A method and system for archiving and storing reminder data based on machine learning

Abstract

The present invention proposes a method and system for archiving, storing and managing reminder data based on machine learning. The method includes obtaining reminder audio, converting the reminder audio into text through ASR, and building a reminder data corpus with the converted text; building a reminder data text model through the reminder data corpus, inputting new reminder data into the reminder data text model to form a customer's first repayment tendency; obtaining the reminder audio data to form a customer's second repayment tendency, weightedly adding the customer's first repayment tendency and the second repayment tendency to form a third repayment tendency, and storing them in different storage media according to the customer's repayment tendency and the time when the reminder data was formed. Through this method and the corresponding system, machine learning can be used to establish an automated management process, automatically allocate storage locations, and build a repayment tendency model through machine collection audio, which can effectively predict the probability that the customer will repay within the specified period.

Description

Technical Field

[0001] The present invention proposes a method and system for archiving, storing and managing reminder data based on machine learning, and relates to the field of machine learning. Background Art

[0002] In recent years, my country's internet finance industry has experienced rapid growth. Financial products such as "Jiebei," "Huabei," and "Bai Tiao" have become deeply ingrained in people's lives. Together with traditional bank lending services, these products have formed a vast financial lending industry. However, due to insufficient business models and limited consumer repayment capacity and willingness, the industry anticipates exponential growth in debt data annually. Traditional manual debt collection services alone are insufficient to meet the enormous market demand.

[0003] Collection data refers to information that records various activities and communications during the collection process. Because collection data may be used in future audits, legal proceedings, or investigations, by properly archiving and managing this data, organizations can quickly find the required information when needed and prove specific activities or transactions.

[0004] In traditional management, debt collection data typically exists in paper form. Companies typically use filing cabinets and folders to store relevant information. These files often require manual organization and maintenance. With the advancement of technology, many companies have converted debt collection data to digital form. This digital storage method often involves manually filing and storing debt collection data on computer hard drives or servers, which is still not intelligent enough. Summary of the Invention

[0005] The present invention provides a machine learning-based archiving, storage and management method and system for reminder data to solve the above-mentioned problems:

[0006] The present invention proposes a method for archiving, storing and managing reminder data based on machine learning, the method comprising:

[0007] Acquire reminder audio, convert the reminder audio into text through ASR, and build a reminder data corpus from the converted text;

[0008] Building a reminder data text model through the reminder data corpus, inputting new reminder data into the reminder data text model to form the customer's first repayment tendency;

[0009] Acquiring data from the reminder audio to form a second repayment propensity of the customer, weightedly adding the first and second repayment propensities of the customer to form a third repayment propensity, and forming a final repayment propensity of the customer based on the third repayment propensity;

[0010] The customer reminder data is classified and stored in different storage media according to the customer's final repayment tendency and the time when the reminder data is formed.

[0011] Furthermore, the reminder audio is obtained, the reminder audio is converted into text through ASR, and the converted text is used to build a reminder data corpus, including:

[0012] An automated voice system for making collection calls is connected to an ASR, and the ASR converts the collection audio into text;

[0013] The converted text is formed into a label, phone number, time and call text format, wherein the label is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call;

[0014] 25% of the call texts are used to construct a collection corpus, and the call texts are segmented. After segmentation, valid words are extracted and saved in a dictionary.

[0015] For each of the valid words, mark its main type and subtype as well as trigger words;

[0016] The classified and identified words will be added to the corresponding call text and labeled, providing a manual review interface.

[0017] Furthermore, a reminder data text model is constructed using the corpus, and new reminder data is passed through the reminder data text model to form the customer's first repayment tendency, including:

[0018] After the manual inspection, the inspected call texts are used to train the LSTM model, and the remaining 75% of the call texts are used to verify the trained LSTM model. Based on the verification results, the LSTM model is then adjusted and optimized to form a reminder data text model.

[0019] The new reminder data is input into the reminder data text model to form the customer's first repayment tendency.

[0020] Furthermore, obtaining the reminder audio data to form the customer's second repayment propensity, weightedly adding the customer's first repayment propensity and the second repayment propensity to form a third repayment propensity, and forming the customer's final repayment propensity based on the third repayment propensity, including:

[0021] Acquiring reminder audio data, the reminder audio data including the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history;

[0022] The reminder audio data is used as the independent variable and the customer's second repayment tendency is used as the dependent variable, and a polynomial regression model is used for training to construct a second repayment tendency model;

[0023] Obtaining new reminder audio data and inputting it into the second repayment propensity model to form the customer's second repayment propensity;

[0024] The customer's first repayment propensity and the second repayment propensity are weightedly added to form the customer's third repayment propensity, and the customer's final repayment propensity is formed based on the customer's third repayment propensity.

