WO2025097891A1 - Pitch recommendation method and apparatus, and device - Google Patents

Abstract

The present application relates to the technical field of data processing, and provides a pitch recommendation method and apparatus, and a device. The method comprises: determining feature information of a first client and feature information of first service personnel in a target dialogue, the target dialogue being a dialogue between the first client and the first service personnel; parsing dialogue data of the target dialogue to determine a service scenario intention corresponding to the target dialogue, the dialogue data comprising dialogue context of the target dialogue; on the basis of target data, obtaining recommended pitches, the target data comprising: the feature information of the first client, the feature information of the first service personnel, and the service scenario intention corresponding to the target dialogue, and the recommended pitches being optional pitches used for replying to the first client; and providing a recommended pitch display interface, the recommended pitch display interface being used for displaying the recommended pitches. According to the present application, the accuracy and relevance of pitch recommendation can be effectively improved, thereby assisting service personnel in completing high-efficiency services. The present application is used for auxiliary recommendation of pitches.

Description

Method, device and equipment for recommending speech skills

This application claims the priority of the Chinese patent application filed with the Chinese Patent Office on November 8, 2023, with application number 202311481562.8 and application name “Speech Recommendation Method, Device, Computing Device Cluster and Storage Medium”, and the priority of the Chinese patent application filed with the Chinese Patent Office on February 29, 2024, with application number 202410231625.2 and application name “Speech Recommendation Method, Device and Equipment”, all contents of which are incorporated by reference in this application.

Technical Field

The present application relates to the field of data processing technology, and in particular to a method, device and equipment for recommending speech skills.

Background Art

When an enterprise provides services to customers, its sales staff need to provide sales information and guidance on corresponding products and services according to the needs and demands of customers. In the general sales process, enterprise sales staff contact customers through telephone, online and other channels, and provide specific sales guidance and product sales. The communication between sales staff and customers often depends on the sales staff's experience and communication skills. Sales staff need to query the enterprise's relevant systems at any time according to the user's questions to supplement key information and answer customer questions. With the increasing amount of data, information and knowledge, some intelligent sales assistance technologies are urgently needed to assist sales staff to improve work efficiency.

In the related art, sales talk recommendations are made to sales staff in the following ways: 1. Searching a database based on keywords in the conversation between the sales staff and the customer to recommend talk; 2. Recommending talk based on the customer profile and historical behavior; 3. Retrieving knowledge based on the knowledge base and knowledge graph to recommend talk; 4. Training a text model based on historical conversations between the customer and the sales staff, and recommending talk based on the text model.

However, there is an ambiguity problem with keywords. The same keyword may represent different meanings in different scenarios, and the weights of different keywords vary greatly, resulting in low recommendation accuracy. Customer portraits and customer historical behaviors are inherent information and cannot be applied to various scenarios, resulting in low recommendation accuracy. In addition, this method involves user information collection and storage, and data security is poor. The cost of building a comprehensive and complete knowledge graph and knowledge base is high. The recommendation accuracy of the method based on text models for speech recommendations depends on the accuracy of the model. When the model accuracy is low, the recommendation accuracy is also low.

Summary of the invention

The present application provides a method, device and equipment for recommending speech techniques, which solve the problems of low accuracy, poor data security and high cost of speech techniques recommendation in related technologies, and can effectively improve the accuracy and relevance of speech techniques recommendation, thereby assisting service personnel to complete efficient services.

In a first aspect, the present application provides a method for recommending words of speech, the method comprising: determining characteristic information of a first customer and characteristic information of a first service staff in a target conversation, the target conversation being a conversation between the first customer and the first service staff; parsing conversation data of the target conversation to determine a business scenario intention corresponding to the target conversation, the conversation data comprising a conversation context of the target conversation; obtaining recommended words of speech based on the target data, the target data comprising: characteristic information of the first customer, characteristic information of the first service staff, and business scenario intention corresponding to the target conversation, the recommended words of speech being an optional word of speech for replying to the first customer; obtaining recommended words of speech based on the target data, the target data comprising: characteristic information of the first customer, characteristic information of the first service staff, and business scenario intention corresponding to the target conversation, the recommended words of speech being an optional word of speech for replying to the first customer.

