Marketing verbal skill generation method and device and electronic equipment

Abstract

The embodiment of the invention provides a marketing verbal skill generation method and device and electronic equipment, and relates to the technical field of financial science and technology. The marketing verbal skill generation method comprises the following steps: acquiring a customer intention label and a marketing stage label in a historical outbound marketing scene; using the fine-tuned large model to generate a marketing verbal skill according to the client intention label and the marketing stage label; and storing the generated marketing verbal skill in a marketing verbal skill library. According to the method, the outbound marketing efficiency can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial technology, and in particular to a marketing script generation method and device and electronic equipment. BACKGROUND

[0002] Outbound marketing, as one of the important ways of modern marketing activities, has a positive effect on improving sales efficiency, reducing costs, and improving customer satisfaction.

[0003] In related technologies, marketing script generation is usually based on rule matching or shallow machine learning models. For example, high-frequency marketing scripts are extracted from historical outbound marketing dialogue data, and a static marketing script library is constructed. During the call, the similarity between the current dialogue content and the marketing scripts in the static marketing script library is analyzed, and the marketing script with the highest similarity is recommended. SUMMARY

[0004] The present application provides a marketing script generation method, device and electronic equipment to improve the efficiency of outbound marketing.

[0005] In a first aspect, the present application provides a marketing script generation method, comprising: obtaining customer intent labels and marketing stage labels in a historical outbound marketing scenario; generating marketing scripts based on the customer intent labels and the marketing stage labels using a fine-tuned large model; and storing the generated marketing scripts in a marketing script library.

[0006] In one possible embodiment, generating marketing scripts based on the customer intent labels and the marketing stage labels using the fine-tuned large model comprises: obtaining a preset prompt template, the prompt template including a combined description of the customer intent labels and the marketing stage labels; and inputting the prompt template into the fine-tuned large model to generate marketing scripts that conform to the customer intent labels and the marketing stage labels.

[0007] In one possible embodiment, after storing the generated marketing scripts in the marketing script library, the marketing script generation method further comprises: during outbound marketing, performing intent recognition on customer dialogue text using the fine-tuned large model to generate customer intent labels for the customer dialogue text; performing marketing stage judgment on the customer dialogue text using the fine-tuned large model to generate marketing stage labels for the customer dialogue text; and matching marketing scripts with the same labels from the marketing script library based on the customer intent labels and the marketing stage labels of the customer dialogue text.

[0008] In a possible implementation, the intent recognition of the customer conversation text by the fine-tuned large model to generate the customer intent label of the customer conversation text comprises: inputting an intent recognition prompt word into the fine-tuned large model to classify the customer conversation text and output the customer intent label of the customer conversation text; and the marketing stage judgment of the customer conversation text by the fine-tuned large model to generate the marketing stage label of the customer conversation text comprises: inputting a marketing stage judgment prompt word into the fine-tuned large model to classify the customer conversation text and output the marketing stage label of the customer conversation text.

[0009] In a possible implementation, after the generated marketing script is stored in the marketing script library, the marketing script generation method further comprises: in the outbound marketing process: performing compliance quality inspection on the agent conversation text by using rule matching and / or a natural language model to generate a quality inspection result; and in a case where the quality inspection result indicates a violation type label of the agent conversation text, matching a remedial script with the same label from the marketing script library according to the violation type label of the agent conversation text.

[0010] In a possible implementation, the compliance quality inspection on the agent conversation text by using rule matching and / or a natural language model comprises at least one of the following: performing rule matching to detect whether the agent conversation text contains a keyword in a violation word library; and performing natural language model to calculate semantic similarity between the agent conversation text and the violation word library to determine whether there is an implicit violation.

[0011] In a possible implementation, before the fine-tuned large model is used to generate the marketing script according to the customer intent label and the marketing stage label, the marketing script generation method further comprises: performing parameter compression on the fine-tuned large model by using a model distillation technology to generate a lightweight model; and storing high-frequency marketing scripts by using a caching mechanism to reduce the number of calls of the fine-tuned large model.

[0012] In a possible implementation, after the generated marketing script is stored in the marketing script library, the marketing script generation method further comprises: collecting customer acceptance rate data of the outbound marketing scenario; and adjusting the strategy of the fine-tuned large model for generating the marketing script based on the customer acceptance rate data by using a reinforcement learning framework.

[0013] In a second aspect, the present application provides a marketing script generation device, comprising: an acquisition module configured to acquire customer intent labels and marketing stage labels in historical outbound marketing scenarios; a marketing script generation module configured to generate marketing scripts according to the customer intent labels and the marketing stage labels by using a fine-tuned large model; and a storage module configured to store the generated marketing scripts in a marketing script library.

[0014] Thirdly, this application provides an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method as described in any of the first aspects.

[0015] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0016] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method of any one of the first aspects.

[0017] In this embodiment, customer intent tags and marketing stage tags are extracted from historical outbound marketing data from historical outbound marketing scenarios. These tags serve as input conditions for generating marketing scripts using a fine-tuned large model. The fine-tuned large model generates dynamic marketing scripts tailored to customer needs based on the combination of customer intent tags and marketing stage tags. The marketing scripts generated by the fine-tuned large model can be stored in a marketing script library for easy retrieval during subsequent outbound marketing calls. The marketing script library is dynamically updated based on customer intent tags and marketing stage tags, and is more personalized.

