An Online Mediation Scheme Generation Method and System

Abstract

The present invention relates to the technical fields of artificial intelligence and natural language processing, and particularly relates to a method and system for generating an online mediation plan. By constructing guiding texts for multiple dialogue rounds to communicate with customers, the customers are guided to input identity information and status information. This application constructs a background knowledge base and invokes a large language model to handle complex dialogue processes. If no useful information is extracted, after the dialogue ends, a general mediation plan is directly output to the customer. If identity information is extracted, the corresponding mediation plan is determined using the identity information and a pre-constructed prediction model and output to the customer. If both identity information and status information are extracted, a more accurate personalized mediation plan is output to the customer using the identity information, status information, and the pre-constructed prediction model. This application adopts the method of online mediation, which has high efficiency. At the same time, a corresponding mediation plan is formulated based on the characteristics of the customer, and the accuracy is relatively high.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and natural language processing, and specifically to an online mediation plan generation method and system. Background Art

[0002] Before filing a lawsuit against some overdue loans and borrowings, banks often entrust relevant units to mediate between the creditor and the debtor.

[0003] Existing mediations generally adopt the method of offline mediation. The mediation process highly depends on the experience of mediators and is prone to intensify the contradictions between the two parties. In addition, the method of generating mediation plans through offline mediation takes up a lot of time of both parties and has low mediation efficiency. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide an online mediation plan generation method and system to solve the problems of low efficiency and poor accuracy in offline mediation in the prior art.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] An online mediation plan generation method of the present invention includes the steps of:

[0007] When receiving a dialogue request from a customer, based on a pre-constructed background knowledge base and invoking a large language model to execute multiple rounds of dialogue with the customer;

[0008] When the dialogue text in multiple rounds of dialogue does not contain any target information, sending a pre-constructed general mediation plan to the customer, where the target information is the identity information or status information of the customer;

[0009] When there is identity information in the dialogue text of multiple rounds of dialogue but no status information, determining the mediation plan with the highest acceptance probability based on the identity information and a pre-constructed prediction model, where the prediction model includes the acceptance probabilities of multiple mediation factors corresponding to multiple identity parameters in multiple states;

[0010] When there is identity information and status information in the dialogue text of multiple rounds of dialogue, determining the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model.

[0011] In an embodiment of the present application, based on a pre-constructed background knowledge base and invoking a large language model to execute multiple rounds of dialogue with the customer includes:

[0012] S1, determining the current dialogue round and sending the pre-constructed guiding text to the customer based on the current dialogue round;

[0013] S2. When receiving a response text from the customer, extract information from the response text; when no response text from the customer is received beyond the target duration, take the next conversation turn as the current conversation turn and return to step S1 until all conversation turns are completed;

[0014] S3. When the target information of the current conversation turn exists in the response text, take the next conversation turn as the current conversation turn and return to step S1 until all conversation turns are completed, where the target information is the customer identity or status feature;

[0015] S4. When the target information of the current conversation turn does not exist in the response text and there is a question entity in the response text, extract the keywords in the response text, match the keywords with the pre-constructed background knowledge base. When there is structured data in the background knowledge base that matches the keywords, return the text in the structured data to the customer; when there is no structured data in the background knowledge base that matches the keywords, forward the response text to the invoked large language model for processing and return to step S2.

[0016] In an embodiment of the present application, when there is no question entity in the response text, resend the pre-constructed guiding text to the customer based on the current conversation turn and return to step S2.

[0017] In an embodiment of the present application, extracting information from the response text includes:

[0018] Segment the response text to obtain multiple words;

[0019] Invoke an identification model to perform entity identification on the multiple words. When an entity matching the current conversation turn is included in the multiple words, it is determined that the target information of the current conversation turn exists in the response text; otherwise, it is determined that the target information of the current conversation turn does not exist in the response text, where the identification model is obtained by training an artificial neural network with identity words or status words marked as training data.

[0020] In an embodiment of the present application, the method for constructing the background knowledge base includes:

[0021] Obtain background knowledge materials, where the background knowledge materials include bank policy documents, loan product information, and relevant legal provisions;

[0022] Segment the background knowledge materials to obtain multiple paragraphs;

[0023] Call a language large model to extract keywords and summarize paragraphs for the multiple paragraphs, and vectorize the keywords, paragraph summaries, texts, and corresponding titles of the multiple paragraphs to obtain structured data;

[0024] Construct a background knowledge base based on the structured processing.

