Intelligent marketing system and method based on large model

Abstract

The invention discloses an intelligent marketing system and method based on a large model, and the method comprises the steps: automatically replying a newest message in a preset time through a self-adaptive chat large model, obtaining the emergency degree of a current customer session according to the recorded message number and message keywords, and obtaining the current customer session according to the emergency degree; the session with the highest emergency degree is screened out to send emergency early warning to the staff; a client database of the current emergency client is connected according to the emergency early warning, and the intention degree of the current emergency client is obtained by using an intention evaluation large model; matching corresponding activities in a marketing activity database according to the intention degree of the current emergency customer, and directionally recommending the corresponding activities to the current emergency customer; and according to the customer levels and the car insurance expiration time in the customer database, acquiring customer IDs of a plurality of customers to be subjected to insurance renewal, and sending an insurance renewal message. The invention relates to the technical field of intelligent insurance marketing, and solves the technical problems that in an existing marketing method, intentional customers are difficult to screen, and stock customer management is incomplete.

Description

Technical Field

[0001] The present invention belongs to the field of insurance marketing, and relates to data analysis and machine learning technology, and specifically is an intelligent marketing system and method based on a large model. Background Art

[0002] In the field of auto insurance marketing, the market presents the characteristics of a large number of customer inquiries, diversified demands and fierce competition. Due to the complexity of auto insurance products themselves, such as different insurance terms, claims procedures, preferential policies, etc., customers have a large number of consultation needs. At the same time, different customers have different focuses when purchasing auto insurance. Some focus on price discounts, some pay attention to claims services, and some have high expectations for additional services. In order to compete for market share, many insurance companies must strive to meet these diverse needs. However, at present, auto insurance marketing is mainly based on manual responses to customer inquiries, supplemented by simple data recording and analysis. Staff answer customer questions through telephone or online customer service, and insurance companies only handle customer data by storing basic information and transaction records, lacking in-depth data mining technology.

[0003] Therefore, traditional marketing methods have many disadvantages. First, in terms of customer consultation responses, manual responses are inefficient. When the number of customer inquiries is large, customers often have to wait for a long time to get a response, which seriously affects the customer experience and may even lead to customer loss. Moreover, it is difficult for staff to quickly and accurately identify important customers among many inquiries, such as customers with high purchasing potential or urgent needs, and they may miss the best marketing opportunities, resulting in potential business loss.

[0004] At the same time, in terms of customer intention identification, traditional marketing methods mainly rely on the subjective experience of marketers, which cannot accurately judge the degree of customer purchase intention and it is difficult to screen out customers with real purchase intention from a large number of customers, which leads to a waste of marketing resources. When marketing to existing customers, customers' insurance expiration dates are often checked manually one by one, which cannot be quickly located and effectively reached, resulting in the loss of existing customers. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an intelligent marketing system and method based on a large model, which are used to solve the technical problems of difficulty in screening potential customers and imperfect management of existing customers in existing marketing methods.

[0006] To achieve the above object, the first aspect of the present invention provides an intelligent marketing system based on a large model, comprising:

[0007] Automatic reply module: used to record the number of messages and message keywords of several customers, and automatically reply to several latest messages using the adaptive chat model when the preset time is reached; and,

[0008] The urgency of the current customer conversation is obtained according to the number of messages and message keywords, and the conversations with the highest urgency are screened out to send emergency warnings to the staff; wherein the adaptive chat model is built based on a deep learning algorithm;

[0009] Conversation analysis module: used to connect to the customer database of the current emergency customer according to the emergency warning, and use the intention evaluation model to obtain the intention degree of the current emergency customer according to the historical conversation records; wherein the intention evaluation model is constructed based on the machine learning algorithm;

[0010] Targeted marketing module: used to match the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommend the corresponding activities to the current urgent customer;

[0011] Customer maintenance module: used to obtain the customer IDs of several customers waiting for renewal based on the customer level and auto insurance expiration time in the customer database, and send renewal messages to several customers waiting for renewal.

[0012] Based on the above technical modules, it is possible to quickly reply to consulting customers and quickly locate urgent customers in intelligent marketing scenarios. Based on the historical conversation records with urgent customers, the big model technology is used to obtain their intention level, and then different marketing activities are promoted to customers with different intention levels. It helps staff focus on high-intent customers to increase the signing rate, and also uses the customer maintenance module to automatically keep in touch with customers awaiting renewal, thereby improving customer retention rate.