[0018] In addition, by fine-tuning the large model to obtain a fine-tuned large model, the training cost of the large model can be reduced, enabling the fine-tuned large model to quickly adapt to the task of generating marketing scripts, thereby enhancing the flexibility and responsiveness of outbound marketing. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 This is a schematic diagram illustrating the application of the marketing script generation method of this application to an outbound marketing scenario of a financial institution;

[0021] Figure 2 This is a flowchart of a marketing script generation method according to an embodiment of this application;

[0022] Figure 3 This is a schematic diagram of a marketing script generation method according to another embodiment of this application;

[0023] Figure 4 This is a schematic diagram of an outbound marketing process for a marketing script generation method according to another embodiment of this application.

[0024] Figure 5A schematic diagram of a marketing language generation method according to another embodiment of the present application;

[0025] Figure 6 A structural schematic diagram of a marketing language generation device according to an embodiment of the present application;

[0026] Figure 7 A structural schematic diagram of an electronic device according to an embodiment of the present application;

[0027] The specific embodiments of the present application have been shown in the above-described drawings, and will be described in more detail hereinafter. These drawings and the written description are not intended to restrict the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0028] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to indicate the same or similar components. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0029] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards of relevant countries and regions, necessary security measures are taken, public order and good customs are not violated, and appropriate operation portals are provided for users to choose authorization or refusal.

[0030] And the present application involves big data analysis of user information (including but not limited to personal biological characteristics, identity data, consumption data, asset data, electronic terminal operation data, etc.), and uses artificial intelligence technology for automatic decision-making, and makes technical solutions based on automatic decision-making results that have a significant impact on personal rights and interests, provides appropriate operation portals for users to choose to agree or refuse automatic decision-making results; if the user chooses to refuse, the expert decision-making process is entered.

[0031] It should be noted that the marketing language generation method, device and electronic device provided by the present application can be applied to the field of financial technology, and can also be applied to any field other than financial technology. The application of the marketing language generation method, device and electronic device is not limited by the present application.

[0032] The marketing script generation method of the embodiments of the present application can be applied to the outbound marketing scenario of a financial institution. Outbound marketing is an important means for a financial institution to promote financial products or loan services, aiming to improve customer conversion rate and optimize service experience.

[0033] Figure 1 A schematic diagram of the marketing script generation method of the embodiments of the present application applied to the outbound marketing scenario of a financial institution. As shown in Figure 1 , the agent of the financial institution calls the terminal device 2 of the customer through the terminal device 1 with a calling function and communicates, and the communication content can be sent to the server 3 of the financial institution through the terminal device 1. The server 3 retrieves the marketing script with the highest similarity from the marketing script library based on the communication content and sends the marketing script to the terminal device 1 of the agent for assisting the agent to accurately market.

[0034] In related technology 1, the agent needs to communicate with the customer by relying on static, pre-set standardized marketing script library. This way at least one of the following problems exists:

[0035] 1. The content of the standardized marketing script library is updated behind schedule, and it is difficult to adapt to the rapidly changing market demand and customer individualized demand.

[0036] 2. The agent needs to manually select marketing scripts in the call, resulting in low response efficiency, and the communication strategy cannot be dynamically adjusted according to the real-time communication content of the dialogue.

[0037] 3. The identification of customer intent in the outbound marketing process is highly dependent on human experience, which is prone to misjudgment or omission, affecting the marketing effect.

[0038] 4. There may be compliance problems in outbound marketing, for example, the agent's illegal language may cause customer complaints and even legal risks.

[0039] The marketing script generation method, device and electronic equipment provided by the present application aim to at least solve one of the above technical problems of related technology.

[0040] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail in the specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the drawings.

[0041] Figure 2 Flowchart of the marketing script generation method of the embodiments of the present application. The marketing script generation method of the embodiments of the present application can be executed by a server, specifically including steps S110 to S130.

[0042] S110, acquire a customer intention label and a marketing stage label in a historical outbound marketing scenario.

[0043] The customer intention label refers to a classification identification of a potential demand or attitude of a customer. The customer intention label includes, for example, no need, customer hesitation, customer repulsion, and the like. For example, when a customer expresses “I will consider it again” in a dialogue, the corresponding label is “customer hesitation”.

[0044] The marketing stage label refers to a classification identification of different links in a marketing process. According to the order of the marketing stage, the marketing stage label includes, for example, opening contact, achieving conversion, objection handling, achieving conversion, and ending, and the like. For example, when a customer expresses a “product introduction” demand in a dialogue, the corresponding label is “product introduction”.

[0045] S120, generate a marketing script according to the customer intention label and the marketing stage label by using the fine-tuned large model.

[0046] The large model refers to a deep learning model with a large number of parameters, which can generate natural language content through context understanding. The large model includes, for example, one of the following: a 13B parameter model (a large language model with a parameter size of 130 billion), a generative pre-trained Transformer (GPT), and a pathways language model (PaLM). In the embodiments of the present application, the large model is taken as the 13B parameter model for example.

[0047] Fine-tuning refers to local adjustment of parameters of a pre-trained model to adapt to specific task requirements. For example, low-rank adaptation (LoRA) technology can be used for parameter compression.