[0025] In an embodiment of the present application, the construction method of the prediction model includes:

[0026] Obtain the questionnaire data of the target population, where the questionnaire data includes the current status, identity information of the interviewee, and the selection of multiple mediation factors, and the mediation factors include repayment methods, repayment periods, and repayment installments;

[0027] Based on the current status and identity information, divide the questionnaire data multiple times to obtain multiple data units, where each data unit includes the selection of multiple mediation factors corresponding to one value or value range of a single identity parameter under one current status;

[0028] Calculate the probability of the selection of each mediation factor in each data unit where k represents the current status serial number, i represents the mediation factor serial number, j=n represents the value n or value range n of the identity parameter j, is the number of selections of the mediation factor i for the group with the value n or value range n of the identity parameter j under the current status k, D is the total number of selections for the group with the value n or value range n of the identity parameter j under the current status k, and the current status includes good, medium, and poor;

[0029] Based on the probability of the selection of each mediation factor in multiple data units Construct a prediction model.

[0030] In an embodiment of the present application, determining the mediation plan with the highest acceptance probability based on the identity information and the pre-constructed prediction model includes:

[0031] Substitute the value of each identity parameter in the identity information into the prediction model to obtain the initial mediation plan for the value of each identity parameter in multiple states, where the initial mediation plan includes the probability of each selection of multiple adjustment factors;

[0032] Merge the initial mediation plans corresponding to multiple identity parameters in multiple states to obtain the intermediate mediation plan corresponding to each identity parameter, where the intermediate mediation plan includes the total probability of each selection of multiple adjustment factors;

[0033] Take the selection with the highest total probability as the predicted selection of the corresponding adjustment factor, and construct a mediation plan based on the predicted selections of multiple adjustment factors.

[0034] In one embodiment of the present application, determining the mediation solution with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model includes:

[0035] Determine the current status in the status information, and substitute the value of each identity parameter in the identity information in the current status into the prediction model to obtain an initial mediation solution corresponding to each identity parameter in the current status, where the initial mediation solution includes the probability of each choice of multiple adjustment factors;

[0036] Merge multiple initial mediation solutions to obtain an intermediate mediation solution, where the intermediate mediation solution includes the total probability of each choice of multiple adjustment factors;

[0037] Take the choice with the highest total probability as the predicted choice for the corresponding adjustment factor, and construct a mediation solution based on the predicted choices of multiple adjustment factors.

[0038] In one embodiment of the present application, it further includes:

[0039] When there is identity information in the dialogue text of multiple rounds of conversations, extract the customer's overdue information based on the identity information, and generate a repayment plan based on the mediation solution and the overdue information;

[0040] Return the repayment plan to the customer.

[0041] The present application also provides an online mediation solution generation system, including:

[0042] A dialogue module, configured to, when receiving a dialogue request from a customer, based on a pre-constructed background knowledge base, and call a large language model to execute multiple rounds of conversations with the customer;

[0043] A first generation module, configured to send a pre-constructed general mediation solution to the customer when there is no target information in the dialogue text of multiple rounds of conversations, where the target information is the customer's identity information or the customer's status information;

[0044] A second generation module, configured to determine the mediation solution with the highest acceptance probability based on the identity information and a pre-constructed prediction model when there is identity information but no status information in the dialogue text of multiple rounds of conversations, where the prediction model includes the acceptance probability of multiple adjustment factors corresponding to multiple identity parameters in multiple states;

[0045] A third generation module, configured to determine the mediation solution with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model when there is both identity information and status information in the dialogue text of multiple rounds of conversations.

[0046] The beneficial effects of the present invention are as follows: An online mediation plan generation method and system of the present invention perform mediation through online conversations. During the online conversation process, by constructing the guiding text for each conversation turn to communicate with the customer, the customer is guided to input identity information and status information. Considering the complexity of the conversation, the present application constructs a background knowledge base and invokes a large language model to handle the complex conversation process. If no useful information is extracted during the conversation, a general mediation plan is directly output to the customer after the conversation ends. If the identity information is extracted, the corresponding mediation plan is determined using the identity information and a pre-constructed prediction model and output to the customer. If the identity information and status information are extracted, a more accurate personalized mediation plan is output to the customer using the identity information, status information, and a pre-constructed prediction model. The present application adopts the method of online mediation, which has high efficiency. At the same time, corresponding mediation plans are formulated based on the characteristics of the customer, and the accuracy is relatively high. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described below in conjunction with the drawings and embodiments:

[0048] Figure 1 FIG. is a diagram of an application scenario of an online mediation plan generation method shown in an embodiment of the present application;

[0049] Figure 2 FIG. is a flowchart of an online mediation plan generation method shown in an embodiment of the present application;

[0050] Figure 3 FIG. is a structural diagram of an online mediation plan generation system shown in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0052] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the layers related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape, and size of the layers in actual implementation. The type, quantity, and ratio of each layer in actual implementation can be arbitrarily changed, and the layer layout type may also be more complex.

[0053] In the following description, a large number of details are explored to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details.

