Credit data analysis and processing method, system and storage medium based on NLP

Abstract

The NLP-based credit data analysis and processing method, system and storage medium provided in the embodiments of the present invention realizes comprehensive analysis of the target enterprise's credit data review text and personalized resource push through data processing, semantic mining, feature splicing, and credit risk control and resource push demand identification. Such collaborative effect can improve the accuracy and security of credit risk management and resource push, thereby improving the technical problems that traditional technologies are difficult to accurately implement credit risk identification and difficult to implement personalized and secure resource push.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a credit data analysis and processing method, system and storage medium based on NLP. Background Art

[0002] Corporate credit text data refers to text information related to corporate credit. These text data may include various forms of text content such as corporate evaluations, comments, reports, announcements, news reports, etc. These data usually involve descriptions, analyses, and evaluations of corporate credit status, operating capabilities, financial health, industry status, and goodwill and reputation.

[0003] Corporate credit text data can come from multiple channels, including but not limited to the following: Corporate credit reports: reports on the credit status of a company issued by professional institutions such as credit rating agencies, financial institutions, and credit reporting agencies, which contain rich text descriptions and evaluations; User comments and evaluations: Text data from consumers, partners or other companies that evaluate and comment on a company, which can reflect information on the company's service quality, business ethics, customer satisfaction, etc.; News reports and media analysis: News reports, industry analysis, market reviews and other text content involving the company can help understand the company's image and reputation in the public eye; Company announcements and annual reports: Announcements, annual financial reports, shareholder letters and other text materials issued by the company provide information on the company's business strategy, financial status and development plans; Social media comments: Comments, discussions and public opinion analysis on the company on social media platforms (such as Weibo, WeChat, etc.) can reflect the public's attitudes and opinions on the company.

[0004] Corporate credit text data is of great value in corporate assessment, risk control, market research, public opinion analysis, etc. By processing, analyzing and mining these text data, we can gain a deeper understanding of the corporate credit status and provide reference and basis for decision-making. In actual application, when faced with the joint analysis needs of credit risk control and resource push, traditional technologies are struggling. Summary of the invention

[0005] In order to improve the above problems, the present invention provides a credit data analysis and processing method, system and storage medium based on NLP.

[0006] In a first aspect, a credit data analysis and processing method based on NLP is provided, which is applied to an artificial intelligence analysis system, and the method comprises:

[0007] Acquire a review text of the enterprise credit data to be analyzed uploaded by the business resource push server, wherein the review text of the enterprise credit data to be analyzed includes a plurality of enterprise credit behavior events;

[0008] Loading the review text of the enterprise credit data to be analyzed into the review text semantic mining model to perform text semantic mining to obtain the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed;

[0009] The original credit report text semantics is loaded into a structured text semantic processing model, and text semantic vectors of different levels in the original credit report text semantics are concatenated to obtain a text semantic concatenation vector;

[0010] The text semantic splicing vector is loaded into the credit risk control discrimination algorithm and the resource push demand discrimination algorithm respectively, and the credit risk control opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the credit risk control discrimination algorithm, and the resource push demand opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the resource push demand discrimination algorithm.

[0011] In some possible embodiments, the review text of the enterprise credit data to be analyzed is loaded into a review text semantic mining model for text semantic mining to obtain the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed. The review text semantic mining model includes multiple state stages, including:

[0012] Loading the review text of the enterprise credit data to be analyzed into the review text semantic mining model;

[0013] The text semantic vectors within the state level are spliced ​​based on the first set splicing strategy, and different state levels are combined based on the second set splicing strategy to obtain the original credit report text semantics of the corporate credit data review text to be analyzed generated by the sub-model of each state level.

[0014] In some possible embodiments, the step of loading the original credit report text semantics into a structured text semantics processing model, and performing text semantic vector splicing on text semantic vectors of different levels in the original credit report text semantics to obtain a text semantic splicing vector includes:

[0015] Performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale among the original credit report text semantics generated by the sub-model of each state order to obtain a text semantic expansion vector;

[0016] According to the text semantic extension vector, adjusting the original credit report text semantics consistent with the feature scale of the text semantic extension vector;

[0017] Jump to the step of performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale in the original credit report text semantics generated by the sub-model of each state order, until the original credit report text semantics generated by the sub-model of each state order are spliced ​​to obtain a text semantic splicing vector.

[0018] In some possible embodiments, adjusting the original credit report text semantics consistent with the feature scale of the text semantics extension vector according to the text semantics extension vector includes:

[0019] Performing a semantic logic-based intersection between the text semantic extension vector and the original credit report text semantics that is consistent with the feature scale of the text semantic extension vector, to obtain a text semantic intersection result;

[0020] The text semantic crossover result is subjected to text semantic vector splicing based on a reversible unit to obtain an original text semantic splicing vector, and the original text semantic splicing vector is adjusted to the original credit report text semantics that is consistent with the feature scale of the text semantic expansion vector.

[0021] In some possible embodiments, the text semantic concatenation vector is loaded into a credit risk control discrimination algorithm, and a credit risk control opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the credit risk control discrimination algorithm. The credit risk control opinion label includes credit risk control keywords output by the credit risk control discrimination algorithm, credit risk control suggestion labels output by the credit risk control discrimination algorithm, and feature scale variables output by the credit risk control discrimination algorithm, including:

[0022] Loading the text semantic concatenation vector into the credit risk control discrimination algorithm;

[0023] Generate, based on the first AI discrimination component in the credit risk control discrimination algorithm, a credit risk control keyword output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed;

[0024] Generating a credit risk control recommendation label output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed according to the first gradient optimization component in the credit risk control discrimination algorithm;

[0025] A feature scale variable output by the credit risk control discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed is generated according to the first feature scale output component in the credit risk control discrimination algorithm.

[0026] In some possible embodiments, the text semantic concatenation vector is loaded into a resource push demand discrimination algorithm, and a resource push demand opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the resource push demand discrimination algorithm, and the resource push demand opinion label includes a resource push demand keyword, a push mode label output by the resource push demand discrimination algorithm, a feature scale variable output by the resource push demand discrimination algorithm, and a push content directory, including:

[0027] Loading the text semantic splicing vector into the resource push demand determination algorithm;

[0028] Generate, based on the second AI discrimination component in the resource push demand discrimination algorithm, a credit risk control keyword output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed;

[0029] Generate a push mode label output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the second gradient optimization component in the resource push demand discrimination algorithm;

[0030] Generate a feature scale variable output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the second feature scale output component in the resource push demand discrimination algorithm;

[0031] Generate frequent item behavior preference thermal data corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the behavior thermal analysis component in the resource push demand discrimination algorithm;

[0032] A push content directory generation component in the resource push demand identification algorithm generates a push content directory for each enterprise credit behavior event in the enterprise credit data review text to be analyzed.

