Telephone collection method, device and equipment based on large language model, and medium

Abstract

The invention discloses a call collection method and device based on a big language model, equipment and a medium, and relates to the technical field of financial credit loan, and the method comprises the steps: obtaining feature information and historical call records corresponding to a target debtor, obtaining a psychological portrait of the target debtor according to the big language model and the target historical call records, and based on an initial cue word; generating a to-be-verified collection strategy based on the initial prompt word and the psychological portrait, and performing compliance rule verification on the to-be-verified collection strategy to obtain a target collection strategy according to a verification result; carrying out collection by using the target collection strategy, recording a collection record in real time, and obtaining a repayment willingness of the target debtor by using the collection record; and judging whether the repayment willingness meets an expected condition or not, and ending the collection process of the target debtor according to a corresponding judgment result. The personalized collection strategy is generated according to the feature information of the debtor, and the collection strategy is verified, so that the compliance and effectiveness of the collection process are ensured.

Description

Technical Field

[0001] The present invention relates to the field of financial credit technology, and in particular to a telephone debt collection method, device, equipment and medium based on a large language model. Background Art

[0002] With the rapid development of artificial intelligence technology, especially the application of natural language processing (NLP) and large language models in dialogue systems, new solutions have been provided for telephone debt collection. Large language models can understand and generate natural language texts and have strong semantic understanding and generation capabilities, which makes it possible to achieve intelligent and automated telephone debt collection.

[0003] The existing AI-based telephone collection method uses standardized collection strategies to collect debts from debtors. This collection method does not make full use of historical call record data and is difficult to make personalized adjustments based on the specific circumstances of the debtor (such as psychological state, willingness to repay, etc.). In addition, the compliance of this collection method is manually inspected, and there is a problem of being unable to accurately check the compliance of the collection strategy, making it difficult to ensure the compliance of the collection strategy and communication process and the collection effect. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a telephone debt collection method, device, equipment and medium based on a large language model, which can generate a personalized debt collection strategy based on the characteristic information of the debtor and verify the debt collection strategy to ensure the compliance and effectiveness of the debt collection process. The specific scheme is as follows:

[0005] In a first aspect, the present application provides a telephone debt collection method based on a large language model, comprising:

[0006] Obtain target feature information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, obtain a psychological portrait of the target debtor based on the target big language model and the target historical call records, and generate an initial prompt word based on the target feature information;

[0007] Generate a collection strategy to be verified based on the initial prompt word and the psychological portrait, and perform compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result;

[0008] Using the target collection strategy to collect debts from the target debtor, recording the collection records generated during the collection process in real time, and using the collection records to obtain the current repayment willingness of the target debtor;

[0009] Determine whether the current repayment willingness of the target debtor meets the expected conditions, and end the collection process for the target debtor according to the corresponding first judgment result.

[0010] Optionally, before acquiring the target characteristic information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, the method further includes:

[0011] Obtaining historical debt collection dialogue data and initial feature information corresponding to any debtor, and performing data cleaning on the historical debt collection dialogue records and the initial feature information to obtain cleaned historical debt collection records and cleaned feature information;

[0012] Annotating the historical debt collection conversation records after data cleaning to obtain annotated historical debt collection conversation records, and performing data format conversion on the annotated historical debt collection conversation records to obtain historical call records;

[0013] The corpus is constructed using the historical call records and the cleaned feature information, and the corpus is stored in the preset big data platform.

[0014] Optionally, generating a collection strategy to be verified based on the initial prompt word and the psychological portrait includes:

[0015] Generating an initial collection strategy corresponding to the target debtor using the initial prompt word, and generating a post-inspection collection strategy based on the initial collection strategy;

[0016] An optimized prompt word is generated according to the post-inspection collection strategy and the psychological portrait, and the collection strategy to be verified is generated using the optimized prompt word.

[0017] Optionally, obtaining a target collection strategy corresponding to the target debtor according to the corresponding verification result includes:

[0018] If the verification result indicates that the collection strategy to be verified has passed the compliance rule verification, the collection strategy to be verified is determined as the target collection strategy corresponding to the target debtor;

[0019] If the verification result indicates that the collection strategy to be verified has not passed the compliance rule verification, then jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor to generate a new target collection strategy.

[0020] Optionally, the step of ending the collection process on the target debtor according to the corresponding first judgment result includes:

[0021] If the target value corresponding to the current repayment willingness reaches the preset expected value, the collection process for the target debtor is terminated;

[0022] If the target value corresponding to the current repayment willingness does not reach the preset expected value, jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor, and after obtaining the new repayment willingness of the target debtor next time, judge whether the willingness value corresponding to the repayment willingness of the target debtor increases, and end the collection process on the target debtor according to the corresponding second judgment result.