For example, the business scenario intent may include: business scenario tags and/or intent tags. Business scenario tags may include sales scenario tags or objection scenario tags. Sales scenario tags include but are not limited to: order promotion, renewal, sales transfer, contract or non-cloud service, etc. Objection scenario tags include but are not limited to: price, comparison of friendly products or complaints, etc. The intent tags of the first customer and the first service staff may refer to the entities and concerns of the first customer's intentions, etc.

Recommendation scripts may include, for example: answers to questions raised by customers, scripts to introduce products, solutions to objections raised by customers, and marketing activities related to recommended products.

The beneficial effect is that it combines the customer feature information and the service personnel feature information at the same time, and can more effectively assist in the recommendation of speech techniques in accordance with the individual characteristics of the service personnel and the customer compared to the related technologies. And the comprehensive use of dynamic parameters (conversation text) and static parameters (feature information of the first customer and feature information of the first service personnel) for speech technique recommendation can effectively improve the accuracy and relevance of speech technique recommendation compared to the related technologies, thereby assisting the service personnel to complete efficient services.

In a possible implementation, the characteristic information of the first service personnel includes at least one of the following: the ability level of the first service personnel, The first service personnel’s business team, the first service personnel’s sales status, and the first service personnel’s sales history.

For example, the ability level includes at least one of the following: elementary, intermediate and advanced, the business team includes: new business team and/or existing business team, the sales status includes: customer satisfaction and/or industry expertise, and the sales history includes at least one of the following: number of customers with orders, number of products with orders, total amount of orders and total cumulative call duration, wherein the higher the ability level of the first service personnel, the higher the sales ability of the first service personnel.

In one possible implementation, the recommendation script is a script for introducing the target product; the target product satisfies at least one of the following characteristics: the target product corresponds to the ability level of the first service personnel, the target product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the target product is greater than a first preset threshold, the target product belongs to the industry that the first service personnel is good at, and the number of sales history orders of the target product by the first service personnel is greater than a second preset threshold.

In a possible implementation, the method further includes: determining at least one preferred product of the first customer based on characteristic information of the first customer and characteristic information of the first service personnel, and the target data further includes at least one preferred product.

In one possible implementation, the process of determining at least one preferred product of the first customer based on the characteristic information of the first customer and the characteristic information of the first service staff includes: inputting the characteristic information of the first customer and the characteristic information of the first service staff into a first model, and obtaining at least one preferred product output by the first model, wherein the first model is obtained by training a first initial model with first training data, and the first training data includes: the characteristic information of the second customer, the characteristic information of the second service staff, and the historical transaction products of the second customer.

The first training data can be regarded as a matrix of customers, sales, customer satisfaction with the product and transaction results, and the first model is obtained by training with successful sales as the goal.

In one possible implementation, the process of parsing the conversation data of a target conversation to determine the business scenario intention corresponding to the target conversation includes: inputting the conversation data into a second model to obtain the business scenario intention corresponding to the target conversation output by the second model, wherein the second model is obtained by training a second initial model with second training data, and the second training data includes: historical conversation data of historical conversations between a third customer and a third service staff and the business scenario intention corresponding to the historical conversations.

For example, the method further includes: acquiring a preset number of continuous dialogue texts from dialogue texts of historical dialogues between the third customer and the third service personnel to obtain historical dialogue data.

For example, a preset number of continuous dialogue texts can be taken from the dialogue texts of historical dialogues between the third customer and the third service personnel through a dynamic window limit range to obtain historical dialogue data. The beneficial effect is that the rolling window dialogue text can maximize the reuse of multiple rounds of information in a dialogue, and can be continuously corrected, which can improve the output accuracy and stability of the second model.

In one possible implementation, the process of obtaining recommended words based on target data includes: inputting the target data into a third model to obtain a model recall result output by the third model, wherein the third model is obtained by training a third initial model with third training data, and the third training data includes: characteristic information of a fourth customer, characteristic information of a fourth service personnel, business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service personnel, and preset words; based on the model recall result, the recommended words are obtained.

For example, the target data may also include at least one of the following: conversation data between the first customer and the first service personnel, at least one preferred product of the first customer, and a first keyword group. The first keyword group may be obtained by extracting keywords from the conversation data between the first customer and the first service personnel, and the first keyword group may include business-related, product-related words and entities, etc. The keyword extraction process may be performed in conjunction with a pre-established keyword library.

