A credit risk control method and device based on dynamic text information extraction

Abstract

The embodiment of the application provides a credit risk control method and equipment based on dynamic text information extraction, the method comprises the following steps: obtaining user information reflecting user credit, and determining the mandatory association level between the user information and the user and the dynamic influence level of the user information on the user according to the category of the user information; performing dynamic text recognition on the user information in the current stage to extract the first credit feature corresponding to the user in the current stage; performing risk assessment on the user to generate a lending strategy for the user according to the evaluation result of the risk assessment; selecting specified user information with a dynamic influence level higher than a preset influence threshold, and when the stage corresponding to the specified user information changes, performing dynamic text recognition on the specified user information in the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage; and modifying the lending strategy according to the compensated evaluation result.

Description

Technical Field

[0001] This application relates to the field of credit risk control technology, specifically to a credit risk control method and device based on dynamic text information extraction. Background Technology

[0002] Credit is a credit activity in which a money holder temporarily lends out an agreed amount of funds at an agreed interest rate, and the borrower repays the principal and interest within an agreed period and under agreed conditions. Credit risk prediction is a set of decision support technologies that assists lending institutions in issuing consumer credit. When making a loan, lending institutions need to assess the user's creditworthiness to determine the appropriate loan amount and repayment strategy.

[0003] However, user information (where users can be individuals or businesses) is dynamic, and changes in certain information are likely to affect the user's risk assessment results, thereby affecting the accuracy of lending strategies. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes a credit risk control method based on dynamic text information extraction, comprising: acquiring user information used to reflect user credit, and determining the mandatory association level between the user information and the user, and the dynamic impact level of the user information on the user, according to the category to which the user information belongs;

[0005] Based on the preset dynamic recognition pattern in the dynamic text information, the user information in the current stage is subjected to dynamic text recognition to extract the first credit feature of the user in the current stage.

[0006] Based on the mandatory association level corresponding to the first credit feature, a risk assessment is performed on the user to generate a lending strategy for the user based on the assessment results of the risk assessment.

[0007] Among all the user information corresponding to the user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage.

[0008] The assessment result is compensated based on the second credit characteristic, and the lending strategy is modified based on the compensated assessment result.

[0009] In one implementation of this application, before performing dynamic text recognition on the user information at the current stage based on a preset dynamic recognition pattern in the dynamic text information, the method further includes:

[0010] Collect several user information entries, and perform keyword recognition on each of the user information entries to obtain the recognition results;

[0011] A training set is generated based on the recognition results, such that the training set includes multiple training samples, each training sample includes at least one triplet, each triplet carries three recognition results, and among the three recognition results, the similarity between the first recognition result and the second recognition result on the first credit assessment benchmark is higher than a first preset threshold, the similarity between the first recognition result and the third recognition result on the second credit assessment benchmark is lower than the second preset threshold, the second preset threshold is lower than the first preset threshold, the first credit assessment benchmark and the second credit assessment benchmark are preset and have corresponding physical meanings;

[0012] The pre-generated model is trained based on the training set until the model converges, resulting in a trained credit feature extraction model. The user information is then identified based on the credit feature extraction model to obtain credit features.

[0013] In one implementation of this application, keyword recognition is performed on the aforementioned user information to obtain recognition results, specifically including:

[0014] For each piece of user information, the user information is segmented into words to obtain multiple keywords after segmentation;

[0015] Based on the attributes of each keyword and the first mapping relationship set in the preset corpus, the positive and negative relationships corresponding to each keyword on each credit rating benchmark are determined, and the positive and negative relationships include positive relationships and negative relationships.

[0016] Based on the frequency of occurrence of the positive and negative relationships of each keyword on each credit rating criterion, keywords that do not meet the preset requirements are filtered out, and the remaining keywords are used as the identification results of keyword recognition.