[0023] Optionally, the step of ending the collection process on the target debtor according to the corresponding second judgment result includes:

[0024] If the willingness value corresponding to the repayment willingness increases, and the willingness value corresponding to the new repayment willingness does not reach the preset expected value, jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor, so as to generate a new target collection strategy in real time to collect debts from the target debtor;

[0025] If the willingness value corresponding to the repayment willingness does not increase, the collection process for the target debtor is terminated.

[0026] Optionally, before obtaining the psychological portrait of the target debtor according to the target large language model and the target historical call records, the method further includes:

[0027] Acquire different types of initial training data, and add corresponding labels to the initial training data to obtain target training data;

[0028] Different training sets are constructed according to the type of the target training data and the target training data, and different types of training are performed on the initial large language model using the different training sets to obtain the target large language model.

[0029] In a second aspect, the present application provides a telephone debt collection device based on a large language model, comprising:

[0030] A prompt word generation module is used to obtain target feature information and target historical call records corresponding to a target debtor from a corpus in a preset big data platform, obtain a psychological portrait of the target debtor according to a target big language model and the target historical call records, and generate an initial prompt word based on the target feature information;

[0031] A collection strategy acquisition module, used to generate a collection strategy to be verified based on the initial prompt word and the psychological portrait, and to perform compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result;

[0032] A repayment willingness acquisition module is used to collect debts from the target debtor using the target collection strategy, record the collection records generated during the collection process in real time, and obtain the current repayment willingness of the target debtor using the collection records;

[0033] The collection termination module is used to determine whether the current repayment willingness of the target debtor meets the expected conditions, and terminate the collection process for the target debtor according to the corresponding first judgment result.

[0034] In a third aspect, the present application provides an electronic device, including:

[0035] Memory, used to store computer programs;

[0036] A processor is used to execute the computer program to implement the aforementioned telephone debt collection method based on a large language model.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the aforementioned telephone collection method based on a large language model.

[0038] The present application first obtains the target feature information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, obtains the psychological portrait of the target debtor according to the target large language model and the target historical call records, and generates an initial prompt word based on the target feature information, and then generates a collection strategy to be verified based on the initial prompt word and the psychological portrait, and performs compliance rule verification on the collection strategy to be verified to obtain the target collection strategy corresponding to the target debtor according to the corresponding verification result, and then uses the target collection strategy to collect debts from the target debtor, records the collection records generated during the collection process in real time, and uses the collection records to obtain the current repayment intention of the target debtor, and finally determines whether the current repayment intention of the target debtor meets the expected conditions, and ends the collection process for the target debtor according to the corresponding first judgment result. It can be seen that this application obtains the psychological portrait of the debtor and uses the psychological portrait to generate the collection strategy to be verified, thereby ensuring the pertinence of the generated collection strategy and thus improving the collection effect; by verifying the collection strategy to be verified, the compliance of the obtained collection strategy is ensured, thereby ensuring the compliance of the collection process; by real-time analysis of the debtor's repayment willingness and dynamically changing the collection process according to the repayment willingness, the effect of telephone collection is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A flowchart of a telephone collection method based on a large language model disclosed in this application;

[0041] Figure 2 This is a structural diagram of a telephone collection system based on a large language model disclosed in this application;

[0042] Figure 3 A flowchart of telephone debt collection based on a large language model disclosed in this application;

[0043] Figure 4 A flow chart of a large language model training method disclosed in this application;

[0044] Figure 5 A large language model training flow chart disclosed in this application;

[0045] Figure 6 This is a schematic diagram of the structure of a telephone collection device based on a large language model disclosed in this application;

[0046] Figure 7 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0047] 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.

[0048] At present, the method of telephone collection is to use standardized collection strategies to collect debts from debtors. This method has a low degree of intelligence and lacks an effective compliance inspection mechanism, making it difficult to ensure the compliance and collection effect of the collection strategy and communication process. To this end, this application provides a telephone collection method based on a large language model, which generates personalized collection strategies based on the characteristic information of the debtor and verifies the collection strategies, thereby ensuring the compliance and effectiveness of the collection process.

[0049] See also Figure 1 As shown, the embodiment of the present invention discloses a telephone debt collection method based on a large language model, comprising:

[0050] Step S11, obtaining target feature information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, obtaining a psychological portrait of the target debtor according to the target large language model and the target historical call records, and generating an initial prompt word based on the target feature information.