The beneficial effect is that the model is trained and fine-tuned by integrating the context data structure, and the obtained large model can recommend the speech in accordance with the context data structure, and at the same time, combined with the customer feature information and sales feature information, it can more effectively fit the personality characteristics of the service staff and customers to assist in the recommendation of speech compared with related technologies. In addition, the first model, the second model and the third model can realize self-closed loop sustainable iterative update, thereby further realizing the real-time and accuracy of speech recommendation.

In a possible implementation, the method also includes: storing first vector data and preset words in a vector database, the first vector data including: a vector of characteristic information of a fourth customer, a vector of characteristic information of a fourth service staff, and a vector of business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff.

In a possible implementation, the method also includes: performing vector retrieval based on second vector data to obtain a vector recall result, the second vector data including a vector of the target data; based on the model recall result, the process of obtaining the recommended words includes: weighting the model recall result and the vector recall result to obtain the recommended words.

The beneficial effect is that the model output results and vector retrieval results are weighted, which can effectively reduce the errors in recommended words caused by data cold start or model errors.

In a possible implementation, the method further includes: acquiring a preset number of continuous dialogue texts from the dialogue text of the target dialogue to obtain dialogue data.

For example, a dynamic window can be used to limit the range, and a preset number of continuous dialogue texts can be taken from the dialogue text of the target dialogue to obtain the dialogue count. The beneficial effect is that the rolling window-style conversation text can realize real-time business scenario intention recognition, which can maximize the reuse of multiple rounds of information in a conversation and enable continuous correction with high accuracy and stability.

In a second aspect, the present application provides a speech recommendation device, which includes: a processing module, which is used to determine the characteristic information of the first customer and the characteristic information of the first service staff in a target conversation, and the target conversation is a conversation between the first customer and the first service staff; the processing module is also used to parse the conversation data of the target conversation to determine the business scenario intention corresponding to the target conversation, and the conversation data includes the conversation context of the target conversation; a recommendation module, which is used to obtain recommended speech based on the target data, and the target data includes: the characteristic information of the first customer, the characteristic information of the first service staff, and the business scenario intention corresponding to the target conversation, and the recommended speech is an optional speech for replying to the first customer; a display module, which is used to provide a recommended speech display interface, and the recommended speech display interface is used to display the recommended speech.

In a possible implementation, the characteristic information of the first service personnel includes at least one of the following: the ability level of the first service personnel, the business team of the first service personnel, the sales status of the first service personnel, and the sales history of the first service personnel.

In one possible implementation, the recommendation script is a script for introducing the target product; the target product satisfies at least one of the following characteristics: the target product corresponds to the ability level of the first service personnel, the target product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the target product is greater than a first preset threshold, the target product belongs to the industry that the first service personnel is good at, and the number of sales history orders of the target product by the first service personnel is greater than a second preset threshold.

In a possible implementation, the processing module is further used to determine at least one preferred product of the first customer based on the characteristic information of the first customer and the characteristic information of the first service personnel, and the target data also includes at least one preferred product.

In one possible implementation, the processing module is specifically used to input the characteristic information of the first customer and the characteristic information of the first service personnel into the first model to obtain at least one preferred product output by the first model; wherein the first model is obtained by training the first initial model with the first training data, and the first training data includes: the characteristic information of the second customer, the characteristic information of the second service personnel and the historical transaction products of the second customer.

In one possible implementation, the processing module is specifically used to input the conversation data into the second model to obtain the business scenario intention corresponding to the target conversation output by the second model; wherein the second model is obtained by training the second initial model with second training data, and the second training data includes: historical conversation data of historical conversations between a third customer and a third service staff, and the business scenario intention corresponding to the historical conversations.

In one possible implementation, the recommendation module is specifically used to: input the target data into a third model to obtain a model recall result output by the third model; and obtain recommended words based on the model recall result; wherein the third model is obtained by training a third initial model with third training data, and the third training data includes: feature information of a fourth customer, feature information of a fourth service staff, business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff, and preset words.

In one possible implementation, the device also includes: a storage module for storing first vector data and preset words in a vector database, the first vector data including: a vector of characteristic information of a fourth customer, a vector of characteristic information of a fourth service staff, and a vector of business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff.

In a possible implementation, the recommendation module is also used to perform vector retrieval based on the second vector data to obtain a vector recall result, and the second vector data includes a vector of the target data; the recommendation module is specifically used to: weight the model recall result and the vector recall result to obtain a recommended speech.