[0048] Exemplarily, the server can periodically evaluate the performance of the fine-tuned large model by using a model evaluation rule to obtain an evaluation result. The evaluation result includes, for example, the accuracy and fluency of the marketing script generated by the fine-tuned large model. In the case where the evaluation result indicates that the performance of the fine-tuned large model does not meet the standard, the LoRA parameters and / or the hyperparameters can be adjusted according to the evaluation result. The LoRA parameters include rank (r), scaling factor (Alpha, a), and LoRA dropout. The hyperparameters include learning rate, batch size, and weight decay value.

[0049] Take adjusting the rank and scaling coefficient to improve the performance of the fine-tuned large model as an example. The rank determines the number of trainable model parameters and the expression ability of the fine-tuned large model. If the evaluation result indicates that the fine-tuned large model is under-fitting, for example, the accuracy of the generated marketing language is low, it may be that the rank is set too small, resulting in insufficient capacity of the fine-tuned large model to capture the task characteristics of generating marketing language. At this time, the rank can be appropriately increased. The scaling coefficient controls the scaling ratio of the low-rank matrix to the original weights of the fine-tuned large model. If the fine-tuned large model is under-fitting, the value of a can be increased to enhance the influence of the low-rank matrix on the original weights of the fine-tuned large model; if it is over-fitting, the value of a is reduced.

[0050] In the embodiments of the present application, the marketing language corresponding to the same customer intention in different marketing stages may be different. The fine-tuned large model is used in combination with the customer intention label and the marketing stage label to avoid customer demand understanding bias and facilitate accurate generation of marketing language. For example, the marketing stage label is opening contact, and the customer intention label is not needed. At this time, the customer demand should be understood as not needing outbound marketing. For another example, the marketing stage label is product introduction, and the customer intention label is not needed. At this time, the customer demand should be understood as not needing to recommend products. In the above two cases, the customer intention label is the same, and the marketing stage label is different, so different marketing language should be generated.

[0051] S130, store the generated marketing language to the marketing language library.

[0052] Take a specific example: the marketing language library includes the marketing language "Structural deposit has long-term yield advantage: …" for the customer intention label of customer hesitation and the marketing stage label of achieving conversion, which is used to enhance the customer's product purchase tendency.

[0053] In the embodiments of the present application, the customer intention label and the marketing stage label are extracted from the historical outbound marketing data in the historical outbound marketing scene as the input condition for the fine-tuned large model to generate marketing language. The fine-tuned large model generates dynamic marketing language that meets customer demand according to the label combination of the customer intention label and the marketing stage label. The marketing language generated by the fine-tuned large model can be stored in the marketing language library for subsequent calling in the outbound marketing process. The marketing language library can be dynamically updated based on the customer intention label and the marketing stage label, and is more personalized.

[0054] In addition, in the case of fine-tuning the large model to obtain the fine-tuned large model, the training cost of the large model can be reduced, so that the fine-tuned large model can quickly adapt to the task of generating marketing language, and enhance the flexibility and response speed of outbound marketing.

[0055] The above embodiments are used to construct a marketing language library including marketing language, Figure 3A schematic diagram of a method for generating marketing scripts for another embodiment of the present application is shown. Unlike the above embodiment, as shown in Figure 3 another embodiment of the present application can also generate remedial scripts and store them in the marketing script library during the process of constructing the marketing script library.

[0056] As shown in Figure 3 The server can obtain violation category labels in historical outbound marketing scenarios; a fine-tuned large model generates remedial scripts according to the violation category labels; and the generated remedial scripts are stored in the marketing script library, so that the marketing script library includes marketing scripts and remedial scripts.

[0057] For example, after generating marketing scripts or remedial scripts, the marketing scripts and remedial scripts can also be audited, and after passing the audit, they can be stored in the marketing script library. The script audit can be manual.

[0058] In the marketing script library, the marketing scripts have customer intent labels and marketing stage labels, and the remedial scripts have violation type labels. The violation type labels include exaggeration, word stealing, inappropriate language, etc.

[0059] In one possible embodiment, the step S120 of generating marketing scripts according to customer intent labels and marketing stage labels using a fine-tuned large model can include: obtaining a preset prompt template; and inputting the prompt template into the fine-tuned large model to generate marketing scripts that meet the customer intent labels and marketing stage labels.

[0060] The prompt template includes a combined description of the customer intent label and the marketing stage label. The prompt template refers to an instruction structure used to guide the fine-tuned large model to generate specific content, and usually contains a formatted description of the input information. For example, the prompt template can include: please generate marketing scripts for customer hesitation-to-achieve conversion.

[0061] The combined description refers to combining the customer intent label and the marketing stage label into a natural language instruction according to a preset logic. For example, the combined description includes: combining the customer intent label as customer hesitation and the marketing stage label as achieving conversion. The combined result is, for example, customer hesitation-achieve conversion.

[0062] In the embodiments of the present application, the server can extract the prompt word template matched with the customer intention label and the marketing stage label from the preset prompt word template library, and input the prompt word template into the fine-tuned large model, which can enhance the understanding ability of the fine-tuned large model for the task of generating marketing language. For example, for the combined description of customer hesitation-achieve conversion, the fine-tuned large model can be instructed by the prompt word template to output content for alleviating customer hesitation and promoting conversion (such as "the long-term benefits of structural deposits can meet your needs"), avoiding the fine-tuned large model from generating generalized or irrelevant content as the sentence of marketing language.