[0051] In this embodiment, before obtaining the psychological portrait of the target debtor, the process also includes: obtaining the historical debt collection dialogue data and initial feature information corresponding to any debtor, performing data cleaning on the historical debt collection dialogue records and the initial feature information to obtain cleaned historical debt collection records and cleaned feature information; annotating the cleaned historical debt collection dialogue records to obtain annotated historical debt collection dialogue records, and performing data format conversion on the annotated historical debt collection dialogue records to obtain historical call records; constructing a corpus using the historical call records and the cleaned feature information, and storing the corpus in a preset big data platform; specifically, Figure 2 As shown, this solution involves a data collection and preprocessing module, a large language model training module, a feature extraction and context understanding module, a speech generation and optimization module, a telephone collection execution module and a compliance inspection module. Among them, the data collection and preprocessing module is responsible for collecting and organizing historical collection dialogue data, debtor basic information, repayment records and other multi-dimensional data, cleaning, labeling and structuring, building a collection dialogue corpus, and storing it on the big data platform. The multi-dimensional data other than the historical collection dialogue data is the initial feature information corresponding to the target debtor; the above process can specifically include: collecting historical collection dialogue records (including the conversation content between the collector and the debtor, the call duration, the debtor's feedback, etc.), the debtor's basic information (including the surname and surname), and the debtor's basic information (including the name and surname of the debtor). The model collects the debt collection results (successful, partial repayment, rejection, lost contact, etc.), collects the debts received, and collects the results of the collection (successful, partial repayment, rejection, lost contact, etc.). The model also collects the debts received, collects the debts received, and collects the results of the collection (successful, partial repayment, rejected ... collects the results of the collection (successful, partial repayment, rejected, lost contact, etc.). The model also collects the debts received, collects the debts received, collects the results of the collection (successful, partial repayment, rejected, lost contact, etc.). The model also collects the debts received, collects the debts received, collects the results of the collection (successful, partial repayment, rejected, lost contact, etc.). The model also collects the debts received, collects the debts received, collects the results of the collection (successful, partial repayment, rejected, lost contact

[0052] In this embodiment, before obtaining the psychological portrait of the target debtor according to the target big language model and the target historical call records, it also includes: obtaining different types of initial training data, adding corresponding labels to the initial training data to obtain the target training data; constructing different training sets according to the type of the target training data and the target training data, and using different training sets to perform different types of training on the initial big language model to obtain the target big language model. It should be noted that in this embodiment, different types of training are required for the big language model, so as to subdivide the big language model into a psychological analysis big model, a compliance detection big model, a repayment willingness big model, and a collection strategy big model, wherein the process of training the big language model is the responsibility of the big language model training module, which is mainly used to train and generate the psychological analysis big model, the compliance detection big model, the repayment willingness big model, and the collection strategy big model. Based on large language models such as the Llama series, Falcon, ChatGLM (Chat Generative Language Model, a large language model) series, and Qwen series, unsupervised training is performed on the preprocessed corpus to enable it to have professional field knowledge (psychology, compliance, debt collection, etc.); supervised learning is then used to optimize the model's understanding of debt collection scenarios and the ability to generate dialogue, ensuring that the final dialogue can be generated in a way that is both efficient and compliant. By performing data cleaning, data labeling, data format conversion and other operations on the collected raw data, the reliability of the data is improved, and the efficiency of data processing and analysis in subsequent processes is improved.

[0053] Step S12: generating a collection strategy to be verified based on the initial prompt word and the psychological portrait, and performing compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result.

[0054] In this embodiment, the process of generating a collection strategy to be verified based on the initial prompt words and the psychological portrait specifically includes: using the initial prompt words to generate an initial collection strategy corresponding to the target debtor, and generating a post-inspection collection strategy based on the initial collection strategy; generating optimized prompt words based on the post-inspection collection strategy and the psychological portrait, and using the optimized prompt words to generate the collection strategy to be verified; specifically, the feature extraction and prompt word generation module obtains the basic information of the debtor (the debtor's arrears, overdue days, historical communication feedback, etc.) and historical context call records from the big data platform; the historical context records are sent to the psychological analysis big model to obtain the debtor's psychological feature portrait, and the initial prompt words are generated in combination with the debtor's basic information, and sent to the speech generation and optimization module; the speech generation and optimization module obtains the initial prompt words from the feature extraction and context understanding module The prompt words are sent to the collection strategy language model to generate an initial collection strategy; the initial collection strategy is sent to the feature extraction and prompt word generation module, the compliance inspection content is added, the prompt words for the compliance inspection are generated, the prompt words are sent to the compliance inspection model, and the collection strategy after the compliance inspection is generated, that is, the above-mentioned collection strategy after inspection; the collection strategy after inspection is sent to the feature extraction and prompt word generation module, the customer psychological content is added, and the prompt words optimized by psychological words are generated, and the prompt words are sent to the psychological model to generate the collection strategy to be verified; it can be understood that the feature extraction and prompt word generation module is mainly used for prompt word generation, and the prompt words required for each large model are generated according to the debtor information, including the arrears, overdue days, historical communication feedback, etc., and context call records; the speech generation and optimization module is committed to generating a collection strategy that is both efficient and compliant.