[0033] In some possible embodiments, before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, the method further includes:

[0034] Obtain a sample of the first enterprise credit data review text for the credit risk control identification task;

[0035] Loading the first enterprise credit data review text sample of the credit risk control judgment task into the original review text semantic mining model for text semantic mining, and obtaining the first credit report text semantic sample of the original credit risk control judgment task corresponding to the first enterprise credit data review text sample of the credit risk control judgment task;

[0036] Inputting the first credit report text semantic sample of the original credit risk control discrimination task into the original structured text semantic processing model, performing text semantic vector splicing on text semantic vectors of different levels in the first credit report text semantic sample of the original credit risk control discrimination task, and obtaining a first text semantic splicing vector sample;

[0037] The first text semantic concatenation vector sample is input into the original credit risk control discrimination algorithm, and a credit risk control prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data review text sample of the credit risk control discrimination task is generated according to the original credit risk control discrimination algorithm.

[0038] In some possible embodiments, before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, the method further includes:

[0039] Obtain a sample text of a review of the second enterprise's credit data;

[0040] Loading the first enterprise credit data review text sample of the resource push demand determination task into the original review text semantic mining model for text semantic mining, and obtaining the first credit report text semantic sample of the original resource push demand determination task corresponding to the first enterprise credit data review text sample of the resource push demand determination task;

[0041] Inputting the first credit report text semantic sample of the original resource push demand discrimination task into the original structured text semantic processing model, performing text semantic vector splicing on text semantic vectors of different levels in the first credit report text semantic sample of the original resource push demand discrimination task, and obtaining a second credit report text semantic sample;

[0042] The second credit report text semantic sample is input into the original resource push demand judgment algorithm, and the resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the resource push demand judgment task is generated based on the original resource push demand judgment algorithm.

[0043] In some possible embodiments, after the resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the resource push demand judgment task is generated according to the original resource push demand judgment algorithm, the first enterprise credit data comment text sample of the credit risk control judgment task carries prior credit risk control judgment knowledge, and the first enterprise credit data comment text sample of the resource push demand judgment task carries prior resource push demand judgment knowledge, further comprising:

[0044] Determine a first algorithm discrimination cost based on the credit risk control prediction result and the prior credit risk control discrimination knowledge, and determine a second algorithm discrimination cost based on the resource push demand prediction result and the prior resource push demand discrimination knowledge;

[0045] The model parameters of the original comment text semantic mining model and the original structured text semantic processing model are improved according to the first algorithm discrimination cost and the second algorithm discrimination cost to obtain the comment text semantic mining model and the structured text semantic processing model.

[0046] In a second aspect, an artificial intelligence analysis system is provided, comprising a processor and a memory communicating with each other, wherein the processor is used to retrieve a computer program from the memory and implement the method described in the first aspect by running the computer program.

[0047] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the method described in the first aspect when running.

[0048] The embodiment of the present invention provides a credit data analysis and processing method, system and storage medium based on NLP, which can obtain the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, and the review text of the enterprise credit data to be analyzed includes multiple enterprise credit behavior events; the review text of the enterprise credit data to be analyzed is loaded into the review text semantic mining model for text semantic mining, and the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed is obtained; the original credit report text semantics is loaded into the structured text semantic processing model, and the text semantic vectors of different levels in the original credit report text semantics are spliced ​​to obtain the text semantic splicing vector; the text semantic splicing vector is loaded into the credit risk control discrimination algorithm and the resource push demand discrimination algorithm respectively, and the credit risk control opinion label corresponding to each enterprise credit behavior event in the review text of the enterprise credit data to be analyzed is predicted based on the credit risk control discrimination algorithm, and the resource push demand opinion label corresponding to each enterprise credit behavior event in the review text of the enterprise credit data to be analyzed is predicted based on the resource push demand discrimination algorithm.

[0049] In step 110, by obtaining the credit data review text of the enterprise to be analyzed, the credit data review information of the target enterprise is obtained. These data will become the basis for subsequent analysis and decision-making. In step 120, the original credit report text is semantically mined through the comment text semantic mining model to extract the implicit meaning and emotional information of the text. This helps to deeply understand the credit status of the target enterprise and the evaluation of related events. In step 130, the original credit report text semantics are processed using the structured text semantic processing model to extract text semantic vectors of different levels and splice them. This can integrate text features at different levels, making subsequent analysis and judgment more comprehensive and accurate. Through the credit risk control discrimination algorithm, corresponding credit risk control viewpoint labels are generated for each enterprise credit behavior event in step 140. These labels can reflect the credit risk level of the target enterprise in business interactions and help identify potential risks and problems. According to the resource push demand discrimination algorithm, corresponding resource push demand viewpoint labels are generated for each enterprise credit behavior event in step 140. These labels can guide the determination of targeted resource push strategies to ensure the effective distribution and utilization of resources. The collaboration between the text semantic mining model, the credit risk control identification algorithm, and the resource push demand identification algorithm makes the overall solution more accurate and practical. By comprehensively considering the semantic information of the evaluation text, the credit risk level, and the resource push demand, we can better understand the status of the target enterprise and take targeted measures for resource push and risk management.

[0050] In summary, the overall process from step 110 to step 140 realizes comprehensive analysis of the target enterprise credit data review text and personalized resource push through data processing, semantic mining, feature splicing, credit risk control and resource push demand identification. Such collaborative effect can improve the accuracy and security of credit risk management and resource push. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0052] Figure 1 The present invention provides a flowchart of a credit data analysis and processing method based on NLP. DETAILED DESCRIPTION

[0053] In order to better understand the above technical scheme, the technical scheme of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical scheme of the present invention, rather than limitations on the technical scheme of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0054] Figure 1 A credit data analysis and processing method based on NLP is shown, which is applied to an artificial intelligence analysis system. The method includes the following steps 110-140.

[0055] Step 110: Obtain the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, wherein the review text of the enterprise credit data to be analyzed includes multiple enterprise credit behavior events.

[0056] The credit data review text of the enterprise to be analyzed is the credit data review text for the target enterprise, and the credit data review text of the enterprise to be analyzed may be a credit evaluation report made by other systems or terminals on the target enterprise. The enterprise credit behavior event is a credit-related behavior event involved in the business interaction process of the target enterprise.

[0057] In this step, by communicating with the business resource push server, the review texts of the credit data of the target enterprise to be analyzed are obtained. These review texts are evaluations and comments on the credit data of the target enterprise. They can come from credit evaluation reports generated by other systems or terminals. These review texts of the credit data of the target enterprise to be analyzed cover multiple corporate credit behavior events, which refer to credit-related behavior events involved in the target enterprise during its business interaction. For example, it may include information on payment behavior, contract performance, credit default, etc. By obtaining these review texts of the credit data of the target enterprise to be analyzed, subsequent text semantic mining and analysis can be performed to understand the credit status and credit risk of the target enterprise.