[0054] Figure 1 It is an application scenario diagram of a method for generating an online mediation solution shown in an embodiment of the present application. As Figure 1 shown, in the present application, the client 110 can conduct online mediation by accessing the server 120. Among them, the server 120 calls the language large model by accessing the NLP (Natural Language Processing) server 130. At the same time, a background database is set inside the server 120 to handle the professional questions of customers through the structured data in the background database and call the language large model to handle other questions of customers. In the present application, the server 120 extracts useful information from multi-round conversations and inputs it into a pre-constructed prediction model to predict a mediation solution with a relatively high probability of customer acceptance.

[0055] Figure 2 It is a flowchart of a method for generating an online mediation solution shown in an embodiment of the present application. As Figure 1 shown: A method for generating an online mediation solution in this embodiment may include steps S110 to S140:

[0056] S110, when receiving a conversation request from a customer, based on a pre-constructed background knowledge base and calling the language large model to execute multi-round conversations with the customer;

[0057] The use of the present application can be to mediate between banks and users with overdue loans, or it can also be used for other mediations;

[0058] Taking the mediation between banks and users with overdue loans as an example, after receiving a notice from the bank, the user can access the online mediation platform by logging in to the relevant server and then conduct multi-round conversations.

[0059] The multi-round conversations in the present application are requested by the customer and responded to by the server, thus officially entering the conversation process.

[0060] Considering the complexity of the conversation, the present application calls the language large model service by linking the API interface to handle the complex and changeable conversation process. In addition, taking the mediation between banks and users with overdue loans as an example, the customer may ask many questions about bank policies, loan product information, and relevant legal provisions. Therefore, the present application also constructs a background knowledge base with bank policies, loan product information, and relevant legal provisions as background materials to fill in the background knowledge.

[0061] Specifically, with the server as the execution end, the process of multi-round conversation is as follows:

[0062] S1. Determine the current conversation turn, and send the pre-constructed guiding text to the customer based on the current conversation turn;

[0063] In this application, for each round of conversation, a corresponding guiding text is preset.

[0064] For example, the guiding text for the first round of conversation is mainly for a simple introduction and to let the customer input personal information. The text example is:

[0065] "Hello, I am the mediation robot XX. Please enter your personal information and press the X key to end."

[0066] The above guiding text can be in text form or in voice form. If the multi-round conversation is executed in voice form, then a voice / text conversion service also needs to be called for conversion. For example, for online business, websocket polls to access the voice, triggering the real-time speech-to-text (ASR) service based on the funAsr model.

[0067] S2. When receiving the reply text from the customer, extract information from the reply text; when the reply text from the customer is not received beyond the target duration, take the next conversation turn as the current conversation turn and return to step S1 until all rounds of conversation are completed;

[0068] The customer's reply text may directly provide corresponding information according to the guiding text, or may input irrelevant information. Therefore, information extraction is also required to determine whether there is the desired text information.

[0069] For example, for the reply text in the first round of conversation, it is necessary to extract whether there is identity information. The identity information can be a name or an ID number, etc. Since the identity information input by the customer is usually incomplete, it is also necessary to rely on the input identity information to perform a retrieval in the information database provided by the bank to obtain the complete personal information.

[0070] Specifically, generally speaking, in order to protect the privacy of information, the bank will encrypt the customer information. In this application, after obtaining partial identity information of the customer, it is matched in the information database. When any customer is hit, the secret key of the customer is requested correspondingly. When the secret key feedback from the bank system is received, decryption is performed to obtain the complete customer information.

[0071] In the second-round conversation, the guiding text usually guides the customer to provide status information to obtain the customer's recent financial situation. Generally speaking, the probability of overdue for customers in normal situations is relatively low. Therefore, obtaining the customer's financial situation has certain reference value for formulating corresponding mediation plans subsequently.

[0072] Therefore, the reply text for the second-round conversation needs to determine whether there is status information.

[0073] In an embodiment of the present application, information extraction is performed on the reply text, including:

[0074] The reply text is segmented to obtain multiple words. The segmentation algorithm in this embodiment can adopt a string-based matching method or a statistical-based segmentation method (such as the N-gram model, Hidden Markov Model (HMM), Conditional Random Field (CRF), etc.);

[0075] An identification model is called to perform entity recognition on the multiple words. When an entity matching the current conversation round exists in the multiple words, it is determined that there is target information for the current conversation round in the reply text; otherwise, it is determined that there is no target information for the current conversation round in the reply text, where the identification model is obtained by training an artificial neural network with identity words or status words marked as training data.

[0076] Specifically, the present application is based on deep learning models: such as Long Short-Term Memory Network (LSTM), BiLSTM+CRF, etc. to identify identity information and status information in the reply text. For example, if the reply text contains a name, ID number, and there is a description of the current status in the reply text, such as "unemployed", "divorced", "business failure", etc., it can be accurately identified. Finally, the customer's status is classified into three grades: good, medium, and poor. Different grades can correspond to different mediation plan strategies.

