System, method and electronic device for generating speech based on causal big model

Abstract

The present application relates to the field of intelligent customer service technology, and provides a system, method, and electronic device for generating speech based on a large causal model. The system includes: a user identification module for identifying the user's basic information; a solution matching module for obtaining a business causal solution that matches the basic information; a product processing module for obtaining product-related information based on the contextual information and the business causal solution; and a speech generation module for obtaining target speech corresponding to the basic information, the business causal solution, and the product-related information. Some embodiments of the present application can accurately identify user intent and provide users with a reasonable product-solution causal chain by generating accurate target speech.

Description

Technical Field

[0001] The present application relates to the field of intelligent customer service technology, and more specifically, to a system, method, and electronic device for generating speech based on a causal large model. Background Art

[0002] With the continuous development of artificial intelligence, intelligent customer service robots are widely used in various business fields.

[0003] Currently, users seeking information or product information in a specific business area typically interact with intelligent customer service bots. These bots can identify the text messages entered by users and then, using their pre-configured responses, search for the most relevant response to the user's text message. However, this approach only mechanically responds to user questions and fails to truly understand user intent, making it impossible to provide valuable product recommendations.

[0004] Therefore, how to provide a technical solution for a method of generating speech with higher accuracy has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The purpose of some embodiments of the present application is to provide a system, method and electronic device for generating speech based on a causal big model. Through the technical solutions of the embodiments of the present application, it is possible to accurately locate the user's intention, and then provide users with effective, high-quality speech content containing business product suggestions, provide users with valuable auxiliary information, and enhance user experience.

[0006] In a first aspect, some embodiments of the present application provide a system for generating speech based on a causal big model, including: a user identification module for identifying basic information of a user, wherein the basic information includes: user questions and contextual information related to the user; a solution matching module for obtaining a business causal solution that matches the basic information; a product processing module for obtaining product-related information based on the contextual information and the business causal solution, wherein the product-related information includes: product explanation text, product business data and product demand forecast information; a speech generation module for obtaining target speech corresponding to the basic information, the business causal solution and the product-related information.

[0007] Some embodiments of this application provide a system that identifies a user's question and contextual information, matches them to a business causal solution, then derives product-related information based on the contextual information and the business causal solution. Finally, the system uses this information to obtain the corresponding target marketing language. This application can accurately identify user intent and provide users with effective, high-quality marketing language content containing business and product recommendations, providing users with valuable auxiliary information and enhancing the user experience.

[0008] In some embodiments, the solution matching module is used to: retrieve multiple product solution causal chains that match the user problem from a preset solution document; annotate the multiple product solution causal chains through user information and product configuration documents to obtain multiple annotated product solution causal chains; and obtain the business causal solution based on the multiple annotated product solution causal chains, the user problem and the target solution matching model.

[0009] Some embodiments of this application retrieve solutions that match the user's problem from pre-set solution documents, annotate them based on user information and product configuration documents, and finally output target recommended products through a target solution matching model. Some embodiments of this application can accurately identify user intent and improve the accuracy of product recommendations, thereby allowing users to easily understand product information, improving user experience and marketing accuracy.

[0010] In some embodiments, the solution matching module is used to: perform matching calculations with the three stage documents in the preset solution document based on the contextual information to obtain multiple matching values; use the solutions corresponding to the matching values ​​located before the preset positions in the multiple matching values ​​as the multiple product solution causal chains; wherein the first stage document in the preset solution document includes: product type and product planning; the second stage document includes: product parameters of each product in the product type; the third stage document includes: product name and product application scenario of each product and product core terms.

[0011] Some embodiments of the present application perform a matching analysis between the contextual information of the user question and the three-stage documents to obtain multiple product solution causal chains, which can be closer to user needs and improve the accuracy of user intent positioning.

[0012] In some embodiments, the solution matching module is used to: if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are consistent with the product elements in the product configuration document, then mark it as a valid solution; if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are inconsistent with the product elements in the product configuration document, then mark it as an invalid solution and generate invalid reason text information; wherein, the valid solution and the invalid solution constitute the multiple marked product solution causal chains.

[0013] Some embodiments of the present application compare the product names, user information and product configuration documents of the product solution causal chain for consistency, and then label multiple product solution causal chains to obtain accurate product information and improve the accuracy of product recommendations.

[0014] In some embodiments, the solution matching module is used to: generate prompt text corresponding to the multiple labeled product solution causal chains and the user questions; input the prompt text into the target solution matching model to obtain the business causal solution; wherein, the target solution matching model is obtained by training a large language model.

[0015] Some embodiments of the present application generate prompt text that meets the target solution matching model and then input it into the model to obtain a business causal solution, thereby ensuring the accuracy of the product solution causal chain recommendation.