[0055] In this embodiment, the process of obtaining the target collection strategy corresponding to the target debtor based on the verification result may specifically include: if the verification result indicates that the collection strategy to be verified has passed the compliance rule verification, then the collection strategy to be verified is determined as the target collection strategy corresponding to the target debtor; if the verification result indicates that the collection strategy to be verified has not passed the compliance rule verification, then jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor to generate a new target collection strategy; the above process is completed by the above-mentioned compliance check module, which is based on fixed compliance rules and keyword detection to ensure that the collection strategy and communication process comply with laws, regulations, industry specifications and other requirements; in a specific implementation method, if the collection strategy to be verified If the collection strategy fails to pass the compliance rule verification, the feature extraction and context understanding module will be used to obtain the initial prompt words, which will be sent to the collection strategy large language model to generate a new initial collection strategy. The new initial collection strategy will be sent to the feature extraction and prompt word generation module, the compliance check content will be added, the compliance check prompt words will be generated, the prompt words will be sent to the compliance check large model to generate a new collection strategy after the compliance check; finally, the new collection strategy after the compliance check will be sent to the feature extraction and prompt word generation module, the customer psychological content will be added, the prompt words optimized for psychological rhetoric will be generated, the prompt words will be sent to the psychological large model to generate a new collection strategy to be verified, and the rhetoric will be verified again until the collection strategy to be verified passes the compliance rule verification. Through in-depth analysis and understanding of debtor information, personalized collection strategies can be generated, which not only improves the targeted nature of collection, but also increases the effectiveness of communication with debtors; by adjusting collection strategies in real time based on information such as the debtor's psychological state, repayment willingness, and historical communication feedback, collection can be carried out in a smarter and more personalized manner; by verifying the collection strategy, it is ensured that the collection strategy and communication process strictly comply with laws, regulations, industry norms and other requirements, which greatly reduces the compliance risks in the collection process, protects the legitimate rights and interests of the debtor, and also enhances the reputation and image of the collection agency.

[0056] Step S13: Use the target collection strategy to collect debts from the target debtor, record the collection records generated during the collection process in real time, and use the collection records to obtain the current repayment willingness of the target debtor.

[0057] In this embodiment, the process of collecting debts from the target debtor using the generated target collection strategy is completed through the above-mentioned telephone collection execution module. The telephone collection execution module is a key link in the intelligent collection system that directly interacts with the debtor. It is responsible for communicating with the debtor, executing the collection strategy, recording all calls, capturing the debtor's voice feedback during the call, converting it into text, and storing it in a safe and reliable database. By obtaining the current repayment willingness of the target debtor in real time based on the collection records generated during the collection process, the collection strategy can be dynamically adjusted according to the repayment willingness of the target debtor, ensuring the flexibility of the collection process and improving the collection efficiency.

[0058] Step S14: determine whether the current repayment willingness of the target debtor meets the expected conditions, and end the collection process for the target debtor according to the corresponding first judgment result.

[0059] In this embodiment, the process of terminating the telephone collection of the target debtor according to the judgment result of the target debtor's current repayment willingness is as follows: Figure 3As shown, it can specifically include: if the target value corresponding to the current repayment willingness reaches the preset expected value, then the collection process for the target debtor is terminated; if the target value corresponding to the current repayment willingness does not reach the preset expected value, then jump to the step of using the initial prompt word to generate the initial collection strategy corresponding to the target debtor, and after obtaining the new repayment willingness of the target debtor next time, judge whether the willingness value corresponding to the repayment willingness of the target debtor increases, and end the collection process for the target debtor according to the corresponding second judgment result; specifically, send the collection record to the repayment willingness large model, and judge whether to terminate the dialogue according to the repayment willingness. When the repayment willingness does not increase, or decreases instead of increases, or reaches the expected value in multiple rounds of dialogue, The conversation is terminated in a timely manner; in a specific implementation manner, if after the target debtor is collected using the first set of target collection strategies, the target debtor's current willingness to repay has reached the expected value, there is no need to continue the collection, and accordingly, the collection process for the target debtor is terminated; in another specific implementation manner, if the target debtor's current willingness to repay does not reach the expected value after the first round of collection, then jump to the step of generating an initial collection strategy corresponding to the target debtor using the initial prompt word, generate a new round of target collection strategies, and use the newly generated target collection strategies to collect, after the collection is completed, obtain the target debtor's new willingness to repay, and terminate the collection in a timely manner based on the new willingness to repay. The above-mentioned process of timely terminating the collection based on the new repayment willingness may specifically include: if the willingness value corresponding to the repayment willingness increases, and the willingness value corresponding to the new repayment willingness does not reach the preset expected value, then jump to the step of using the initial prompt word to generate the initial collection strategy corresponding to the target debtor, so as to generate a new target collection strategy in real time to collect from the target debtor; if the willingness value corresponding to the repayment willingness does not increase, then the collection process for the target debtor is terminated; in a specific implementation, if the new repayment willingness of the target debtor reaches the expected value, the collection process is terminated; in a second specific implementation, if the new current repayment willingness of the target debtor increases compared with the previous repayment willingness, and the new current repayment willingness does not reach the expected value, then a new target collection strategy is generated to conduct a new round of collection on the target debtor; in a third specific implementation, if the new repayment willingness of the target debtor decreases compared with the previous repayment willingness, it indicates that the target debtor has no repayment willingness. In order to avoid further reduction of the repayment willingness of the target debtor, the collection of the target debtor is stopped, wherein the calculation formula of the debtor's repayment willingness is as follows,