[0025] Furthermore, the customer reminder data is classified and stored in different storage media according to the final repayment tendency and the time when the reminder data was generated, including:

[0026] Customers' final repayment tendency is divided into those who are willing to repay and those who are not willing to repay.

[0027] The reminder data is stored in different storage media according to the time and the final repayment tendency. If the time is less than three months from the deposit time, it will be stored in a high-speed, low-latency storage medium. If the time is between three months and one year from the deposit time, it will be stored in a delayed access layer. If the time is more than one year from the deposit time, it will be stored in a long-term storage layer.

[0028] A machine learning-based reminder data archiving and storage management system, the system comprising:

[0029] Constructing a corpus module for obtaining reminder audio, converting the reminder audio into text through ASR, and building a reminder data corpus from the converted text;

[0030] a module for forming a first repayment tendency, configured to construct a reminder data text model using the reminder data corpus, input new reminder data into the reminder data text model, and form a customer's first repayment tendency;

[0031] a module for forming a final repayment propensity, configured to obtain the reminder audio data, form a second repayment propensity of the customer, perform a weighted addition of the first repayment propensity and the second repayment propensity of the customer to form a third repayment propensity, and form a total repayment propensity of the customer based on the third repayment propensity;

[0032] The storage module is used to store the customer reminder data into different storage media according to the customer's final repayment tendency and the time when the reminder data was generated.

[0033] Furthermore, in a machine learning-based reminder data archiving and storage management system, the corpus building module includes:

[0034] A text conversion module is used to connect the automatic voice system for making collection calls to the ASR, and convert the collection audio into text through the ASR;

[0035] A visualization module is used to convert the converted text into a label, phone number, time, and call text format, wherein the label is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call;

[0036] The word segmentation module is used to construct a debt collection corpus by taking 25% of the call texts, performing word segmentation on the call texts, and extracting valid words from the texts after word segmentation. The valid words will be saved in a dictionary;

[0037] a marking module, configured to mark, for each of the valid words, its main type, subtype and trigger word;

[0038] An interface module is provided to add the classified and identified words to the corresponding call text, add labels, and provide a manual inspection interface.

[0039] Furthermore, in a machine learning-based reminder data archiving and storage management system, the module for forming the first repayment tendency includes:

[0040] The module for training the reminder data text model is used to train the LSTM model using the checked call text after manual inspection, and to verify the trained LSTM model using the remaining 75% of the call text. The module then adjusts and optimizes the LSTM model based on the verification results to form the reminder data text model.

[0041] The new data input module is used to input new reminder data into the reminder data text model to form the customer's first repayment tendency.

[0042] Furthermore, in a machine learning-based reminder data archiving and storage management system, the module for forming the final repayment tendency includes:

[0043] a module for obtaining reminder audio data, configured to obtain reminder audio data, wherein the reminder audio data includes the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history;

[0044] Constructing a second repayment propensity model module, which is used to use the reminder audio data as an independent variable and the customer's second repayment propensity as a dependent variable, and to train a polynomial regression model to construct a second repayment propensity model;

[0045] A second repayment propensity module is formed, which is used to obtain new reminder audio data and input it into the second repayment propensity model to form the customer's second repayment propensity;

[0046] A module for forming a third repayment tendency is used to perform a weighted addition of the customer's first repayment tendency and the second repayment tendency to form the customer's third repayment tendency, and to form the customer's final repayment tendency based on the customer's third repayment tendency.

[0047] Furthermore, a machine learning-based reminder data archiving storage management system, the storage module includes:

[0048] The repayment tendency classification module is used to classify customers' final repayment tendency into those who are willing to repay and those who are not willing to repay.

[0049] The storage medium module is used to store the reminder data in different storage media according to the time and final repayment tendency. If the time is less than three months from the deposit time, it will be stored in a high-speed, low-latency storage medium; if the time is between three months and one year from the deposit time, it will be stored in a delayed access layer; if the time is more than one year from the deposit time, it will be stored in a long-term storage layer.

[0050] The beneficial effects of the present invention are as follows: by using machine learning to establish an automated management process, such as automatically allocating storage locations, and building a repayment tendency model through machine collection audio, the probability that a customer will repay within the specified period can be effectively predicted. For customers who are predicted to have the intention to repay, that is, customers who can achieve the repayment purpose without manual collection methods, the machine can further automatically remind them to achieve the repayment purpose, which greatly liberates manpower. In addition, through the repayment tendency model, customer reminder information is classified and stored, and infrequently used reminder information is placed on a high-latency storage medium, which can effectively save storage costs; placing frequently used information on a fast storage medium can improve data access and response speed; by distinguishing data with different access frequencies and storing them on the most suitable medium, data management can be simplified, ensuring that critical data can be processed and responded to at a faster speed, while non-critical data can be stored and accessed at a lower cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a schematic diagram of a machine learning-based archiving, storage and management method for reminder data according to the present invention. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0053] One embodiment of the present invention provides a method for archiving, storing, and managing reminder data based on machine learning, the method comprising:

[0054] Acquire reminder audio, convert the reminder audio into text through ASR, and build a reminder data corpus from the converted text;

[0055] Building a reminder data text model through the reminder data corpus, inputting new reminder data into the reminder data text model to form the customer's first repayment tendency;

[0056] Acquiring data from the reminder audio to form a second repayment propensity of the customer, weightedly adding the first and second repayment propensities of the customer to form a third repayment propensity, and forming a final repayment propensity of the customer based on the third repayment propensity;

[0057] The customer reminder data is classified and stored in different storage media according to the customer's final repayment tendency and the time when the reminder data is formed.

[0058] The working principle of the above technical solution is as follows: obtain and convert the reminder audio, and use automatic speech recognition (ASR) to convert the reminder audio into text; use the reminder data corpus to establish a reminder data text model; input new reminder data into the reminder data text model, and form the customer's first repayment tendency according to the model; obtain the reminder audio data to form the customer's second repayment tendency; based on the customer's final repayment tendency and the time when the reminder data was formed, the customer's reminder data is classified and stored in different storage media.

[0059] The effects of the above technical solution are as follows: New debt collection data is automatically classified through speech-to-text conversion and modeling, enabling the automatic processing of large amounts of debt collection information and saving labor costs; based on the constructed debt collection data text model, a relatively accurate prediction of the customer's repayment propensity can be made; a weighted sum of the customer's first and second repayment propensities is calculated to form the customer's final repayment propensity. Based on this final repayment propensity, personalized services more tailored to the customer's actual situation are provided, rather than a blanket approach of manual debt collection. This can greatly improve debt collection efficiency, while machine debt collection is used for those with a repayment propensity, significantly freeing up manpower; customer debt collection data is categorized and stored in different storage media as needed, ensuring both access speed and full utilization of storage resources. The above solution can effectively improve the analysis and management of customer repayment propensities, and achieve more flexible management of data storage, while providing customers with more personalized services.

[0060] One embodiment of the present invention provides a machine learning-based archiving, storage, and management method for reminder data, which obtains reminder audio, converts the reminder audio into text using ASR, and constructs a reminder data corpus from the converted text, including:

[0061] An automated voice system for making collection calls is connected to an ASR, and the ASR converts the collection audio into text;

[0062] The converted text is formed into a label, phone number, time and call text format, wherein the label is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call;

[0063] 25% of the call texts are used to construct a collection corpus, and the call texts are segmented. After segmentation, valid words are extracted and saved in a dictionary.

[0064] For each of the valid words, mark its main type and subtype as well as trigger words;

[0065] The classified and identified words will be added to the corresponding call text and labeled, providing a manual review interface.

[0066] For example, words such as "OK", "You can", "OK", and "I understand" are marked as the main type of willingness to repay, and the subtype is "Promise to repay". Trigger words are "OK", "You can", "OK", and "I understand". Words such as "Will handle it later", "Process it tomorrow", and "Process it later" are marked as the main type of willingness to repay, and the subtype is "Delayed repayment". Words such as "Wrong number", "No loan", and "No money borrowed" are marked as the main type of no willingness to repay, and the subtype is "Denial".

[0067] The working principle of the above technical solution is as follows: the audio of the collection call is converted into text through automatic speech recognition (ASR); the text will be labeled, including a label identifying a unique call, the customer's phone number, the call time and the call content; 25% of the call texts will be used to build a collection corpus, valid words will be extracted from these call texts, and then a dictionary will be built to store these valid words; the valid words will be segmented and their main types, subtypes and trigger words will be marked for further identification and classification; the classified and identified words will be added to the corresponding call text and labeled so that they can be provided to the manual inspection interface.

[0068] The effects of the above technical solution are: automatic textualization, converting the audio of the collection call into text form through ASR, so that the call content can be digitized and then processed by a computer program; annotating the call text and building a collection corpus to facilitate subsequent vocabulary processing and classification recognition; performing word segmentation on the call text, extracting valid words from it, and forming a dictionary to help extract and analyze key information; annotating each valid word with the main type, subtype and trigger word, which facilitates the further use and processing of the word; providing a manual inspection interface, through which the recognized words and labels can be reviewed to ensure accuracy and completeness; the above solution can help the machine system better understand the content of the collection call and perform structured processing on the call information, providing a basis for subsequent analysis and management.

[0069] One embodiment of the present invention provides a machine learning-based archiving, storage, and management method for reminder data. The method constructs a reminder data text model using the corpus, and applies new reminder data to the reminder data text model to form a customer's first repayment tendency, including:

[0070] After the manual inspection, the inspected call texts are used to train the LSTM model, and the remaining 75% of the call texts are used to verify the trained LSTM model. Based on the verification results, the LSTM model is then adjusted and optimized to form a reminder data text model.