[0013] Furthermore, the use of the adaptive chat model to automatically reply to a number of latest messages includes:

[0014] A11, download the pre-trained open source language model as the initial model;

[0015] A12: Collect the historical conversation records of the current staff member, use the initial model to perform self-supervised learning on the historical conversation records, and obtain the adaptive chat model of the current staff member;

[0016] A13, extracting keywords of the latest news, calculating the matching degree between the keywords of the latest news and a number of annotated keywords in the speech database, and obtaining a number of matching degrees;

[0017] A14, comparing whether the matching degrees are all lower than a preset threshold; if yes, setting the urgency of the current conversation to the highest; if no, filtering out the reply sentences corresponding to the marked keywords with the highest matching degree, and obtaining several reply sentences;

[0018] A15, inputs a plurality of reply statements into the adaptive chat model of the current staff member, obtains a plurality of adjusted reply statements and automatically replies to the corresponding dialog box.

[0019] Using the pre-trained open source language model as the basis for further training avoids the huge cost and long cycle of building a model from scratch. At the same time, self-supervised learning is carried out on the historical conversation records of staff members, fully tapping the value of internal data, making the model more targeted and effective in dealing with current business problems.

[0020] Furthermore, obtaining the urgency of the current customer conversation according to the number of messages and message keywords includes:

[0021] A21, using a counter to accumulate the number of messages sent by the customer, obtains the number of messages, and sets an initial urgency level;

[0022] A22, extracting message keywords from the latest message, calculating the matching degree between the message keywords and the preset important words, and saving the message keywords with matching degrees higher than a threshold to the important information list;

[0023] A23, comparing whether the number of messages is equal to the preset number; if yes, the urgency of the current customer conversation is increased by one level and the process jumps to A24; if no, the process jumps to A24;

[0024] A24, comparing whether the number of elements in the important message list is equal to the preset number of important words; if yes, the urgency of the current customer conversation is increased by one level; if no, the current urgency is maintained.

[0025] This method realizes dynamic monitoring of the urgency of customer conversations by using a counter to accumulate the number of messages sent by customers and combining the matching degree between the keywords in the latest message and the preset important words to comprehensively evaluate the urgency of the current customer conversation. For example, when a customer sends messages multiple times or there are many important keywords in the message, the system can detect and increase the urgency of the conversation in time, ensuring that the staff can give priority to these customer needs that may be of higher importance.

[0026] Furthermore, the intention evaluation model is used to obtain the current degree of intention of the urgent customer according to the historical conversation records, including:

[0027] B1, extracting the transaction information of the current urgent customer from the customer database to obtain historical transaction data; wherein the historical transaction data includes transaction time, transaction amount, discount rate, and customer evaluation;

[0028] B2, according to the formula Calculate the current emergency customer's grade score; where Ai represents the transaction time impact factor, B i represents the transaction amount influencing factor, C i represents the preferential rate influencing factor, D i represents the customer evaluation influencing factor, and the customer evaluation value is 1-5, n represents the number of transactions in the customer database, T i Indicates the interval between the i-th transaction time and the current time, M i represents the transaction amount of the i-th transaction, R i represents the preferential rate of the i-th transaction, S i represents the customer evaluation score of the i-th transaction;

[0029] B3, counting the interaction index data of the current urgent customer in the historical conversation records to obtain customer interaction index data;

[0030] B4, input the customer interaction index data and customer level score into the intention assessment model to obtain the transaction probability, and mark the transaction probability as the intention level according to the preset degree division range to obtain the intention level of the current urgent customer; among them, the historical interaction indicators include: active question rate, average message reply speed, question level, and the intention level includes low intention, medium intention and high intention.

[0031] On the basis of calculating the customer grade score based on the comprehensive historical transaction data, we further count the customer's interaction indicator data in the historical conversation records, and input the two into the intention assessment model to obtain the transaction probability and intention degree. This method avoids the one-sidedness of evaluating customer intentions based on a single factor, and can more accurately grasp the customer's current purchase intention. For example, although a customer's historical transaction data shows that the grade is average, he has recently shown a high rate of active questioning and attention to high-level issues in the conversation. It is possible that his purchase intention is increasing. The intention assessment model can capture this signal in time and take corresponding measures.