[0077] In some cases, some customers are unwilling to disclose their own status information. In the present application, the emotional state of the reply text can also be used to analyze the customer's status. For example, a pre-constructed emotional word library is used to match the multiple words, where the emotional word library contains multiple emotional tendency words, and the emotional weight of each emotional tendency word is pre-marked. For example, the weight of a negative word is -1, the weight of a neutral word is 0, and the weight of a positive word is 1. The emotional tendency weight of each word in the reply text is determined using the emotional word library, and finally the average value is calculated. If the average value is less than -0.3, the status is determined to be poor; if the average value is greater than 0.3, the status is determined to be good; if the average value is between -0.3 and 0.3, it is determined to be medium.

[0078] S3. When there is target information of the current conversation turn in the reply text, take the next conversation turn as the current conversation turn and go back to step S1 until all turns of the conversation are completed, where the target information is the customer identity or status feature;

[0079] If there is corresponding target information in the reply text of the current conversation turn, enter the next round of conversation. For example, after obtaining the customer identity information, enter the conversation turn for obtaining the customer status information. At this time, go back to step S1 and send the guiding text of the turn for obtaining the customer status information to the customer.

[0080] S4. When there is no target information of the current conversation turn in the reply text and there is a question entity in the reply text, extract the keywords in the reply text, match the keywords with the pre-constructed background knowledge base. When there is structured data in the background knowledge base that matches the keywords, return the text in the structured data to the customer. When there is no structured data in the background knowledge base that matches the keywords, forward the reply text to the called large language model for processing and go back to step S2.

[0081] When there is no corresponding target information in the reply text, it is necessary to determine whether the customer is asking a rhetorical question. The customer may ask rhetorical questions about the loan product itself, the bank's repayment policy, and relevant legal knowledge. At this time, further determine whether there is a question entity in the reply text, such as "how", "why", "what", "?", etc. In this application, the existence of a question entity is determined by character matching. If it exists, it is confirmed that the customer is asking a rhetorical question at this time. When there is no question entity in the reply text, it means that there is neither target information nor a rhetorical question in the reply text at this time, and it is determined as an invalid text. At this time, based on the current conversation turn, send the pre-constructed guiding text to the customer again and go back to step S2 to prompt the customer.

[0082] When the customer asks a rhetorical question, this application needs to determine what the customer's reply text wants to ask. This application determines the key entity in the reply text by keyword extraction.

[0083] The extraction of the key entity can adopt the method of character matching or the method of training an identification model for extraction.

[0084] For example, the reply text is "How is the quota of XXX given?" Here, the keywords "XXX" and "quota" can be extracted. Therefore, corresponding knowledge points can be matched according to the keywords in the background knowledge base and provided to the customer.

[0085] Among them, the construction method of the background knowledge base includes:

[0086] (1) Obtain background knowledge materials, where the background knowledge materials include bank policy documents, loan product information, and relevant legal provisions; the background knowledge materials can be in formats such as docx, pdf, html, etc., and the above formats can quickly obtain text content.

[0087] (2) Segment the background knowledge materials to obtain multiple paragraphs;

[0088] When segmenting paragraphs, it can be based on paragraph marks. However, in this embodiment, for the coherence of paragraph content, machine learning is used for segmentation.

[0089] (3) Invoke a large language model to extract keywords and summarize paragraphs for the multiple paragraphs, and vectorize the keywords, paragraph summaries, texts, and corresponding titles of the multiple paragraphs to obtain structured data;

[0090] (4) Build a background knowledge base based on the structured processing. Specifically, paragraph summary embedding, title embedding, text embedding, keywords, summaries are stored in elasticsearch, the original file path of the text is stored in mysql, and the embedding type data is stored in milvus.

[0091] For example, the following paragraph:

[0092] Document: How is the quota of XXX given? The quota of XXX is automatically given by the system after comprehensive evaluation according to multi-dimensional evaluation criteria.

[0093] The title is XXX common questions;

[0094] Extract keywords, XXX, quota, evaluation;

[0095] Summary: The quota of Du Xiaoman is a multi-dimensional evaluation system, and the system automatically evaluates;

[0096] Then use the existing embedding model shaw / dmeta-embedding-zh to calculate the embedding vector values of keywords, titles, summaries, and texts, and store them in the corresponding databases in turn.

[0097] Taking the question "How is the quota of XXX given?" as an example, the keywords "XXX" and "quota" are extracted from the previous text. Through phrase matching, the text content corresponding to "Extract keywords, XXX, quota, evaluation" can be obtained, that is, "The quota of XXX is automatically given by the system after comprehensive evaluation according to multi-dimensional evaluation criteria."

