Credit review settlement generation method and system

Abstract

The invention discloses a credit check settlement generation method and system, and the method comprises the steps: constructing sample data, taking a credit check data text as a feature of a sample, and taking a credit check settlement as a label; carrying out first training on the model based on the training samples of all the sample data, and after the training is completed, carrying out second training on the model based on a training set constructed by a specified type of samples; and after the second training is completed, collecting credit auditing data of the credit auditing user, converting the credit auditing data into a credit auditing data text, inputting the credit auditing data text into the trained model, and outputting a credit auditing statement corresponding to the predicted credit auditing data by the model. According to the method, the model is trained for the first time through the credit check knots of different styles, the adaptability of the model is improved, then the model is trained again through the excellent credit check knots (the style of the specified credit check knots is specified), fine adjustment of model parameters is achieved, and the credit check knots output by the trained model are close to the excellent credit check knots; according to the method, the automatic generation of the credit check notes is realized, and the quality of the generated credit check notes is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of credit review technology, and more specifically, the present invention relates to a credit review summary generation method and system. Background Art

[0002] In the credit approval process, completing the credit review summary is a key step. The credit review summary needs to provide a comprehensive and accurate summary of the customer's basic situation and loan qualifications.

[0003] The traditional credit review summary relies on the credit reviewer's personal experience and subjective judgment. During the electronic review process, the credit reviewer needs to ask questions and verify the customer's application materials while recording the customer's answers and his own analysis. After the electronic review, the credit reviewer completes the credit review summary. The entire process relies on the credit reviewer's personal experience and subjective judgment, which not only cannot achieve a unified style, but also easily leads to information omissions, and seriously affects the timeliness of approval. In addition, with the rapid development of the credit market and the diversification of customer groups, the traditional credit review summary method is increasingly unable to meet the requirements of approval efficiency.

[0004] The generation of credit review summary is an application scenario of text summarization technology. Traditional text summarization technology adopts the template filling method: the template to be output is pre-configured, the corresponding field information is extracted from the customer's information according to the template, and the information is directly filled into the slot of the set template. The details are as follows: Figure 3 As shown in the figure. Due to the huge differences in the clients’ historical work, qualifications and other information, the credit reviewers will ask different questions during the electronic review, and the clients will give different answers. These questions and answers cannot be predicted in advance, which will result in some important questions and answers not having slots in the template and unable to be filled in the template, resulting in missing information. Summary of the invention

[0005] The present invention provides a credit review summary generation method, aiming to improve at least one of the above problems.

[0006] The present invention is implemented in this way: a method for generating a credit review summary, the method is specifically as follows:

[0007] (1) Construct sample data, where the sample uses the credit review data text as features and the credit review summary as a label;

[0008] (2) The model is trained for the first time based on the training samples of all sample data. After the training is completed, the model is trained for the second time based on the training set constructed by the specified type of samples;

[0009] (3) After the second training is completed, the credit review data of the credit review users is collected, the credit review data text is constructed, and the trained model is input. The model outputs the credit review summary corresponding to the predicted credit review data.

[0010] Furthermore, the credit review data text is composed of a credit summary generated based on a personal credit report, an application information summary generated based on an application form, a driving information summary generated based on extracted driver's license information, and the voice credit review dialogue is converted into a credit review dialogue text.

[0011] Furthermore, the model adopts the Qwen2-72B model.

[0012] Furthermore, the loss function language_loss(θ 72B ) are as follows:

[0013] language_loss(θ 72B )=∑ i logP 72B (s i |c i ,...,c i+k-1 θ 72B );

[0014] Among them, P 72B (s i |c i ,...,c i+k-1 θ 72B ) means that the string c in the sample feature i ,...,c i+k-1 The input model parameters are θ 72B Qwen2-72B model, Qwen2-72B model output character s i The predicted probability of character s i is the i-th character in the label.

[0015] Furthermore, the training process is as follows:

[0016] (21) construct a sliding window with a window size of k, and the sliding window traverses from the first character in the feature text;

[0017] (22) The k characters c in the sliding window i ,...,c i+k-1 and its corresponding label character s i Input the model and adjust the model parameters based on the model loss;

[0018] (23) The sliding window moves backward one character and executes step (22) until the entire feature text is traversed.