[0071] The new reminder data is input into the reminder data text model to form the customer's first repayment tendency.

[0072] New debt collection data is processed through text segmentation and valid data extraction to form an input sequence (x1, x2, ..., xn). After processing it through the recurrent LSTM neural network, the model output predicted probabilities can be obtained using the following function. Assume that in a specific scenario, we have two node outputs z0 and z1, representing "no intention to repay" and "willingness to repay," respectively:

[0073] P （还款倾向1）=

[0074] To give a concrete numerical example, let's assume z0 = 1.5 and z1 = 3.2. Substituting these values ​​into the above function, we can calculate the probability: P (willingness to repay) = 0.845, which means that the probability that the customer is willing to repay is approximately 84.5%.

[0075] The above technical solution works as follows: After manual inspection, the call transcripts are used to train the LSTM model. This process learns patterns based on the call transcript's features and contextual information, and constructs a text model for the debt collection data. The trained LSTM model is then validated using the remaining 75% of the call transcripts, and the validation results are used to tune and optimize the LSTM model. This involves adjusting hyperparameters and regularization to improve model performance. New debt collection data is then processed through text segmentation and valid data extraction to form an input sequence (x1, x2, ..., xn). This is then processed by the tuned and optimized LSTM model to obtain a predicted probability of repayment willingness.

[0076] The effects of the above technical solution are: improved accuracy. By using data to train, verify, and tune the model, the model has higher accuracy and reliability when predicting customers' repayment willingness; the optimized model can receive new reminder data and quickly predict customers' repayment tendency, providing timely support for further decision-making in actual business; parameter adjustment and optimization ensure that the model better adapts to actual data and can more accurately predict customers' repayment willingness in various scenarios.

[0077] Debt collection data text models can provide insights into customers' initial repayment propensities, providing crucial information for machine-assisted decision-making, thereby improving collection efficiency and quality. Through the above methods, a more accurate and practical debt collection data text model can be established, providing stronger support for managers in debt collection decisions.

[0078] One embodiment of the present invention provides a machine learning-based archiving, storage, and management method for debt reminder data, which obtains the debt reminder audio data to form a customer's second repayment propensity, performs a weighted addition of the customer's first and second repayment propensities to form a third repayment propensity, and forms the customer's final repayment propensity based on the third repayment propensity, including:

[0079] Acquiring reminder audio data, the reminder audio data including the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history;

[0080] The reminder audio data is used as the independent variable and the customer's second repayment tendency is used as the dependent variable, and a polynomial regression model is used for training to construct a second repayment tendency model;

[0081] Obtaining new reminder audio data and inputting it into the second repayment propensity model to form the customer's second repayment propensity;

[0082] The customer's first repayment propensity and the second repayment propensity are weightedly added to form the customer's third repayment propensity, and the customer's final repayment propensity is formed based on the customer's third repayment propensity.

[0083] For a given reminder audio data feature ( x_1, x_2, ..., x_n ), polynomial regression can combine them with each other and introduce nonlinear relationships:

[0084]

[0085] in, is the second repayment propensity, w0, w1, ..., w nn are the parameters (weights) that the second repayment propensity model needs to learn to capture the nonlinear relationship between input features.

[0086] P (还款倾向3) =α*P (还款倾向1) + (1-α) * P (还款倾向2)

[0087] Among them, α is the weight coefficient, ranging from 0 to 1, P (还款倾向3) The third repayment tendency.

[0088] The working principle of the above technical solution is as follows: audio data of debt reminders is obtained and features are extracted, including the maximum number of calls made on the day, the maximum number of calls made in history, the maximum number of calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history. A polynomial regression model is trained using the above debt reminder audio data as the independent variable and the customer's second repayment propensity as the dependent variable. During model training, a second repayment propensity prediction model is constructed by learning the complex relationships between various features. New debt reminder audio data is input into the trained polynomial regression model to obtain a prediction result for the customer's second repayment propensity. The customer's first and second repayment propensities are weighted and added together to form the customer's third repayment propensity. The weights need to be adjusted according to actual circumstances. The customer's third repayment propensity is the customer's final repayment propensity.

[0089] The effects of the above technical solution are as follows: Research has shown that factors such as the maximum number of calls made on the day, the maximum number of calls made in history, the maximum number of calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history also affect a customer's repayment propensity. By using reminder audio data and extracting the above features, a polynomial regression model is used to more comprehensively capture the complex relationships between the features, thereby improving the model's prediction accuracy. By integrating reminder audio data with the customer's first repayment propensity, a personalized third repayment propensity can be derived, helping to better understand customer behavior and improve the accuracy of repayment predictions. The third repayment propensity is formed through weighted addition, and ultimately the customer's final repayment propensity is formed, improving the robustness and reliability of the overall prediction. This technical solution can be applied to real-time data processing to provide rapid feedback on the customer's final repayment propensity, supporting actual business decisions. By integrating the polynomial regression model, reminder audio data, and the weighted addition method, a customer repayment propensity prediction system that is more realistic can be constructed, providing more accurate and competitive prediction results.