[0016] In some embodiments, the product processing module includes: a product explanation sub-module, a data processing sub-module and a demand prediction sub-module; wherein the product explanation sub-module is used to determine the standard question corresponding to the user question; retrieve the product terms content that matches the standard question from the product terms document; generate the product explanation text according to the standard question, the product terms content and the target large language model; the data processing sub-module is used to calculate the product business data corresponding to the business causal scheme when it is confirmed that the user needs the product business data; the demand prediction sub-module is used to predict the product demand prediction information associated with the user's conversation data.

[0017] Some embodiments of the present application generate and predict relevant needs of users through sub-modules in the product processing module, which can achieve accurate positioning of users and subsequently provide users with valuable speech content information.

[0018] In some embodiments, the product explanation submodule is used to: use a trained semantic vector model to calculate the user question and the preset standard sentence to obtain a vector distance set; obtain the target sentence in the vector distance set that is less than a set threshold; and use the sentence corresponding to the minimum value of the vector distance in the target sentence as the standard question.

[0019] Some embodiments of the present application process and analyze user questions and preset standard sentences through a trained semantic vector model to determine corresponding standard questions, so as to accurately retrieve product terms and content related to user needs.

[0020] In some embodiments, the product explanation submodule is used to: confirm that the vector distances in the vector distance set are not less than the set threshold, then record the user question; set a standard sentence corresponding to the user question, and iteratively update the trained semantic vector model.

[0021] Some embodiments of the present application improve the accuracy and breadth of the semantic vector model by recording user questions and updating the semantic vector model.

[0022] In some embodiments, the product explanation submodule is used to: input preset standard sentences and original product terms content into a large language model to obtain actual reference product terms; verify and mark the actual reference product terms to determine the product terms reference content; and construct the product terms document based on the preset standard sentences, the product terms reference content, and the product information of each product.

[0023] Some embodiments of the present application construct corresponding product terms documents after verifying and marking the retrieved actual reference product terms, thereby improving the accuracy of the final constructed document.

[0024] In some embodiments, the product explanation submodule is used to: generate prompt information corresponding to the standard questions, the product terms content, the user questions, the conversation context data and the situational information; input the prompt information into the target large language model, and output the product explanation text; wherein, the target large language model is obtained by training the initial large model through a training data set; the training data set includes: user question samples, product demand samples, user situation samples, product terms samples and explanation text samples.

[0025] Some embodiments of this application generate product explanation text by inputting relevant prompt information into a target large language model, which is both efficient and convenient, and can also improve the user experience. The target large language model is trained with a training dataset, allowing it to output different product explanation texts for different scenarios and products, improving the model's practicality.

[0026] In some embodiments, the speech generation module is used to: generate a model input text corresponding to the basic information, the business causal scheme and the product association information; input the model input text into a trained causal model to obtain the target speech.

[0027] Some embodiments of the present application obtain target speech by inputting user-related information into a trained causal model, which can provide users with high-quality and valuable product-related speech content and provide effective auxiliary suggestions for users to select business products.

[0028] On the second aspect, some embodiments of the present application provide a method for generating speech based on a causal big model, including: identifying basic information of a user, wherein the basic information includes: user questions and contextual information related to the user; obtaining a business causal plan that matches the basic information; obtaining product-related information based on the contextual information and the business causal plan, wherein the product-related information includes: product explanation text, product business data and product demand forecast information; obtaining target speech corresponding to the basic information, the business causal plan and the product-related information.

[0029] In a third aspect, some embodiments of the present application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any embodiment of the second aspect.

[0030] In a fourth aspect, some embodiments of the present application provide an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor can implement a method as described in any embodiment of the second aspect when executing the program.

[0031] In a fifth aspect, some embodiments of the present application provide a computer program product, comprising a computer program, wherein the computer program, when executed by a processor, can implement the method described in any embodiment of the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of some embodiments of the present application, the following is a brief introduction to the drawings required for use in some embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 A system diagram for generating speech based on a causal model provided in some embodiments of the present application;

[0034] Figure 2 A flow chart of a method for generating speech based on a causal model provided in some embodiments of the present application;

[0035] Figure 3 A schematic diagram of an electronic device is provided for some embodiments of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in some embodiments of the present application will be described below in conjunction with the drawings in some embodiments of the present application.