[0019] Furthermore, the credit review summary prediction method corresponding to the credit review data is as follows:

[0020] Get the credit review data text of the current user to be reviewed, and start the sliding window from the first character of the credit review data text, and slide the k characters c in the sliding window i ,...,c i+k-1 The trained model is input, and the model outputs the predicted characters and their corresponding prediction probabilities. After traversing the entire credit review data text, the character string with the highest prediction probability is output as the credit review summary.

[0021] The present invention is implemented as follows: a credit review summary generation system, the system comprising:

[0022] The information collection unit, input unit, trained model and output unit are connected in sequence;

[0023] Among them, the information collection unit is used to collect credit review data and convert the credit review data into credit review data text; the input unit traverses the credit review data text in sequence based on the sliding window, inputs the k character strings in the current sliding window into the trained model, and the model outputs the predicted characters and their predicted probabilities. After completing the traversal of the entire credit review data text, the output unit outputs the character string with the highest prediction probability as the credit review summary.

[0024] Furthermore, the credit review data text is composed of a credit summary generated based on a personal credit report, an application information summary generated based on an application form, a driving information summary generated based on extracted driver's license information, and the voice credit review dialogue is converted into a credit review dialogue text.

[0025] Furthermore, the model adopts the Qwen2-72B model.

[0026] The present invention uses credit review summaries of different styles as samples to perform the first training on the Qwen2-72B model to improve the adaptability of the model. Then, the model trained for the first time is trained again with excellent credit review summaries (specified credit review summary styles) to achieve fine-tuning of model parameters. The credit review summary output by the trained model is close to the excellent credit review summary, which not only realizes the automatic generation of the credit review summary, but also improves the quality of the generated credit review summary. In addition, the generation process of the credit review summary fully learns the credit review user's credit summary, application information summary, driving information summary, and credit review dialogue text, and the generated credit review summary information is relatively comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of a method for generating a credit review summary provided by an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of the structure of a credit review summary generation system provided by an embodiment of the present invention;

[0029] Figure 3An example of text summary generation based on a template filling method provided in an embodiment of the present invention. Implementation

[0031] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0032] The present invention will use the Qwen2-72B model as the basic model and use the preprocessed sample data for training. The Qwen2 model is based on the Transformer architecture, which processes text-type sequence data through a self-attention mechanism, can be calculated in parallel, effectively improves training efficiency, and can capture long-distance dependencies. Its main training corpus comes from the Chinese Internet, which is more suitable for Chinese task scenarios, and helps the model learn rich language patterns and knowledge. It has shown excellent capabilities in multiple fields such as dialogue understanding, text generation, code understanding and generation.

[0033] Figure 1 A flowchart of a method for generating a credit review summary provided by an embodiment of the present invention, the method comprising the following steps:

[0034] (1) Construct sample data;

[0035] In an embodiment of the present invention, each credit reviewer gives a credit review summary based on the credit review data of the credit review user. The credit review data includes: personal credit report, application form, driver's license and voice credit review dialogue. Sample data is constructed based on each credit review substructure. A credit summary is first generated based on the personal credit report, an application information summary is generated based on the application form, and a driving information summary is generated based on the extracted driver's license information. The voice credit review dialogue is converted into a credit review dialogue text. The credit summary, application information summary, driving information summary and credit review dialogue text constitute the credit review data text. The credit review data text is used as the feature of the sample, and the credit review summary is used as the label of the sample. The credit review data text and the credit review summary are composed of a character string, a separator, and the start and end identifiers of the text.

[0036] The following is a detailed description of the processing of credit data, application information, driver's license information, and voice credit review dialogue:

[0037] (a) Generate a credit summary based on the personal credit report of the credit review user provided by the bank. The credit summary includes: name, previous work, previous address, credit inquiry record in the past two years (number of times, institution, time), various information of loans held (credit card, non-revolving loan, revolving loan, etc.) (credit institution, time, amount, repayment status of each period, overdue status, etc.).