[0098] The above problems are about background knowledge. If the question in the response text does not target background knowledge and no keywords can be extracted, it is transferred to the language large model for processing. The existing language large model is used to answer the customer's question, and then it returns to step S2 to extract relevant information from the customer's response again.

[0099] The above process is repeated, which can effectively handle the conversations of customers in various situations and extract effective information as much as possible.

[0100] S120, when the conversation text in the multi-round conversation does not contain any target information, send a pre-constructed general mediation plan to the customer, where the target information is the customer's identity information or the customer's status information;

[0101] If the customer never replies or always replies with invalid text, after the conversation ends, it is impossible to predict the plan through relevant feature information. Then directly output the pre-constructed general mediation plan, for example: 15,000 yuan is recommended to be paid in installments for 6 to 12 periods. The above installment plan is determined in advance with the bank. That is to say, if the customer accepts the plan, it means the mediation is successful.

[0102] S130, when there is identity information in the conversation text of the multi-round conversation but no status information, determine the mediation plan with the highest acceptance probability based on the identity information and the pre-constructed prediction model, where the prediction model includes the acceptance probabilities of multiple mediation factors corresponding to multiple identity parameters in multiple states;

[0103] S140, when there is identity information and status information in the conversation text of the multi-round conversation, determine the mediation plan with the highest acceptance probability based on the identity information, the status information, and the pre-constructed prediction model.

[0104] In step S130 and step S140, if identity information is extracted from the multi-round conversation, or both identity information and status information are extracted. Then the mediation plan with the highest probability can be determined according to the pre-constructed prediction model.

[0105] The prediction model in this application is constructed based on questionnaire data, and the mediation selection tendencies of multiple groups of people are expounded from a statistical perspective. For example, people with a relatively good current economic status generally tend to choose to accept the mediation plan more. The young group is more inclined to repay in installments. Compared with self-employed or individual business owners, office workers tend to choose more installments.

[0106] The big data in the questionnaire data is used to reflect the selection tendencies of multiple groups of people. Finally, the customer's multiple identity parameters are substituted into the model to obtain the mediation plan with the highest probability corresponding to each identity parameter. Finally, from multiple mediation plans, determine the choice with the highest probability for each.

[0107] In one embodiment of the present application, the method for constructing the prediction model includes:

[0108] (1) Obtain the questionnaire data of the target population, where the questionnaire data includes the current status, identity information of the respondents, and the selection of multiple mediation factors. The mediation factors include repayment methods, repayment periods, and the number of repayment installments.

[0109] Among them, questionnaires are conducted on the target population that has had overdue behavior or is in an overdue status. A large amount of questionnaire data is obtained. The questionnaire includes the current status, identity information, and the selection of multiple mediation factors. For example, the selection of repayment methods includes lump-sum repayment and installment repayment. When repaying in installments, the repayment period can be selected as half a year, one year, or more, and the number of repayment installments can be selected as 3 periods, 6 periods, 12 periods, or more.

[0110] (2) Based on the current status and identity information, divide the questionnaire data multiple times to obtain multiple data units, where each data unit includes the selection of multiple mediation factors corresponding to one value or value range of a single identity parameter under one current status.

[0111] For example, the current status is divided into good, medium, and poor. In the identity information, gender is divided into male and female, occupation is divided into working and self-employed, and age is divided into three age groups: 20 - 30, 30 - 50, and over 50, etc.

[0112] Based on the current status being divided into good, medium, and poor, the questionnaire data is divided into three data sets. Then, based on male and female, each data set is divided into 2 data units. Based on the three age groups of 20 - 30, 30 - 50, and over 50, each data set is divided into three data units, etc.

[0113] (3) Calculate the probability of the selection of each mediation factor in each data unit where k represents the serial number of the current status, i represents the serial number of the mediation factor, j = n represents the value n or value range n of the identity parameter j, is the number of selections of the mediation factor i for the group with the value n or value range n of the identity parameter j under the current status k, D is the total number of selections for the group with the value n or value range n of the identity parameter j under the current status k, and the current status includes good, medium, and poor;

[0114] Each data unit represents the selection preference of a group of people in one of the states. The present application uses the probability of the selection of each mediation factor to reflect the selection preference of the population.

[0115] (4) Based on the probability of the selection of each mediation factor in multiple data units Construct a prediction model.

[0116] Finally, use the selection biases of multiple populations to construct a prediction model.

[0117] Therefore, the process of predicting the mediation plan for customers based on the prediction model is as follows.