[0037] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0038] With the emergence and development of artificial intelligence and big model technologies, current approaches, such as general-purpose big models, can directly provide specific business solutions (such as insurance solutions or causal chains for financial product solutions) related to user questions to meet user needs and improve product marketing efficiency. For example, big models in the insurance field can provide answers to basic FAQs or popular product recommendations based on user questions. However, these answers are based on the model's pre-defined response logic and fail to provide users with precise, supportive recommendations. When matching user needs with products, static matching logic is often employed. This requires users to provide a large amount of cumbersome information at once, leaving them to delve deeper and provide detailed information before their specific needs can be further confirmed. The resulting solutions are also one-off solutions. Adjustments require users to modify the information form and resubmit it to generate a matching solution. This fails to proactively and dynamically identify user needs and provide more appropriate planning and sales pitches. Furthermore, when users want to understand the terms and conditions of a particular product, current intelligent customer service agents often provide general explanations or simply output lengthy, verbose textual descriptions of the terms and conditions, which can easily confuse users and make them confused. On the other hand, directly generating corresponding broad copy can easily lead to misunderstandings, fabricated terms, and incorrect product explanations that mislead users.

[0039] It can be seen from the above-mentioned related technologies that the existing technology has poor accuracy in locating user intentions, and the reply content given cannot provide effective suggestions to users, cannot plan products for users, and the user experience is poor.

[0040] In view of this, some embodiments of the present application provide a method for generating speech based on a causal model. This method can obtain a matching business causal solution after identifying the user's user question and contextual information. It can then obtain product-related information associated with the user based on this information, and finally generate a corresponding target speech based on the above-obtained content. In some embodiments of the present application, it is possible to accurately locate the user's intention, and the target speech provided can provide the user with effective product planning suggestions, thereby improving the quality of the response and user experience.

[0041] The following is combined with Figure 1 The overall composition structure of the speech generation system based on the causal big model provided by some embodiments of the present application is exemplified.

[0042] like Figure 1 As shown, some embodiments of the present application provide a system for generating speech based on a causal model. The system for generating speech based on a causal model may include: a user identification module 110, a solution matching module 120, a product processing module 130 and a speech generation module 140. Among them, the product processing module 130 also includes: a product explanation submodule 131, a data processing submodule 132, a demand prediction submodule 133 and an information collection module 134. Through this application Figure 1 The mutual interaction between the modules in the system shown can provide personalized product planning for users and provide users with high-quality target scripts, so that users can choose corresponding products in a targeted manner based on the recommended information in the target scripts.

[0043] The following is an example of Figure 1 The specific functions of each module.

[0044] In some embodiments of the present application, the user identification module 110 is used to identify basic information of the user, wherein the basic information includes: user questions and contextual information related to the user.

[0045] For example, in some embodiments of the present application, a user can interact with an intelligent customer service robot in a business field through a terminal device, and the large model agent combination deployed by the intelligent customer service robot can recognize the conversation content input by the user. The user question currently expressed by the user is identified through the user question agent; the user context agent can identify the current context information of the user. For example, taking the insurance business field as an example, the user question is "buy insurance for my daughter", and the context information can be "number of children: 1, gender of the child: female, stage of life: marriage and childbearing period", etc. Among them, the large model agent combination can be obtained by training a specific data set in a specific scenario (for example, the insurance business field). It can be understood that in actual applications, text recognition models or other models in natural language processing can also be used to recognize the content of the user's conversation, and the embodiments of the present application are not limited to this.

[0046] In some embodiments of the present application, the solution matching module 120 is used to obtain a business causal solution that matches the basic information.

[0047] For example, in some embodiments of the present application, the solution matching module 120 can first search the preset solution document in the large model based on the user question understood by the pre-set agent, and extract the N solutions that are most relevant to the user question (as a specific example of a causal chain of multiple product solutions). Then, the N solutions obtained above are annotated in combination with the product configuration document to obtain multiple annotated product solution causal chains. Taking the insurance business as an example, the insurance product configuration document can provide the insurance requirement elements of each insurance product (for example, the elements can be three parts: insured age range, insured occupation range, health requirement range), as well as whether it is on sale. It should be understood that the content of the product configuration document can be adaptively adjusted for different business scenarios, and the embodiments of the present application are not limited to this. Finally, the N annotated product solution causal chains are combined with the user question and input into the target solution matching model to obtain the business causal solution output by the model. For example, taking the above-mentioned example of buying insurance for the daughter, the business causal solution is: buy insurance for the daughter -> medical insurance, critical illness insurance, accident insurance.

[0048] In some embodiments of the present application, the solution matching module 120 is used to perform matching calculations with the three stage documents in the preset solution document based on the contextual information to obtain multiple matching values; and use the solutions corresponding to the matching values ​​located before the preset positions in the multiple matching values ​​as the multiple product solution causal chains; wherein the first stage document in the preset solution document includes: product type and product planning; the second stage document includes: product parameters of each product in the product type; the third stage document includes: product name and product application scenario of each product and product core terms.