In a possible implementation, the processing module is further configured to obtain a preset number of continuous dialogue texts from the dialogue text of the target dialogue to obtain dialogue data.

In a third aspect, the present application provides a computing device cluster, comprising at least one computing device, each computing device comprising a processor and a memory; the processor of at least one computing device is used to execute instructions stored in the memory of at least one computing device, so that the computing device cluster performs a method as described in any one of the first aspects.

In a fourth aspect, the present application provides a computer program product comprising instructions, wherein when the instructions are executed by a computing device cluster, the computing device cluster executes any method in the first aspect.

In a fifth aspect, the present application provides a computer-readable storage medium, comprising computer program instructions. When the computer program instructions are executed by a computing device cluster, the computing device cluster executes any method in the first aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the structure of a speech recommendation system provided in an embodiment of the present application;

FIG2 is a flow chart of a method for recommending speech skills according to an embodiment of the present application;

FIG3 is a flow chart of a model training method provided in an embodiment of the present application;

FIG4 is a schematic diagram of a model training process provided in an embodiment of the present application;

FIG5 is a schematic diagram of a speech recommendation process provided by an embodiment of the present application;

FIG6 is a business flow chart of a speech recommendation provided in an embodiment of the present application;

FIG7 is a block diagram of a speech recommendation device provided in an embodiment of the present application;

FIG8 is a block diagram of another speech recommendation device provided in an embodiment of the present application;

FIG9 is a schematic diagram of the structure of a computing device provided in an embodiment of the present application;

FIG10 is a schematic diagram of the structure of a computing device cluster provided in an embodiment of the present application;

FIG. 11 is a schematic diagram of the structure of another computing device cluster provided in an embodiment of the present application.

DETAILED DESCRIPTION

In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

The terms "first", "second", etc. in the specification embodiments, claims, and drawings of the present application are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, including a series of steps or units. The method, system, product, or device is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And/or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and/or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character "/" generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

Please refer to Figure 1, which is a structural diagram of a speech recommendation system provided in an embodiment of the present application. The system can be a product form for internal services, and can also be used for commercial speech assistance. The system includes: an online service module, a data management module, a model management module, and an annotation module. Among them, the data management module is used to centrally manage the data used in the embodiment of the present application, and it can include a real-time data storage submodule and an offline data storage submodule. The online service module can be integrated by other systems, which is used to provide an online application programming interface (application programming interface, API) to serve the speech recommendation service, and store the real-time data in the speech recommendation process to the real-time data storage submodule. For example, when the embodiment of the present application recommends speech through a model, the online service module can provide an online service API to serve the deployed model. The model management module is used to design a model algorithm, and use training data for model training and model deployment. For example, it can include a language model management submodule, a large model management submodule, and a vector data management submodule. The annotation module is used to annotate the selected corpus data.

It should be noted that the structure of the system shown in FIG. 1 is for exemplary purposes only, and the modules included in the structure can be determined based on the speech recommendation method provided in the embodiment of the present application, and the embodiment of the present application does not limit this.

The present application embodiment provides a method for recommending speech skills, which can be used to assist service personnel in recommending speech skills, and can also be used in customer support scenarios. Service personnel can include sales personnel or customer service personnel. Please refer to Figure 2, which is a flow chart of a method for recommending speech skills provided by the present application embodiment, which can be applied to a speech skill recommendation system, such as the speech skill recommendation system shown in Figure 1. The method can include the following process:

101. Determine characteristic information of a first customer and characteristic information of a first service staff in a target dialogue, where the target dialogue is a dialogue between the first customer and the first service staff.

The first customer and the first service staff are in a communication scenario. The characteristic information of the first customer can be obtained from the customer portrait library of the first customer, which may include at least one of the following: identity information, scenario status, behavior records, business information, etc. Among them, the identity information may include, for example, at least one of the following: Vx (such as V1, V2, V3, etc.) members and customized labels of customers by service staff (such as top customers and specialization, etc.). The scenario status may include, for example, at least one of the following: activity preferences, connection status, and return visit status. Behavior records may include: service access preferences (such as cloud service access preferences). Business information may include at least one of the following: grouping of the first customer by the service staff, cloud service usage, and cancellation status, etc.