[0118] In an embodiment of the present application, determining the mediation plan with the highest acceptance probability based on the identity information and the pre-constructed prediction model includes:

[0119] (1) Substitute the value of each identity parameter in the identity information into the prediction model to obtain the initial mediation plan for each identity parameter value in multiple states, where the initial mediation plan includes the probability of each choice of multiple adjustment factors;

[0120] The initial mediation plan includes the selection probabilities of multiple adjustment factors in multiple states;

[0121] For example, for people in the age group of 20 - 30 years old, in the case of poor economic status, there is a 20% probability of choosing more than 12 installments, while in the case of good economic conditions, there is only a 5% probability of choosing more than 12 installments.

[0122] (2) Combine the initial mediation plans corresponding to multiple identity parameters in multiple states to obtain the intermediate mediation plan corresponding to each identity parameter, where the intermediate mediation plan includes the total probability of each choice of multiple adjustment factors;

[0123] The combined intermediate mediation plan sums the selection probabilities of multiple adjustment factors in three states and also sums the selection probabilities of multiple adjustment factors corresponding to each identity parameter. For example, customer A, 25 years old, male. Belonging to the population in the age group of 20 - 30 years old, the male ethnic group, the respective initial mediation plans include:

[0124] Population in the age group of 20 - 30 years old: Good: Factor 1: Option A: 20%, Option B: 80%;

[0125] Medium: Factor 1: Option A: 40%, Option B: 60%;

[0126] Poor: Factor 1: Option A: 60%, Option B: 40%;

[0127] Male group: Good: Factor 1: Option A: 35%, Option B: 65%;

[0128] Medium: Factor 1: Option A: 40%, Option B: 60%;

[0129] Poor: Factor 1: Option A: 45%, Option B: 55%;

[0130] After merging, in the intermediate mediation plan, for mediation factor one: Option A: 240%, Option B: 360%; other mediation factors are calculated according to the above method.

[0131] (3) Select the option with the highest total probability as the predicted selection for the corresponding adjustment factor, and construct a mediation plan based on the predicted selections of multiple adjustment factors.

[0132] Finally, select the option with the highest probability as the predicted selection and form the final mediation plan.

[0133] In an embodiment of the present application, determining the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model includes:

[0134] (1) Determine the current status in the status information, and substitute the value of each identity parameter in the identity information in the current status into the prediction model to obtain an initial mediation plan corresponding to each identity parameter in the current status, where the initial mediation plan includes the probability of each selection of multiple adjustment factors;

[0135] (2) Merge multiple initial mediation plans to obtain an intermediate mediation plan, where the intermediate mediation plan includes the total probability of each selection of multiple adjustment factors;

[0136] (3) Select the option with the highest total probability as the predicted selection for the corresponding adjustment factor, and construct a mediation plan based on the predicted selections of multiple adjustment factors.

[0137] Similar to the above process, but no longer merge the mediation plans in different states. For example:

[0138] For people aged 20 - 30: Good: For factor one, Option A: 20%, Option B: 80%;

[0139] For the male group: Good: For factor one, Option A: 35%, Option B: 65%;

[0140] In the final plan, for factor one, Option A: 55%, Option B: 145%; it is speculated that for mediation factor one, the customer will choose Option B.

[0141] Finally, when there is identity information in the dialogue text of multiple rounds of conversations, extract the customer's overdue information based on the identity information, generate a repayment plan based on the mediation plan and the overdue information; and return the repayment plan to the customer.

[0142] In addition, if the client hangs up or loses connection for any reason, the business large model interrupts the current program to save computing resources. Or if the business large model detects the identifier of the termination of the current conversation, after playing the last polite phrase, end the program.

[0143] An online mediation plan generation method of the present invention performs mediation through an online conversation. During the online conversation, by constructing the guiding text for each conversation turn to communicate with the customer, the customer is guided to input identity information and status information. Considering the complexity of the conversation, the present application constructs a background knowledge base and invokes a large language model to handle the complex conversation process. If no useful information is extracted during the conversation, after the conversation ends, a general mediation plan is directly output to the customer. If the identity information is extracted, the corresponding mediation plan is determined using the identity information and a pre-constructed prediction model and output to the customer. If the identity information and status information are extracted, a more accurate personalized mediation plan is output to the customer using the identity information, status information, and a pre-constructed prediction model. The present application adopts the method of online mediation, which has high efficiency. At the same time, a corresponding mediation plan is formulated based on the characteristics of the customer, and the accuracy is relatively high.

[0144] As Figure 3 shown, the present application also provides an online mediation plan generation system, including:

[0145] A conversation module, configured to, when receiving a conversation request from a customer, based on a pre-constructed background knowledge base and invoke a large language model to perform a multi-turn conversation with the customer;

[0146] A first generation module, configured to send a pre-constructed general mediation plan to the customer when the conversation text of the multi-turn conversation does not contain any target information, where the target information is the identity information of the customer or the status information of the customer;

[0147] A second generation module, configured to, when there is identity information but no status information in the conversation text of the multi-turn conversation, determine the mediation plan with the highest acceptance probability based on the identity information and a pre-constructed prediction model, where the prediction model includes the acceptance probabilities of multiple mediation factors corresponding to multiple identity parameters in multiple states;

[0148] A third generation module, configured to, when there is identity information and status information in the conversation text of the multi-turn conversation, determine the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model.