[0049] For example, in some embodiments of the present application, the insurance business field is taken as an example for explanation, and the insurance business content is divided into three-stage solution documents according to the planning stage of the intelligent insurance planner. The solution documents are all constructed in the form of a causal chain. The left side of the causal chain is the user demand, the right side is the solution, and the middle is connected by '->' (for example, the user hopes to renew the insurance after the claim -> Changxiang'an Long-term Medical Insurance (insurance product name) -> guaranteed renewal for 20 years). Among them, the first stage document may include: the initial planning (as a specific example of product planning) and the major types of insurance explained to the user (as a specific example of the product type). The second stage document may include: after determining the insurance type, giving the recommended logic of the main parameters under the specific insurance type, such as the insurance amount, premium range, etc. (as a specific example of product parameters). The third stage document may include: the specific product name and the problem to be solved (as a specific example of the product application scenario), as well as the core content of the product terms (as a specific example of the core terms of the product) to recommend specific products and match the specific preferences of users. For example, consider a solution called Darwin 8 Pilot Edition Critical Illness Insurance. The product's application scenario involves the need for income loss protection due to critical illness. The core product terms and conditions stipulate that the first major illness benefit for each of the 120 types of critical illnesses is 100% of the basic insured amount. It is understood that the first, second, and third stage documents can be designed and generated as needed, and the embodiments of this application are not limited thereto.

[0050] Specifically, the solution matching module 120 can perform matching calculations through the user context involved in the user problem (as a specific example of context information) and the user context set in the preset solution document. For example, if the user context is consistent, the item in the document will be added with 0.5 points, and if it is inconsistent, the item will be deducted with 0.2 points. Finally, the sum is calculated to obtain the final matching value for each item in the three-stage document, where the matching values ​​are sorted from large to small, and the solutions in the first N positions (as a specific example of the preset position) are the N product solution causal chains (that is, N solutions) that are most relevant to the user problem. It should be noted that N is a positive integer, and its value can be set as needed. The matching calculation method can also be flexibly adjusted according to the actual application scenario, and the embodiments of the present application are not limited to this.

[0051] In some embodiments of the present application, the solution matching module 120 is used to mark a solution as a valid solution if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are consistent with the product elements in the product configuration document; if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are inconsistent with the product elements in the product configuration document, it is used to mark it as an invalid solution and generate invalid reason text information; wherein, the valid solution and the invalid solution constitute the multiple marked product solution causal chains.

[0052] For example, in some embodiments of the present application, the solution matching module 120 can extract the product names in N product solution causal chains, and then check and match them according to the product configuration document and the collected user context (such as user age, occupation, and health) to confirm whether the content elements (as a specific example of product elements) are consistent. If there are any inconsistencies (such as age inconsistency, occupation inconsistency, health inconsistency, or the product has been removed from the shelves, etc.), then "invalid solution" will be noted on the basis of the product solution causal chain, and text information of the invalid reason will be given (such as age inconsistency, occupation inconsistency, health inconsistency, or the product has been removed from the shelves). Otherwise, "valid solution" will be noted on the basis of the product solution causal chain. After the N product solution causal chains are labeled, N labeled product solution causal chains can be obtained.

[0053] In some embodiments of the present application, the solution matching module 120 is used to generate prompt text corresponding to the multiple labeled product solution causal chains and the user question; input the prompt text into the target solution matching model to obtain the business causal solution; wherein, the target solution matching model is obtained by training a large language model.

[0054] For example, in some embodiments of the present application, the solution matching module 120 writes the processed N annotated product solution causal chains and the user question into a prompt (as a specific example of prompt text), then calls the target solution matching model and inputs the prompt into the target solution matching model to obtain a business causal solution. For example, the user question is: What insurance should an infant buy? The business causal solution ultimately output from the N annotated product solution causal chains is a causal diagram supplemented with the user's needs: Infancy -> Medical Insurance (Insurance Type); Infancy -> Critical Illness Insurance (Insurance Type); Infancy -> Accident Insurance (Insurance Type); Infancy -> Education Fund Insurance (Insurance Type).

[0055] In some embodiments of the present application, the product processing module 130 is used to obtain product-related information based on the contextual information and the business causal scheme, wherein the product-related information includes: product explanation text, product business data and product demand forecast information.

[0056] For example, in some embodiments of the present application, the product processing module 130 may formulate the next communication strategy (as a specific example of product association information) based on the business cause-effect scenario and the user context.

[0057] In some embodiments of the present application, the data processing submodule 132 is configured to calculate the product business data corresponding to the business causal scheme when it is confirmed that the user needs the product business data.

[0058] For example, if it is determined that the insured amount and premium need to be calculated (as a specific example of product business data), the data processing submodule 132 is scheduled to calculate the insured amount and premium of the corresponding product.