[0063] The fine-tuned large model generates marketing language that conforms to the context according to the context information (such as the scene of customer hesitation and the goal of achieving conversion) in the prompt word template. In the embodiments of the present application, the combined description of the customer intention label and the marketing stage label is used as the input of the fine-tuned large model through the prompt word template, which can improve the accuracy and quality of the fine-tuned large model for generating marketing language, and also improve the adaptability of the fine-tuned large model for generating marketing language to customer needs, reducing the burden of manual review of marketing language.

[0064] Figure 4 The out-bound marketing process schematic diagram of the marketing language generation method of another embodiment of the present application is shown.

[0065] As shown in Figure 4 After the step S130 of storing the generated marketing language into the marketing language library, the marketing language generation method can further include: in the out-bound marketing process: using the fine-tuned large model to perform intention recognition on the customer dialogue text to generate a customer intention label; using the fine-tuned large model to perform marketing stage judgment on the customer dialogue text to generate a marketing stage label of the customer dialogue text; and matching the marketing language with the same label from the marketing language library according to the customer intention label and the marketing stage label of the customer dialogue text.

[0066] For example, the voice stream of the dialogue between the agent and the customer can be collected after authorization, and the voice stream of the dialogue between the agent and the customer is converted into customer dialogue text and agent dialogue text through automatic speech recognition (ASR).

[0067] Intention recognition refers to extracting the classification label of the potential needs or attitudes of the customer from the customer dialogue text through natural language processing technology. For example, the customer dialogue text includes "I will consider it again", and through intention recognition, the customer intention label can be determined as customer hesitation.

[0068] The marketing stage judgment refers to extracting a link classification label of a current marketing process from the customer dialogue text by using a natural language processing technology. For example, the customer dialogue text includes "I am really happy to hear your interest in xxx", and through the marketing stage judgment, it can be determined that the marketing stage label is product introduction.

[0069] In the embodiments of the present application, before generating the marketing dialogue, the fine-tuned large model is used to perform intent recognition and marketing stage judgment on the customer dialogue text to automatically generate the customer intent label and the marketing stage label of the customer dialogue text. In addition, the semantic analysis capability of the fine-tuned large model can improve the accuracy of intent recognition and marketing stage judgment.

[0070] In a possible embodiment, the fine-tuned large model is used to perform intent recognition on the customer dialogue text to generate the customer intent label of the customer dialogue text, which includes inputting the intent recognition prompt word into the fine-tuned large model to classify the customer dialogue text and output the customer intent label of the customer dialogue text.

[0071] The fine-tuned large model is used to perform marketing stage judgment on the customer dialogue text to generate the marketing stage label of the customer dialogue text, which includes inputting the marketing stage judgment prompt word into the fine-tuned large model to classify the customer dialogue text and output the marketing stage label of the customer dialogue text.

[0072] The intent recognition prompt word refers to an instruction template for guiding the fine-tuned large model to extract the customer intent. For example, the intent recognition prompt word can include: please identify the customer intent from the input text, and output the customer intent label of the customer dialogue text: no need / customer hesitation / customer repulsion.

[0073] The marketing stage judgment prompt word refers to an instruction template for guiding the fine-tuned large model to extract the marketing stage. For example, the marketing stage judgment prompt word can include: please identify the marketing stage from the input text, and output the marketing stage label of the customer dialogue text: opening contact / achieving conversion / disagreement processing.

[0074] In the embodiments of the present application, through the double-prompt-word mechanism of the intent recognition prompt word and the marketing stage judgment prompt word, the fine-tuned large model can accurately obtain the relevant information of the corresponding task when processing the intent recognition task and the marketing stage judgment task, which helps to improve the accuracy of intent recognition and marketing stage judgment, and also facilitates to constrain the fine-tuned large model to output compliant customer intent labels and marketing stage labels.

[0075] In addition, in the case where the number of fine-tuned large models is multiple, the intent recognition prompt word and the marketing stage judgment prompt word can be parallel to improve the efficiency and response speed of generating the customer intent label and the marketing stage label.

[0076] Further, as the marketing business dynamically changes, new added intent recognition prompt words and marketing stage judgment prompt words can be added to support new marketing businesses and scenarios, and good scalability is achieved.

[0077] Exemplarily, the server can acquire a violation category label in a historical outbound marketing scenario, and generating a remediation script according to the violation category label by using the fine-tuned large model can include: inputting the remediation prompt word into the fine-tuned large model to generate a remediation script conforming to the violation type label.

[0078] The remediation prompt word template refers to an instruction template for guiding the fine-tuned large model to perform script remediation.

[0079] As Figure 4 After storing the generated marketing script in the marketing script library in step S130, the marketing script generation method can further include: in the outbound marketing process: using rule matching and / or a natural language model to perform compliance quality inspection on the agent conversation text to generate a quality inspection result; in the case where the quality inspection result indicates a violation type label of the agent conversation text, matching a remediation script with the same label from the marketing script library according to the violation type label of the agent conversation text.