[0017] In one implementation of this application, after using the remaining keywords as the recognition result of keyword recognition, the method further includes:

[0018] Determine the first-dimensional matrix and the second-dimensional matrix corresponding to the first recognition result and the second recognition result;

[0019] Based on the first dimension matrix, the second norm of the dimensions in the second dimension matrix that are located at the same position as the first dimension matrix is ​​calculated, so as to obtain the similarity of the dimensions based on the second norm; the same position means that the rows and columns corresponding to the dimensions are the same;

[0020] Construct a dimension similarity matrix based on the similarity of the dimensions;

[0021] The number of zero values ​​in the dimensional similarity matrix is ​​determined, and the similarity between the first identification result and the second identification result on the credit rating benchmark is determined based on the number of zero values.

[0022] In one implementation of this application, risk assessment of the user is performed based on the mandatory association level corresponding to the first credit feature, specifically including:

[0023] The range to which the forced association level belongs is determined. The range includes at least three ranges: a first range that is higher than the first preset association level, a second range that is lower than the first preset association level but higher than the second preset association level, and a third range that is lower than the second preset association level.

[0024] Select the first credit feature that belongs to the first range of the forced association level, and construct the user model corresponding to the user based on it;

[0025] Using the first credit feature as a risk assessment parameter, a risk assessment model is generated, wherein the weight of the first credit feature corresponding to the second range is higher than that of the first credit feature corresponding to the first range and the third range.

[0026] The risk assessment model is fused with the user model to perform risk assessment on the user using the fused user model.

[0027] In one implementation of this application, determining the mandatory association level between the user information and the user based on the category to which the user information belongs specifically includes:

[0028] Determine the category to which the user information belongs;

[0029] If the category is identity information, then a corresponding mandatory association level is obtained based on the identity information, and the mandatory association level belongs to the first range;

[0030] If the category is transaction information, then the transaction information is analyzed to obtain the commercial transaction information therein;

[0031] Based on the transaction partners in the business transaction information, determine the upstream and downstream nodes corresponding to the user;

[0032] Using the operating status of the upstream and downstream nodes as the influence coefficient and the transaction viscosity between the user and the upstream and downstream nodes as the influence factor, the forced association level corresponding to the business transaction information is obtained, and the forced association level belongs to the second range or the third range.

[0033] In one implementation of this application, determining the dynamic impact level of the user information on the user based on the category to which the user information belongs specifically includes:

[0034] Determine the category to which the user information belongs;

[0035] If the category is identity information, then the dynamic impact level of the identity information is determined to be lower than a preset impact threshold;

[0036] If the category is transaction information, the dynamic impact level is determined according to the scope corresponding to the mandatory association level of the transaction information, wherein the mandatory association level is positively correlated with the dynamic impact level.

[0037] In one implementation of this application, when the stage corresponding to the specified user information changes, dynamic text recognition is performed on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage, specifically including:

[0038] Determine the dynamic recognition pattern generated based on different stages of user information, and perform dynamic text recognition on the user information at different stages based on the recognition frequency contained in the dynamic recognition pattern.

[0039] If the stage corresponding to the specified user information changes and meets the recognition frequency, then dynamic text recognition is performed on the specified user information under the changed stage to obtain the specified keywords that change in the changed stage.

[0040] Based on the attributes of the specified keywords, the preset degree of difference required for the specified keywords is determined through the second mapping relationship in the preset corpus;

[0041] The specified keywords are analyzed to determine the actual degree of difference they bring before and after the change.

[0042] If the actual difference reaches the preset difference, the specified keyword before the change is replaced with the specified keyword after the change, and the second credit feature is obtained based on the specified keyword after the change.

[0043] If the actual difference does not reach the preset difference, then both the specified keywords before and after the change are retained, and a second credit feature is obtained based on the specified keywords before and after the change.

[0044] In one implementation of this application, after modifying the lending strategy based on the compensated evaluation result, the method further includes:

[0045] Obtain the user's credit records, and filter out the user's short-term credit records and long-term credit records from the credit records;

[0046] Calculate the user's default threshold based on the short-term credit record and the long-term credit record;

[0047] Based on the user's lending strategy, the default threshold is adjusted so that collection efforts can be initiated against the user according to the adjusted default threshold.