[0058] When it comes to corporate credit behavior events, here are some detailed examples:

[0059] (1) Payment behavior: including whether the target company pays its suppliers or creditors’ bills on time, or whether it is late for payment or defaults;

[0060] (2) Contract performance: whether the target enterprise performs business cooperation in accordance with the terms and conditions of the contract, including the punctuality of product delivery or service provision and the quality compliance with requirements;

[0061] (3) Credit default: refers to the target enterprise's violation of contractual agreements or legal provisions, such as termination of the contract without reason, failure to perform contractual obligations, false advertising, infringement of intellectual property rights, etc.;

[0062] (4) Repayment capacity: This involves whether the target enterprise has sufficient liquidity and financial capacity to repay debts, pay interest, and meet other economic obligations;

[0063] (5) Customer complaints: reflects whether the target company has complaint problems in customer relationship management, such as frequent customer complaints, dissatisfaction with services, etc.

[0064] (6) Legal proceedings: including legal disputes, lawsuits and judicial decisions between the target company and other entities, which may have an impact on the company's credit.

[0065] Step 120: Load the review text of the enterprise credit data to be analyzed into the review text semantic mining model to perform text semantic mining to obtain the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed.

[0066] Among them, the comment text semantic mining model can be a deep structured semantic network.

[0067] In this step, the review text semantic mining model is used to process the review text of the enterprise credit data to be analyzed. This model can be a deep structured semantic network that can understand the semantic information in the text. For example, suppose there is a review text of the enterprise credit data to be analyzed: "The enterprise often overdue payments, resulting in tension in supplier relationships." Load this text into the review text semantic mining model, and the model will parse and semantically analyze the text through deep learning algorithms and natural language processing technology. As a result, the corresponding original credit report text semantics will be obtained. In this example, the mining model may identify overdue payment as an enterprise credit behavior event and extract relevant semantic information, such as the target enterprise's payment delay frequency, the impact on supplier relationships, etc. In this way, the original credit report text semantics of the enterprise credit behavior event is obtained. These semantic information will provide a basis for subsequent analysis and judgment.

[0068] The following are some examples when loading the review text of the corporate credit data to be analyzed into the review text semantic mining model for text semantic mining.

[0069] Example 1: "The company has continuously breached several contracts in the past year."

[0070] Semantic mining results: The target enterprise’s credit behavior event is identified as “default contract” and relevant semantic information is extracted, such as the number of defaults, time range, etc.

[0071] Example 2: "Disputes between the business and its customers have resulted in multiple legal actions."

[0072] Semantic mining results: The target company's credit behavior events are identified as "legal proceedings" and "customer disputes", and relevant semantic information is extracted, such as the number of lawsuits, related customers, etc.

[0073] Example 3: "The business has an excellent payment record and pays its vendor bills on time."

[0074] Semantic mining results: The target enterprise’s credit behavior events are identified as “good payment record” and “paying supplier bills on time”, and relevant semantic information is extracted, such as payment punctuality, supplier relationship positivity, etc.

[0075] Example 4: "The company frequently changes partners, which affects business stability."

[0076] Semantic mining results: The target enterprise’s credit behavior event is identified as “changing partners”, and relevant semantic information is extracted, such as frequency, impact on business stability, etc.

[0077] The above example shows that when the review text of the credit data of the enterprise to be analyzed is loaded into the review text semantic mining model, the model can automatically analyze and understand the meaning of the text based on the semantic network and deep learning algorithm, and extract relevant semantic information. This helps to have a deeper understanding and evaluation of the credit status and credit behavior events of the target enterprise.

[0078] Step 130: Load the original credit report text semantics into a structured text semantic processing model, perform text semantic vector splicing on text semantic vectors of different levels in the original credit report text semantics, and obtain a text semantic splicing vector.

[0079] Among them, text semantic vectors of different levels can be understood as text features at different levels.

[0080] In this step, the semantics of the original credit report text is loaded into the structured text semantic processing model. This model can process and analyze the text according to its semantic information. For example, suppose there are two text semantic vectors of different levels, representing the target company's "good payment record" and "default contract" respectively. These text semantic vectors are loaded into the structured text semantic processing model. The model will splice these text semantic vectors of different levels to form a comprehensive text semantic splicing vector. By splicing these vectors, more comprehensive and integrated text features can be obtained. For example, the spliced ​​text semantic splicing vector may contain comprehensive information about the target company's payment record, default situation, and other credit behavior events. This information will provide more accurate and comprehensive input data for the subsequent credit risk control judgment algorithm and resource push demand judgment algorithm.

[0081] Step 140: load the text semantic splicing vector into the credit risk control discrimination algorithm and the resource push demand discrimination algorithm respectively, and generate a credit risk control opinion label corresponding to each of the corporate credit behavior events in the corporate credit data review text to be analyzed according to the credit risk control discrimination algorithm, and generate a resource push demand opinion label corresponding to each of the corporate credit behavior events in the corporate credit data review text to be analyzed according to the resource push demand discrimination algorithm.

[0082] Among them, the credit risk control discrimination algorithm corresponds to the credit risk control discrimination subnet, and the resource push demand discrimination algorithm corresponds to the resource push demand discrimination subnet. The credit risk control opinion label can characterize the credit risk level of the target business. Based on this, the resource push demand opinion label corresponding to each enterprise credit behavior event can be determined in combination with the credit risk control opinion label, so as to combine the credit risk control opinion label and the resource push demand opinion label to push targeted business resources, thereby realizing personalized and secure resource push distribution. This can improve the technical problems that traditional technologies are difficult to accurately realize credit risk discrimination and difficult to realize personalized and secure resource push.

[0083] Suppose there is a corporate credit data review text to be analyzed, which contains three corporate credit behavior events: "Event A", "Event B" and "Event C".

[0084] In step 120, the semantics of the original credit report text is obtained by loading the comment text semantic mining model. In step 130, the semantics of the original credit report text is loaded into the structured text semantic processing model, and the text semantic vectors of different levels are spliced ​​to obtain the text semantic splicing vector. Now, in step 140, the text semantic splicing vector is loaded into the credit risk control judgment algorithm and the resource push demand judgment algorithm respectively.

[0085] For the credit risk control discrimination algorithm, it will generate a credit risk control opinion label for each enterprise credit behavior event based on predefined rules or models. For example, for "Event A", the credit risk control discrimination algorithm may generate a "high risk" opinion label, for "Event B", it may generate a "medium risk" opinion label, and for "Event C", it may generate a "low risk" opinion label. For the resource push demand discrimination algorithm, it will determine the resource push demand opinion label corresponding to each enterprise credit behavior event based on the credit risk control opinion label and predefined rules or models. For example, if "Event A" is marked as "high risk", the resource push demand discrimination algorithm can decide to push relevant resources to the target enterprise to reduce risks. Combining credit risk control opinion labels and resource push demand opinion labels can achieve targeted business resource push and ensure personalized and secure resource push distribution.