[0080] Rule matching refers to directly detecting whether the agent conversation text contains explicit violation keywords through a pre-set violation word library. For example, rule matching is used to detect whether the agent conversation text contains keywords such as “immediate effect”.

[0081] In the embodiments of the present application, for example, a natural language model with smaller model specifications than the fine-tuned large model can be selected for compliance quality inspection to improve the speed of compliance quality inspection on the agent conversation text. For example, a deep bidirectional encoding model (Bidirectional Encoder Representations from Transformers, BERT) based on the Transformer architecture can be selected for compliance quality inspection.

[0082] The BERT model can be used for semantic feature extraction and similarity calculation. For example, the semantic similarity between the agent conversation text and the violation word library can be calculated by the BERT model, and then it is judged whether the agent conversation text is implicitly violated.

[0083] Exemplarily, before calculating the semantic similarity between the agent conversation text and the violation word library by the BERT model, for example, the agent conversation text can be subjected to vectorization processing to facilitate subsequent calculation of semantic similarity.

[0084] In the embodiments of the present application, the server can perform quality inspection on the agent conversation text through rule matching and / or a natural language model, thereby improving the coverage and accuracy of compliance quality inspection on the agent conversation text. For example, rule matching can be used to directly detect whether there are explicit keywords in the agent conversation text that violate rules, and a natural language model can be used to perform semantic similarity analysis to detect whether there are implicit violations in the agent conversation text. The combination of the two can cover different violation scenarios, while avoiding the problem of missed detection due to excessive reliance on a single compliance quality inspection, thereby improving the compliance of the agent outbound marketing.

[0085] For example, in the case where the server performs double quality inspection on the agent conversation text through rule matching and a natural language model, the priority of compliance quality inspection based on rule matching can be set to be higher than the priority of compliance quality inspection based on the natural language model.

[0086] Figure 5 A schematic diagram of a marketing script generation method according to another embodiment of the present application is shown in FIG. 3. As shown in FIG. 3, in one possible embodiment, the compliance quality inspection on the agent conversation text using rule matching and / or a natural language model includes at least one of the following: using rule matching to detect whether the agent conversation text contains keywords in a violation keyword library; and using a natural language model to calculate the semantic similarity between the agent conversation text and the violation keyword library to determine whether there is an implicit violation. Figure 5

[0087] For example, before detecting whether the agent conversation text contains keywords in the violation keyword library or calculating the semantic similarity between the agent conversation text and the violation keyword library, the agent conversation text can be preprocessed to remove non-text content and retain only the pure text for compliance quality inspection.

[0088] The violation keyword library refers to a set of keywords including prohibited content according to laws and regulations, restricted vocabulary according to industry standards, and inappropriate language, etc. The violation keyword library can be constructed and maintained by business personnel according to business scenarios.

[0089] The semantic similarity refers to the semantic correlation between the agent conversation text and the violation keyword library calculated by a vector space. For example, if the semantic similarity between the agent conversation text "Today, you may miss the opportunity" and the keyword "Miss today and there will be no discount" in the violation keyword library exceeds a threshold value, it can be determined that the agent conversation text has an implicit violation. For example, the threshold value can be 0.8 in the range of [0, 1].

[0090] ​Exemplarily, in a case where the seat conversation text contains a keyword in the violation keyword library, it is determined that the quality inspection result indicates that the violation type label of the seat conversation text is the same as the violation type label of the keyword. Alternatively, in a case where the semantic similarity between the seat conversation text and the violation keyword library exceeds a threshold, it is determined that the quality inspection result indicates that the violation type label of the seat conversation text is the same as the violation type label of the keyword whose semantic similarity exceeds the threshold. In a case where both of the following conditions are met: the seat conversation text does not contain a keyword in the violation keyword library, and the semantic similarity between the seat conversation text and the violation keyword library is less than a threshold, it is determined that the quality inspection result is that the seat conversation text is compliant.

[0091] Exemplarily, in the marketing script generation method of an embodiment of the present application, the keywords in the violation keyword library may, for example, include original keywords and enhanced keywords. The enhanced keywords may, for example, be obtained by data enhancement based on the original keywords or combination of the original keywords according to a preset rule. For example, the preset rule restricts at least one of the following: the order of the original keywords, the number of characters between the original keywords, and the logical relationship between the original keywords, which may, for example, include simultaneous occurrence or parallel relationship. The marketing script generation method of an embodiment of the present application can provide accurate and sufficient keyword references through the violation keyword library, facilitating accurate detection of the compliance of the marketing script.

[0092] In addition, the above-mentioned preset rule can be configured to be updated, for example, by a business personnel configuring and updating the preset rule to adjust the keywords in the violation keyword library, so as to provide flexible and dynamic keyword references that can adapt to various flexible and dynamic marketing scenarios.

[0093] In a possible embodiment, before the fine-tuned large model generates the marketing script according to the customer intention label and the marketing stage label in step S120, the marketing script generation method may further include: performing parameter compression on the fine-tuned large model by using a model distillation technology to generate a lightweight model; and storing high-frequency marketing scripts by using a caching mechanism to reduce the number of calls to the fine-tuned large model.

