An automated post-loan management method and device based on deep learning

The deep learning-based method automates loan contract analysis to extract structured data from unstructured text and tables, enabling real-time monitoring and reducing manual oversight, thus enhancing post-loan management efficiency and risk control.

CN118587718BActive Publication Date: 2024-12-03CHINA ZHESHANG BANK +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411069775.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-12-03
Estimated Expiration
2044-08-06

AI Technical Summary

Technical Problem

Conventional loan contracts with phased repayment conditions are difficult to automate due to varied expression, embedded positions, and diverse logical relationships, making it challenging to extract and supervise loan accounts efficiently.

Method used

An automated post-loan management method using deep learning models for parsing and identifying text and table information, converting it into structured data, and monitoring repayment conditions through SQL queries, with models like BERT, BiLSTM+CRF+PointerNetwork, and Transformer for named entity recognition and relationship extraction.

Benefits of technology

Automates the extraction of unstructured data, reduces manual review costs, enhances post-loan management efficiency, and provides real-time monitoring with early warnings for repayment defaults, improving financial risk control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118587718B_ABST
    Figure CN118587718B_ABST
Patent Text Reader

Abstract

The present invention discloses an automated post-loan management method and device based on deep learning. Through deep learning technology, important information and quantitative relationships are extracted from loan decision documents in various formats, so as to realize automated real-time supervision of loan accounts. Machine vision and text recognition technology are used to realize the reading of file data in various formats such as pictures, docx documents, and pdf documents. The named entity recognition model is used to effectively extract the logic of table data and text data, and the generative question-answering model is used to convert it into an operation logic that can be understood by the computer, so as to realize real-time query of the database for default comparison, realize automated post-loan management, and batch monitor loan accounts. The present invention provides a batch management method for converting unstructured data in loan decision documents into structured data and extracting operation logic, completing end-to-end automated supervision, saving labor costs and time, and realizing real-time monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing, machine vision and other deep learning fields, and in particular to an automated post-loan management method and device based on deep learning. Background Art

[0002] Conventional loan contracts include one or more loan conditions, and most of them are phased repayment agreements, such as: 30% of the loan must be repaid after six months; 40 million must be repaid after the sales amount reaches 50 million, etc. This feature makes automated supervision of loan accounts face the following problems:

[0003] 1. The repayment conditions in the text data are expressed in different ways, and it is difficult to extract them through rules, such as 6 months, half a year, etc.;

[0004] 2. The repayment conditions and their corresponding repayment amounts are embedded in long texts, their corresponding positions are not fixed, and their logical relationships are diverse;

[0005] 3. In tabular data, the table style is varied and cannot be directly determined by rules. Summary of the invention

[0006] The purpose of the present invention is to provide a set of information extraction methods based on deep learning models to solve the problem that unstructured text data such as loan decision documents cannot be directly written into the database, and loan accounts cannot be automatically identified and supervised. There are problems such as untimely and time-consuming manual supervision. This method realizes the integration of multiple deep learning models to automatically read multiple text data files and extract text and table data as structured data that can be recognized by the database.

[0007] The objective of the present invention is achieved through the following technical solution: an automated post-loan management method based on deep learning, the method comprising the following steps:

[0008] (1) Parse and identify the text and table information in the loan contract, extract the entities of repayment conditions and repayment amount through named entity recognition, and obtain structured data of the quantitative relationship mapping between repayment conditions and repayment amount through relationship recognition;

[0009] (2) Convert the extracted structured data into SQL query statements to monitor whether the loan account meets the repayment conditions. When the loan account does not meet the repayment conditions, obtain the current violated repayment terms and issue an early warning.

[0010] Furthermore, in step (1), the text and table information in the loan contract are parsed and identified through a layout analysis algorithm and an optical character recognition (OCR) deep learning model; the obtained text information is converted into sentence vectors through the semantic representation model BERT and input into a bidirectional long short-term memory network BiLSTM+conditional random field CRF+pointer network PointerNetwork model for named entity prediction; for the obtained table information, named entity recognition is used to confirm the category of the elements in the cell, thereby inferring the table style and the quantitative relationship mapping between the elements.

[0011] Furthermore, the OCR deep learning model is implemented based on the DBnet and SVTR models, in which a path aggregation network PAN module is added to the DBnet, that is, a bottom-up feature pyramid is added after the feature pyramid network FPN module. The FPN layer conveys semantic features from top to bottom, and the PAN module feature pyramid conveys positioning features from bottom to top, thereby improving text detection capabilities.