[0090] One embodiment of the present invention provides a machine learning-based archiving, storage, and management method for reminder data, which categorizes customer reminder data into different storage media based on final repayment trends and the time when the reminder data was generated, including:

[0091] Customers' final repayment tendency is divided into those who are willing to repay and those who are not willing to repay.

[0092] The reminder data is stored in different storage media according to the time and the final repayment tendency. If the time is less than three months from the deposit time, it will be stored in a high-speed, low-latency storage medium. If the time is between three months and one year from the deposit time, it will be stored in a delayed access layer. If the time is more than one year from the deposit time, it will be stored in a long-term storage layer.

[0093] The working principle of the above technical solution is as follows: data storage is divided according to time, and reminder data is stored in different storage media according to the said time and final repayment tendency. This process is to better manage data and reasonably store it according to the access frequency and importance of the data; dynamic storage medium selection, according to the length of time from the storage time, choose to store data in high-speed and low-latency storage media, lagging access layer or long-term storage layer to balance data access speed and cost-effectiveness; data access optimization, by storing data at different levels, can optimize data access efficiency according to the frequency of data use, ensure that commonly used data can be quickly accessed, and at the same time reduce storage costs.

[0094] The effects of the above technical solutions are: improving data access efficiency, high-speed and low-latency storage media can meet the needs of rapid access to recently active data, helping to improve the overall system response speed and user experience; reducing data storage costs, and effectively controlling data storage costs by using delayed access layers and long-term storage layers. For data that is rarely accessed, a balance is found between data cost and access efficiency; reasonable data management, storing data in layers according to access patterns and time, helps to better manage and utilize data, maximizes the application value of data while ensuring data security, and improves overall data management efficiency; through such technical solutions, data access speed can be made more efficient, effectively reducing data storage costs, while rationally managing and utilizing data, thereby providing more solid data support for business operations and decision-making.

[0095] One embodiment of the present invention provides a machine learning-based reminder data archiving and storage management system, the system comprising:

[0096] Constructing a corpus module for obtaining reminder audio, converting the reminder audio into text through ASR, and building a reminder data corpus from the converted text;

[0097] a module for forming a first repayment tendency, configured to construct a reminder data text model using the reminder data corpus, input new reminder data into the reminder data text model, and form a customer's first repayment tendency;

[0098] a module for forming a final repayment propensity, configured to obtain the reminder audio data, form a second repayment propensity of the customer, perform a weighted addition of the first repayment propensity and the second repayment propensity of the customer to form a third repayment propensity, and form a total repayment propensity of the customer based on the third repayment propensity;

[0099] The storage module is used to store the customer reminder data into different storage media according to the customer's final repayment tendency and the time when the reminder data was generated.

[0100] The working principle of the above technical solution is as follows: obtain and convert the reminder audio, and use automatic speech recognition (ASR) to convert the reminder audio into text; use the reminder data corpus to establish a reminder data text model; input new reminder data into the reminder data text model, and form the customer's first repayment tendency according to the model; obtain the reminder audio data to form the customer's second repayment tendency; based on the customer's final repayment tendency and the time when the reminder data was formed, the customer's reminder data is classified and stored in different storage media.

[0101] The effects of the above technical solution are as follows: New debt collection data is automatically classified through speech-to-text conversion and modeling, enabling the automatic processing of large amounts of debt collection information and saving labor costs; based on the constructed debt collection data text model, a relatively accurate prediction of the customer's repayment propensity can be made; a customer's first and second repayment propensities are weighted and added together to form the customer's final repayment propensity. Based on this final repayment propensity, personalized services more tailored to the customer's actual situation are provided, rather than a blanket policy of manual collection for all debt collection cases, which can greatly improve collection efficiency. For those with a repayment propensity, machine collection is used for debt collection, which greatly frees up manpower; customer debt collection data is categorized and stored in different storage media as needed, ensuring both access speed and full utilization of storage resources. The above solution can effectively improve the analysis and management of customer repayment propensities, and achieve more flexible management of data storage, while providing customers with more personalized services.

[0102] One embodiment of the present invention provides a machine learning-based reminder data archiving and storage management system, wherein the corpus building module includes:

[0103] A text conversion module is used to connect the automatic voice system for making collection calls to the ASR, and convert the collection audio into text through the ASR;

[0104] A visualization module is used to convert the converted text into a label, phone number, time, and call text format, wherein the label is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call;

[0105] The word segmentation module is used to construct a debt collection corpus by taking 25% of the call texts, performing word segmentation on the call texts, and extracting valid words from the texts after word segmentation. The valid words will be saved in a dictionary;

[0106] a marking module, configured to mark, for each of the valid words, its main type, subtype and trigger word;

[0107] An interface module is provided to add the classified and identified words to the corresponding call text, add labels, and provide a manual inspection interface.