[0032] Furthermore, the statistical analysis of the interaction index data of the current urgent customer in the historical session records includes:

[0033] B3-1, traverse the historical conversation records and obtain customer messages and staff messages through the sender ID of the message;

[0034] B3-2, calculate the ratio of the number of question messages to the total number of messages in customer messages to obtain the active questioning rate;

[0035] B3-3, extract the timestamps of the customer message and the staff message, calculate the time interval between each staff message and the corresponding customer message according to the timestamp, obtain several time intervals, and calculate the average of several time intervals to obtain the average message reply speed;

[0036] B3-4, extracting keywords of several question contents in customer messages through natural language processing technology, matching them with the annotated keywords in the speech database, and obtaining corresponding question levels;

[0037] B3-5, record the highest level among several problem levels to obtain the problem level;

[0038] B3-6, marking the active question rate, the average message reply speed, and the question level as interaction indicators, and obtaining the interaction indicator data of the current urgent customer in the historical conversation record.

[0039] By using natural language processing technology to extract the key words in customer messages and matching them with the annotated keywords in the speech database to obtain the question level, the depth and complexity of customer needs can be revealed. Questions of different levels reflect that customers are at different stages of purchase decision-making. For example, customers who focus on primary questions may still be in the initial stage of understanding the product, while customers who ask high-level questions may have already considered the purchase details in depth and may have a higher purchase intention.

[0040] Furthermore, the intention assessment model is constructed based on a machine learning algorithm, including:

[0041] C1, collect historical conversation records and customer databases of several customers, calculate the customer levels of several customers, and obtain the interaction index data of several customers by statistics, and obtain several independent variable data;

[0042] C2, binarize the transaction status of several customers to obtain several label data, where the transaction is 1 and the abandonment is 0;

[0043] C3, build an initial model based on the machine learning algorithm, input some independent variable data and some label data into the initial model for training and verification, and iteratively optimize the initial model according to the verification results;

[0044] C4, selects the initial model with the best verification result as the final pre-training model, and obtains an intention evaluation model whose input is customer level and customer interaction indicator data and output is customer transaction probability.

[0045] A second aspect of the present invention provides an intelligent marketing method based on a large model, comprising:

[0046] S1, records the number of messages and message keywords of several customers, and automatically replies to several latest messages using the adaptive chat model when the preset time is reached; and

[0047] Get the urgency of the current customer conversation based on the number of messages and message keywords, and filter out the most urgent conversations to send emergency warnings to staff;

[0048] S2, connects to the customer database of the current emergency customer according to the emergency warning, and uses the intention evaluation model to obtain the intention level of the current emergency customer according to the historical conversation records;

[0049] S3, matching the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommending the corresponding activities to the current urgent customer;

[0050] S4, according to the customer level and the expiration date of the car insurance in the customer database, obtain the customer IDs of several customers to be renewed, and send renewal messages to the several customers to be renewed.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] Through the automatic reply module, the adaptive chat model is used to automatically reply to the customer's latest message when the preset time arrives, solving the problem of delayed manual response when the number of customer consultations is large. At the same time, the module determines the urgency of the conversation based on the number of customer messages and keywords, and sends emergency warnings to staff, so that staff can quickly identify high-urgency customer needs, reasonably allocate work energy, and thus improve work efficiency;

[0053] Through the conversation analysis module, the intention evaluation model is combined with historical conversation records to determine the degree of customer intention. Unlike the traditional judgment based on marketers' subjective experience, this module comprehensively considers historical transaction data and interaction indicator data in historical conversation records to evaluate customer purchase intention, thus improving the accuracy of customer intention identification;

[0054] Through the targeted marketing module, according to the customer intention level evaluated by the big model, the corresponding activities are matched from the marketing activity database and recommended to customers in a targeted manner, avoiding the blind push of marketing activities and effectively mining and converting potential buyers among existing customers;

[0055] Through the customer maintenance module, based on the customer level and auto insurance expiration date in the customer database, customers waiting for renewal are automatically screened and renewal messages are sent, which reduces the omission of renewal customers, helps stabilize the existing customer base, and improves customer retention rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1A framework diagram of an intelligent marketing system based on a large model provided by the present invention;

[0058] Figure 2 A technical flow chart of automatic reply using an adaptive chat model provided by the present invention;

[0059] Figure 3 A flow chart of a method for calculating the urgency of a current conversation provided by the present invention;

[0060] Figure 4 A schematic diagram of a flow chart of a large model-based intelligent marketing method provided by the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. 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.