[0086] It can be seen that according to the credit risk control opinion label and resource push demand discrimination algorithm in step 140, the resource push demand opinion label corresponding to different credit risk control opinion labels can be determined, and different resource pushes can be performed.

[0087] For example, event A is marked as "high risk" by the credit risk control discrimination algorithm. Based on predefined rules or models, the resource push demand discrimination algorithm can determine the resource push demand opinion label related to "high risk". For example, it may be determined that the event needs to provide resources such as risk management consulting services, legal support or financial audits to help the target enterprise cope with potential credit risks. Event B is marked as "medium risk" by the credit risk control discrimination algorithm. The resource push demand discrimination algorithm can determine the resource push demand opinion label related to "medium risk". For example, it may be determined that the event needs to provide resources such as training courses, business process improvement suggestions or market research data to help the target enterprise reduce credit risk and improve business efficiency. Event C is marked as "low risk" by the credit risk control discrimination algorithm. The resource push demand discrimination algorithm can determine the resource push demand opinion label related to "low risk". For example, it may be determined that the event only needs to provide general business support, marketing recommendations or industry insight reports to help the target enterprise maintain a good credit status and promote business growth. By determining the resource push demand opinion label according to different credit risk levels, the required business resource support is provided in a targeted manner.

[0088] In step 110, by obtaining the credit data review text of the enterprise to be analyzed, the credit data review information of the target enterprise is obtained. These data will become the basis for subsequent analysis and decision-making. In step 120, the original credit report text is semantically mined through the comment text semantic mining model to extract the implicit meaning and emotional information of the text. This helps to deeply understand the credit status of the target enterprise and the evaluation of related events. In step 130, the original credit report text semantics are processed using the structured text semantic processing model to extract text semantic vectors of different levels and splice them. This can integrate text features at different levels, making subsequent analysis and judgment more comprehensive and accurate. Through the credit risk control discrimination algorithm, corresponding credit risk control viewpoint labels are generated for each enterprise credit behavior event in step 140. These labels can reflect the credit risk level of the target enterprise in business interactions and help identify potential risks and problems. According to the resource push demand discrimination algorithm, corresponding resource push demand viewpoint labels are generated for each enterprise credit behavior event in step 140. These labels can guide the determination of targeted resource push strategies to ensure the effective distribution and utilization of resources. The collaboration between the text semantic mining model, the credit risk control identification algorithm, and the resource push demand identification algorithm makes the overall solution more accurate and practical. By comprehensively considering the semantic information of the evaluation text, the credit risk level, and the resource push demand, we can better understand the status of the target enterprise and take targeted measures for resource push and risk management.

[0089] In summary, the overall process from step 110 to step 140 realizes comprehensive analysis of the target enterprise credit data review text and personalized resource push through data processing, semantic mining, feature splicing, credit risk control and resource push demand identification. Such collaborative effect can improve the accuracy and security of credit risk management and resource push.

[0090] In some optional embodiments, the comment text semantic mining model includes multiple state stages. Based on this, the step 120 of loading the comment text of the enterprise credit data to be analyzed into the comment text semantic mining model for text semantic mining to obtain the original credit report text semantics corresponding to the comment text of the enterprise credit data to be analyzed includes steps 121 and 122.

[0091] Step 121: Load the review text of the enterprise credit data to be analyzed into the review text semantic mining model.

[0092] Step 122: perform text semantic vector splicing within a state level based on a first set splicing strategy, and perform combination between different state levels based on a second set splicing strategy, to obtain the original credit report text semantics of the corporate credit data review text to be analyzed generated by a sub-model of each state level.

[0093] The comment text semantic mining model can be a deep structured semantic network under the NLP technology framework, including multiple levels (i.e., the above-mentioned state level). On this basis, the first setting splicing strategy can be a layered deep splicing strategy, and the second setting splicing strategy can be an iterative deep splicing strategy.

[0094] In step 121, loading the review text of the enterprise credit data to be analyzed into the review text semantic mining model means inputting the review text into the model for processing. For example, in the case of using a deep structured semantic network as a review text semantic mining model, the review text can be input into the network for hierarchical processing to obtain the text semantics generated by the sub-model of each state order.

[0095] In step 122, performing text semantic vector splicing within a state level based on the first set splicing strategy means using a specific splicing strategy to integrate different text semantic vectors within each state level. For example, if a hierarchical deep splicing strategy is used, the text semantic vectors within each state level can be spliced ​​hierarchically to form a comprehensive representation for the state level.

[0096] The combination between different state levels based on the second setting splicing strategy is to use another splicing strategy to combine different state levels. Taking the iterative deep splicing strategy as an example, information transmission and combination can be achieved by iteratively splicing the output of the previous state level with the input of the current state level. In this way, the text semantics generated by the sub-model of each state level will be gradually integrated and enriched.

[0097] In general, the combined beneficial effect of step 121 and step 122 is based on the multi-state design of the semantic mining model. By loading the review text into the semantic mining model for processing, and then integrating the text semantic vectors within and between different state levels according to the splicing strategy, a more comprehensive and accurate original credit report text semantics can be obtained. This method can make full use of the semantic information of multiple state levels, and improve the understanding and analysis capabilities of the enterprise credit data review text in a comprehensive and iterative manner, thereby improving the overall effect.

[0098] Under some preferred design ideas, the step 130 of loading the original credit report text semantics into a structured text semantic processing model, performing text semantic vector splicing on text semantic vectors of different levels in the original credit report text semantics, and obtaining a text semantic splicing vector may include steps 131 to 133.

[0099] Step 131: Perform semantic feature expansion processing on the original credit report text semantics with the smallest feature scale among the original credit report text semantics generated by the sub-model of each state order to obtain a text semantic expansion vector.

[0100] Step 132: Based on the text semantic extension vector, adjust the original credit report text semantics to be consistent with the feature scale of the text semantic extension vector.

[0101] Step 133, jump to the step of performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale in the original credit report text semantics generated by the sub-model of each state order (i.e., step 131), until the original credit report text semantics generated by the sub-model of each state order are spliced ​​to obtain a text semantic splicing vector.

[0102] In step 131, semantic feature expansion processing is performed on the vector with the smallest feature scale in the original credit report text semantics generated by the sub-model of each state order. This means that the text semantic vector with a smaller dimension is processed to expand its semantic expression ability. For example, when processing an original credit report, if the text semantic vector generated by a certain state order has a small dimension, the semantic representation of the vector can be expanded by applying technical methods (such as word embedding, sequence model, etc.) to obtain richer and more accurate semantic information.