[0059] In some embodiments of the present application, the demand prediction submodule 133 is used to predict the product demand prediction information associated with the user's conversation data.

[0060] For example, based on the current multiple rounds of communication content with the user and the extracted user questions, user scenarios, and business causal solutions, the user's possible further potential needs (as a specific example of product demand forecast information) are inferred. For example, the user has a 5-year-old daughter who may have more colds and fevers and needs a medical insurance product that covers outpatient care. It is understandable that the demand forecasting submodule 133 can store lists of different types of products corresponding to the user's potential needs, and the user's potential needs can be predicted through matching and searching. Other inference methods can also be used in actual applications, and the embodiments of the present application are not limited to this.

[0061] In some embodiments of the present application, the information collection module 134 may need to provide more information based on the current contextual information and business cause-and-effect scheme, so that the AI ​​planner (that is, the intelligent customer service robot) can further converge the user's explanation of the selection of specific products or product parameters. At this time, the information collection module 134 can be called for processing. For example, in the insurance business, when calculating product premiums and insured amounts, users are required to provide age information; selecting product types (mid-range medical care or million-dollar medical care) requires users to provide premium budgets, etc. The information collection module 134 can collect information after obtaining the user's authorization, so as to more accurately locate the user's intentions.

[0062] In addition, if the user is matched with a specific product and wants to obtain the terms and conditions of the product, the product explanation submodule 131 can be scheduled to output the product explanation text, which is relatively concise and easy for the user to understand.

[0063] The specific implementation functions of the product explanation submodule 131 are exemplarily described below.

[0064] In some embodiments of the present application, the product explanation submodule 131 is used to determine the standard question corresponding to the user question; retrieve the product terms content that matches the standard question from the product terms document; and generate the product explanation text based on the standard question, the product terms content and the target large language model.

[0065] For example, in some embodiments of the present application, in order to effectively retrieve the matching product solution causal chain and corresponding product terms for users, the product explanation submodule 131 first normalizes the user questions collected from users, that is, it is necessary to analyze and process the user questions and determine standard questions (which can be referred to as standard questions). Afterwards, the above-determined standard questions are used as search parameters to retrieve the matching product terms content in the pre-deployed product terms document. Through the trained target large language model, combined with the above-mentioned user questions, standard questions and product terms content, a concise and easy-to-understand product explanation text can be generated for users.

[0066] In some embodiments of the present application, the product explanation submodule 131 is used to use a trained semantic vector model to calculate the user question and the preset standard sentence to obtain a vector distance set; obtain the target sentence in the vector distance set that is less than a set threshold; and use the sentence corresponding to the minimum value in the vector distance in the target sentence as the standard question.

[0067] For example, in some embodiments of the present application, a trained semantic vector model is used to calculate the distance between the user question and the pre-planned preset standard question (as a preset standard sentence) (wherein, the smaller the vector distance, the more similar they are), and all the preset standard questions that are less than the set threshold are taken out, and then the one with the smallest vector distance is selected as the standard question of this user question.

[0068] It should be noted that the trained semantic vector model is obtained by training with raw data. Raw data includes user question samples and their corresponding standard question samples. For example, a user question sample might be: "Can I get insurance tomorrow if I buy it today?"; a standard question sample might be: "What is the waiting period?" The literal meanings of these two questions differ significantly, making it unreasonable to directly use an off-the-shelf semantic model. Therefore, we prepare possible user questions and standard questions for each product and input them into the semantic model training to ultimately obtain a trained semantic vector model that can be adapted to different products.

[0069] In some embodiments of the present application, the product explanation submodule 131 is used to confirm that the vector distances in the vector distance set are not less than the set threshold, then record the user question; set the standard sentence corresponding to the user question, and iteratively update the trained semantic vector model.

[0070] For example, in some embodiments of the present application, if there is no preset standard question that is less than a set threshold in the trained semantic vector model, the user question is used as the standard question, and an exception record log is generated. Subsequently, the corresponding standard question is reset based on this user question, and the semantic vector model is iteratively updated and trained to improve the accuracy of the standard question matching.

[0071] In some embodiments of the present application, the product explanation submodule 131 is used to input preset standard sentences and original product terms content into a large language model to obtain actual reference product terms; verify and mark the actual reference product terms to determine the product terms reference content; and construct the product terms document based on the preset standard sentences, the product terms reference content and the product information of each product.