[0108] The working principle of the above technical solution is as follows: the audio of the collection call is converted into text through automatic speech recognition (ASR); the text will be labeled, including a label identifying a unique call, the customer's phone number, the call time and the call content; 25% of the call texts will be used to build a collection corpus, valid words will be extracted from these call texts, and then a dictionary will be built to store these valid words; the valid words will be segmented and their main types, subtypes and trigger words will be marked for further identification and classification; the classified and identified words will be added to the corresponding call text and labeled so that they can be provided to the manual inspection interface.

[0109] The effects of the above technical solution are: automatic textualization, converting the audio of the collection call into text form through ASR, so that the call content can be digitized and then processed by a computer program; annotating the call text and building a collection corpus to facilitate subsequent vocabulary processing and classification recognition; performing word segmentation on the call text, extracting valid words from it, and forming a dictionary to help extract and analyze key information; annotating each valid word with the main type, subtype and trigger word, which facilitates the further use and processing of the word; providing a manual inspection interface, through which the recognized words and labels can be reviewed to ensure accuracy and completeness; the above solution can help the machine system better understand the content of the collection call and perform structured processing on the call information, providing a basis for subsequent analysis and management.

[0110] In one embodiment of the present invention, a machine learning-based reminder data archiving and storage management system, the module for forming a first repayment tendency includes:

[0111] The module for training the reminder data text model is used to train the LSTM model using the checked call text after manual inspection, and to verify the trained LSTM model using the remaining 75% of the call text. The module then adjusts and optimizes the LSTM model based on the verification results to form the reminder data text model.

[0112] The new data input module is used to input new reminder data into the reminder data text model to form the customer's first repayment tendency.

[0113] The above technical solution works as follows: After manual inspection, the call transcripts are used to train the LSTM model. This process learns patterns based on the call transcript's features and contextual information, and constructs a text model for the debt collection data. The trained LSTM model is then validated using the remaining 75% of the call transcripts, and the validation results are used to tune and optimize the LSTM model. This involves adjusting hyperparameters and regularization to improve model performance. New debt collection data is then processed through text segmentation and valid data extraction to form an input sequence (x1, x2, ..., xn). This is then processed by the tuned and optimized LSTM model to obtain a predicted probability of repayment willingness.

[0114] The effects of the above technical solution are: improved accuracy. By using data to train, verify, and tune the model, the model has higher accuracy and reliability when predicting customers' repayment willingness; the optimized model can receive new reminder data and quickly predict customers' repayment tendency, providing timely support for further decision-making in actual business; parameter adjustment and optimization ensure that the model better adapts to actual data and can more accurately predict customers' repayment willingness in various scenarios.

[0115] Debt collection data text models can provide insights into customers' initial repayment propensities, providing crucial information for machine-assisted decision-making, thereby improving collection efficiency and quality. Through the above methods, a more accurate and practical debt collection data text model can be established, providing stronger support for managers in debt collection decisions.

[0116] In one embodiment of the present invention, a machine learning-based reminder data archiving and storage management system, the module for forming a final repayment tendency includes:

[0117] a module for obtaining reminder audio data, configured to obtain reminder audio data, wherein the reminder audio data includes the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history;

[0118] Constructing a second repayment propensity model module, which is used to use the reminder audio data as an independent variable and the customer's second repayment propensity as a dependent variable, and to train a polynomial regression model to construct a second repayment propensity model;

[0119] A second repayment propensity module is formed, which is used to obtain new reminder audio data and input it into the second repayment propensity model to form the customer's second repayment propensity;

[0120] A module for forming a third repayment tendency is used to perform a weighted addition of the customer's first repayment tendency and the second repayment tendency to form the customer's third repayment tendency, and to form the customer's final repayment tendency based on the customer's third repayment tendency.

[0121] The working principle of the above technical solution is as follows: audio data of debt reminders is obtained and features are extracted, including the maximum number of calls made on the day, the maximum number of calls made in history, the maximum number of calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history. A polynomial regression model is trained using the above debt reminder audio data as the independent variable and the customer's second repayment propensity as the dependent variable. During model training, a second repayment propensity prediction model is constructed by learning the complex relationships between various features. New debt reminder audio data is input into the trained polynomial regression model to obtain a prediction result for the customer's second repayment propensity. The customer's first and second repayment propensities are weighted and added together to form the customer's third repayment propensity. The weights need to be adjusted according to actual circumstances. The customer's third repayment propensity is the customer's final repayment propensity.