[0149] An online mediation solution generation system of the present invention performs mediation through online conversations. During the online conversation process, by constructing the guiding text for each conversation turn to communicate with the customer, guiding the customer to input identity information and status information. Considering the complexity of the conversation, this application responds to the complex conversation process by constructing a background knowledge base and invoking a large language model. If no useful information is extracted during the conversation, after the conversation ends, a general mediation solution is directly output to the customer. If identity information is extracted, the corresponding mediation solution is determined using the identity information and a pre-constructed prediction model and output to the customer. If both identity information and status information are extracted, a more accurate personalized mediation solution is output to the customer using the identity information, status information, and a pre-constructed prediction model. This application adopts the method of online mediation, which has high efficiency. At the same time, corresponding mediation solutions are formulated based on the characteristics of the customer, and the accuracy is relatively high.

[0150] This embodiment also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements any one of the methods in this embodiment, where the method is the execution logic of this system.

[0151] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0152] The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory so that the terminal executes any one of the methods in this embodiment.

[0153] For the computer-readable storage medium in this embodiment, those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to the computer program. The foregoing computer program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: various media such as ROM, RAM, magnetic disk, or optical disk that can store program codes.

[0154] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication with each other. The memory is used to store a computer program, the communication interface is used for communication, and the processor and the transceiver are used to run the computer program so that the electronic terminal executes each step of the above method.

[0155] In this embodiment, the memory may include a random access memory (Random Access Memory, abbreviated as RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0156] The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU for short), a Network Processor (NP for short), etc.; it may also be a Digital Signal Processor (DSP for short), an Application Specific Integrated Circuit (ASIC for short), a Field-Programmable Gate Array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0157] In the above embodiments, although the present invention has been described in conjunction with specific embodiments of the present invention, many substitutions, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art based on the foregoing description. The embodiments of the present invention are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims.

[0158] The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. An online mediation solution generation method, characterized in that, Including the steps: When receiving a dialogue request from a customer, based on a pre-constructed background knowledge base, and invoking a large language model to execute multi-round dialogues with the customer; When the dialogue text in the multi-round dialogue does not contain any target information, sending a pre-constructed general mediation plan to the customer, where the target information is the customer's identity information or the customer's status information; When there is identity information but no status information in the dialogue text of a multi-round dialogue, determine the mediation plan with the highest acceptance probability based on the identity information and a pre-constructed prediction model, where the prediction model includes the acceptance probabilities of multiple mediation factors corresponding to multiple identity parameters in multiple states; the construction method of the prediction model includes: obtaining the questionnaire data of the target population, where the questionnaire data includes the current status, identity information, and the selection of multiple mediation factors of the respondents, and the mediation factors include the repayment method, repayment period, and number of repayment installments; based on the current status and identity information, divide the questionnaire data multiple times to obtain multiple data units, where each data unit includes the selection of multiple mediation factors corresponding to a value or value range of a single identity parameter under a certain current status; calculate the probability of the selection of each mediation factor in each data unit , , where represents the current status serial number, represents the mediation factor serial number, represents the identity parameter value or value range , is the current status under which the identity parameter value or value range of the group's mediation factor selection quantity, is the current status under which the identity parameter value or value range of the group's total selection quantity, and the current status includes good, medium, and poor; based on the probability of the selection of each mediation factor in multiple data units construct a prediction model; When there is identity information and status information in the dialogue text of the multi-round dialogue, determining the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model.

2. The method for generating an online mediation solution according to claim 1, wherein, Based on a pre-constructed background knowledge base, and invoking a large language model to execute multi-round dialogues with the customer, including: S1, determining the current dialogue turn, and sending a pre-constructed guiding text to the customer based on the current dialogue turn; S2, when receiving a response text from the customer, extracting information from the response text; when not receiving a response text from the customer beyond the target duration, taking the next dialogue turn as the current dialogue turn, and returning to step S1 until all rounds of the dialogue are completed; S3, when the target information of the current dialogue turn exists in the response text, taking the next dialogue turn as the current dialogue turn, and returning to step S1 until all rounds of the dialogue are completed, where the target information is the customer identity or status feature; S4, when the target information of the current dialogue turn does not exist in the response text, and there is a question entity in the response text, extracting keywords from the response text, matching the keywords with the pre-constructed background knowledge base, when there is structured data in the background knowledge base that matches the keywords, returning the text in the structured data to the customer, when there is no structured data in the background knowledge base that matches the keywords, forwarding the response text to the invoked large language model for processing, and returning to step S2.