[0048] This application provides a credit risk control device based on dynamic text information extraction, including:

[0049] At least one processor; and,

[0050] A memory communicatively connected to the at least one processor; wherein,

[0051] The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to:

[0052] Obtain user information used to reflect user credit, and determine the mandatory association level between the user information and the user, as well as the dynamic impact level of the user information on the user, based on the category to which the user information belongs;

[0053] Based on the preset dynamic recognition pattern in the dynamic text information, the user information in the current stage is subjected to dynamic text recognition to extract the first credit feature of the user in the current stage.

[0054] Based on the mandatory association level corresponding to the first credit feature, a risk assessment is performed on the user to generate a lending strategy for the user based on the assessment results of the risk assessment.

[0055] Among all the user information corresponding to the user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage.

[0056] The assessment result is compensated based on the second credit characteristic, and the lending strategy is modified based on the compensated assessment result.

[0057] The credit risk control method based on dynamic text information extraction proposed in this application can bring the following beneficial effects:

[0058] Determining the mandatory association level and dynamic impact level between user information and users considers both the degree of association between user information and users, as well as the impact of dynamic changes in user information on risk assessment results, enabling a more comprehensive risk assessment of users. By conducting risk assessments based on the mandatory association level, the impact of different categories of user information on the risk assessment results varies, leading to more accurate results. For designated user information whose dynamic impact equals or exceeds a preset impact threshold, dynamic text recognition is re-performed on the designated user information when a change in its corresponding stage is detected. This reduces the negative impact of user information changes on the risk assessment results, thereby ensuring the accuracy of lending strategies. Attached Figure Description

[0059] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0060] Figure 1 A flowchart illustrating a credit risk control method based on dynamic text information extraction, provided for an embodiment of this application;

[0061] Figure 2 This is a schematic diagram of the structure of a credit risk control device based on dynamic text information extraction, provided as an embodiment of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0064] like Figure 1 As shown in the embodiment of this application, a credit risk control method based on dynamic text information extraction is provided, including:

[0065] S101: Obtain user information used to reflect user credit, and determine the mandatory association level between user information and user, as well as the dynamic impact level of user information on user, based on the category to which the user information belongs.

[0066] When making a loan, the lending institution first needs to conduct an access investigation on the user's credit to obtain user information that reflects the user's credit. This user information includes identity information, transaction information, behavioral information, and other information that can reflect the user's basic personal credit.

[0067] In one embodiment, the mandatory association level between different categories of user information and users varies. It is understood that the mandatory association level is related to the degree of association between the user information and the user; the higher the mandatory association level, the more closely the user information and the user are associated. Determining the mandatory association level between user information and users can be achieved through the following steps:

[0068] Determine the category to which the user information belongs. If the category is identity information, such as ID number or name, then the user's identity can be directly identified based on the user information, and its corresponding mandatory association level is the highest. In this case, the mandatory association level is classified as the first range.

[0069] If the category is transaction information, further analysis is needed to extract the commercial transaction details. Commercial transaction information represents records of business transactions, such as a company's sales and production activities. Small-scale purchases in daily life are not considered, thus narrowing down the scope of user information identification and improving accuracy. After obtaining the commercial transaction information, the upstream and downstream nodes corresponding to the user are determined based on the transaction partners within the information.

[0070] Taking enterprise users as an example, upstream and downstream nodes refer to the companies upstream and downstream of the current user. For companies with a complete supply chain, upstream nodes can be manufacturers, midstream nodes are the current user (which can be intermediaries), and downstream nodes can be retailers. For the current user, the higher the transaction stickiness between them and upstream and downstream nodes, the more business transactions the user has with these nodes, and correspondingly, the stronger the mandatory association between the business transaction information and the user. For upstream and downstream nodes, their operating conditions can affect the current user's business transaction information. If upstream and downstream nodes are not performing well, the current user's supply chain will collapse, making it impossible to continue business transactions. Therefore, using the operating status of upstream and downstream nodes as an influence coefficient and the transaction stickiness between the user and these nodes as an influence factor, the mandatory association level of the business transaction information can be obtained by multiplying the coefficient and the influence factor. This mandatory association level represents the degree of correlation between the user's business transaction information and the user themselves.