[0012] Furthermore, the BiLSTM+CRF+PointerNetwork model is specifically implemented as follows: discrete sentence vectors are mapped through the BERT model to obtain low-dimensional dense vectors, and then the dense sentence vectors are input into the BiLSTM bidirectional recurrent neural network for training. The output results are used to extract the relationship between entities through CRF+PointerNetwork multi-task learning, thereby predicting the entity information in the text and establishing the decision conditional logic in the text.

[0013] Furthermore, for tabular data, named entity recognition is first performed on the table content, and then the entity labels are input into the neural network classification model to obtain the table layout style. According to the output of the classification model, it is determined whether there is a decision condition mapping between the two cells of the table, and the logical semantics of the mutual links are obtained to extract information from the tabular data.

[0014] Furthermore, relationship identification adopts the Transformer generative question-answering model, which decomposes the original encoder and decoder modules in the Transformer into an entity selection module and a quantitative relationship output module. The entity selection module divides the input entity into two parts: repayment conditions and repayment amount, and uses them as the input of the quantitative relationship output module to generate a series of candidate combinations, select the optimal combination through training parameters, and finally output the quantitative relationship.

[0015] Furthermore, for the models used in named entity recognition and relationship recognition, the saved input data is regularly annotated, and sampled and combined from the original training samples to enhance the diversity of the input data to improve the training data set and enhance the model recognition capability.

[0016] In a second aspect, the present invention also provides an automated post-loan management device based on deep learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, the automated post-loan management method based on deep learning is implemented.

[0017] In a third aspect, the present invention further provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the automated post-loan management method based on deep learning.

[0018] In a fourth aspect, the present invention further provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the automated post-loan management method based on deep learning.

[0019] The beneficial effects of the present invention are as follows:

[0020] 1. Automatically extract unstructured text data and complex table data to reduce manual review costs and errors caused by manual calculations;

[0021] 2. Make paper repayment agreements electronic, provide data basis for real-time monitoring of whether repayment accounts are in default, improve post-loan management efficiency and reduce risks;

[0022] 3. This method realizes real-time monitoring of repayment accounts, that is, monitoring the repayment conditions triggered by the current account and the actual repayment behavior at a specified frequency, and immediately returns an early warning when a default behavior is found;

[0023] 5. The method supports deep learning model maintenance, regularly combines newly generated online data with original training data to retrain the model and gradually improve the model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the method structure and technical points in the invention will be illustrated below. Obviously, the drawings in the following description are some embodiments of the present invention.

[0025] Figure 1 This is a schematic diagram of the implementation principle of an automated post-loan management method based on deep learning in one embodiment of the present invention.

[0026] Figure 2 It is a schematic diagram of the principle of the PAN structure added to the text detection model of the present invention.

[0027] Figure 3 The network framework diagram of the entity naming model training process of the present invention.

[0028] Figure 4 A structural diagram of an automated post-loan management device based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Aiming at the problem that paper repayment contracts are digitized and the mapping between phased repayment conditions and repayment amounts is not convenient for direct storage and query in a database, the method of the present invention establishes a series of deep learning models to realize a full-process automated post-loan management method from document reading, key information extraction, quantitative relationship mapping extraction, real-time query of repayment account status, and whether the currently triggered repayment conditions are met.

[0031] like Figure 1 As shown, the pre-verified pre-loan decision opinion file has a file format including scanned images, pdf documents, docx documents, etc., and the content usually includes both plain text and table data. The text and table in the image or PDF file are read and parsed through the layout analysis algorithm based on yolov2+FPN and the OCR deep learning algorithm based on DBnet and SVTR model, and stored as a string. In the text detection part, the present invention optimizes DBnet, such as Figure 2 As shown in the figure, a PAN module is added, that is, a bottom-up feature pyramid is added after the FPN module. The FPN layer conveys strong semantic features from top to bottom, and the PAN module feature pyramid conveys strong positioning features from bottom to top, thereby improving the text detection capability.