[0060] ;

[0061] Where N is the total number of conversations, is the repayment willingness score of the ith dialogue, and S is the repayment willingness of the debtor. The calculation formula of the change value of the repayment willingness of the target debtor is as follows:

[0062] ;

[0063] in, is the change in the debtor's willingness to repay, is the debtor's willingness to repay this time, is the debtor’s previous willingness to repay.

[0064] A specific implementation plan of this embodiment is as follows: 1.1. Obtain the basic information of the debtor as follows: Name: Zhang San; Debt situation: Personal loan arrears of 8,000 yuan; Overdue days: 45 days; Historical communication feedback: During the first three collections, he said that he would raise funds as soon as possible, but he has not paid it back so far, and his attitude has gradually become cold. 1.2. Input the historical call records into the psychological analysis model to obtain the psychological analysis results of the debtor, such as: Psychological analysis: Zhang San is an emotional customer, who is very sensitive to the topic of relatives. Relatives should be avoided during the collection process. Initial prompt words: Zhang San, personal loan arrears of 8,000 yuan, overdue for 45 days, emotional customer. 2. Generate an opening statement through the collection strategy model: "Hello, Mr. Zhang San, I am a collection specialist of XX Loan Company. Regarding your current personal loan arrears of 8,000 yuan, it has been overdue for 45 days. We hope to find a solution with you. Is it convenient for you to talk now?" 3. Compliance check: Through the compliance check model, confirm that the collection strategy complies with relevant laws and regulations, does not mention relatives, and does not have inappropriate words or threats. Collection strategy after compliance: the same as the initial collection strategy, because it has met the requirements of compliance and emotional sensitivity. 4. Optimized prompts for psychological speech: emotional resonance, personal difficulties, and positive coping. Collection strategy after psychological optimization: "Mr. Zhang San, I am a collection specialist from XX Loan Company. Regarding your current personal loan arrears of 8,000 yuan, it has been overdue for 45 days. We are well aware that life sometimes encounters unforeseen difficulties, which may put everyone under financial pressure. We understand your current situation and are willing to actively face this problem with you. Please believe that through our joint efforts, we will be able to find a suitable solution." 5. Compliance verification: Confirm that the optimized collection strategy meets the compliance requirements and does not contain any abusive language; Result: Compliance verification passed. 6. Telephone collection execution: The collection specialist uses the optimized collection strategy to communicate with Li Hua by phone, paying attention to maintaining emotional resonance and avoiding mentioning relatives; Real-time recording: Record the conversation content during the collection process, Li Hua's feedback, etc. 7. Collection records: send the collection records to the repayment willingness model; Repayment willingness evaluation: the model evaluates that Zhang San’s repayment willingness has increased, but it has not reached the expected value, and repeats steps 2 to 6; after multiple rounds of dialogue, if the debtor’s repayment willingness remains unchanged or decreases instead of increasing, the call is ended. Through the training and optimization of the large language model, the effectiveness and compliance of the collection strategy can be continuously improved. At the same time, this solution can also record the collection process in real time, and dynamically adjust and optimize it according to the debtor’s feedback and repayment willingness, thereby forming a closed-loop collection process. This efficient and continuously optimized feature continuously improves the collection effect, while also reducing the collection cost and time cost.

[0065] It can be seen that the embodiment of the present application can dynamically adjust the collection strategy according to the target debtor's repayment willingness by obtaining the current repayment willingness of the target debtor in real time based on the collection records generated during the collection process, thereby ensuring the flexibility of the collection process and improving the collection efficiency; by verifying the collection strategy, it is ensured that the collection strategy and communication process strictly comply with laws, regulations, industry norms and other requirements, thereby greatly reducing the compliance risks in the collection process, protecting the legitimate rights and interests of the debtor, and also enhancing the reputation and image of the collection agency.

