflict is identified. This anomaly triggers the review mechanism, activating the review channel.

[0077] After the verification channel is activated, the verification and certification phase begins. During this phase, the conflicting data undergoes further manual or rule-based verification. Specifically, manual review or rule-driven verification methods are used to recheck the conflicting parts. For example, if a significant discrepancy is found between the time in the first and second recognition results, the accuracy of the relevant data sources is verified manually or by consulting other reliable external data sources (such as industry reports and publicly available financial data). Data correction methods (such as correcting erroneous data points based on external data sources) can also be used to correct the conflicting parts.

[0078] After completing the review and certification, update the collaborative certification analysis results.

[0079] Step S600: reporting the enterprise credit risk based on the collaborative authentication analysis results.

[0080] In an embodiment of the present application, when reporting the credit risk of an enterprise based on the collaborative authentication analysis results, the credit risk of the enterprise is first evaluated based on the first identification result, the second identification result, and the third identification result. Specifically, based on the information in these three parts of the identification results, such as enterprise behavior analysis, time evolution trend, and behavior correction after review, the risks that the enterprise may face are identified. For example, if the first identification result indicates that the financial behavior of the enterprise is abnormal, the second identification result shows that the behavior pattern is unstable, and the third identification result predicts that the future financial situation of the enterprise will deteriorate, combined with this information, it is analyzed whether the enterprise has a credit default risk, a capital chain rupture risk, or a risk of inaccurate financial statements. Through this process, the credit risk of the enterprise is reported based on these comprehensive data, where the credit risk of the enterprise can include multiple aspects such as credit default risk, capital chain rupture risk, and inaccurate financial statement risk.

[0081] Furthermore, in the method provided in the embodiment of the application, the reporting of the enterprise credit risk based on the collaborative authentication analysis results further includes:

[0082] Perform risk level matching based on the enterprise credit risk and establish a risk level warning; configure a warning signal using the risk level warning and execute a warning alarm.

[0083] In this embodiment, risk level matching is first performed based on the results of the enterprise credit risk assessment. Specifically, different risk types are mapped to different levels according to a preset rule table. For example, credit default risk is mapped to a medium risk level, capital chain disruption risk is mapped to a high risk level, and financial statement inaccuracy risk is mapped to a low risk level.

[0084] Next, we configure corresponding early warning signals for each risk level. For example, for high-risk enterprises, we configure a red early warning signal, indicating that the enterprise faces serious risks and may need to take immediate emergency measures; for medium-risk enterprises, we configure a yellow early warning signal, indicating that there is a certain potential risk and further monitoring is required; for low-risk enterprises, we configure a green early warning signal, indicating that the enterprise is currently stable and the risk is low. Immediate action is not required, but monitoring is still required.

[0085] Finally, based on the configured risk level alerts, early warnings are issued. During this process, alert notifications are sent to relevant personnel (such as the risk management team and finance department) based on different warning signals. For example, if a company's risk level is high, a red alert is issued, alerting management to serious defaults or financial problems and the need for immediate risk control measures.

[0086] In the embodiments of the present application, the above-mentioned embodiments have at least the following technical effects:

[0087] The present application configures the standard enterprise name mapping of the target enterprise, sets the mapping trust degree, establishes the enterprise basic knowledge base including network storage data and network question and answer data, initializes the multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition and graph prediction, sends the data preprocessed enterprise basic knowledge base to the initialized multi-model collaborative recognition framework, performs multi-source data collaborative authentication based on the multi-model collaborative recognition framework, generates collaborative authentication analysis results, and reports enterprise credit risks according to the collaborative authentication analysis results. The present application solves the technical problems that the existing enterprise credit risk analysis depends on a single data source and model, cannot fully integrate various information sources and recognize the complexity of enterprise behaviors, and achieves the technical effect of improving the accuracy of enterprise credit risk analysis results by using the multi-model collaborative framework of lightweight recognition, deep recognition, behavior evolution recognition and graph prediction.

[0088] Embodiment two, based on the same inventive concept as the enterprise credit risk analysis method of multi-model collaboration in the foregoing embodiments, as shown in the present application, a multi-model collaborative enterprise credit risk analysis system is provided, and the system and method embodiments in the present application are based on the same inventive concept. Wherein, the system comprises: Figure 2

[0089] The configuration module 11 is configured to configure the standard enterprise name mapping of the target enterprise, and the standard enterprise name mapping is set with a mapping trust degree; the knowledge base establishment module 12 is configured to establish an enterprise basic knowledge base, and the enterprise basic knowledge base includes network storage data and network question and answer data; the initialization module 13 is configured to initialize a multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition and graph prediction; the preprocessing module 14 is configured to send the data preprocessed enterprise basic knowledge base to the initialized multi-model collaborative recognition framework; the collaborative authentication module 15 is configured to perform multi-source data collaborative authentication based on the multi-model collaborative recognition framework, and generate collaborative authentication analysis results; and the credit risk reporting module 16 is configured to report enterprise credit risks according to the collaborative authentication analysis results.