[0071] The dynamic impact level indicates the stability between user information and user credit. For example, for user information such as name, changing a user's name will not affect their credit status. However, for consumption records, once a user's consumption records are changed, their corresponding credit status is very likely to change accordingly. Therefore, the dynamic impact level of name is low, while the dynamic impact level of consumption records is high.

[0072] In one embodiment, the category to which the user information belongs is determined. If the category to which the user information belongs is identity information, then the dynamic impact level corresponding to the user information is lower than a preset impact threshold.

[0073] If the category is transaction information, the mandatory association level corresponding to the transaction information can directly affect the user's credit. In this case, the dynamic impact level of the user can be determined based on the scope corresponding to the mandatory association level of the transaction information. It should be noted that the mandatory association level and the dynamic impact level are positively correlated. The higher the mandatory association level, the stronger the correlation between the current transaction information and the user. And once the transaction information changes, the greater the impact on the user.

[0074] Both mandatory association level and dynamic impact level can affect a user's credit. Therefore, after obtaining user information, it is necessary to determine the mandatory association level between the user information and the user, as well as the dynamic impact level of the user information, based on its category. This will allow for the determination of the user's lending strategy based on the obtained mandatory association level and dynamic impact level.

[0075] S102: Based on the preset dynamic recognition pattern in the dynamic text information, perform dynamic text recognition on the user information at the current stage to extract the first credit feature corresponding to the user at the current stage.

[0076] User information is categorized into different stages. It's important to note that a stage refers to the duration of time a user information exists in a particular state. In other words, if user information changes, its stage will also change accordingly. As the stage of user information changes, the corresponding dynamic recognition mode will also change, and each recognition mode has a different recognition frequency.

[0077] Before performing dynamic text recognition on user information, the specific dynamic recognition mode needs to be determined.

[0078] Specifically, several user information sets are collected, and keyword recognition is performed on each set to obtain recognition results. A training set is then generated based on these results. The training set includes multiple training samples, each containing at least one triplet. Each triplet carries three recognition results, where the similarity between the first and second recognition results on a first credit rating benchmark is higher than a first preset threshold, and the similarity between the first and third recognition results on a second credit rating benchmark is lower than a second preset threshold. The second preset threshold is lower than the first preset threshold. The first and second credit rating benchmarks can be the same in some cases, but typically they have different physical meanings; for example, the first credit rating benchmark could be the delinquency rate, and the second credit rating benchmark could be income status.

[0079] After generating the training set, the pre-generated model is trained using the training set until training is complete, resulting in a converged credit feature extraction model. This model can identify user information, thereby extracting credit features from textual information. It should be noted that while the first and second credit assessment benchmarks have actual physical meaning, the credit features ultimately identified during model training may be difficult to represent with physical meaning. For example, the keyword might be "Type A user," but the final identified credit feature might be simply "A." These numerical symbols, while lacking actual physical meaning, can still represent certain information from the keywords.

[0080] In this way, after selecting the dynamic recognition mode, the user information at the current stage is dynamically recognized by using the recognition frequency corresponding to the dynamic recognition mode, based on the trained credit feature extraction model.

[0081] In one embodiment, keyword identification of user information can be achieved through the following steps: First, the user information is segmented into words to obtain multiple keywords. After obtaining multiple keywords, the attributes of the keywords are determined. Attributes refer to the categories to which the keywords belong, such as identity, transaction, behavior, etc. Then, based on the attributes of the keywords and the mapping relationships set in the preset corpus, the positive and negative relationships corresponding to each keyword on each credit assessment benchmark are determined. Positive and negative relationships include positive and negative relationships. A positive relationship indicates that the keyword can affect the credit assessment benchmark, while a negative relationship indicates that the keyword cannot affect the credit assessment benchmark. For example, if the credit assessment benchmark is the delinquency rate, and the keyword category is behavior, the corresponding positive and negative relationships are determined according to the preset mapping relationships. If the keyword is "debt," then it can affect the delinquency rate, and it has a positive relationship on the credit assessment benchmark. If the keyword is "car," it is impossible to determine the degree of its impact on the delinquency rate, and it has a negative relationship on the credit assessment benchmark.