[0066] Based on the above embodiments, this application describes how to generate a personalized collection strategy based on the characteristic information of the debtor and use the collection strategy to dynamically perform the overall process of collection. Next, this application will explain the process of training a large language model. Figure 4 As shown, the embodiment of the present application discloses a large language model training method, including:

[0067] Step S21: Acquire different types of initial training data, and add corresponding labels to the initial training data to obtain target training data.

[0068] In this embodiment, the process of training the initial large language model is as follows: Figure 5 As shown, different types of training need to be performed on the initial large language model so that the trained target large language model can be divided into a psychological analysis large model, a compliance detection large model, a repayment willingness large model, and a collection strategy large model, where each large model has different functions; accordingly, the above four large models need to use different training data to construct different training sets, and use the corresponding training sets to train the initial large language model to obtain the corresponding large model.

[0069] Step S22: construct different training sets according to the type of the target training data and the target training data, and use the different training sets to perform different types of training on the initial large language model to obtain the target large language model.

[0070] In this embodiment, in order to obtain a large psychological analysis model, it is first necessary to perform unsupervised pre-training on the initial large language model, and then perform supervised fine-tuning on the initial large language model; the unsupervised pre-training process includes: selecting an open source large language model containing rich psychological knowledge as the basis. Input a large amount of psychology literature, psychological counseling conversation records, etc., so that the model has preliminary psychological analysis capabilities; the supervised fine-tuning process includes: on the basis of pre-training, obtain the labeled psychological portrait data in the historical call records from the big data platform to fine-tune the model. These data include the debtor's speech patterns, emotional reactions, etc., and corresponding psychological portrait labels (such as anxiety, avoidance, cooperation, etc.). Through supervised learning, the model can more accurately judge the debtor's psychological state; the negative log-likelihood function formula used in the above pre-training process is as follows,

[0071] ;

[0072] Where N is the total number of training samples, For the target word, Indicates the context before the word. is the corresponding negative log-likelihood loss function. The formula of the cross entropy loss function used in the above fine-tuning process is as follows:

[0073] ;

[0074] Where C is the number of categories, is the actual label (one-hot encoding), is the probability of this class predicted by the model, is the corresponding cross entropy loss function.

[0075] Similarly, the unsupervised pre-training process of the compliance detection model includes: collecting comprehensive compliance requirements, laws and regulations, industry norms and other text materials, and inputting them into the large language model for unsupervised learning. By learning these texts, the model gradually establishes the semantic space of compliance knowledge; the supervised fine-tuning process includes: further fine-tuning the model using labeled compliant and non-compliant collection dialogue samples. The samples include various compliant and non-compliant speech, as well as corresponding compliance labels. Through supervised learning, the model can accurately identify and judge whether the collection strategy meets the compliance requirements; the data preparation process of the repayment willingness model includes: extracting the debtor's repayment behavior data from the historical collection dialogue records, including the repayment status after collection, repayment amount, repayment time, etc., as well as the corresponding repayment willingness labels (such as high, medium, and low). The model training process includes: using these data to conduct supervised training on the large language model, so that it can predict the debtor's repayment willingness based on his speech and behavior characteristics. By learning the debtor's repayment pattern and communication method, the model can more accurately assess his willingness to repay; the process of unsupervised pre-training of the large model of collection strategy includes: inputting text data such as collection manuals, internal training materials, and excellent collection cases to perform unsupervised pre-training on the large language model. These data cover various aspects of collection strategies, techniques, and speech, so that the model has preliminary collection expertise; the process of supervised fine-tuning includes: on the basis of pre-training, obtaining labeled collection dialogue data from the big data platform to fine-tune the model. These data include successful collection dialogues and failed collection dialogues, as well as corresponding collection effect labels. Through supervised learning, the model can learn how to generate effective collection strategies based on debtor characteristics and historical interactions.

[0076] It can be seen that this embodiment constructs a training set by using different training data, and uses different training sets to train the preset large language models respectively, so that the trained target large language model has a variety of different functions, thereby ensuring the reliability of each process in the telephone collection process.

[0077] See also Figure 6 As shown, the embodiment of the present invention discloses a telephone collection device based on a large language model, comprising:

[0078] The prompt word generation module 11 is used to obtain target feature information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, obtain the psychological portrait of the target debtor according to the target big language model and the target historical call records, and generate initial prompt words based on the target feature information;

[0079] A collection strategy acquisition module 12 is used to generate a collection strategy to be verified based on the initial prompt word and the psychological portrait, and to perform compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result;

[0080] The repayment willingness acquisition module 13 is used to collect debts from the target debtor using the target collection strategy, record the collection records generated during the collection process in real time, and obtain the current repayment willingness of the target debtor using the collection records;

[0081] The collection termination module 14 is used to determine whether the current repayment willingness of the target debtor meets the expected conditions, and terminate the collection process for the target debtor according to the corresponding first judgment result.