[0103] In step 132, the original credit report text semantics is adjusted to keep the feature scale of the extended vector consistent with that of the text semantics extension vector. This means that the original credit report text semantics is transformed or adjusted accordingly to match or adapt to the feature scale of the extended vector. For example, a linear transformation or other adjustment method may be used to adjust the original credit report text semantics in dimension so that it has the same feature scale as the extended vector.

[0104] In step 133, the process jumps to the step of performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale (i.e., step 131). This means that the text semantic vector generated by the sub-model of each state level is processed in a loop until the original credit report text semantics generated by all state levels are expanded. In this way, a text semantic concatenation vector containing all state levels can be obtained in the end, in which the semantic information of each state level has been expanded and adjusted.

[0105] In general, the design idea of ​​steps 131 to 133 is to use the structured text semantic processing model to splice text semantic vectors of different levels, and to obtain a more comprehensive and accurate text semantic expression through semantic feature expansion and adjustment. By expanding the semantics of the original credit report text generated at different state levels, the ability to understand each state level can be improved and integrated into a more informative text semantic splicing vector. This method helps to improve the semantic analysis ability of the original credit report, thereby enhancing the overall effect and accuracy.

[0106] Under some exemplary design ideas, the step 132 described in which the original credit report text semantics is adjusted to be consistent with the feature scale of the text semantic extension vector based on the text semantic extension vector includes steps 1321 and 1322.

[0107] Step 1321: perform a semantic logic-based intersection on the text semantic expansion vector and the original credit report text semantics that is consistent with the feature scale of the text semantic expansion vector to obtain a text semantic intersection result.

[0108] Step 1322: perform text semantic vector splicing on the text semantic cross-over result based on the reversible unit to obtain the original text semantic splicing vector, and adjust the original text semantic splicing vector to the original credit report text semantics that is consistent with the feature scale of the text semantic extension vector.

[0109] In step 1321, the text semantics expansion vector and the original credit report text semantics with the same feature scale are cross-linked based on semantic logic according to the feature scale of the text semantics expansion vector. This means that the two vectors are cross-linked by applying semantic logic operations (such as multiplication, addition, element-by-element operations, etc.) to obtain the association and complementary information between them. For example, assuming that the text semantics expansion vector is [0.8, 0.5], and the original credit report text semantics is [0.6, 0.3], the two vectors can be element-by-element multiplied or added to obtain a cross-linked result such as [0.48, 0.15], where the value at each position represents the cross-linked result of the semantic feature at the corresponding position.

[0110] In step 1322, the text semantics vectors of the text semantic crossover results are spliced ​​based on the reversible unit. This means that the text semantic vector after the crossover processing is spliced ​​with the original credit report text semantic vector to form a more complete original text semantic splicing vector. At the same time, the feature scale of this splicing vector is adjusted to be consistent with the feature scale of the text semantic extension vector. For example, if the text semantic crossover result is [0.48, 0.15], and the feature scale of the text semantic extension vector is 2, the two vectors can be spliced ​​to obtain the original text semantic splicing vector such as [0.48, 0.15, 0.6, 0.3]. Then, the dimension of this splicing vector can be adjusted to match the feature scale of the text semantic extension vector, for example, by truncation or zero padding.

[0111] In summary, the design idea of ​​step 1321-step 1322 is to fuse and expand the text semantic extension vector with the original credit report text semantics through the splicing of cross- and reversible units based on semantic logic, and generate a more comprehensive and accurate original text semantic splicing vector. This method can use semantic logic operations to capture the correlation between different vectors, and integrate multiple vectors into a unified representation through splicing operations. By adjusting the feature scale of the splicing vector to keep it consistent with the text semantic extension vector, the consistency and compatibility between vectors are further enhanced. In this way, when further analysis and processing are performed, the text semantic information can be more effectively utilized to improve the overall effect and accuracy.

[0112] In some possible design ideas, the credit risk control opinion label includes the credit risk control keywords output by the credit risk control discrimination algorithm, the credit risk control suggestion label output by the credit risk control discrimination algorithm, and the feature scale variable output by the credit risk control discrimination algorithm. Based on this, step 140 loads the text semantic concatenation vector into the credit risk control discrimination algorithm, and generates the credit risk control opinion label corresponding to each of the corporate credit behavior events in the corporate credit data review text to be analyzed according to the credit risk control discrimination algorithm, including steps 141-144.

[0113] Step 141: Load the text semantic concatenation vector into a credit risk control discrimination algorithm.

[0114] Step 142: Generate credit risk control keywords output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed based on the first AI discrimination component in the credit risk control discrimination algorithm.

[0115] Step 143: Generate a credit risk control recommendation label output by the credit risk control discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the first gradient optimization component in the credit risk control discrimination algorithm.

[0116] Step 144: Generate, based on the first feature scale output component in the credit risk control discrimination algorithm, a feature scale variable output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed.

[0117] Among them, the AI ​​discrimination component includes at least a classification branch, the gradient optimization component includes at least a residual branch, and the feature scale output component includes at least a scale branch.

[0118] In step 141, the text semantic concatenation vector is loaded into the credit risk control discrimination algorithm. This means that the text semantic concatenation vector is provided as input to the credit risk control discrimination algorithm so that the algorithm can use the vector to analyze and evaluate the credit risk of the enterprise.

[0119] In step 142, based on the first AI identification component of the credit risk control identification algorithm, a credit risk control keyword corresponding to each enterprise credit behavior event is generated. For example, assuming that the credit behavior event of a certain enterprise is "overdue repayment", the credit risk control identification algorithm may output keywords related to this event, such as "default, arrears, bad credit", etc. These keywords can help further understand the credit status and potential risks of the enterprise.

[0120] In step 143, based on the first gradient optimization component of the credit risk control discrimination algorithm, a credit risk control suggestion label corresponding to each corporate credit behavior event is generated. Based on the output of the algorithm, credit risk control suggestions for each event may be generated, such as "strengthening collection measures, limiting credit limits, increasing risk reserves, etc." These suggestion labels help guide the adoption of appropriate risk control measures to reduce potential credit losses.

[0121] In step 144, based on the first feature scale output component of the credit risk control discrimination algorithm, feature scale variables corresponding to each enterprise credit behavior event are generated. These feature scale variables may provide numerical measurements or other relevant information about the enterprise credit behavior event. For example, for the "overdue repayment" event, the feature scale variables may include the number of overdue days, the overdue amount, etc. These variables can help further evaluate and quantify the credit risk level of the enterprise.