[0094] Model distillation refers to transferring the parameters and knowledge of the fine-tuned large model to a lightweight model by using a knowledge transfer technology to reduce the computational complexity. For example, the fine-tuned 13B large model can be distilled into a 3B lightweight model.

[0095] The caching mechanism refers to storing high-frequency marketing scripts locally to reduce the computational cost of repeatedly generating high-frequency marketing scripts. For example, the high-frequency marketing script of “customer hesitation - conversion” can be cached.

[0096] Exemplarily, to balance the parameter compression of the fine-tuned large model and the generation effect of the marketing language, in the embodiment of the present application, the layers (such as the bottom feature extraction layer) that have less influence on the generation of the marketing language in the fine-tuned large model can be compressed at a high proportion, and the layers (such as the voice generation layer) that have greater influence on the generation of the marketing language can be compressed at a low proportion, so as to avoid a significant decline in the generation effect of the marketing language of the large model. In addition, a dynamic compression threshold can also be set, and by verifying the real-time marketing language generation accuracy, conversion rate and other indicators, when the indicators decrease to the preset threshold, the compression rate for corresponding proportion parameter compression of the fine-tuned large model can be stopped increasing.

[0097] In the embodiment of the present application, the fine-tuned large model is compressed into a lightweight model (such as a 3B parameter model) through the model distillation technology, and the high-frequency marketing language is stored by using the caching mechanism, so as to avoid repeated calling of the fine-tuned large model and reduce the calculation load of the fine-tuned large model. For example, the lightweight model can be deployed on the edge device to reduce the cloud calling delay, and the caching mechanism directly returns the high-frequency marketing language to avoid the increase of the calculation cost caused by repeated generation of the high-frequency marketing language. Through the cooperative optimization of hardware (caching mechanism) and algorithm (model distillation technology), the embodiment of the present application improves the real-time response capability of the marketing language generation method in the high-concurrency scene and reduces the waiting time of the outbound marketing.

[0098] In a possible embodiment, after the step S140 of storing the generated marketing language into the marketing language library, the marketing language generation method can further include: collecting customer acceptance rate data of the outbound marketing scene; and based on the customer acceptance rate data, adjusting the strategy of the fine-tuned large model for generating the marketing language by using a reinforcement learning framework.

[0099] Exemplarily, the collection of the customer acceptance rate data of the outbound marketing scene and the adjustment of the strategy of the fine-tuned large model for generating the marketing language by using the reinforcement learning framework can be performed before the parameter compression of the fine-tuned large model by using the model distillation technology to generate the lightweight model and the storage of the high-frequency marketing language by using the caching mechanism in the above embodiment.

[0100] The customer acceptance rate data refers to a statistical index of the acceptance degree of customers to the marketing script in an outbound marketing scenario. The customer acceptance rate data can be obtained by the ratio of the total number of customer acceptances in a statistical period to the total number of effective calls. The total number of effective calls can be the total number of calls in which the agent and the customer complete effective communication. The total number of customer acceptances can be the number of calls in which the acceptance behavior is completed. The acceptance behavior can be one of the following: the customer expresses willingness to further understand or leaves a contact method for subsequent connection, the customer confirms the purchase intention or agrees to participate in use / experience, the customer completes the signing of the agreement, and the like. The specific behavior of the acceptance behavior can be selected by the business personnel according to the business needs. The statistical period can be, for example, a preset number of days, and the preset number of days can be, for example, in the range of 1 day to 30 days.

[0101] The reinforcement learning framework refers to a machine learning method for driving the iteration of the model parameters of the fine-tuned large model through a reward function. The fine-tuned large model serves as an agent and is used to generate a marketing script according to the customer intention label and the marketing stage label. The environment is the actual interaction scenario of outbound marketing, such as customer acceptance rate. The state is the input information before the fine-tuned large model generates the marketing script, such as the customer intention label and the marketing stage label. The reward is used to drive the optimization of the fine-tuned large model.

[0102] The reward function can be set as the sum of the basic reward and the cumulative reward. The basic reward: in a single outbound marketing, if the customer produces an acceptance behavior, the reward value is set to a positive value, if the customer does not produce an acceptance behavior, the reward value is set to a negative value, and the outbound failure is not counted as a reward. The cumulative reward: if the cumulative customer acceptance rate in a certain statistical period is higher than the historical average, additional rewards are added, and if it is lower than the historical average, penalties are added. The sum of the basic reward and the cumulative reward can be normalized to the interval [-1, 1].

[0103] Thus, the fine-tuned large model generates an initial marketing script according to the current state (customer intention label and marketing stage label). The server records the customer acceptance rate data after outbound marketing, and calculates the reward value of the initial marketing script generated this time based on the reward function. The reward value is used to evaluate the value of the initial marketing script generated by the fine-tuned large model according to the current state, and the model parameters of the fine-tuned large model are adjusted according to the value, so as to adjust the strategy of the fine-tuned large model for generating the marketing script.

[0104] In this embodiment, by collecting customer acceptance rate data during outbound marketing calls and adjusting the strategy for generating marketing scripts in the fine-tuned large model based on a reinforcement learning framework, the marketing scripts generated by the fine-tuned large model are made more aligned with customer needs. Furthermore, by using customer acceptance rate data to drive adjustments to the strategy for generating marketing scripts in the fine-tuned large model, a closed-loop optimization can be achieved from generating marketing scripts to whether customers accept the scripts, thereby improving marketing success rates and adapting to market changes and shifts in customer bias.