[0090] Further, the system is also used to realize the following functions:

[0091] ​Activate the lightweight model in the multi-model collaborative recognition framework, use the lightweight model to classify and identify the sentence complexity of the data preprocessing results, and establish a classification identification result; use the lightweight model to call the standard enterprise name mapping to locate the main entity of the data preprocessing results, and establish a positioning result, and the positioning result is provided with a trust identifier based on the mapping trust; quickly identify simple sentences in the classification identification results based on the positioning result to generate a first recognition result; synchronize the first recognition result, the classification identification result, and the positioning result to the multi-model collaborative recognition framework.

[0092] Furthermore, the system is also used to implement the following functions:

[0093] The complex sentence is located using the classification identification result, and the deep recognition model in the multi-model collaborative recognition framework is activated to update the location result to the deep recognition model; the behavior trigger words and limiting conditions of the complex sentence are extracted to establish the extraction result; the semantic motivation analyzer of the deep recognition model is called to perform behavior-driven semantic analysis of the extraction result to establish a first temporary result; the dependency analyzer of the deep recognition model is called to perform multi-layer nested modification chain splitting of the complex sentence, reconstruct the semantic link according to the splitting result, and recycle the backbone structure, and use the recycled backbone structure to perform semantic recognition to establish a second temporary result; the feature retrieval engine of the deep recognition model is called, and the feature retrieval engine is used to perform enterprise graph reverse query reasoning based on behavior characteristics, industry characteristics, and time characteristics to establish a third temporary result; a second recognition result is generated based on the first temporary result, the second temporary result, and the third temporary result; the second recognition result is synchronized to the multi-model collaborative recognition framework.

[0094] Furthermore, the system is also used to implement the following functions:

[0095] Perform temporary authentication of the same statement based on the first temporary storage result, the second temporary storage result, and the third temporary storage result, and update the statement temporary storage result; use the updated statement temporary storage result to perform cross-statement association analysis, and use the association analysis result to perform cross-statement authentication to establish the second recognition result.

[0096] Furthermore, the system is also used to implement the following functions:

[0097] Activate the time evolution model of the multi-model collaborative recognition framework; after marking the positioning results and time nodes of the preprocessing results, establish a data time series; use the first recognition result and the second recognition result to mark the data time series, and input it into the time evolution model; establish a third recognition result based on the time evolution result, and upload the third recognition result to the multi-model collaborative recognition framework.

[0098] Furthermore, the system is also used to implement the following functions:

[0099] Perform identification conflict analysis on the first identification result, the second identification result, and the third identification result; if the identification conflict analysis result triggers a preset abnormality threshold, activate the review channel; perform review authentication according to the review channel to update the collaborative authentication analysis result.

[0100] Furthermore, the system is also used to implement the following functions:

[0101] After obtaining access rights, the associated network data of the target enterprise is accessed and data download is performed, wherein the associated network data is provided with a data source identifier; the data download result includes network storage data and network question and answer data, and the enterprise basic knowledge base is established according to the data download result.

[0102] Furthermore, the system is also used to implement the following functions:

[0103] Obtain a name collection database of the target enterprise, wherein the name collection database integrates name data of multiple target enterprises; perform enterprise identification matching using the names in the name collection database as matching data, and establish an identification matching result; perform identification risk analysis based on the identification matching result and the target enterprise, and establish the mapping trust; complete standard enterprise name mapping based on the name collection database and the mapping trust.

[0104] Furthermore, the system is also used to implement the following functions:

[0105] Perform risk level matching based on the enterprise credit risk and establish a risk level warning; configure a warning signal using the risk level warning and execute a warning alarm.