The characteristic information of the first service personnel can be obtained from the service portrait library of the first service personnel. For example, the first service personnel is a salesperson. The characteristic information of the first service personnel is a collection of the service personnel's capability attribute fields and labels, which is used to characterize the service personnel's capabilities. It may include at least one of the following: the capability level of the first service personnel, the business team of the first service personnel, the sales status of the first service personnel, the sales history of the first service personnel, etc. Among them, the capability level may, for example, include at least one of the following: elementary, intermediate, and advanced. The higher the capability level of the first service personnel, the higher the sales capability of the first service personnel. The business team may, for example, include: a new development team and/or an existing team, etc. The sales status may, for example, include: customer satisfaction and/or industry expertise, etc. The sales history may, for example, include at least one of the following: the number of customers who have completed orders, the number of cloud services that have completed orders, the total amount of orders, and the total cumulative call duration, etc. The aforementioned characteristic information is only for illustrative purposes and does not constitute a limitation.

For example, at least one preferred product of the first customer can also be determined based on the characteristic information of the first customer and the characteristic information of the first service staff. In one example, the characteristic information of the first customer and the characteristic information of the first service staff can be input into the first model to obtain at least one preferred product of the first customer output by the first model. As shown in Figure 1, the first model can be called through the online service module. The first model is obtained by training with the first training data, and the first training data includes: the characteristic information of the second customer, the characteristic information of the second service staff, and the historical transaction products of the second customer. Among them, the characteristic information of the second customer can be expressed as <C-Tags′> (custom tags), the characteristic information of the second service staff can be expressed as <S-Tags′> (salesperson tags), and the historical transaction products of the second customer can be expressed as <P-list′> (product list).

The second customer and the second service staff are persons who have communicated with each other before, and the historical transaction products refer to the list of products purchased by the second customer through communication with the second service staff. It should be noted that the second customer and the second service staff are not specific, and a second customer and a second service staff are regarded as a combination, and the number of such combinations is multiple. The first training data includes data corresponding to multiple combinations, that is, the first training data includes multiple groups of the following data: <C-Tags′>, <S-Tags′>, <P-list′>.

In another example, a vector search can be performed in a first database (e.g., a first vector database) based on the first vector data to obtain at least one preferred product of the first customer recalled by the vector, and the first vector data includes a vector of feature information of the first customer and a vector of feature information of the first service personnel. The first database may store vector data corresponding to the first training data, and a vector search is performed based on the stored vector data. The vector data corresponding to the first training data includes multiple groups of the following data: a vector of <C-Tags′>, a vector of <S-Tags′>, and a vector of <P-list′>.

In another example, the two examples described above can be weighted. The feature information of the first customer and the feature information of the first service personnel are input into the first model to obtain the first model output result of the first model output. A vector search is performed in the first database based on the first vector data to obtain a first vector recall result. The first model output result and the first vector recall result are then weighted to obtain at least one preferred product of the first customer. Weighting the model output result and the vector retrieval result can effectively reduce the error of at least one preferred product of the first customer caused by data cold start or model error.

As shown in FIG. 1 , the online service module may store the characteristic information of the first customer, the characteristic information of the first service personnel, and at least one determined preferred product of the first customer in the real-time data storage submodule.

102. Parse the conversation data of the target conversation to determine the business scenario intention corresponding to the target conversation, the conversation data including the conversation context of the target conversation.

Business scenario intent may include: business scenario tags and/or intent tags. Business scenario tags may include sales scenario tags or objection scenario tags. Sales scenario tags include but are not limited to: order promotion, renewal, sales transfer, contract or non-cloud services, etc. Objection scenario tags include but are not limited to: price, comparison of friendly products or complaints, etc. The intent tags of the first customer and the first service staff may refer to the entities and concerns of the first customer's intentions, etc.

For example, a preset number of continuous conversation texts can be obtained from the conversation text of the target conversation to obtain conversation data. For example, a preset number of continuous conversation texts can be obtained from the conversation text of the target conversation by limiting the range through a dynamic window to obtain conversation data. Acquiring conversation data by rolling window mode can realize real-time business scenario intention recognition, which can maximize the reuse of multiple rounds of information in a conversation, and can continuously correct, with high accuracy and stability.

If the first customer and the first service personnel have a conversation via voice, that is, the target conversation is voice data, the target conversation can be first converted into a conversation text, and then the conversation data can be obtained based on the conversation text. For example, the target conversation can be converted into a conversation text by automatic speech recognition (ASR) technology.