[0082] It can be seen that this application obtains the psychological portrait of the debtor and uses the psychological portrait to generate the collection strategy to be verified, thereby ensuring the pertinence of the generated collection strategy and thus improving the collection effect; by verifying the collection strategy to be verified, the compliance of the obtained collection strategy is ensured, thereby ensuring the compliance of the collection process; by real-time analysis of the debtor's repayment willingness and dynamically changing the collection process according to the repayment willingness, the effect of telephone collection is further improved.

[0083] In some specific embodiments, the prompt word generating module 11 further includes:

[0084] A data cleaning unit, used to obtain historical debt collection dialogue data and initial feature information corresponding to any debtor, and perform data cleaning on the historical debt collection dialogue records and the initial feature information to obtain cleaned historical debt collection records and cleaned feature information;

[0085] a format conversion unit, configured to annotate the historical debt collection dialogue records after data cleaning to obtain annotated historical debt collection dialogue records, and to perform data format conversion on the annotated historical debt collection dialogue records to obtain historical call records;

[0086] A data storage unit is used to construct the corpus using the historical call records and the cleaned feature information, and store the corpus in the preset big data platform.

[0087] In some specific embodiments, the prompt word generating module 11 further includes:

[0088] A training data acquisition unit, used to acquire different types of initial training data and add corresponding labels to the initial training data to obtain target training data;

[0089] A model training unit is used to construct different training sets according to the type of the target training data and the target training data, and use the different training sets to perform different types of training on the initial large language model to obtain the target large language model.

[0090] In some specific embodiments, the collection strategy acquisition module 12 specifically includes:

[0091] An initial speech generating unit, used to generate an initial collection strategy corresponding to the target debtor using the initial prompt words, and generate a post-inspection collection strategy based on the initial collection strategy;

[0092] A collection strategy generating unit is used to generate optimized prompt words according to the post-check collection strategy and the psychological portrait, and use the optimized prompt words to generate the collection strategy to be verified.

[0093] In some specific embodiments, the collection strategy acquisition module 12 specifically includes:

[0094] A collection strategy determination unit, configured to determine the collection strategy to be verified as the target collection strategy corresponding to the target debtor if the verification result indicates that the collection strategy to be verified passes the compliance rule verification;

[0095] The first step is a jump unit, which is used to jump to the step of generating an initial collection strategy corresponding to the target debtor using the initial prompt word if the verification result indicates that the collection strategy to be verified has not passed the compliance rule verification, so as to generate a new target collection strategy.

[0096] In some specific embodiments, the collection end module 14 specifically includes:

[0097] A first collection termination unit, configured to terminate the collection process for the target debtor if the target value corresponding to the current repayment willingness reaches a preset expected value;

[0098] The collection termination submodule is used to jump to the step of generating an initial collection strategy corresponding to the target debtor using the initial prompt word if the target value corresponding to the current repayment willingness does not reach the preset expected value, and judge whether the willingness value corresponding to the repayment willingness of the target debtor increases after obtaining the new repayment willingness of the target debtor next time, and terminate the collection process for the target debtor according to the corresponding second judgment result.

[0099] In some specific embodiments, the collection end submodule specifically includes:

[0100] The second step is a jump unit, which is used to jump to the step of generating an initial collection strategy corresponding to the target debtor using the initial prompt word if the willingness value corresponding to the repayment willingness increases and the willingness value corresponding to the new repayment willingness does not reach the preset expected value, so as to generate a new target collection strategy in real time to collect debts from the target debtor;

[0101] The second collection ending unit is used to end the collection process for the target debtor if the willingness value corresponding to the repayment willingness has not increased.

[0102] Furthermore, the present application also discloses an electronic device. Figure 7 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.

[0103] Figure 7 A schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the telephone collection method based on a large language model disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0104] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0105] In addition, the memory 22 as a carrier for storing resources may be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon may include an operating system 221, a computer program 222, etc., and the storage method may be temporary storage or permanent storage.

[0106] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program that can be used to complete the telephone collection method based on the large language model performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs that can be used to complete other specific tasks.

[0107] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the telephone collection method based on a large language model disclosed above is implemented. For the specific steps of the method, reference may be made to the corresponding contents disclosed in the aforementioned embodiments, and no further description will be given here.