[0106] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0107] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0108] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. The multi-model collaborative enterprise credit risk analysis method is characterized by: The method comprises: Configuring a standard enterprise name mapping for a target enterprise, wherein the standard enterprise name mapping is set with a mapping trust level; Establishing an enterprise basic knowledge base, wherein the enterprise basic knowledge base includes network storage data and network question and answer data; Initializing a multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction; After data preprocessing is performed on the enterprise basic knowledge base, the data is sent to the initialized multi-model collaborative recognition framework; Performing collaborative authentication of multi-source data based on the multi-model collaborative recognition framework to generate collaborative authentication analysis results; Reporting the enterprise credit risk based on the collaborative authentication analysis results; The multi-source data collaborative authentication based on the multi-model collaborative recognition framework includes: Activate the lightweight model in the multi-model collaborative recognition framework, use the lightweight model to classify and identify the sentence complexity of the data preprocessing results, and establish a classification identification result; Using the lightweight model to call the standard enterprise name mapping to locate the main entity of the data preprocessing result, and establish a positioning result, wherein the positioning result is provided with a trust identifier based on the mapping trust degree; Quickly identifying the simple sentences in the classification identification result based on the positioning result to generate a first recognition result; Synchronizing the first recognition result, the classification identification result, and the positioning result to the multi-model collaborative recognition framework; After synchronizing the first recognition result, the classification identification result, and the positioning result to the multi-model collaborative recognition framework, the method includes: Locating complex sentences using the classification identification results, activating a deep recognition model in a multi-model collaborative recognition framework, and updating the positioning results to the deep recognition model; Extracting behavioral trigger words and limiting conditions from the complex sentence and establishing extraction results; Calling the semantic motivation analyzer of the deep recognition model to perform behavior-driven semantic analysis of the extraction result and establish a first temporary result; The dependency analyzer of the deep recognition model is called to split the multi-layer nested modification chains of complex sentences. The semantic links are reconstructed based on the splitting results, and the trunk structure is recovered. The recovered trunk structure is used for semantic recognition to create a second temporary result. Invoke a feature retrieval engine of the deep recognition model, use the feature retrieval engine to perform reverse query reasoning on the enterprise graph based on behavioral features, industry features, and time features, and establish a third temporary result; generating a second recognition result according to the first temporarily stored result, the second temporarily stored result, and the third temporarily stored result; The second recognition result is synchronized to the multi-model collaborative recognition framework.

2. The multi-model collaborative enterprise credit risk analysis method according to claim 1, characterized in that: Generating a second recognition result according to the first temporarily stored result, the second temporarily stored result, and the third temporarily stored result includes: Perform temporary storage authentication on the same statement according to the first temporary storage result, the second temporary storage result, and the third temporary storage result, and update the temporary storage result of the statement; The updated temporary statement storage result is used to perform cross-statement association analysis, and the association analysis result is used to perform cross-statement authentication to establish the second recognition result.

3. The multi-model collaborative enterprise credit risk analysis method according to claim 1, characterized in that: After synchronizing the second recognition result to the multi-model collaborative recognition framework, the method includes: activating a time evolution model of the multi-model collaborative recognition framework; After marking the positioning result and time node on the preprocessing result, a data time series is established; After labeling the data time series using the first recognition result and the second recognition result, the data is input into the time evolution model; A third recognition result is established according to the time evolution result, and the third recognition result is uploaded to the multi-model collaborative recognition framework.

4. The multi-model collaborative enterprise credit risk analysis method according to claim 3, characterized in that: The uploading of the third recognition result to the multi-model collaborative recognition framework includes: performing recognition conflict analysis on the first recognition result, the second recognition result, and the third recognition result; If the conflict analysis result triggers the preset abnormality threshold, the review channel is activated; The review authentication is performed according to the review channel to update the collaborative authentication analysis result.

5. The multi-model collaborative enterprise credit risk analysis method according to claim 1, characterized in that: The establishment of the enterprise basic knowledge base includes: After obtaining access rights, access the target enterprise's associated network data and perform data downloading, wherein the associated network data is provided with a data source identifier; The data download results include network storage data and network question and answer data, and the enterprise basic knowledge base is established based on the data download results.

6. The multi-model collaborative enterprise credit risk analysis method according to claim 1, characterized in that: The configuration of the target enterprise's standard enterprise name mapping includes: Acquire a name collection database of target enterprises, wherein the name collection database integrates name data of a plurality of target enterprises; Collecting names in a database using the name as matching data, performing enterprise identification matching, and establishing an identification matching result; Performing identification risk analysis based on the identification and matching results and the target enterprise to establish the mapping trust level; The standard enterprise name mapping is completed according to the name collection database and the mapping trust.

7. The multi-model collaborative enterprise credit risk analysis method according to claim 1, characterized in that: The reporting of the enterprise credit risk according to the collaborative authentication analysis results includes: Perform risk level matching based on the credit risk of the enterprise and establish risk level warning; The risk level warning is used to configure a warning signal and execute a pre-alarm.

8. The multi-model collaborative enterprise credit risk analysis system is characterized by: The system is used to execute the multi-model collaborative enterprise credit risk analysis method according to any one of claims 1 to 7, and the system includes: A configuration module, configured to configure a standard enterprise name mapping of a target enterprise, wherein the standard enterprise name mapping is provided with a mapping trust level; A knowledge base building module is used to build an enterprise basic knowledge base, which includes network storage data and network question and answer data; An initialization module, configured to initialize a multi-model collaborative recognition framework using the standard enterprise name mapping, wherein the recognition behaviors of the multi-model collaborative recognition framework include lightweight recognition, deep recognition, behavior evolution recognition, and graph prediction; A preprocessing module, configured to preprocess the data of the enterprise basic knowledge base and then send the data to the initialized multi-model collaborative recognition framework; A collaborative authentication module, configured to perform collaborative authentication of multi-source data based on the multi-model collaborative recognition framework and generate collaborative authentication analysis results; The credit risk reporting module is used to report the enterprise credit risk based on the collaborative authentication analysis results.

Images