[0122] The design idea of ​​steps 141-144 is to use the text semantic concatenation vector to load into the credit risk control discrimination algorithm, and generate credit risk control opinion labels related to corporate credit behavior events through different components of the algorithm. These labels include keywords, recommended labels, and feature scale variables, providing multi-dimensional assessment and insight into corporate credit risk. By combining text semantic information with the credit risk control algorithm, a more comprehensive understanding of the credit status and risk characteristics of the enterprise can be achieved, and more targeted credit risk control recommendations and predictions can be provided to decision makers. This can improve the accuracy and efficiency of credit risk control and reduce potential losses caused by credit risks.

[0123] In some preferred embodiments, the resource push demand viewpoint label includes resource push demand keywords, push mode labels output by the resource push demand discrimination algorithm, feature scale variables output by the resource push demand discrimination algorithm, and push content directory. Based on this, the step 140 of loading the text semantic concatenation vector into the resource push demand discrimination algorithm, and generating the resource push demand viewpoint label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed according to the resource push demand discrimination algorithm, includes steps 140a-140f.

[0124] Step 140a: Load the text semantic concatenation vector into a resource push demand determination algorithm.

[0125] Step 140b: Generate credit risk control keywords output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the second AI discrimination component in the resource push demand discrimination algorithm.

[0126] Step 140c: Generate a push mode label output by the resource push demand identification algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the second gradient optimization component in the resource push demand identification algorithm.

[0127] Step 140d: Generate a feature scale variable output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the second feature scale output component in the resource push demand discrimination algorithm.

[0128] Step 140e: Generate frequent item behavior preference thermal data corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the behavior thermal analysis component in the resource push demand discrimination algorithm.

[0129] Step 140f: Generate a push content directory for each enterprise credit behavior event in the enterprise credit data review text to be analyzed based on the push content directory generation component in the resource push demand identification algorithm.

[0130] In step 140a, the text semantic concatenation vector is loaded into the resource push demand identification algorithm, which means that the text semantic concatenation vector is provided as input to the resource push demand identification algorithm so that the algorithm can use the vector to analyze and evaluate the resource push demand of the enterprise.

[0131] In step 140b, based on the second AI discrimination component of the resource push demand discrimination algorithm, resource push demand keywords corresponding to each enterprise credit behavior event are generated. For example, assuming that the credit behavior event of an enterprise is "increased capital demand", the resource push demand discrimination algorithm may output keywords related to the event, such as "financing, loan, investment", etc. These keywords can help further understand the resource demand of the enterprise.

[0132] In step 140c, based on the second gradient optimization component of the resource push demand discrimination algorithm, a push mode label corresponding to each enterprise credit behavior event is generated. According to the output of the algorithm, a push mode label for each event may be generated, such as "targeted push, precise recommendation, full network advertising", etc. These labels can guide the method and strategy of resource push to meet the specific needs of enterprises.

[0133] In step 140d, based on the second feature scale output component of the resource push demand discrimination algorithm, feature scale variables corresponding to each enterprise credit behavior event are generated. These feature scale variables may provide numerical measurements or other relevant information about the enterprise credit behavior event. For example, for the "increase in capital demand" event, the feature scale variables may include the amount of capital, the financing period, etc. These variables can help further evaluate and quantify the level of resource demand of the enterprise.

[0134] In step 140e, the behavior thermal analysis component based on the resource push demand discrimination algorithm generates frequent behavior preference thermal data corresponding to each enterprise credit behavior event. These data can reveal the behavior preferences and behavior patterns of enterprises under specific credit behavior events. For example, in the event of "increased capital demand", the component may analyze which financing method the enterprise is more inclined to use or which investment opportunities it chooses.

[0135] In step 140f, the push content directory generation component based on the resource push demand discrimination algorithm generates a push content directory for each enterprise credit behavior event. This means that according to the output of the algorithm, a corresponding push content directory is generated for each event, which lists the available resource push content. This helps to provide personalized resource push options to meet the specific needs of the enterprise.

[0136] The design idea of ​​steps 140a-140f is to load the text semantic concatenation vector into the resource push demand identification algorithm, and generate resource push demand opinion labels related to corporate credit behavior events through different components of the algorithm. These labels include keywords, push mode labels, feature scale variables, frequent item behavior preference thermal data, and push content directory. This can provide a more comprehensive understanding of the resource needs and preferences of the enterprise, and provide customized solutions and options for resource push. By combining text semantic information with the resource push demand identification algorithm, the accuracy and effect of resource push can be improved, and the matching degree between enterprises and resource suppliers can be enhanced, thereby achieving overall efficiency improvement.

[0137] In some optional embodiments, before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server as described in step 110, the method also includes steps 011 to 014.

[0138] Step 011: Obtain a first corporate credit data review text sample for a credit risk control identification task.

[0139] Step 012: Load the first enterprise credit data review text sample of the credit risk control judgment task into the original review text semantic mining model for text semantic mining to obtain the first credit report text semantic sample of the original credit risk control judgment task corresponding to the first enterprise credit data review text sample of the credit risk control judgment task.

[0140] Step 013: Input the first credit report text semantic sample of the original credit risk control judgment task into the original structured text semantic processing model, perform text semantic vector splicing on text semantic vectors of different levels in the first credit report text semantic sample of the original credit risk control judgment task, and obtain a first text semantic splicing vector sample.

[0141] Step 014: input the first text semantic concatenation vector sample into the original credit risk control discrimination algorithm, and generate a credit risk control prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the credit risk control discrimination task based on the original credit risk control discrimination algorithm.

[0142] In step 011, the first corporate credit data review text sample of the credit risk control identification task is obtained. This means obtaining a representative corporate credit data review text sample from the business resource push server, which will be used for subsequent processing and analysis.

[0143] In step 012, the first corporate credit data review text sample is loaded into the original review text semantic mining model for text semantic mining. By performing semantic mining on the sample, the text semantic information contained in the sample can be extracted. For example, if the sample is a comment: "This company has poor service", the semantic mining model may identify key information therein, such as "poor service quality".

[0144] In step 013, the first credit report text semantic sample of the original credit risk control discrimination task is input into the original structured text semantic processing model. In this step, the text semantic vectors of different levels in the first credit report text semantic sample are spliced ​​to obtain an overall text semantic splicing vector sample. The purpose of this is to fuse and retain semantic information at different levels in one vector so as to more comprehensively represent the characteristics of the credit risk control task.

[0145] In step 014, the first text semantic concatenation vector sample is input into the original credit risk control discrimination algorithm, and a credit risk control prediction result is generated according to the algorithm to evaluate the credit risk of each corporate credit behavior event in the first corporate credit data comment text sample. For example, if a company's comment mentions "default behavior", the credit risk control discrimination algorithm may predict that the company has credit risks.