[0062] See also Figure 1 - Figure 3 The first aspect of the present invention provides an intelligent marketing system based on a large model, including:

[0063] Automatic reply module: used to record the number of messages and message keywords of several customers, and automatically reply to several latest messages using the adaptive chat model when the preset time is reached; and,

[0064] The urgency of the current customer conversation is obtained based on the number of messages and message keywords, and the conversations with the highest urgency are screened out to send emergency warnings to the staff, thereby obtaining urgent customers. The adaptive chat model is built based on a deep learning algorithm.

[0065] Conversation analysis module: used to connect to the customer database of the current emergency customer based on the emergency warning, and use the intention evaluation model to obtain the intention level of the current emergency customer based on the historical conversation records; the intention evaluation model is built based on the machine learning algorithm;

[0066] Targeted marketing module: used to match the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommend the corresponding activities to the current urgent customer;

[0067] Customer maintenance module: used to obtain the customer IDs of several customers waiting for renewal according to the customer level and car insurance expiration date in the customer database, and send renewal messages to several customers waiting for renewal;

[0068] Among them, the customer database in this system is in communication or electrical connection with the automatic reply module, the dialogue analysis module, and the targeted marketing module, and the customer database is in communication or electrical connection with the dialogue analysis module and the customer maintenance module.

[0069] In this embodiment, in order to allow customers to experience a more natural and humanized service experience when calling the automatic reply module, an adaptive chat model is constructed, which can adjust the responses in the dialogue database in a stylized manner according to the staff's language style, obtain an automatic response that is more in line with the staff's language habits, and reduce the alienation of the customer caused by the machine response. Specifically, the adaptive chat model in this embodiment is constructed based on a deep learning algorithm, and the following processing needs to be done before calling:

[0070] In this embodiment, in order to allow customers to experience a more natural and humanized service experience when calling the automatic reply module, an adaptive chat model is constructed, which can adjust the responses in the dialogue database in a stylized manner according to the staff's language style, obtain an automatic response that is more in line with the staff's language habits, and reduce the alienation of the customer caused by the machine response. Specifically, the adaptive chat model in this embodiment is constructed based on a deep learning algorithm, and the following processing needs to be done before calling:

[0071] First, select a suitable large language model from the open source platform for download. The selection criteria are based on the versatility and performance of the model in natural language processing tasks; for example, select models that have been pre-trained on large-scale text datasets and have good performance in semantic understanding, text generation, etc., taking advantage of their rich language knowledge and semantic relationships, which can provide a solid foundation for subsequent personalized training;

[0072] Next, the historical conversation records of the current staff are collected to ensure the integrity and accuracy of the data. These historical conversation records cover various scenarios in which the staff communicate with customers, including customer consultation, problem solving, business recommendations, etc. The collected historical conversation records are then input into the initial model for self-supervised learning;

[0073] In the process of self-supervised learning, the model learns the staff's language style by predicting the next word in the conversation, reconstructing sentences, and other tasks. For example, the model can gradually master the common sentence structure and vocabulary choices of staff when explaining auto insurance terms. At the same time, the model can also learn the staff's response strategies in different business scenarios, such as soothing words and problem-solving steps when facing customer complaints. In this way, the model continuously adjusts its own parameters and gradually forms an adaptive chat model that conforms to the current staff's language habits.

[0074] After obtaining the adaptive large model that meets the needs of the current staff, use natural language processing technology to extract keywords from the customer's latest news. For example, for the message in which the customer asks "What materials are needed for auto insurance claims?", keywords such as "auto insurance claims" and "material preparation" can be accurately extracted; then the extracted keywords are compared with the annotated keywords in the speech database. Among them, the annotated keywords in the speech database are sorted and classified, covering all aspects of the auto insurance business dialogue process. By calculating methods such as semantic similarity, the degree of match between the latest message keywords and several annotated keywords can be determined to obtain several degrees of match; for example, the word vector model can be used to calculate the distance between two keywords in the semantic space. The smaller the distance, the higher the degree of match;

[0075] Next, a reasonable threshold is preset to determine the match between the customer's latest message and the speech database. If the matching degree of all keywords is lower than this threshold, it means that the customer's question may be special or complicated. At this time, the urgency of the current session is set to the highest so that the staff can intervene and handle it in time; if there are keywords with a matching degree higher than the threshold, the reply statements with the highest matching degree with these keywords are filtered out from the speech database; for example, when the keywords asked by the customer have a high matching degree with the keywords about car insurance prices in the speech database, the reply statements related to car insurance prices will be filtered out;

[0076] Finally, the adaptive chat model is called to input the selected reply sentences into the trained model. The model adjusts the reply sentences according to the staff's language style that it has learned. For example, if the staff is accustomed to using more euphemistic expressions when replying to customers about price questions, the model will adjust the selected reply sentences to sentences that conform to this style; finally, the adjusted reply sentences are automatically sent to the dialog box corresponding to the customer, ensuring that the customer can receive a reply that conforms to the staff's language style in a timely manner, thereby improving the customer experience.