[0072] For example, in some embodiments of the present application, in order to ensure the effectiveness of product terms retrieval, taking the insurance business field as an example, the product terms document is constructed through the following terms extraction scheme:

[0073] Based on all pre-prepared standard questions, the original product terms (here, dozens of pages of original terms) are input into the large model (as a specific example of a large language model) for retrieval, outputting the actual reference product terms. Insurance experts then annotate the standard questions and answers (i.e., the actual reference product terms), marking any that make sense as acceptable and providing reference content for any that do not. These two steps yield meaningful product terms reference content and corresponding standard questions. All of these product terms reference content and corresponding standard questions are integrated into a single document, resulting in a product terms document. Subsequently, targeted product terms content can be retrieved from the product terms document based on the standard questions.

[0074] In addition, considering that the standard questions may not be comprehensive, this application will also uniformly enter or read from relevant terminals the types of terms that must be included in each product (for example, basic protection content, optional protection content, insurance requirements: occupation, gender, age, health, etc., underwriting company, waiting period, claims method, cooling-off period, etc.) to improve the accuracy of subsequent retrieval.

[0075] In some embodiments of the present application, the product explanation submodule 131 is used to generate prompt information corresponding to the standard questions, the product terms content, the user questions, the conversation context data and the situational information; input the prompt information into the target large language model, and output the product explanation text; wherein, the target large language model is obtained by training the initial large model through a training data set; the training data set includes: user question samples, product demand samples, user situation samples, product terms samples and explanation text samples.

[0076] For example, in some embodiments of the present application, the product explanation submodule 131 can prepare a training data set for training the initial large model, wherein the training data set includes: user question samples, currently concerned products (as a specific example of product demand samples), user context samples (for example, user occupation, usage scenarios, etc.), product terms samples and explanation text samples. Among them, the product terms samples contain general product knowledge and core documents corresponding to each product, as well as related FAQs, key content, etc. The text composed of user question samples, currently concerned products, user contexts and product terms samples is used as the input data of the initial large model, and the explanation text samples are used as the output data of the initial large model for training and fine-tuning to obtain the target large language model for the product of the corresponding scenario.

[0077] Afterwards, the retrieved product terms are combined with the solution previously given by the agent, the user's question, the current user situation (i.e., situational information), and the user's conversation context data to form a prompt (as a specific example of the prompt information in the product explanation submodule 131); the prompt is input into the target large language model trained above to finally generate the product explanation text.

[0078] Through the product explanation submodule 131 in the above-mentioned embodiment of the present application, it is possible to provide product explanations of the solutions required by users based on their specific situations and problems, and the explanations are targeted and concise, allowing users to clearly understand the suggestions and solutions given by the planners, thereby improving the user experience.

[0079] In some embodiments of the present application, the speech generation module 140 is used to obtain target speech relative to the basic information, the business cause-effect scheme and the product association information.

[0080] For example, in some embodiments of the present application, based on the information output by all the previous modules, the target words that are finally output to the user can be obtained, and the target words are easy to understand, which can enable the user to fully understand the planned product content and provide the user with valuable auxiliary suggestions.

[0081] In some embodiments of the present application, the speech generation module 140 is used to generate a model input text corresponding to the basic information, the business causal scheme and the product association information; and input the model input text into a trained causal model to obtain the target speech.

[0082] For example, in some embodiments of the present application, the speech generation module 140 can write user questions, user scenarios, business causal solutions, and product-related information into a final prompt (as a specific example of model input text). The final prompt is input into the trained causal big model to output the target speech. Among them, the trained causal big model is obtained by training the big model (such as GPT4) on a specific data set in a specific scenario. The trained causal big model can better serve the future in a specific scenario and provide users with target speech containing reasonable product suggestions.

[0083] The following is combined with Figure 2 The implementation process of generating speech based on the causal big model provided by some embodiments of the present application is exemplified.

[0084] Please see the attached Figure 2 , Figure 2 A flow chart of a method for generating speech based on a causal big model is provided for some embodiments of the present application. The method for generating speech based on a causal big model may include: S210, identifying basic information of the user, wherein the basic information includes: user questions and contextual information related to the user. S210, obtaining a business causal solution that matches the basic information. S230, obtaining product-related information based on the contextual information and the business causal solution, wherein the product-related information includes: product explanation text, product business data and product demand forecast information. S240, obtaining target speech relative to the basic information, the business causal solution and the product-related information.

[0085] The above process is described below as an example.

[0086] In some embodiments of the present application, S220 may include: S221, retrieving multiple product solution causal chains that match the user problem from a preset solution document; S222, annotating the multiple product solution causal chains through user information and product configuration documents to obtain multiple annotated product solution causal chains; S223, obtaining the business causal solution based on the multiple annotated product solution causal chains, the user problem and the target solution matching model.