[0038] The data in personal credit reports are in the format of key-value pairs and can be directly input into the large model for processing. However, the data in personal credit reports are complex, with many fields, and the meaning of the keys is not explained, which is not conducive to the understanding of the large model, so a pre-processing solution is required first. Although there are many credit data fields involved in personal credit reports, the format is fixed, and the range of fields involved is a finite set. The template filling method can be used directly to generate the credit summary of the required credit data. This processing can not only add meaning to each field that is easy for the model to understand, but also reduce meaningless input and reduce the difficulty of model training and learning.

[0039] (b) An application form provided by the credit review user, including name, ID card, mobile phone number, mobile phone number of relatives, friends or spouse, work unit, monthly income, down payment ratio, interest rate, loan amount, monthly payment, vehicle type, vehicle price, loan reason, etc., using a template filling method to generate an application information summary;

[0040] (c) The credit review user uploads his / her own driver's license image. After confirming that the driver's license holder is the applicant in the application information, the driver's license information is captured from the driver's license image. The driving information includes: name, gender, address, date of birth, first license date, permitted vehicle type, effective start date and effective period, and a driving information summary is generated using a template filling method;

[0041] (d) The credit reviewer has a conversation with the credit review user. The credit reviewer asks confirmatory questions about the information used to form the credit review summary, and the credit review user answers. The above conversation is recorded, and the recorded conversation voice is converted into a conversation text, which is the credit review conversation text. The core processing is the speech-to-text (ASR) capability. Due to the complexity of the noise and accent in the voice recording scene, the converted text needs to be corrected, the text is normalized, and similar sentences are merged. These processing processes all use existing means, which will not be described in detail in the present invention.

[0042] (2) The model is trained for the first time based on the training samples constructed from all samples. After the training is completed, the model is trained for the second time based on the samples of the specified type. After the second training is completed, the model training is completed, and the credit review summary is generated based on the trained model.

[0043] There are differences in the style of the credit review summary formed by each credit reviewer, such as different language expressions, different levels of complexity of the summary, etc. First, all styles of credit review summaries are put into the first sample data set, and the standard style of credit review summaries are selected and put into the second sample data set.

[0044] The samples in the first sample data and the second sample data are divided into training samples and test samples, the model is trained based on the training samples, and the trained model is tested based on the test samples. During the test process, if the model outputs a credit review summary close to the set standard, the current model training is completed.

[0045] The first and second training processes of the model are basically the same, except that the sample types used for the two trainings are different. The first training uses all samples (of all styles) to train the model, aiming to improve the adaptability of the model. The second training uses samples formed by the credit review summary based on the target style for training, aiming to focus on style transfer and improve generalization ability.

[0046] During the training process, the model will output the predicted credit review summary based on the input sample features and labels, compare it with the credit review summary in the label, calculate the model loss, and adjust the model parameter θ based on the model loss. 72B , the loss function language_loss(θ 72B ) are as follows:

[0047] language_loss(θ 72B )=∑ i logP 72B (s i |c i ,...,c i+k-1 θ 72B );

[0048] Among them, P 72B (s i |c i ,...,c i+k-1 θ 72B ) means that the string c in the sample feature i ,...,c i+k-1 The input model parameters are θ 72B Qwen2-72B model, Qwen2-72B model output character s i The predicted probability of character s i is the i-th character in the label.

[0049] In the embodiment of the present invention, the training process is as follows: (21) constructing a sliding window with a window size of k, and the sliding window traverses from the first character in the feature text; (22) i ,...,c i+k-1 and its corresponding label character s iInput the model and adjust the model parameters; (23) the sliding window moves backward one character and executes step (22) until the entire feature text is traversed, completing the model's learning of the current sample.

[0050] (3) Collect the personal credit report, application form, driver's license and voice credit review dialogue of the credit review user, convert it into credit review data text, input it into the trained model, and the model outputs the credit review summary of the corresponding credit review user.

[0051] In the embodiment of the present invention, the personal credit report, application form, driver's license and voice credit review dialogue of the current user to be reviewed are collected to construct the credit review data text of the credit review user. The sliding window starts from the first character of the credit review data text and slides k characters c in the sliding window. i ,...,c i+k-1 And the trained model, the model outputs the predicted characters and their corresponding predicted probabilities. After traversing the entire credit review data text, the character string with the highest prediction probability is output as the credit review summary. The credit review summary gives the conclusion of the loan qualification review, whether it meets the requirements or not. If it does not meet the requirements, the reasons for non-compliance are given, such as income does not match the job type, basic information is not true, loan reasons are not true, etc.