[0077] When the model can accurately identify customer questions and give appropriate responses, customers can get effective responses in a short period of time. Whether it is solving their doubts or providing directions for further consultation, customers can feel that the service is professional and efficient. This positive experience helps to enhance customers' favorability and trust in the company, thereby improving customer loyalty.

[0078] At the same time, during the automatic reply process, the number of messages sent by customers will be counted in real time, and the urgency of the current customer conversation will be judged based on the keywords extracted each time. The most urgent conversations will be screened out to send emergency warnings to the staff to remind them to respond to urgent customers in a timely manner.

[0079] Specifically, the system will first set up a special counter function. Whenever a new message from a customer comes in, the counter will automatically increase by 1 to record the total number of messages sent by the customer since the first message was sent.

[0080] Then, natural language processing technology is also used to extract keywords from the customer's latest message; and the extracted keywords are matched with the important words preset by the system. The preset important words are determined based on business priorities. For example, in the auto insurance marketing business, "buy", "want to buy", "price discount" and other purchase-related words are all important words; semantic analysis technology is used to calculate the semantic similarity between the keyword and the preset important words; if the similarity exceeds the preset threshold, for example, the matching degree is greater than 0.7, the keyword is added to the important information list. Therefore, the important information list is dynamically updated, and with the arrival of each new message from the customer, keyword extraction and matching operations will be performed again;

[0081] Then the system pre-sets a message quantity threshold, which is determined based on business experience and customer service strategies. For example, the preset number is 6. When the cumulative number of messages sent by a customer reaches 6, it indicates that the customer may have a high demand for car insurance purchases, and the system will automatically increase the urgency of the customer's conversation by one level.

[0082] At the same time, the system also sets a threshold for the number of important words, assuming it is 3; that is, when the number of keywords in the important information list is greater than 3, it means that the customer's intention to purchase auto insurance may be higher, and the system will increase the urgency of the customer's conversation by one level. If it does not exceed the threshold, the current urgency will remain unchanged and the subsequent messages will continue to be monitored.

[0083] When the number of customer inquiries is large, manual responses often lag behind, and it is difficult to quickly and accurately filter out urgent needs from numerous customer conversations. Therefore, through the automatic reply module, the system can efficiently and accurately filter out customer conversations that have a high probability of purchasing auto insurance products, and notify staff in a timely manner, which helps staff focus on customers who are most likely to make transactions and avoid missing important business opportunities. At the same time, this judgment method based on data and technology is more objective and comprehensive than manual judgment, and reduces the loss of potential customers due to human negligence or subjective misjudgment. In practical applications, it can greatly improve the efficiency and success rate of auto insurance marketing business.

[0084] After the current urgent customers are screened, the conversation analysis module will connect to the customer database of the current urgent customers based on the emergency warning, and use the intention assessment model to assess the intention of the current urgent customers, so as to further screen out high-intention customers;

[0085] Specifically, the transaction information of the current urgent customers will be extracted from the customer database, including transaction time, transaction amount, discount rate, customer evaluation, and then according to the formula Calculate the current emergency customer's grade score; where A i represents the transaction time impact factor, B i represents the transaction amount influencing factor, C i represents the preferential rate influencing factor, D i represents the customer evaluation influencing factor, and the customer evaluation value is 1-5, n represents the number of transactions in the customer database, T i Indicates the interval between the i-th transaction time and the current time, M i represents the transaction amount of the i-th transaction, R i represents the preferential rate of the i-th transaction, S i represents the customer evaluation score of the i-th transaction;

[0086] For example, suppose a customer has 3 transaction records:

[0087] First transaction: The transaction time is 3 years ago, that is, T1=3, the transaction amount is 800 yuan, that is, M1=800, the discount rate is 0.15, that is, R1=0.15, and the customer evaluation is 3 points, that is, S1=3;

[0088] The second transaction: The transaction time is 2 years ago, that is, T2 = 2, the transaction amount is 1,200 yuan, that is, M2 = 1,200, the discount rate is 0.1, that is, R2 = 0.1, and the customer evaluation is 4 points, that is, S2 = 4;