[0032] For the obtained text information, the sentence vector is converted through Bert and input into the bidirectional long short-term memory network BiLSTM+conditional random field CRF+pointer network PointerNetwork model for named entity prediction to obtain entities with specific meanings in the text, so as to extract effective information, such as the object of pre-loan decision, the type of loan conditions, the content of loan conditions, including the borrowing company, repayment time, repayment amount, repayment conditions, quantity relationship, etc., to complete the information purification of the text. Figure 3As shown in the figure, when training the BiLSTM+CRF+PointerNetwork model, two branches are added after the Bi-LSTM Layer, namely the CRF layer and the PointerNetwork layer. Through multi-task learning, on the one hand, the overall accuracy of entity recognition is improved, and on the other hand, the problem of entity span breakage is solved by using the advantage of PointerNetwork's strong ability to recognize long text entities. When predicting, only the CRF layer is used for prediction. The specific implementation of BiLSTM+CRF+PointerNetwork is as follows: discrete sentence vectors are mapped through the BERT model to obtain low-dimensional dense vectors, and then the dense sentence vectors are input into the BiLSTM bidirectional recurrent neural network for training. The output results are used to extract the relationship between entities through the CRF+PointerNetwork multi-task learning method, so as to predict the entity information in the text and establish the decision condition logic in the text. For example, when the sales amount reaches a certain amount, the repayment amount must also reach a certain amount. Through the prediction of the model, the sales amount and the repayment amount will get corresponding entity labels, so as to obtain the required extraction content by screening the entity labels, and input the entity labels into the deep learning model to extract the quantitative relationship between entities.

[0033] For the tabular information with rich typeset formats, named entity recognition is used to confirm the categories of elements in the cells, and then the entity labels are input into the neural network classification model to obtain the table layout style (for example: horizontal, vertical, etc.). According to the results output by the classification model, it is determined whether there is a decision condition mapping between the two cells in the table, and the logical semantics of the mutual links are obtained. Information is extracted from the tabular data and extracted into structured data that can be used in the database.

[0034] For the deep learning model for extracting quantitative relationships, the present invention adopts the Transformer generative model. After converting the special entities identified by the previous model into sentence vectors, the model is input into the Transformer model for training to obtain the subordinate relationship between entities. The output result is one or more pairs of elements of repayment conditions and repayment amounts, which represent the amount of the amount to be repaid when the repayment situation reaches the repayment conditions. In this way, the quantitative relationship is extracted from the unstructured data, and the implementation method automatically calculates the current amount to be repaid, and compares it with the actual repayment amount in real time by querying the database, so as to achieve the effect of real-time monitoring of whether the repayment account is in default, and realize efficient automatic detection.

[0035] The BERT-BiLSTM-CRF-PointerNetwork model involved in the present invention does not use traditional entities such as names of people, places, and institution names as output targets during training, but focuses on entities such as time, quantity, and amount in the text, and performs sample annotation and data enhancement on them. The reason is that in order to extract the quantitative logical relationship in the text, the loan conditions and the corresponding repayment amount need to be screened out through the above model, and changing the entity type makes the output information more accurate and effective. In order to improve the accuracy of entity recognition, the present invention also uses a neural network classification model to combine the context to predict the category of the entity to improve the credibility of the result.

[0036] For the use of Transformer generative question-answering model, the original Encoder and Decoder modules are decomposed into entity selection module and quantity relationship output module. The entity selection module can divide the input entity into two parts: repayment conditions and repayment amount, and use them as the input of the quantity relationship output module. Then the module generates a series of candidate combinations, converts the target into multiple binary classification problems, and selects the optimal combination through training parameters, so as to finally output the quantity relationship.

[0037] The present invention converts the extracted structured data into SQL query statements to monitor whether the loan account meets the repayment conditions, and sets up timed queries to achieve real-time supervision of the lender's repayment situation. When the loan account does not meet the repayment conditions, the method will return the currently violated repayment agreement, send an alarm to the management personnel, and record the default data as a basis for future pre-loan decisions and as a feature for application in financial risk control scenarios.

[0038] The model deployment design principles of the method are:

[0039] The layout analysis algorithm and OCR deep learning model parse text and tables and identify text and table information. The BERT-BiLSTM-CRF-PointerNetwork model performs named entity recognition. The Transformer model extracts quantitative relationships. The named entity model is used to annotate individual words, effectively screening out information containing calculation logic. After the deep learning model is used to extract quantitative relationships, the accuracy is significantly improved. The accuracy of a single model must be maintained above 95%. The accuracy of the three models can be improved by enhancing the quantity and quality of the data set. The model parameters can be regularly updated with incremental data, and the saved input data can be regularly annotated and sampled and combined from the original training samples to enhance the diversity of the input text to improve the training data set and enhance the model recognition ability.