[0122] The effects of the above technical solution are as follows: Research has shown that factors such as the maximum number of calls made on the day, the maximum number of calls made in history, the maximum number of calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history also affect a customer's repayment propensity. By using reminder audio data and extracting the above features, a polynomial regression model is used to more comprehensively capture the complex relationships between the features, thereby improving the model's prediction accuracy. By integrating reminder audio data with the customer's first repayment propensity, a personalized third repayment propensity can be derived, helping to better understand customer behavior and improve the accuracy of repayment predictions. The third repayment propensity is formed through weighted addition, and ultimately the customer's final repayment propensity is formed, improving the robustness and reliability of the overall prediction. This technical solution can be applied to real-time data processing to provide rapid feedback on the customer's final repayment propensity, supporting actual business decisions. By integrating the polynomial regression model, reminder audio data, and the weighted addition method, a customer repayment propensity prediction system that is more realistic can be constructed, providing more accurate and competitive prediction results.

[0123] One embodiment of the present invention provides a machine learning-based reminder data archiving and storage management system, wherein the storage module includes:

[0124] The repayment tendency classification module is used to classify customers' final repayment tendency into those who are willing to repay and those who are not willing to repay.

[0125] The storage medium module is used to store the reminder data in different storage media according to the time and final repayment tendency. If the time is less than three months from the deposit time, it will be stored in a high-speed, low-latency storage medium; if the time is between three months and one year from the deposit time, it will be stored in a delayed access layer; if the time is more than one year from the deposit time, it will be stored in a long-term storage layer.

[0126] The working principle of the above technical solution is as follows: data storage is divided according to time, and reminder data is stored in different storage media according to the said time and final repayment tendency. This process is to better manage data and reasonably store it according to the access frequency and importance of the data; dynamic storage medium selection, according to the length of time from the storage time, choose to store data in high-speed and low-latency storage media, lagging access layer or long-term storage layer to balance data access speed and cost-effectiveness; data access optimization, by storing data at different levels, can optimize data access efficiency according to the frequency of data use, ensure that commonly used data can be quickly accessed, and at the same time reduce storage costs.

[0127] The effects of the above technical solutions are: improving data access efficiency, high-speed and low-latency storage media can meet the needs of rapid access to recently active data, helping to improve the overall system response speed and user experience; reducing data storage costs, and effectively controlling data storage costs by using delayed access layers and long-term storage layers. For data that is rarely accessed, a balance is found between data cost and access efficiency; reasonable data management, storing data in layers according to access patterns and time, helps to better manage and utilize data, maximizes the application value of data while ensuring data security, and improves overall data management efficiency; through such technical solutions, data access speed can be made more efficient, effectively reducing data storage costs, while rationally managing and utilizing data, thereby providing more solid data support for business operations and decision-making.

[0128] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for archiving and storing reminder data based on machine learning, characterized in that: The method comprises: Acquire reminder audio, convert the reminder audio into text through ASR, and build a reminder data corpus with the converted text; Building a reminder data text model through the reminder data corpus, inputting new reminder data into the reminder data text model, and forming the customer's first repayment tendency; Acquire the data of the reminder audio to form the second repayment tendency of the customer, weightedly add the first repayment tendency and the second repayment tendency of the customer to form a third repayment tendency, and form the final repayment tendency of the customer according to the third repayment tendency; The customer reminder data is classified and stored in different storage media according to the customer's final repayment tendency and the time when the reminder data is formed; The method of obtaining the data of the reminder audio to form the second repayment tendency of the customer, weighting and adding the first repayment tendency and the second repayment tendency of the customer to form the third repayment tendency, and forming the final repayment tendency of the customer according to the third repayment tendency includes: Acquire reminder audio data, wherein the reminder audio data includes the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history; The reminder audio data is used as an independent variable, and the customer's second repayment tendency is used as a dependent variable, and a polynomial regression model is used for training to construct a second repayment tendency model; Acquire new reminder audio data, and input it into the second repayment tendency model to form the customer's second repayment tendency; The customer's first repayment tendency and the second repayment tendency are weightedly added to form the customer's third repayment tendency, and the customer's final repayment tendency is formed based on the customer's third repayment tendency.

2. According to the method for archiving, storing and managing reminder data based on machine learning in claim 1, it is characterized in that: Acquire reminder audio, convert the reminder audio into text through ASR, and build a reminder data corpus with the converted text, including: An automatic voice system for making collection calls is connected to an ASR, and the ASR converts the collection audio into text; The converted text is formed into a number, phone number, time and call text form, wherein the number is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call; 25% of the call texts are taken to construct a collection corpus, and the call texts are segmented. After the segmentation, valid words are extracted and saved in a dictionary. For each of the valid words, annotate its main type and subtype as well as trigger words; The classified and identified words will be added to the corresponding call text and labeled, providing a manual review interface.