[0089] The third transaction: The transaction time is one year ago, that is, T3 = 1, the transaction amount is 900 yuan, that is, M3 = 900, the discount rate is 0.05, that is, R3 = 0.05, and the customer evaluation is 5 points, that is, S3 = 5;

[0090] Substituting the above formulas, we get: A1 = 0.05, B1 = 1.337, C1 = 0.7225, D1 = 0.5; A2 = 0.135, B2 = 1.418, C2 = 0.81, D2 = 0.75; A3 = 0.368, B3 = 1.36, C3 = 0.9025, D3 = 1; L = 0.66; then the customer registration score of the current customer is 0.66;

[0091] Next, collect statistics on the interaction indicators of the current urgent customers in the historical conversation records, including the proactive questioning rate, average message reply speed, and question level;

[0092] Among them, the active question rate is obtained by dividing the number of messages asked by customers by the total number of messages sent by customers; the average message reply speed is calculated by the interval between the last message sent by the staff and the first message sent by the customer in each group of conversations, and then the average of several groups of intervals is the average message reply speed, which is used to evaluate the customer's interest in the current topic, which can be used as a disguised indicator of customer intention judgment; the question level is obtained by matching the extracted question keywords with the annotated keywords in the speech database, that is, when the questions raised by customers remain at the primary level, such as asking about the basic types of auto insurance products, the approximate price range, etc., it usually indicates that the customer may be in the stage of preliminary understanding of the product, and the corresponding question level is low; when the questions raised by customers involve more in-depth and professional issues such as the specific process of auto insurance claims and the determination of liability in complex situations, these questions have a higher level, which often means that the customer has a deeper understanding of the auto insurance business and may have a higher purchase intention;

[0093] Then, the interaction indicator data and customer level score are input into the intention assessment model to obtain the transaction probability, and the transaction probability is divided into ranges according to the preset degree and marked as the intention degree to obtain the intention degree of the current urgent customer, such as low intention, medium intention or high intention.

[0094] In traditional marketing, judging customer purchase intention mainly relies on the subjective experience of marketers, which lacks precision. The conversation analysis module can use the intention evaluation model combined with historical conversation records to determine the customer's intention level, helping marketers to carry out key marketing activities for high-intention customers according to the intention level of different customers, reasonably allocate marketing resources, improve marketing efficiency and success rate, and avoid excessive marketing to low-intention customers that may cause customer disgust, thereby optimizing the entire auto insurance marketing process.

[0095] The core of the conversation analysis module is the intention assessment model, which is also built based on the machine learning algorithm. Specifically, first, the historical conversation records and customer database of several customers are collected from the enterprise customer service system, the customer level is calculated according to the above process, and the customer interaction index data is obtained by analyzing the historical conversation records. Then, these customer levels and interaction index data are used as independent variable data, and the transaction status of these customers is binarized, and the transaction is marked as 1 and the non-purchase is marked as 0, so as to obtain the label data. Next, an initial model is built based on a suitable machine learning algorithm, and the independent variable data and label data are input for training. After training, the data that did not participate in the training is used for verification, and according to the verification results, such as accuracy, recall rate, F1-score and other evaluation indicators, the model parameters are adjusted for iterative optimization. Finally, the initial model with the best verification result is selected as the final pre-training model, that is, the intention assessment model. The model takes the customer level and customer interaction index data as input, and outputs the customer transaction probability. In the present invention, the transaction probability is defined as the intention degree to assist the staff in targeted marketing to customers.

[0096] After obtaining the user's intention level, the targeted marketing module will be used to send targeted marketing activities to customers with different intention levels. First, after obtaining the intention level of the current urgent customer evaluated by the conversation analysis module, it will search and match in the marketing activity database, and then list the activities that match the customer, prompting the staff to carry out targeted marketing for him; the marketing activity database contains various types of marketing activities, such as preferential packages for new customers, renewal discounts for old customers, and value-added car insurance services for specific models;

[0097] Assuming that the conversation analysis module determines that customer A is a high-intent customer, and that he or she shows great concern for the speed of auto insurance claims settlement and service quality in communication with customer service, the targeted marketing module searches for a marketing campaign of "Quick Claims Service Upgrade Package" in the marketing campaign database and recommends this marketing campaign to customer A.

[0098] At the same time, the system also includes a customer maintenance module, which is used to maintain the relationship with customers waiting for renewal. The customer maintenance module focuses on two key factors: customer level and car insurance expiration time; among them, customer level reflects the customer's value and loyalty to the company, and car insurance expiration time helps to increase customer renewal rate; by screening the customer database, the system obtains the customer ID whose car insurance is about to expire, and automatically sends renewal messages to these customers waiting for renewal, reminding them to renew their car insurance.