[0082] After obtaining the positive and negative relationships of each keyword, in order to ensure that the identified keywords have actual credit impact, it is necessary to calculate the frequency of occurrence of each keyword's positive and negative relationships on each credit assessment benchmark. After obtaining the frequency, keywords that do not meet the preset requirements are screened out based on the frequency. That is, if the frequency of occurrence of a certain keyword in the negative relationship is higher than the preset frequency, it means that the identification results of the keyword on multiple credit assessment benchmarks are not credit impact. In this case, the keyword has low reference value when identifying credit features and needs to be deleted from the multiple keywords obtained by the above word segmentation. Thus, the remaining keywords are the final keyword identification results.

[0083] In one embodiment, calculating the similarity between the first identification result and the second identification can be achieved through the following steps:

[0084] Determine the first dimension matrix corresponding to the first recognition result X and the second recognition result Y. , , ] and second dimension matrix [ , , Based on the first-dimensional matrix, calculate the second-order norm of the dimensions in the second-dimensional matrix that occupy the same position as those in the first-dimensional matrix. Then, use the second-order norm to obtain the similarity of the dimensions. The dimension is... , , ... The same position indicates that the rows and columns corresponding to the dimension are identical. Based on the calculated similarity between each dimension, a dimension similarity matrix is ​​constructed. Then, the number of zero values ​​in this matrix is ​​determined; zero values ​​indicate similarity between dimensions. The similarity between the first and second identification results on the credit rating benchmark is determined based on the ratio between the number of zero values ​​and the number of elements in the dimension similarity matrix. Furthermore, triples can be constructed based on the similarity between different identification results.

[0085] S103: Based on the mandatory association level corresponding to the first credit feature, conduct a risk assessment on the user, and generate a lending strategy for the user based on the assessment results.

[0086] After extracting the primary credit feature, a risk assessment is conducted on the user based on the mandatory association level of this feature. The stronger the mandatory association level, the greater its impact on the user's risk assessment result. Furthermore, based on the risk assessment results, a corresponding lending strategy is generated. If the assessment indicates that the user's risk is high, the lending strategy can involve reducing the loan amount and shortening the repayment period to reduce the user's repayment risk. Conversely, if the user's risk is low, the lending amount can be increased and the repayment period extended.

[0087] Specifically, the forced association level corresponds to three ranges: the first range, which is higher than the first preset association level; the second range, which is lower than the first preset association level but higher than the second preset association level; and the third range, which is lower than the second preset association level. The association relationship with the user decreases sequentially from the first range to the third range.

[0088] After determining the mandatory association level, it is necessary to determine the scope of the mandatory key level, then select the first credit feature that belongs to the first scope of the mandatory association level, and construct the user model corresponding to the user based on it. It should be noted that the user model can be a pre-built user profile model or a digital model constructed based on the correlation between the first credit features.

[0089] After constructing the user model, the first credit feature is used as a risk assessment parameter to generate the corresponding risk assessment model. It should be noted that the weight of the first credit feature varies depending on its range when generating the risk assessment model. The weight of the first credit feature corresponding to the second range is higher than that corresponding to the first and third ranges. For first credit features with a mandatory association level in the first range, these are mostly features closely related to the user, such as ID number and name, which, although highly associated with the user, are difficult to reflect the user's credit risk. For first credit features with a mandatory association level in the third range, these are mostly features with a low degree of association with the user, such as the school attended, and similarly, are difficult to reflect the user's credit risk. Therefore, when generating the risk assessment model, the first credit feature in the second range has the strongest reference value for risk assessment and its corresponding weight is the highest.

[0090] Understandably, the user model is generated using the primary credit features within the first range most strongly associated with the user. In other words, the user model primarily identifies the user, while the risk assessment model assesses the user's credit risk. Neither model alone can achieve optimal risk assessment results. Therefore, the user model and risk assessment model need to be merged to perform risk assessment based on the merged model. The merged user model balances user identification and risk assessment, effectively improving the accuracy of risk assessment.

[0091] S104: Among all the user information corresponding to the user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage.