[0040] The present invention makes up for the mechanical nature of string matching rules to a certain extent, flexibly enhances the accuracy of information extraction by identifying entity types, and iteratively optimizes the model through batch data enhancement, reduces the method's dependence on rules, and improves its robustness. The present invention combines a series of natural language processing technologies and deep learning models to complete the automatic identification of pre-loan decision documents and the automatic monitoring of loan accounts. In the present invention, named entity recognition technology realizes the automatic identification of different types of entities, such as different types of texts such as institutions, places, and time. Semantic labels are provided for unstructured text data, wherein semantic labels can be used to analyze the relationship between entities and provide a data basis for extracting the quantitative relationship between repayment conditions and repayment amounts. In the present invention, the open source model Bert-BiLSTM-CRF is used, and the PointerNetwork branch is added: the text embedding word vector is obtained through the Bert network, and is input as a feature into the BiLSTM+CRF+PointerNetwork model to obtain the entity label. To address the problem of quantitative relationship extraction, the open source model Transformer has been widely used in generative question answering, translation and other tasks. In this invention, the Transformer model realizes the extraction and output of quantitative relationships of named entities, and converts them into structured data and writes them into a database, realizing end-to-end application.

[0041] Corresponding to the aforementioned embodiment of an automated post-loan management method based on deep learning, the present invention also provides an embodiment of an automated post-loan management device based on deep learning.

[0042] See also Figure 4 An embodiment of the present invention provides an automated post-loan management device based on deep learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement an automated post-loan management method based on deep learning in the above embodiment.

[0043] An embodiment of an automated post-loan management device based on deep learning provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Figure 4 As shown, it is a hardware structure diagram of any device with data processing capability where an automatic post-loan management device based on deep learning provided by the present invention is located, except Figure 4In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0044] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0045] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0046] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, an automated post-loan management method based on deep learning in the above embodiment is implemented.

[0047] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0048] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the automated post-loan management method based on deep learning.

[0049] The above is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above disclosed methods and technical contents without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. An automated post-loan management method based on deep learning, characterized in that: The method comprises the following steps: (1) Parse and identify the text and table information in the loan contract, extract the entities of repayment conditions and repayment amount through named entity recognition, and obtain structured data of the quantitative relationship mapping between repayment conditions and repayment amount through relationship recognition; Parsing and identifying text and table information in loan contracts is achieved through layout analysis algorithms and optical character recognition (OCR) deep learning models; For the obtained text information, the sentence vector is converted through the semantic representation model BERT, and then input into the bidirectional long short-term memory network BiLSTM+conditional random field CRF+pointer network PointerNetwork model for named entity prediction; the BiLSTM+CRF+PointerNetwork model is specifically implemented as follows: the discrete sentence vector is mapped through the BERT model to obtain a low-dimensional dense vector, and then the dense sentence vector is input into the BiLSTM bidirectional recurrent neural network for training. The output result extracts the relationship between entities through the CRF+PointerNetwork multi-task learning method, thereby predicting the entity information in the text and establishing the decision condition logic in the text; For the obtained table information, named entity recognition is used to confirm the category of the elements in the cell, so as to infer the table style and the quantitative relationship mapping between the elements; for the table data, named entity recognition is first performed on the table content, and then the entity label is input into the neural network classification model to obtain the typesetting style of the table. According to the output of the classification model, it is determined whether there is a decision condition mapping between the two cells of the table, and the logical semantics of the mutual links are obtained to extract information from the table data; Relationship identification uses the Transformer generative question-answering model, which decomposes the original encoder and decoder modules in the Transformer into an entity selection module and a quantity relationship output module. The entity selection module divides the input entity into two parts: repayment conditions and repayment amount. The module uses these two parts as the input of the quantity relationship output module to generate a series of candidate combinations, selects the optimal combination through training parameters, and finally outputs the quantity relationship. (2) Convert the extracted structured data into SQL query statements to monitor whether the loan account meets the repayment conditions. When the loan account does not meet the repayment conditions, obtain the current violated repayment terms and issue an early warning.

2. The method for automated post-loan management based on deep learning according to claim 1, characterized in that: The OCR deep learning model is implemented based on the text detection network DBnet and the text recognition network SVTR model. A path aggregation network PAN module is added to DBnet, that is, a bottom-up feature pyramid is added after the feature pyramid network FPN module. The FPN layer conveys semantic features from top to bottom, and the PAN module feature pyramid conveys positioning features from bottom to top, thereby improving text detection capabilities.

3. The method for automated post-loan management based on deep learning according to claim 1, characterized in that: For the models used for named entity recognition and relationship recognition, the saved input data is regularly labeled, sampled and combined from the original training samples to enhance the diversity of the input data to improve the training data set and enhance the model recognition capability.

4. An automated post-loan management device based on deep learning, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, an automated post-loan management method based on deep learning as described in any one of claims 1-2 is implemented.

5. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by the processor, an automated post-loan management method based on deep learning as described in any one of claims 1-2 is implemented.

6. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by the processor, an automated post-loan management method based on deep learning as described in any one of claims 1-2 is implemented.