3. The method for generating an online mediation solution according to claim 2, wherein When there is no question entity in the response text, sending the pre-constructed guiding text to the customer again based on the current dialogue turn, and returning to step S2.

4. The method for generating an online mediation solution according to claim 2, characterized in that, Extracting information from the response text, including: Segmenting the response text to obtain multiple words; Invoking an identification model to perform entity identification on the multiple words, when there is an entity in the multiple words that matches the current dialogue turn, determining that the target information of the current dialogue turn exists in the response text; otherwise, determining that the target information of the current dialogue turn does not exist in the response text, where the identification model is obtained by training an artificial neural network with identity words or status words marked as training data.

5. The method for generating an online mediation solution according to claim 1, wherein The method for constructing the background knowledge base includes: Obtaining background knowledge materials, where the background knowledge materials include bank policy documents, loan product information, and relevant legal provisions; Segmenting the background knowledge materials to obtain multiple paragraphs; Invoking a large language model to extract keywords and summarize paragraphs from the multiple paragraphs, and performing vectorization processing on the keywords, paragraph summaries, texts, and corresponding titles of the multiple paragraphs to obtain structured data; Constructing a background knowledge base based on the structured processing.

6. The method for generating an online mediation solution according to claim 1, characterized in that, Determining the mediation plan with the highest acceptance probability based on the identity information and a pre-constructed prediction model, including: Substituting the value of each identity parameter in the identity information into the prediction model to obtain the initial mediation plan for each identity parameter value in multiple states, where the initial mediation plan includes the probability of each choice of multiple adjustment factors; Merging the initial mediation plans corresponding to multiple identity parameters in multiple states to obtain the intermediate mediation plan corresponding to each identity parameter, where the intermediate mediation plan includes the total probability of each choice of multiple adjustment factors; Taking the choice with the highest total probability as the predicted choice of the corresponding adjustment factor and constructing a mediation plan based on the predicted choices of multiple adjustment factors.

7. The method for generating an online mediation solution according to claim 1, wherein Determining the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model, including: Determining the current state in the status information and substituting the value of each identity parameter in the identity information in the current state into the prediction model to obtain the initial mediation plan corresponding to each identity parameter in the current state, where the initial mediation plan includes the probability of each choice of multiple adjustment factors; Merging multiple initial mediation plans to obtain an intermediate mediation plan, where the intermediate mediation plan includes the total probability of each choice of multiple adjustment factors; Taking the choice with the highest total probability as the predicted choice of the corresponding adjustment factor and constructing a mediation plan based on the predicted choices of multiple adjustment factors.

8. The method for generating an online mediation solution according to claim 1, characterized in that, It also includes: When there is identity information in the conversation text of multiple rounds of conversations, extracting the overdue information of the customer based on the identity information and generating a repayment plan based on the mediation plan and the overdue information; Returning the repayment plan to the customer.

9. An online mediation plan generation system, characterized in that, It includes: A dialogue module for, when receiving a dialogue request from a customer, based on a pre-constructed background knowledge base and invoking a large language model to perform multiple rounds of conversations with the customer; A first generation module for, when the conversation text of multiple rounds of conversations does not contain any target information, sending a pre-constructed general mediation plan to the customer, where the target information is the identity information of the customer or the status information of the customer; A second generation module, configured to, when there is identity information but no status information in the conversation text of a multi-round conversation, determine a mediation solution with the highest acceptance probability based on the identity information and a pre-constructed prediction model, where the prediction model includes the acceptance probabilities of multiple mediation factors corresponding to multiple identity parameters in multiple states; the method for constructing the prediction model includes: obtaining questionnaire data of a target population, where the questionnaire data includes the current status of the interviewee, identity information, and the selection of multiple mediation factors, and the mediation factors include repayment methods, repayment periods, and the number of repayment installments; based on the current status and identity information, dividing the questionnaire data multiple times to obtain multiple data units, where each data unit includes the selection of multiple mediation factors corresponding to a value or value range of a single identity parameter under a certain current state; calculating the probability of the selection of each mediation factor in each data unit , , where represents the current status serial number, represents the mediation factor serial number, represents the identity parameter value or value range , is the current status under which the identity parameter value or value range of the group's mediation factor selection quantity, is the current status under which the identity parameter value or value range of the group's total selection quantity, and the current status includes good, medium, and poor; based on the probability of the selection of each mediation factor in multiple data units construct a prediction model; A third generation module for, when there is identity information and status information in the conversation text of multiple rounds of conversations, determining the mediation plan with the highest acceptance probability based on the identity information, the status information, and a pre-constructed prediction model.

Images