[0105] Figure 6 This is a schematic diagram of a marketing script generation device 200 provided in an embodiment of this application. Figure 6 As shown, the marketing script generation device 200 provided in this application embodiment includes: an acquisition module 210, a marketing script generation module 220, and a storage module 230.

[0106] The acquisition module is used to obtain customer intent tags and marketing stage tags from historical outbound marketing scenarios.

[0107] The marketing script generation module 220 is used to generate marketing scripts based on customer intent tags and marketing stage tags using a finely tuned large model.

[0108] Storage module 230 is used to store the generated marketing scripts to the marketing script library.

[0109] In one possible embodiment, the marketing script generation module 220 includes: a prompt word template acquisition submodule, used to acquire a preset prompt word template, the prompt word template including a combined description of customer intent tags and marketing stage tags; and a first input submodule, used to input the prompt word template into a fine-tuned large model to generate marketing scripts that conform to customer intent tags and marketing stage tags.

[0110] In one possible embodiment, the marketing script generation device further includes: an intent recognition module, a marketing stage determination module, and a first matching module. The intent recognition module is used during outbound marketing calls to: use a fine-tuned large model to perform intent recognition on the customer dialogue text to generate customer intent tags for the customer dialogue text; the marketing stage determination module is used during outbound marketing calls to: use a fine-tuned large model to determine the marketing stage of the customer dialogue text to generate marketing stage tags for the customer dialogue text. The first matching module is used to match marketing scripts with the same tags from the marketing script library based on the customer intent tags and marketing stage tags of the customer dialogue text.

[0111] In a possible embodiment, the intent recognition module is configured to input the intent recognition prompt word into the fine-tuned large model to classify the customer dialogue text and output a customer intent label of the customer dialogue text. The marketing stage judgment module is configured to input the marketing stage judgment prompt word into the fine-tuned large model to classify the customer dialogue text and output a marketing stage label of the customer dialogue text.

[0112] In a possible embodiment, the marketing script generation apparatus further includes a compliance quality inspection module and a second matching module. The compliance quality inspection module is configured to, in the outbound marketing process: perform compliance quality inspection on the agent dialogue text by using rule matching and / or a natural language model to generate a quality inspection result. The second matching module is configured to, in the outbound marketing process: in a case where the quality inspection result indicates a rule violation type label of the agent dialogue text, match a remedial script with the same label from the marketing script library according to the rule violation type label of the agent dialogue text.

[0113] In a possible embodiment, the compliance quality inspection module includes at least one of: a first quality inspection submodule configured to detect whether the agent dialogue text contains a keyword in a rule violation keyword library by using rule matching; and a second quality inspection submodule configured to calculate semantic similarity between the agent dialogue text and the rule violation keyword library by using a natural language model to determine whether there is an implicit rule violation.

[0114] In a possible embodiment, the marketing script generation apparatus further includes a parameter compression module configured to perform parameter compression on the fine-tuned large model by using a model distillation technique to generate a lightweight model; and a cache mechanism module configured to store high-frequency marketing scripts by using a cache mechanism to reduce the number of calls to the fine-tuned large model.

[0115] In a possible embodiment, the marketing script generation apparatus further includes a customer acceptance rate data collection module configured to collect customer acceptance rate data of the outbound marketing scenario; and an adjustment module configured to adjust a strategy of the fine-tuned large model for generating marketing scripts based on the customer acceptance rate data by using a reinforcement learning framework.

[0116] Figure 7 A structural schematic diagram of an electronic device 300 is provided in an embodiment of the present application. As shown in the figure, an embodiment of the present application provides an electronic device 300 including a processor 301 and a memory 302. Optionally, the device further includes a communication component 303. The processor 301, the memory 302, and the communication component 303 are connected through a bus 304. Figure 7

[0117] In the specific implementation process, the memory 302 stores codes, and the processor 301 executes the codes stored in the memory 302 to perform the method of the above method embodiments.

[0118] ​The specific implementation process of the processor 301 can refer to the method embodiments described above, which have similar implementation principles and technical effects, and will not be described here again in this embodiment.

[0119] In the above Figure 7 In the embodiment shown, it should be understood that the processor 301 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in combination with the application can be directly embodied as hardware processor execution, or executed by a combination of hardware and software modules in the processor.

[0120] The memory 302 can include a high-speed RAM memory, and can also include a non-volatile storage NVM, such as at least one disk memory.

[0121] The bus 304 can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, or an extended industry standard architecture (EISA) bus, etc. The bus 304 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus 304 in the drawings of the present application does not limit to only one bus or one type of bus.

[0122] The embodiment of the present application provides a computer readable storage medium, and the computer readable storage medium stores computer execution instructions. When the processor executes the computer execution instructions, the computer execution instructions are used to implement the method of the above method embodiments.

[0123] The computer readable storage medium described above can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0124] An example readable storage medium is coupled to the processor such that the processor can read information from the readable storage medium and can write information to the readable storage medium. Of course, the readable storage medium can be a part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can exist as discrete components in the device.