[0087] In some embodiments of the present application, S221 may include: performing matching calculations with the three stage documents in the preset solution document based on the contextual information to obtain multiple matching values; using the solutions corresponding to the matching values ​​located before the preset positions in the multiple matching values ​​as the multiple product solution causal chains; wherein the first stage document in the preset solution document includes: product type and product planning; the second stage document includes: product parameters of each product in the product type; the third stage document includes: product name and product application scenario of each product and product core terms.

[0088] In some embodiments of the present application, S222 may include: if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are consistent with the product elements in the product configuration document, then it is marked as a valid solution; if it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are inconsistent with the product elements in the product configuration document, then it is marked as an invalid solution, and invalid reason text information is generated; wherein, the valid solution and the invalid solution constitute the multiple marked product solution causal chains.

[0089] In some embodiments of the present application, S223 may include: generating prompt text corresponding to the multiple labeled product solution causal chains and the user question; inputting the prompt text into the target solution matching model to obtain the business causal solution; wherein, the target solution matching model is obtained by training a large language model.

[0090] In some embodiments of the present application, S230 may include: S231, the product explanation sub-module determines the standard question corresponding to the user question; retrieves the product terms content matching the standard question from the product terms document; generates the product explanation text based on the standard question, the product terms content and the target large language model; S232, the data processing sub-module calculates the product business data corresponding to the business causal scheme when confirming that the user needs the product business data; S233, the demand prediction sub-module predicts the product demand prediction information associated with the user's conversation data.

[0091] In some embodiments of the present application, S231 may include: using a trained semantic vector model to calculate the user question and the preset standard statement to obtain a vector distance set; obtaining a target statement in the vector distance set that is less than a set threshold; and taking the statement corresponding to the minimum value in the vector distance in the target statement as the standard question.

[0092] In some embodiments of the present application, S231 may include: confirming that the vector distances in the vector distance set are not less than the set threshold, then recording the user question; setting a standard sentence corresponding to the user question, and iteratively updating the trained semantic vector model.

[0093] In some embodiments of the present application, S231 may include: inputting preset standard statements and original product terms content into a large language model to obtain actual reference product terms; verifying and marking the actual reference product terms to determine the product terms reference content; and constructing the product terms document based on the preset standard statements, the product terms reference content, and the product information of each product.

[0094] In some embodiments of the present application, S231 may include: generating prompt information corresponding to the standard question, the product terms content, the user question, the conversation context data and the situational information; inputting the prompt information into the target large language model and outputting the product explanation text; wherein, the target large language model is obtained by training the initial large model through a training data set; the training data set includes: user question samples, product demand samples, user situation samples, product terms samples and explanation text samples.

[0095] In some embodiments of the present application, S240 may include: generating a model input text corresponding to the basic information, the business causal scheme and the product association information; inputting the model input text into a trained causal model to obtain the target speech.

[0096] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific implementation process of the above-described method S210 to S240 can refer to the corresponding process in the aforementioned system and will not be elaborated here.

[0097] Some embodiments of the present application further provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the operations corresponding to any of the above methods provided in the above embodiments.

[0098] Some embodiments of the present application further provide a computer program product, which includes a computer program, wherein when the computer program is executed by a processor, it can implement the operations corresponding to any of the above methods provided in the above embodiments.

[0099] like Figure 3As shown, some embodiments of the present application provide an electronic device 300, which includes: a memory 310, a processor 320, and a computer program stored in the memory 310 and executable on the processor 320, wherein the processor 320 reads the program from the memory 310 through the bus 330 and executes the program to implement a method as in any of the above embodiments.

[0100] Processor 320 can process digital signals and can include various computing architectures, such as a complex instruction set computer architecture, a reduced instruction set computer architecture, or an architecture that implements a combination of multiple instruction sets. In some examples, processor 320 can be a microprocessor.

[0101] The memory 310 can be used to store instructions executed by the processor 320 or data related to the execution of instructions. These instructions and / or data may include code for implementing some or all functions of one or more modules described in the embodiments of this application. The processor 320 of the embodiment of the present disclosure can be used to execute the instructions in the memory 310 to implement the method shown above. The memory 310 includes dynamic random access memory, static random access memory, flash memory, optical memory, or other memory known to those skilled in the art.

[0102] The foregoing is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application. It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures.