[0092] After generating a lending strategy for a user, it is necessary to monitor the stage of the user information in real time. Once the stage of the user information changes, it indicates that the current user information has changed. At this time, dynamic text recognition needs to be performed on the user information again to determine whether the current lending strategy needs to be adjusted. This is to avoid the risk control results changing accordingly after the user information changes, which would lead to a decrease in the accuracy of the lending strategy.

[0093] Because some user information, such as name and address, has a relatively small dynamic impact on user credit, even if it changes, the associated transaction records and other information remain unchanged, making it difficult to affect the user's risk assessment results. Therefore, this part of user information does not require real-time monitoring of its current stage. However, for designated user information with a dynamic impact level higher than a preset threshold, such as marital relationships, if a user's marital relationship changes from non-existent to established and the spouse's credit status is poor, then the user's credit risk has substantially increased. In other words, dynamic changes in designated user information are highly likely to affect the user's risk assessment results.

[0094] Therefore, for designated user information whose dynamic impact level is higher than the preset threshold, when a change is detected in the corresponding stage, text recognition needs to be performed again to extract the changed second credit feature, so that the lending strategy can be adjusted later based on the second credit feature.

[0095] Specifically, a dynamic recognition pattern is generated based on different stages of user information. Based on the recognition frequency contained in the dynamic recognition pattern, dynamic text recognition is performed on the user information at different stages. If the stage corresponding to the specified user information changes and meets the recognition frequency, then dynamic text recognition is performed on the specified user information at the changed stage to obtain the specified keywords that have changed in the changed stage.

[0096] Furthermore, the specified keywords can be newly generated keywords or keywords that replace existing keywords. For a new round of dynamic text recognition of user information, it's necessary to ensure that the specified keywords have indeed undergone significant changes. In this case, further judgment is needed on the degree of difference before and after the change in the specified keywords. Based on the attributes of the specified keywords, the required preset degree of difference is determined through the second mapping relationship in the preset corpus. The preset degree of difference indicates how much difference a keyword under different attributes needs to produce to determine that it has changed from an existing keyword. For example, for identity information, only by ensuring a significant difference can it be considered that it has undergone dynamic change, while for transaction information, the corresponding preset degree of difference may be smaller than that for identity information.

[0097] Furthermore, the specified keywords are analyzed to determine the actual degree of difference they bring before and after the change. The actual degree of difference is the actual difference generated by the specified keywords, which can be obtained by calculating the cosine similarity between the specified keywords and keywords in a pre-set corpus.

[0098] If the actual difference reaches the preset difference, the specified keyword before the change will be replaced with the specified keyword after the change, and the second credit feature will be obtained based on the specified keyword after the change.

[0099] If the actual difference does not reach the preset difference, both the specified keywords before and after the change are retained, and the second credit feature is obtained based on the specified keywords before and after the change.

[0100] S105: Compensate the assessment results based on the second credit characteristics, and revise the lending strategy based on the compensated assessment results.

[0101] After extracting the second credit feature after the stage change, the second credit feature replaces the original first credit feature, and a new round of risk assessment is conducted to compensate for the assessment results obtained based on the first credit feature. Then, the lending strategy is modified based on the new assessment results. At this time, the lending strategy can be applied to the user information after the current stage change, avoiding the situation where the risk assessment results are inaccurate due to changes in user information.

[0102] After disbursing a loan to a user, it's necessary to obtain the user's credit history. Based on the loan duration, short-term and long-term credit records are identified. The corresponding repayment periods for these records are then determined, and the user's default threshold is calculated based on these periods. This default threshold is based on the initial lending strategy. After adjusting the strategy, the user's short-term and long-term credit records must be reassessed, and the default threshold adjusted accordingly. Collection efforts will then be initiated based on the adjusted default threshold.

[0103] Figure 2 This is a schematic diagram of a credit risk control device based on dynamic text information extraction, provided as an embodiment of this application. Figure 2 As shown, it includes:

[0104] At least one processor; and,

[0105] At least one processor-communication-connected memory; wherein,

[0106] The memory stores instructions that can be executed by at least one processor, and the instructions, when executed by at least one processor, enable at least one processor to:

[0107] Obtain user information used to reflect user credit, and determine the mandatory association level between user information and user, as well as the dynamic impact level of user information on user, based on the category to which the user information belongs;

[0108] Based on the preset dynamic recognition pattern in the dynamic text information, dynamic text recognition is performed on the user information in the current stage to extract the first credit feature corresponding to the user in the current stage.