[0125] The embodiment of the present application provides a computer program product, including a computer program, and the computer program is executed by a processor to implement the method provided by any of the above embodiments of the present application.

[0126] It should be noted that, for the above-mentioned method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action order described, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0127] Further, it should be noted that although each step in the flowchart is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed with at least part of other steps or other steps. Sub-steps or stages of the stage are alternately or alternately executed.

[0128] It should be understood that the above-mentioned device embodiments are only schematic, and the device of the present application can also be realized by other ways. For example, the division of units / modules in the above-mentioned embodiments is only a logical function division, and another division method can be used in actual implementation. For example, multiple units, modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0129] In addition, each functional unit / module in the embodiments of the present application can be integrated in one unit / module, or each unit / module can exist physically, or two or more units / modules can be integrated together. The integrated unit / module can be realized in the form of hardware or in the form of a software program module.

[0130] If the integrated unit / module is realized in the form of hardware, the hardware can be a digital circuit, an analog circuit, etc. The physical implementation of the hardware structure includes but is not limited to transistors, memristors, etc. Unless otherwise specified, the processor can be any appropriate hardware processor, such as a CPU, a GPU, an FPGA, a DSP, an ASIC, etc. Unless otherwise specified, the storage unit can be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory (RRAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), an enhanced dynamic random access memory (EDRAM), a high-bandwidth memory (HBM), a hybrid memory cube (HMC), etc.

[0131] If the integrated unit / module is realized in the form of a software program module and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the embodiments of the method of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various program codes that can be stored in the medium.

[0132] In the above embodiments, the description of each of the embodiments focuses on different aspects of the embodiments. The parts not described in detail in a certain embodiment can be seen in the relevant description of the other embodiments. The technical features of the above embodiments can be combined in any manner. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as falling within the scope of the disclosure.

[0133] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0134] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the appended claims.

Claims

1. A method for generating marketing scripts, characterized in that, include: Obtain customer intent tags and marketing stage tags from historical outbound marketing scenarios; The finely tuned large model generates marketing scripts based on the customer intent tags and the marketing stage tags; Store the generated marketing scripts in the marketing script library.

2. The method according to claim 1, characterized in that, The process of generating marketing scripts using the finely tuned large model based on the customer intent tags and marketing stage tags includes: Obtain a preset prompt word template, which includes a combined description of customer intent tags and marketing stage tags; The prompt word template is input into the fine-tuned large model to generate marketing scripts that match the customer intent tags and marketing stage tags.

3. The method according to claim 1, characterized in that, After storing the generated marketing scripts in the marketing script library, the method further includes: during the outbound marketing process: The finely tuned large model is used to perform intent recognition on the customer dialogue text in order to generate customer intent tags for the customer dialogue text. The finely tuned large model is used to determine the marketing stage of the customer dialogue text in order to generate marketing stage tags for the customer dialogue text. Based on the customer intent tags and marketing stage tags in the customer dialogue text, match marketing messages with the same tags from the marketing message library.

4. The method according to claim 3, characterized in that, The step of using a fine-tuned large model to perform intent recognition on customer dialogue text in order to generate customer intent tags for the customer dialogue text includes: inputting intent recognition prompts into the fine-tuned large model to classify the customer dialogue text and output customer intent tags for the customer dialogue text. The step of using a finely tuned large model to determine the marketing stage of the customer dialogue text and generate marketing stage tags for the customer dialogue text includes: inputting marketing stage determination prompts into the finely tuned large model to classify the customer dialogue text and output marketing stage tags for the customer dialogue text.

5. The method according to claim 1, characterized in that, After storing the generated marketing scripts in the marketing script library, the method further includes: during the outbound marketing process: The system employs rule matching and / or natural language models to perform compliance checks on agent dialogue text in order to generate quality inspection results. If the quality inspection result indicates a violation type tag in the agent's dialogue text, a remedial script with the same tag is matched from the marketing script library based on the violation type tag in the agent's dialogue text.

6. The method according to claim 5, characterized in that, The compliance check of the agent dialogue text using rule matching and / or natural language models includes at least one of the following: The rules are used for matching to detect whether the agent's dialogue text contains keywords from the violation word library; The natural language model is used to calculate the semantic similarity between the agent dialogue text and the violation lexicon to determine whether there is any implicit violation.

7. The method according to claim 1, characterized in that, Before generating marketing scripts based on the customer intent tags and marketing stage tags using the fine-tuned large model, the method further includes: The fine-tuned large model is compressed using model distillation technology to generate a lightweight model; A caching mechanism is used to store high-frequency marketing scripts in order to reduce the number of times the fine-tuned large model is called.

8. The method according to claim 1, characterized in that, After storing the generated marketing scripts in the marketing script library, the method further includes: Collect customer acceptance rate data for the aforementioned outbound marketing scenarios; Based on the customer acceptance rate data, a reinforcement learning framework is used to adjust the strategy for generating marketing messages in the fine-tuned large model.

9. A marketing script generation device, characterized in that, include: The acquisition module is used to acquire customer intent tags and marketing stage tags from historical outbound marketing scenarios; The marketing script generation module is used to generate marketing scripts based on the customer intent tags and the marketing stage tags using a finely tuned large model. The storage module is used to store the generated marketing scripts into the marketing script library.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.

Images