[0099] For example, customer C has a high ranking in the customer database because he has purchased high-value auto insurance from the company for many consecutive years and has never had any claims disputes. When customer C's auto insurance is about to expire, for example, when it expires in 30 days, the customer maintenance module will obtain customer C's ID and send him a renewal reminder message. The customer maintenance module can help staff keep in touch with customers who are waiting for renewal in a targeted manner, enhance customers' favorability towards the company, increase customers' renewal rate, and thus consolidate the company's customer base.

[0100] See also Figure 4 The second aspect of the present invention provides an intelligent marketing method based on a large model, including:

[0101] S1, records the number of messages and message keywords of several customers, and automatically replies to several latest messages using the adaptive chat model when the preset time is reached; and

[0102] Get the urgency of the current customer conversation based on the number of messages and message keywords, and filter out the most urgent conversations to send emergency warnings to staff;

[0103] S2, connects to the customer database of the current emergency customer according to the emergency warning, and uses the intention evaluation model to obtain the intention level of the current emergency customer according to the historical conversation records;

[0104] S3, matching the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommending the corresponding activities to the current urgent customer;

[0105] S4, according to the customer level and the expiration date of the car insurance in the customer database, obtain the customer IDs of several customers to be renewed, and send renewal messages to the several customers to be renewed.

[0106] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0107] Working principle of the present invention:

[0108] The automatic reply module extracts keywords from the customer's latest message and matches them with the keywords in the speech database. The urgency of the conversation is set according to the degree of match, and the reply sentences are filtered. After adjustment by the adaptive chat model, the reply is automatically replied. At the same time, the number of customer messages is accumulated and the matching degree between the message keywords and the preset important words is calculated to evaluate the urgency. If the conditions are met, an emergency warning is sent to the staff; then, after receiving the emergency warning, the conversation analysis module connects to the customer database, obtains historical transaction data and calculates the customer level score, and then obtains customer interaction indicator data by analyzing historical conversation records. Both are input into the intention evaluation model built based on the machine learning algorithm to obtain the transaction probability and divide the intention degree; then, the targeted marketing module matches the corresponding activities from the marketing activity database according to the customer's intention degree and recommends them to the customer; finally, the customer retention module filters the customers to be renewed according to the customer level and the expiration time of the auto insurance in the customer database and sends renewal messages.

[0109] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An intelligent marketing system based on a large model, characterized in that: include: Automatic reply module: used to record the number of messages and message keywords of several customers, and automatically reply to several latest messages using the adaptive chat model when the preset time is reached; as well as, The urgency of the current customer conversation is obtained according to the number of messages and message keywords, and the conversations with the highest urgency are screened out to send emergency warnings to the staff; wherein the adaptive chat model is built based on a deep learning algorithm; Conversation analysis module: used to connect to the customer database of the current emergency customer according to the emergency warning, and use the intention evaluation model to obtain the intention degree of the current emergency customer according to the historical conversation records; wherein the intention evaluation model is constructed based on the machine learning algorithm; Targeted marketing module: used to match the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommend the corresponding activities to the current urgent customer; Customer maintenance module: used to obtain the customer IDs of several customers waiting for renewal based on the customer level and auto insurance expiration time in the customer database, and send renewal messages to several customers waiting for renewal.

2. According to claim 1, a large model-based intelligent marketing system is characterized in that: The adaptive chat model is used to automatically reply to several latest messages, including: A11, download the pre-trained open source language model as the initial model; A12: Collect the historical conversation records of the current staff member, use the initial model to perform self-supervised learning on the historical conversation records, and obtain the adaptive chat model of the current staff member; A13, extracting keywords of the latest news, calculating the matching degree between the keywords of the latest news and a number of annotated keywords in the speech database, and obtaining a number of matching degrees; A14, comparing whether the matching degrees are all lower than a preset threshold; if yes, setting the urgency of the current conversation to the highest; if no, filtering out the reply sentences corresponding to the marked keywords with the highest matching degree, and obtaining several reply sentences; A15, inputs a plurality of reply statements into the adaptive chat model of the current staff member, obtains a plurality of adjusted reply statements and automatically replies to the corresponding dialog box.