3. According to the method for archiving, storing and managing reminder data based on machine learning in claim 1, it is characterized in that: The reminder data text model is constructed by using the corpus, and the new reminder data is passed through the reminder data text model to form the customer's first repayment tendency, including: After the manual inspection is completed, the LSTM model is trained with the inspected call texts, and the trained LSTM model is verified with the remaining 75% of the call texts. The LSTM model is then adjusted and optimized based on the verification results to form a reminder data text model. The new reminder data is input into the reminder data text model to form the customer's first repayment tendency.

4. According to the method for archiving, storing and managing reminder data based on machine learning as claimed in claim 1, it is characterized in that: According to the final repayment tendency and the time when the reminder data was formed, the customer reminder data is classified and stored in different storage media, including: Customers' final repayment tendency is divided into those who are willing to repay and those who are not willing to repay. The reminder data is stored in different storage media according to the time and the final repayment tendency. If the time is less than three months from the deposit time, it will be stored in a high-speed and low-latency storage medium. If the time is three months to one year from the deposit time, it will be stored in a delayed access layer. If the time is more than one year from the deposit time, it will be stored in a long-term storage layer.

5. A machine learning-based reminder data archiving storage management system, characterized in that: The system comprises: Constructing a corpus module, for obtaining reminder audio, converting the reminder audio into text through ASR, and constructing a reminder data corpus from the converted text; A module for forming a first repayment tendency is used to construct a reminder data text model through the reminder data corpus, input new reminder data into the reminder data text model, and form a customer's first repayment tendency; A module for forming a final repayment tendency is used to obtain the data of the reminder audio to form a second repayment tendency of the customer, weightedly add the first repayment tendency and the second repayment tendency of the customer to form a third repayment tendency, and form a total repayment tendency of the customer based on the third repayment tendency; A storage module, used to store the customer reminder data in different storage media according to the customer's final repayment tendency and the time when the reminder data was formed; Wherein, the module for forming the final repayment tendency includes: A module for obtaining reminder audio data is used to obtain reminder audio data, wherein the reminder audio data includes the maximum number of phone calls made on the day, the maximum number of phone calls made in history, the maximum number of phone calls made per day in history, the maximum call duration in history, the minimum call duration in history, the average call duration in history, the maximum call duration in history, the minimum call duration in history, and the average call duration in history; Constructing a second repayment tendency model module, which is used to use the reminder audio data as an independent variable and the customer's second repayment tendency as a dependent variable, and to train using a polynomial regression model to construct a second repayment tendency model; A second repayment tendency module is formed, which is used to obtain new reminder audio data and input it into the second repayment tendency model to form the customer's second repayment tendency; A module for forming a third repayment tendency is used to perform weighted addition of the first repayment tendency and the second repayment tendency of the customer to form the third repayment tendency of the customer, and to form the final repayment tendency of the customer based on the third repayment tendency of the customer.

6. According to the machine learning-based reminder data archiving storage management system of claim 5, it is characterized in that: The corpus building module includes: A text conversion module, used to connect the automatic voice system for collecting debts to the ASR, and convert the debt collection audio into text through the ASR; A visualization module is used to convert the converted text into a number, phone number, time and call text form, wherein the number is used to uniquely identify the collection call, the phone number is used to identify the customer's phone number, the time is used to identify the time when the collection call was made, and the call text is used to identify the text content of the collection call; The word segmentation module is used to take 25% of the call texts to build a collection corpus, perform word segmentation on the call texts, extract valid words from the texts after word segmentation, and save the valid words into a dictionary; A marking module, used for marking the main type and subtype and trigger word of each valid word; An interface module is provided to add the classified and identified words to the corresponding call text, add labels, and provide a manual inspection interface.

7. According to claim 5, a machine learning-based reminder data archiving storage management system is characterized in that: The forming of the first repayment tendency module comprises: The module for training the reminder data text model is used to train the LSTM model with the checked call text after the manual inspection is completed, and to verify the trained LSTM model with the remaining 75% of the call text, and then adjust and optimize the LSTM model according to the verification result to form the reminder data text model; The new data input module is used to input new reminder data into the reminder data text model to form the customer's first repayment tendency.

8. According to the machine learning-based reminder data archiving storage management system of claim 5, it is characterized in that: The storage module comprises: The repayment tendency classification module is used to classify customers' final repayment tendency into those who are willing to repay and those who are not willing to repay. The storage medium module is used to store the reminder data in different storage media according to the time and the final repayment tendency. If the time is less than three months away from the deposit time, it will be stored in a high-speed and low-latency storage medium. If the time is three months to one year away from the deposit time, it will be stored in a delayed access layer. If the time is more than one year away from the deposit time, it will be stored in a long-term storage layer.

Images