[0108] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0109] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0110] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0111] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0112] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A telephone debt collection method based on a large language model, characterized in that: include: Obtain target feature information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, obtain a psychological portrait of the target debtor based on the target big language model and the target historical call records, and generate an initial prompt word based on the target feature information; Generate a collection strategy to be verified based on the initial prompt word and the psychological portrait, and perform compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result; Using the target collection strategy to collect debts from the target debtor, recording the collection records generated during the collection process in real time, and using the collection records to obtain the current repayment willingness of the target debtor; Determine whether the current repayment willingness of the target debtor meets the expected conditions, and end the collection process for the target debtor according to the corresponding first judgment result.

2. The telephone debt collection method based on a large language model according to claim 1 is characterized in that: Before obtaining the target characteristic information and target historical call records corresponding to the target debtor from the corpus in the preset big data platform, the method further includes: Obtaining historical debt collection dialogue data and initial feature information corresponding to any debtor, and performing data cleaning on the historical debt collection dialogue records and the initial feature information to obtain cleaned historical debt collection records and cleaned feature information; Annotating the historical debt collection conversation records after data cleaning to obtain annotated historical debt collection conversation records, and performing data format conversion on the annotated historical debt collection conversation records to obtain historical call records; The corpus is constructed using the historical call records and the cleaned feature information, and the corpus is stored in the preset big data platform.

3. The telephone debt collection method based on a large language model according to claim 1 is characterized in that: The generating of a collection strategy to be verified based on the initial prompt word and the psychological portrait includes: Generating an initial collection strategy corresponding to the target debtor using the initial prompt word, and generating a post-inspection collection strategy based on the initial collection strategy; An optimized prompt word is generated according to the post-inspection collection strategy and the psychological portrait, and the collection strategy to be verified is generated using the optimized prompt word.

4. The telephone debt collection method based on a large language model according to claim 3 is characterized in that: The obtaining of the target collection strategy corresponding to the target debtor according to the corresponding verification result includes: If the verification result indicates that the collection strategy to be verified has passed the compliance rule verification, the collection strategy to be verified is determined as the target collection strategy corresponding to the target debtor; If the verification result indicates that the collection strategy to be verified has not passed the compliance rule verification, then jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor to generate a new target collection strategy.

5. The telephone debt collection method based on a large language model according to claim 3 is characterized in that: The step of ending the collection process on the target debtor according to the corresponding first judgment result includes: If the target value corresponding to the current repayment willingness reaches the preset expected value, the collection process for the target debtor is terminated; If the target value corresponding to the current repayment willingness does not reach the preset expected value, jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor, and after obtaining the new repayment willingness of the target debtor next time, judge whether the willingness value corresponding to the repayment willingness of the target debtor increases, and end the collection process on the target debtor according to the corresponding second judgment result.

6. The telephone debt collection method based on a large language model according to claim 5 is characterized in that: The step of ending the collection process on the target debtor according to the corresponding second judgment result includes: If the willingness value corresponding to the repayment willingness increases, and the willingness value corresponding to the new repayment willingness does not reach the preset expected value, jump to the step of using the initial prompt word to generate an initial collection strategy corresponding to the target debtor, so as to generate a new target collection strategy in real time to collect debts from the target debtor; If the willingness value corresponding to the repayment willingness does not increase, the collection process for the target debtor is terminated.

7. The telephone debt collection method based on a large language model according to any one of claims 1 to 6, characterized in that: Before obtaining the psychological portrait of the target debtor according to the target large language model and the target historical call records, the method further includes: Acquire different types of initial training data, and add corresponding labels to the initial training data to obtain target training data; Different training sets are constructed according to the type of the target training data and the target training data, and different types of training are performed on the initial large language model using the different training sets to obtain the target large language model.

8. A telephone debt collection device based on a large language model, characterized in that: include: A prompt word generation module is used to obtain target feature information and target historical call records corresponding to a target debtor from a corpus in a preset big data platform, obtain a psychological portrait of the target debtor according to a target big language model and the target historical call records, and generate an initial prompt word based on the target feature information; A collection strategy acquisition module, used to generate a collection strategy to be verified based on the initial prompt word and the psychological portrait, and to perform compliance rule verification on the collection strategy to be verified, so as to obtain a target collection strategy corresponding to the target debtor according to the corresponding verification result; A repayment willingness acquisition module is used to collect debts from the target debtor using the target collection strategy, record the collection records generated during the collection process in real time, and obtain the current repayment willingness of the target debtor using the collection records; The collection termination module is used to determine whether the current repayment willingness of the target debtor meets the expected conditions, and terminate the collection process for the target debtor according to the corresponding first judgment result.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the telephone debt collection method based on a large language model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Used to store a computer program, which, when executed by a processor, implements the telephone debt collection method based on a large language model as described in any one of claims 1 to 7.

Images