3. The intelligent marketing system based on a large model according to claim 1, characterized in that: The step of obtaining the urgency of the current customer conversation according to the number of messages and the message keywords includes: A21, using a counter to accumulate the number of messages sent by the customer, obtains the number of messages, and sets an initial urgency level; A22, extracting message keywords from the latest message, calculating the matching degree between the message keywords and the preset important words, and saving the message keywords with matching degrees higher than a threshold to the important information list; A23, comparing whether the number of messages is equal to the preset number; if yes, the urgency of the current customer conversation is increased by one level and the process jumps to A24; if no, the process jumps to A24; A24, comparing whether the number of elements in the important message list is equal to the preset number of important words; if yes, the urgency of the current customer conversation is increased by one level; if no, the current urgency is maintained.

4. The intelligent marketing system based on a large model according to claim 1, characterized in that: The intention evaluation model is used to obtain the current emergency customer's intention level according to historical conversation records, including: B1, extracting the transaction information of the current urgent customer from the customer database to obtain historical transaction data; wherein the historical transaction data includes transaction time, transaction amount, discount rate, and customer evaluation; B2, according to the formula Calculate the current emergency customer's grade score; where A i represents the transaction time impact factor, B i represents the transaction amount influencing factor, C i represents the preferential rate influencing factor, D i represents the customer evaluation influencing factor, and the customer evaluation value is 1-5, n represents the number of transactions in the customer database, T i Indicates the interval between the i-th transaction time and the current time, M i represents the transaction amount of the i-th transaction, R i represents the preferential rate of the i-th transaction, S i represents the customer evaluation score of the i-th transaction; B3, counting the interaction index data of the current urgent customer in the historical conversation records to obtain customer interaction index data; B4, input the customer interaction index data and customer level score into the intention assessment model to obtain the transaction probability, and mark the transaction probability as the intention level according to the preset degree division range to obtain the intention level of the current urgent customer; among them, the historical interaction indicators include: active question rate, average message reply speed, question level, and the intention level includes low intention, medium intention and high intention.

5. The intelligent marketing system based on a large model according to claim 4, characterized in that: The statistics of the interaction indicator data of the current urgent customer in the historical session records include: B3-1, traverse the historical conversation records and obtain customer messages and staff messages through the sender ID of the message; B3-2, calculate the ratio of the number of question messages to the total number of messages in customer messages to obtain the active questioning rate; B3-3, extract the timestamps of the customer message and the staff message, calculate the time interval between each staff message and the corresponding customer message according to the timestamp, obtain several time intervals, and calculate the average of several time intervals to obtain the average message reply speed; B3-4, extracting keywords of several question contents in customer messages through natural language processing technology, matching them with the annotated keywords in the speech database, and obtaining corresponding question levels; B3-5, record the highest level among several problem levels to obtain the problem level; B3-6, marking the active question rate, the average message reply speed, and the question level as interaction indicators, and obtaining the interaction indicator data of the current urgent customer in the historical conversation record.

6. The intelligent marketing system based on a large model according to claim 1, characterized in that: The intention assessment model is built based on a machine learning algorithm, including: C1, collect historical conversation records and customer databases of several customers, calculate the customer levels of several customers, and obtain the interaction index data of several customers by statistics, and obtain several independent variable data; C2, binarize the transaction status of several customers to obtain several label data, where the transaction is 1 and the abandonment is 0; C3, build an initial model based on the machine learning algorithm, input some independent variable data and some label data into the initial model for training and verification, and iteratively optimize the initial model according to the verification results; C4, selects the initial model with the best verification result as the final pre-training model, and obtains an intention evaluation model whose input is customer level and customer interaction indicator data and output is customer transaction probability.

7. An intelligent marketing method based on a large model, applied to an intelligent marketing system based on a large model as claimed in any one of claims 1 to 6, characterized in that: include: S1, records the number of messages and message keywords of several customers, and automatically replies to several latest messages using the adaptive chat model when the preset time is reached; as well as, Get the urgency of the current customer conversation based on the number of messages and message keywords, and filter out the most urgent conversations to send emergency warnings to staff; S2, connects to the customer database of the current emergency customer according to the emergency warning, and uses the intention evaluation model to obtain the intention level of the current emergency customer according to the historical conversation records; S3, matching the corresponding activities in the marketing activity database according to the current urgent customer's intention level, and recommending the corresponding activities to the current urgent customer; S4, according to the customer level and the expiration date of the car insurance in the customer database, obtain the customer IDs of several customers to be renewed, and send renewal messages to the several customers to be renewed.

Images