[0109] Based on the mandatory association level corresponding to the first credit feature, a risk assessment is conducted on the user, and a lending strategy for the user is generated based on the assessment results.

[0110] Among all the user information corresponding to a user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage.

[0111] The assessment results are compensated based on the second credit characteristic, and the lending strategy is adjusted based on the compensated assessment results.

[0112] The application embodiment also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows:

[0113] Obtain user information used to reflect user credit, and determine the mandatory association level between the user information and the user, as well as the dynamic impact level of the user information on the user, based on the category to which the user information belongs;

[0114] Based on the preset dynamic recognition pattern in the dynamic text information, the user information in the current stage is subjected to dynamic text recognition to extract the first credit feature of the user in the current stage.

[0115] Based on the mandatory association level corresponding to the first credit feature, a risk assessment is performed on the user to generate a lending strategy for the user based on the assessment results of the risk assessment.

[0116] Among all the user information corresponding to the user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage.

[0117] The assessment result is compensated based on the second credit characteristic, and the lending strategy is modified based on the compensated assessment result.

[0118] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0119] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0120] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0125] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0126] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0127] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0128] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A credit risk control method based on dynamic text information extraction, characterized in that, The method includes: Obtain user information used to reflect user credit, and determine the mandatory association level between the user information and the user, as well as the dynamic impact level of the user information on the user, based on the category to which the user information belongs; Based on the preset dynamic recognition pattern in the dynamic text information, the user information in the current stage is subjected to dynamic text recognition to extract the first credit feature of the user in the current stage. Based on the mandatory association level corresponding to the first credit feature, a risk assessment is performed on the user to generate a lending strategy for the user based on the assessment results of the risk assessment. Among all the user information corresponding to the user, select the specified user information whose dynamic influence level is higher than the preset influence threshold, and when the stage corresponding to the specified user information changes, perform dynamic text recognition on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage. The assessment result is compensated based on the second credit characteristic, and the lending strategy is modified based on the compensated assessment result; Before performing dynamic text recognition on the user information at the current stage based on the preset dynamic recognition pattern in the dynamic text information, the method further includes: Collect several user information entries, and perform keyword recognition on each of the user information entries to obtain the recognition results; A training set is generated based on the recognition results, such that the training set includes multiple training samples, each training sample includes at least one triplet, each triplet carries three recognition results, and among the three recognition results, the similarity between the first recognition result and the second recognition result on the first credit assessment benchmark is higher than a first preset threshold, the similarity between the first recognition result and the third recognition result on the second credit assessment benchmark is lower than the second preset threshold, the second preset threshold is lower than the first preset threshold, the first credit assessment benchmark and the second credit assessment benchmark are preset and have corresponding physical meanings; The pre-generated model is trained based on the training set until the model converges, resulting in a trained credit feature extraction model. The user information is then identified based on the credit feature extraction model to obtain credit features. Keyword recognition is performed on the aforementioned user information to obtain recognition results, specifically including: For each piece of user information, the user information is segmented into words to obtain multiple keywords after segmentation; Based on the attributes of each keyword and the first mapping relationship set in the preset corpus, the positive and negative relationships corresponding to each keyword on each credit rating benchmark are determined, and the positive and negative relationships include positive relationships and negative relationships. Based on the frequency of occurrence of the positive and negative relationships of each keyword on each credit rating criterion, keywords that do not meet the preset requirements are filtered out, and the remaining keywords are used as the identification results of keyword identification. Based on the mandatory association level corresponding to the first credit feature, a risk assessment is performed on the user, specifically including: The range to which the forced association level belongs is determined. The range includes at least three ranges: a first range that is higher than the first preset association level, a second range that is lower than the first preset association level but higher than the second preset association level, and a third range that is lower than the second preset association level. Select the first credit feature that belongs to the first range of the forced association level, and construct the user model corresponding to the user based on it; Using the first credit feature as a risk assessment parameter, a risk assessment model is generated, wherein the weight of the first credit feature corresponding to the second range is higher than that of the first credit feature corresponding to the first range and the third range. The risk assessment model is fused with the user model to perform risk assessment on the user using the fused user model.