[0052] Figure 2 The structural diagram of the credit review summary generation system provided by the embodiment of the present invention only shows the parts related to the embodiment of the present invention for the convenience of description. The system includes:

[0053] The information acquisition unit, input unit, Qwen2-72B model and output unit are connected in sequence;

[0054] The information collection unit is used to collect credit review data, including personal credit reports, application forms, driver's licenses, and voice credit review dialogues, and convert the credit review data into credit review data text;

[0055] The input unit traverses the credit review data text in sequence based on a sliding window, and inputs the k character strings in the current sliding window into the trained Qwen2-72B model. The Qwen2-72B model outputs the predicted characters and their predicted probabilities. After completing the traversal of the entire credit review data text, the output unit outputs the character string with the highest prediction probability as the credit review summary.

[0056] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for generating a credit review summary, characterized in that: The method is specifically as follows: (1) Construct sample data, where the sample uses the credit review data text as features and the credit review summary as a label; (2) The model is trained for the first time based on the training samples of all sample data. After the training is completed, the model is trained for the second time based on the training set constructed by the specified type of samples; (3) After the second training is completed, the credit review data of the credit review users is collected, converted into credit review data text, and input into the trained model. The model outputs the credit review summary corresponding to the predicted credit review data.

2. The credit review summary generation method as claimed in claim 1, characterized in that: The credit review data text consists of a credit summary generated based on a personal credit report, an application information summary generated based on an application form, a driving information summary generated based on extracted driver's license information, and the conversion of voice credit review dialogues into credit review dialogue text.

3. The credit review summary generation method according to claim 1, characterized in that: The model adopts Qwen2-72B model.

4. The credit review summary generation method as claimed in claim 3, characterized in that: The loss function of the Qwen2-72B model is language_loss(θ 72B ) are as follows: language_loss(θ 72B )=∑ i logP 72B (s i |c i ,...,c i+k-1 ;θ 72B ); Among them, P 72B (s i |c i , ..., c i+k-1 θ 72B ) means that the string c in the sample feature i , ..., c i+k-1 The input model parameters are θ 72B Qwen2-72B model, Qwen2-72B model output character s i The predicted probability of character s i is the i-th character in the label.

5. The credit review summary generation method as claimed in claim 4, characterized in that: The training process is as follows: (21) construct a sliding window with a window size of k, and the sliding window traverses from the first character in the feature text; (22) The k characters c in the sliding window i ,,c i+k-1 and its corresponding label character s i Input the model and adjust the model parameters based on the model loss; (23) The sliding window moves backward one character and executes step (22) until the entire feature text is traversed.

6. The credit review summary generation method according to claim 1, characterized in that: The specific method for predicting the credit review summary corresponding to the credit review data is as follows: Get the credit review data text of the current user to be reviewed, and start the sliding window from the first character of the credit review data text, and slide the k characters c in the sliding window i ,,c i+k-1 The trained model is input, and the model outputs the predicted characters and their corresponding prediction probabilities. After traversing the entire credit review data text, the character string with the highest prediction probability is output as the credit review summary.

7. A credit review summary generation system, characterized in that: The system comprises: The information collection unit, input unit, trained model and output unit are connected in sequence; Among them, the information collection unit is used to collect credit review data and convert the credit review data into credit review data text; the input unit traverses the credit review data text in sequence based on the sliding window, inputs the k character strings in the current sliding window into the trained model, and the model outputs the predicted characters and their predicted probabilities. After completing the traversal of the entire credit review data text, the output unit outputs the character string with the highest prediction probability as the credit review summary.

8. The credit review summary generation system as claimed in claim 7, characterized in that: The credit review data text consists of a credit summary generated based on a personal credit report, an application information summary generated based on an application form, a driving information summary generated based on extracted driver's license information, and the conversion of voice credit review dialogues into credit review dialogue text.

9. The credit review summary generation system as claimed in claim 7, characterized in that: The model adopts Qwen2-72B model.

Images