[0146] Through the processing flow of step 011-step 014, the enterprise credit data comment text can be processed and analyzed, key semantic information can be extracted from it, and predictions can be made using the credit risk control algorithm. This helps to fully understand the credit status of the enterprise, identify potential risk factors, and provide a reference for subsequent resource push demand analysis and decision-making. Through accurate identification and risk prediction of enterprise credit data, the efficiency and accuracy of enterprise credit management can be improved, helping institutions to better manage and allocate resources, reduce losses caused by credit risks, and enhance the reliability of credit-related decisions.

[0147] In some optional embodiments, before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server as described in step 110, steps 021 to 024 are also included.

[0148] Step 021: Obtain a text sample of the second enterprise credit data review.

[0149] Step 022: load the first enterprise credit data review text sample of the resource push demand identification task into the original review text semantic mining model for text semantic mining, and obtain the first credit report text semantic sample of the original resource push demand identification task corresponding to the first enterprise credit data review text sample of the resource push demand identification task.

[0150] Step 023: Input the first credit report text semantic sample of the original resource push demand identification task into the original structured text semantic processing model, perform text semantic vector splicing on the text semantic vectors of different levels in the first credit report text semantic sample of the original resource push demand identification task, and obtain a second credit report text semantic sample.

[0151] Step 024: input the second credit report text semantic sample into the original resource push demand judgment algorithm, and generate the resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the resource push demand judgment task based on the original resource push demand judgment algorithm.

[0152] In step 021, a second corporate credit data review text sample is obtained, which means obtaining another representative corporate credit data review text sample from the business resource push server, and the sample will be used for resource push demand analysis.

[0153] In step 022, the first corporate credit data review text sample is loaded into the original review text semantic mining model for text semantic mining. By performing semantic mining on the sample, the text semantic information contained in the sample can be extracted. For example, if the sample is a comment: "This company has poor service", the semantic mining model may identify key information therein, such as "poor service quality".

[0154] In step 023, the first credit report text semantic sample of the original resource push demand identification task is input into the original structured text semantic processing model. In this step, the text semantic vectors of different levels in the first credit report text semantic sample are spliced ​​to obtain an overall credit report text semantic sample. The purpose of this is to fuse and retain semantic information at different levels in one vector so as to more comprehensively represent the characteristics of resource push demand.

[0155] In step 024, the second credit report text semantic sample is input into the original resource push demand discrimination algorithm, and a resource push demand prediction result is generated according to the algorithm to evaluate the resource push demand of each enterprise credit behavior event in the first enterprise credit data comment text sample. For example, if a certain enterprise's comment mentions "expanding production scale", the resource push demand discrimination algorithm may predict that the enterprise has a large demand for production equipment and human resources.

[0156] Through the processing flow of step 021 to step 024, the resource push demand analysis can be performed based on the second enterprise credit data comment text sample. This helps to deeply understand the enterprise's resource demand situation and provide the enterprise with a corresponding resource push solution based on the results of semantic mining and demand prediction. By comprehensively analyzing the semantic information and demand prediction results of multiple samples, the enterprise's resource demand can be understood more comprehensively and accurately, thereby improving resource utilization efficiency, accuracy and personalization.

[0157] In some examples, the first enterprise credit data comment text sample of the credit risk control judgment task carries prior credit risk control judgment knowledge, and the first enterprise credit data comment text sample of the resource push demand judgment task carries prior resource push demand judgment knowledge. Based on this, after the resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the resource push demand judgment task is generated according to the original resource push demand judgment algorithm as described in step 024, steps 025 and 026 are also included.

[0158] Step 025: determine a first algorithm discrimination cost based on the credit risk control prediction result and the prior credit risk control discrimination knowledge, and determine a second algorithm discrimination cost based on the resource push demand prediction result and the prior resource push demand discrimination knowledge.

[0159] Step 026: Improve the model parameters of the original comment text semantic mining model and the original structured text semantic processing model based on the first algorithm discrimination cost and the second algorithm discrimination cost to obtain the comment text semantic mining model and the structured text semantic processing model.

[0160] In step 025, the discrimination cost of the first algorithm is determined based on the credit risk control prediction result and the prior credit risk control discrimination knowledge, and the discrimination cost of the second algorithm is determined based on the resource push demand prediction result and the prior resource push demand discrimination knowledge. These discrimination costs can be used to measure the cost or benefit of the two algorithms when making discrimination decisions. For example, in terms of credit risk control, if the prediction result of a certain credit behavior event is consistent with the prior knowledge, the discrimination cost may be low; if the prediction result is contrary to the prior knowledge, the discrimination cost may be high.

[0161] In step 026, based on the discrimination cost of the first algorithm and the discrimination cost of the second algorithm, the model parameters of the original comment text semantic mining model and the original structured text semantic processing model are improved. The purpose of this is to adjust the model parameters so that it can better adapt to the requirements of prior knowledge and discrimination cost. This will obtain an improved comment text semantic mining model and structured text semantic processing model, which will have better performance in credit risk control and resource push demand discrimination tasks.

[0162] Through the processing flow of step 025 and step 026, the model can be improved according to prior knowledge and discrimination cost to improve the effect of credit risk control and resource push demand discrimination tasks. By considering prior knowledge, domain expertise or empirical rules can be introduced to improve the accuracy and interpretability of the model. At the same time, adjusting the model parameters according to the discrimination cost can better balance the costs and benefits of different discrimination decisions, making the system more reliable and cost-effective in practical applications. Such improvements help optimize the decision-making process, improve resource utilization efficiency, reduce decision-making risks, and enhance the adaptability and operability of the system.

[0163] On the basis of the above, an artificial intelligence analysis system is provided, comprising a processor and a memory communicating with each other, wherein the processor is used to retrieve a computer program from the memory and implement the above method by running the computer program.

[0164] Based on the above, a computer-readable storage medium is provided, on which a computer program is stored, and the computer program implements the above method when running.