2. The credit risk control method based on dynamic text information extraction according to claim 1, characterized in that, After using the remaining keywords as the keyword recognition result, the method further includes: Determine the first-dimensional matrix and the second-dimensional matrix corresponding to the first recognition result and the second recognition result; Based on the first dimension matrix, the second norm of the dimensions in the second dimension matrix that are located at the same position as the first dimension matrix is ​​calculated, so as to obtain the similarity of the dimensions based on the second norm; the same position means that the rows and columns corresponding to the dimensions are the same; Construct a dimension similarity matrix based on the similarity of the dimensions; The number of zero values ​​in the dimensional similarity matrix is ​​determined, and the similarity between the first identification result and the second identification result on the credit rating benchmark is determined based on the number of zero values.

3. The credit risk control method based on dynamic text information extraction according to claim 1, characterized in that, Based on the category to which the user information belongs, the mandatory association level between the user information and the user is determined, specifically including: Determine the category to which the user information belongs; If the category is identity information, then a corresponding mandatory association level is obtained based on the identity information, and the mandatory association level belongs to the first range; If the category is transaction information, then the transaction information is analyzed to obtain the commercial transaction information therein; Based on the transaction partners in the business transaction information, determine the upstream and downstream nodes corresponding to the user; Using the operating status of the upstream and downstream nodes as the influence coefficient and the transaction viscosity between the user and the upstream and downstream nodes as the influence factor, the forced association level corresponding to the business transaction information is obtained, and the forced association level belongs to the second range or the third range.

4. The credit risk control method based on dynamic text information extraction according to claim 1, characterized in that, Based on the category to which the user information belongs, the dynamic impact level of the user information on the user is determined, specifically including: Determine the category to which the user information belongs; If the category is identity information, then the dynamic impact level of the identity information is determined to be lower than a preset impact threshold; If the category is transaction information, the dynamic impact level is determined according to the scope corresponding to the mandatory association level of the transaction information, wherein the mandatory association level is positively correlated with the dynamic impact level.

5. A credit risk control method based on dynamic text information extraction according to claim 1, characterized in that, When the stage corresponding to the specified user information changes, dynamic text recognition is performed on the specified user information under the changed stage according to the dynamic recognition mode to obtain the second credit feature corresponding to the changed stage, specifically including: Determine the dynamic recognition pattern generated based on different stages of user information, and perform dynamic text recognition on the user information at different stages based on the recognition frequency contained in the dynamic recognition pattern. If the stage corresponding to the specified user information changes and meets the recognition frequency, then dynamic text recognition is performed on the specified user information under the changed stage to obtain the specified keywords that change in the changed stage. Based on the attributes of the specified keywords, the preset degree of difference required for the specified keywords is determined through the second mapping relationship in the preset corpus; The specified keywords are analyzed to determine the actual degree of difference they bring before and after the change. If the actual difference reaches the preset difference, the specified keyword before the change is replaced with the specified keyword after the change, and the second credit feature is obtained based on the specified keyword after the change. If the actual difference does not reach the preset difference, then both the specified keywords before and after the change are retained, and a second credit feature is obtained based on the specified keywords before and after the change.

6. The credit risk control method based on dynamic text information extraction according to claim 1, characterized in that, After adjusting the lending strategy based on the compensated assessment results, the method further includes: Obtain the user's credit records, and filter out the user's short-term credit records and long-term credit records from the credit records; Calculate the user's default threshold based on the short-term credit record and the long-term credit record; Based on the user's lending strategy, the default threshold is adjusted so that collection efforts can be initiated against the user according to the adjusted default threshold.

7. A credit risk control device based on dynamic text information extraction, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a credit risk control method based on dynamic text information extraction as described in any one of claims 1-6.

Images