[0103] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0104] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

Claims

1. A system for generating speech based on a large causal model, characterized by: include: A user identification module, configured to identify basic information of a user, wherein the basic information includes: user questions and contextual information related to the user; A solution matching module, configured to obtain a business cause-effect solution that matches the basic information; A product processing module is configured to obtain product-related information based on the context information and the business causal scheme, wherein the product-related information includes: product explanation text, product business data, and product demand forecast information; A speech generation module is used to obtain target speech corresponding to the basic information, the business cause-effect scheme and the product association information; The solution matching module is configured to: calculate the matching degree with the three-stage documents in the preset solution document based on the context information, and obtain multiple matching degree values; and use the solutions corresponding to the matching degree values ​​located before the preset positions in the multiple matching degree values ​​as multiple product solution causal chains; wherein the first-stage document in the preset solution document includes: product type and product planning; the second-stage document includes: product parameters of each product in the product type; and the third-stage document includes: product name, product application scenario, and product core terms of each product; Annotate the multiple product solution causal chains using user information and product configuration documents to obtain multiple annotated product solution causal chains; The business causal solution is obtained according to the multiple labeled product solution causal chains, the user problem and the target solution matching model.

2. The system according to claim 1, wherein The solution matching module is used to: If it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains and the user information are consistent with the product elements in the product configuration document, then it is marked as a valid solution; If it is confirmed that the product name of each product solution causal chain in the multiple product solution causal chains, the user information and the product elements in the product configuration document are inconsistent, it is marked as an invalid solution and an invalid reason text message is generated; The valid solutions and the invalid solutions constitute the causal chains of the multiple labeled product solutions.

3. The system according to claim 1, wherein: The solution matching module is used to: Generate prompt text corresponding to the multiple annotated product solution causal chains and the user question; The prompt text is input into the target solution matching model to obtain the business causal solution; wherein the target solution matching model is obtained by training a large language model.

4. The system according to claim 1, wherein: The product processing module includes: a product explanation submodule, a data processing submodule and a demand forecasting submodule; wherein, The product explanation submodule is configured to determine a standard question corresponding to the user question; retrieve product terms and conditions that match the standard question from a product terms and conditions document; and generate the product explanation text based on the standard question, the product terms and conditions, and the target large language model. The data processing submodule is configured to calculate the product business data corresponding to the business causal scheme when confirming that the user needs the product business data; The demand prediction submodule is used to predict the product demand prediction information associated with the user's conversation data.

5. The system according to claim 4, wherein: The product explanation submodule is used to: Calculate the user question and the preset standard sentence using the trained semantic vector model to obtain a vector distance set; Obtain target sentences in the vector distance set that are less than a set threshold; The sentence corresponding to the minimum value of the vector distance in the target sentence is used as the standard question.

6. The system according to claim 5, wherein: The product explanation submodule is used to: Confirming that all vector distances in the vector distance set are not less than the set threshold, then recording the user question; A standard sentence corresponding to the user question is set, and the trained semantic vector model is iteratively updated.

7. The system according to claim 4, wherein: The product explanation submodule is used to: Input the preset standard sentences and original product terms into the large language model to obtain the actual reference product terms; Verify and mark the actual reference product terms to determine the reference content of the product terms; The product terms document is constructed based on the preset standard statement, the product terms reference content and the product information of each product.

8. The system according to claim 4, wherein: The product explanation submodule is used to: Generate prompt information corresponding to the standard question, the product terms content, the user question, the conversation context data and the situational information; Inputting the prompt information into the target large language model and outputting the product explanation text; The target large language model is obtained by training the initial large model through a training data set; the training data set includes: user question samples, product demand samples, user scenario samples, product terms samples and explanation text samples.

9. The system according to claim 1, wherein: The speech generation module is used to: generating a model input text corresponding to the basic information, the business cause-effect scheme, and the product association information; The model input text is input into the trained causal model to obtain the target speech.

10. A method for generating speech based on a causal model, characterized in that: include: Identifying basic information of the user, wherein the basic information includes: user questions and contextual information related to the user; Obtaining a business cause-and-effect solution that matches the basic information; Acquire product-related information based on the context information and the business causal scheme, wherein the product-related information includes: product explanation text, product business data, and product demand forecast information; Obtaining target speech corresponding to the basic information, the business cause-effect scheme, and the product association information; The obtaining of a business causal solution matching the basic information includes: Based on the context information, a matching degree is calculated with the three stage documents in the preset solution document to obtain multiple matching degree values; solutions corresponding to matching degree values ​​located before the preset position in the multiple matching degree values ​​are used as multiple product solution causal chains; wherein the first stage document in the preset solution document includes: product type and product planning; the second stage document includes: product parameters of each product in the product type; the third stage document includes: product name, product application scenario and product core terms of each product; Annotate the multiple product solution causal chains using user information and product configuration documents to obtain multiple annotated product solution causal chains; The business causal solution is obtained according to the multiple labeled product solution causal chains, the user problem and the target solution matching model.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program is executed by a processor to perform the method according to claim 10 .

12. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program performs the method according to claim 10 when being run by the processor.

13. A computer program product, characterized in that The computer program product comprises a computer program, wherein the computer program is configured to execute the method according to claim 10 when executed by a processor.

Images