[0165] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A credit data analysis and processing method based on NLP, characterized in that: Applied to an artificial intelligence analysis system, the method comprises: Acquire a review text of the enterprise credit data to be analyzed uploaded by the business resource push server, wherein the review text of the enterprise credit data to be analyzed includes a plurality of enterprise credit behavior events; Loading the review text of the enterprise credit data to be analyzed into the review text semantic mining model to perform text semantic mining to obtain the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed; The original credit report text semantics is loaded into a structured text semantic processing model, and text semantic vectors of different levels in the original credit report text semantics are concatenated to obtain a text semantic concatenation vector; The text semantic concatenation vector is loaded into the credit risk control discrimination algorithm and the resource push demand discrimination algorithm respectively, and the credit risk control opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the credit risk control discrimination algorithm, and the resource push demand opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the resource push demand discrimination algorithm; The review text of the enterprise credit data to be analyzed is loaded into the review text semantic mining model for text semantic mining to obtain the original credit report text semantics corresponding to the review text of the enterprise credit data to be analyzed. The review text semantic mining model includes multiple state stages, including: Loading the review text of the enterprise credit data to be analyzed into the review text semantic mining model; Based on the first set splicing strategy, the text semantic vectors within the state level are spliced, and based on the second set splicing strategy, different state levels are combined to obtain the original credit report text semantics of the corporate credit data review text to be analyzed generated by the sub-model of each state level; The step of loading the original credit report text semantics into a structured text semantics processing model, and performing text semantic vector splicing on text semantic vectors of different levels in the original credit report text semantics to obtain a text semantic splicing vector includes: Performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale among the original credit report text semantics generated by the sub-model of each state order to obtain a text semantic expansion vector; According to the text semantic extension vector, adjusting the original credit report text semantics consistent with the feature scale of the text semantic extension vector; Jump to the step of performing semantic feature expansion processing on the original credit report text semantics with the smallest feature scale in the original credit report text semantics generated by the sub-model of each state order, until the original credit report text semantics generated by the sub-model of each state order are spliced ​​to obtain a text semantic splicing vector.

2. The method according to claim 1, characterized in that The step of adjusting the original credit report text semantics consistent with the feature scale of the text semantics extension vector according to the text semantics extension vector includes: Performing a semantic logic-based intersection between the text semantic extension vector and the original credit report text semantics that is consistent with the feature scale of the text semantic extension vector, to obtain a text semantic intersection result; The text semantic crossover result is subjected to text semantic vector splicing based on a reversible unit to obtain an original text semantic splicing vector, and the original text semantic splicing vector is adjusted to the original credit report text semantics that is consistent with the feature scale of the text semantic expansion vector.

3. The method according to claim 1, characterized in that The text semantic concatenation vector is loaded into the credit risk control discrimination algorithm, and a credit risk control opinion label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the credit risk control discrimination algorithm. The credit risk control opinion label includes the credit risk control keywords output by the credit risk control discrimination algorithm, the credit risk control suggestion label output by the credit risk control discrimination algorithm, and the feature scale variable output by the credit risk control discrimination algorithm, including: Loading the text semantic concatenation vector into the credit risk control discrimination algorithm; Generate, based on the first AI discrimination component in the credit risk control discrimination algorithm, a credit risk control keyword output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed; Generating a credit risk control recommendation label output by the credit risk control discrimination algorithm corresponding to each corporate credit behavior event in the corporate credit data review text to be analyzed according to the first gradient optimization component in the credit risk control discrimination algorithm; A feature scale variable output by the credit risk control discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed is generated according to the first feature scale output component in the credit risk control discrimination algorithm.

4. The method according to claim 1, characterized in that The text semantic concatenation vector is loaded into the resource push demand discrimination algorithm, and a resource push demand viewpoint label corresponding to each of the enterprise credit behavior events in the enterprise credit data review text to be analyzed is generated according to the resource push demand discrimination algorithm, wherein the resource push demand viewpoint label includes resource push demand keywords, a push mode label output by the resource push demand discrimination algorithm, a feature scale variable output by the resource push demand discrimination algorithm, and a push content directory, including: Loading the text semantic splicing vector into the resource push demand determination algorithm; Generate, based on the second AI discrimination component in the resource push demand discrimination algorithm, a credit risk control keyword output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed; Generate a push mode label output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the second gradient optimization component in the resource push demand discrimination algorithm; Generate a feature scale variable output by the resource push demand discrimination algorithm corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the second feature scale output component in the resource push demand discrimination algorithm; Generate frequent item behavior preference thermal data corresponding to each enterprise credit behavior event in the enterprise credit data review text to be analyzed according to the behavior thermal analysis component in the resource push demand discrimination algorithm; A push content directory generation component in the resource push demand identification algorithm generates a push content directory for each enterprise credit behavior event in the enterprise credit data review text to be analyzed.

5. The method according to claim 1, characterized in that Before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, the method further includes: Obtain a sample of the first enterprise credit data review text for the credit risk control identification task; Loading the first enterprise credit data review text sample of the credit risk control judgment task into the original review text semantic mining model for text semantic mining, and obtaining the first credit report text semantic sample of the original credit risk control judgment task corresponding to the first enterprise credit data review text sample of the credit risk control judgment task; Inputting the first credit report text semantic sample of the original credit risk control discrimination task into the original structured text semantic processing model, performing text semantic vector splicing on text semantic vectors of different levels in the first credit report text semantic sample of the original credit risk control discrimination task, and obtaining a first text semantic splicing vector sample; The first text semantic concatenation vector sample is input into the original credit risk control discrimination algorithm, and a credit risk control prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data review text sample of the credit risk control discrimination task is generated according to the original credit risk control discrimination algorithm.

6. The method according to claim 5, characterized in that Before obtaining the review text of the enterprise credit data to be analyzed uploaded by the business resource push server, the method further includes: Obtain a sample of the second enterprise credit data review text; Loading the first enterprise credit data review text sample of the resource push demand determination task into the original review text semantic mining model for text semantic mining, and obtaining the first credit report text semantic sample of the original resource push demand determination task corresponding to the first enterprise credit data review text sample of the resource push demand determination task; Inputting the first credit report text semantic sample of the original resource push demand discrimination task into the original structured text semantic processing model, performing text semantic vector splicing on text semantic vectors of different levels in the first credit report text semantic sample of the original resource push demand discrimination task, and obtaining a second credit report text semantic sample; Inputting the second credit report text semantic sample into the original resource push demand discrimination algorithm, and generating a resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data review text sample of the resource push demand discrimination task according to the original resource push demand discrimination algorithm; Among them, after the resource push demand prediction result corresponding to each enterprise credit behavior event in the first enterprise credit data comment text sample of the resource push demand judgment task is generated according to the original resource push demand judgment algorithm, the first enterprise credit data comment text sample of the credit risk control judgment task carries prior credit risk control judgment knowledge, and the first enterprise credit data comment text sample of the resource push demand judgment task carries prior resource push demand judgment knowledge, further comprising: Determine a first algorithm discrimination cost based on the credit risk control prediction result and the prior credit risk control discrimination knowledge, and determine a second algorithm discrimination cost based on the resource push demand prediction result and the prior resource push demand discrimination knowledge; The model parameters of the original comment text semantic mining model and the original structured text semantic processing model are improved according to the first algorithm discrimination cost and the second algorithm discrimination cost to obtain the comment text semantic mining model and the structured text semantic processing model.

7. An artificial intelligence analysis system, characterized in that: The artificial intelligence analysis system includes a processor and a memory that communicate with each other, and the processor is used to call a computer program from the memory and implement the method described in any one of claims 1 to 6 by running the computer program.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program implements the method according to any one of claims 1 to 6 when running.

Images