In one example, the conversation data between the first customer and the first service personnel can be input into the second model to obtain the business scenario intention corresponding to the target conversation output by the second model. As shown in Figure 1, the second model can be called through the online service module. The second model is obtained by training with the second training data, and the second training data includes: the historical conversation data of the historical conversation between the third customer and the third service personnel and the business scenario intention corresponding to the historical conversation. The specific content of the business scenario intention can be referred to the above description, and the embodiments of the present application will not be repeated here.

Referring to the description of the first training data in the aforementioned process 101, the third customer and the third service personnel are not specific, and a third customer and a third service personnel are regarded as a combination, and the number of such combinations is multiple. The training data includes historical conversation data of multiple combinations of historical conversations and the business scenario intentions corresponding to each historical conversation.

In another example, vector retrieval can be performed in a second database (e.g., a second vector database) based on the second vector data to obtain the business scenario intent corresponding to the target conversation recalled by the vector, and the second vector data includes the vector of the conversation data between the first customer and the first service personnel. The second database may store vector data corresponding to the second training data, and vector retrieval is performed based on the vector data corresponding to the second training data. The vector data corresponding to the second training data includes multiple groups of the following data: vectors of historical conversation data and vectors of business scenario intent corresponding to historical conversations.

In another example, the two examples described above can be weighted. The conversation data between the first customer and the first service personnel is input into the second model to obtain the second model output result output by the second model. A vector search is performed in the second database based on the second vector data to obtain a second vector recall result. The second model output result and the second vector recall result are then weighted to obtain the business scenario intent corresponding to the target conversation. Weighted processing of the model output result and the vector retrieval result can effectively reduce the error in the business scenario intent corresponding to the target conversation caused by data cold start or model error.

As shown in Figure 1, the conversation data between the first customer and the first service personnel can be stored in the real-time data storage submodule. The online service module can store the business scenario intention corresponding to the determined target conversation in the real-time data storage submodule.

For example, keyword extraction can also be performed on the conversation data to obtain a first keyword group (Words). Based on the conversation data and the first keyword group, the business scenario intention corresponding to the target conversation is determined. The first keyword group can be obtained by keyword extraction of the conversation data between the first customer and the first service personnel. The first keyword group can include business-related, product-related words and entities. The keyword extraction process can be performed in conjunction with a pre-established keyword library.

At this time, in the above example, the conversation data and the first keyword group can be input into the second model to obtain the business scenario intent corresponding to the target conversation output by the second model. The second training data also includes: a second keyword group, which is obtained by keyword extraction of historical conversation data. The second vector data also includes: a vector of the second keyword group. The vector data corresponding to the second training data includes multiple groups of the following data: a vector of historical conversation data, a vector of the second keyword group, and a vector of the business scenario intent corresponding to the historical conversation.

103. Obtain recommended words of speech based on the target data, where the target data includes: characteristic information of the first customer, characteristic information of the first service personnel, and business scenario intention corresponding to the target conversation, and the recommended words of speech are optional words of speech for replying to the first customer.

As shown in Figure 1, the target data may be stored in the real-time data storage submodule. The explanation of each data included in the target data may refer to the aforementioned processes 101 and 102, and the present embodiment will not be described in detail here.

Recommendation scripts may include, for example: answers to questions raised by customers, scripts to introduce products, solutions to objections raised by customers, and marketing activities related to recommended products.

For example, for service personnel with different characteristic information, the recommended words also meet different characteristics. Taking the words introducing products as an example, the recommended words of the first service personnel are words introducing the first product, and the number of the first product is at least one. The first product meets at least one of the following characteristics: the first product corresponds to the ability level of the first service personnel, the first product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the first product is greater than the first preset threshold, the first product belongs to the industry that the first service personnel is good at, and the number of sales history of the first product by the first service personnel is greater than the second preset threshold.

For example, the target data may also include at least one of the following: conversation data between the first customer and the first service personnel, at least one preferred product obtained in the aforementioned process 101, and a first keyword group (Words). The first keyword group may be obtained by extracting keywords from the conversation data between the first customer and the first service personnel, and the first keyword group may include business-related, product-related words and entities, etc. The keyword extraction process may be performed in conjunction with a pre-established keyword library.

In one example, the target data can be input into the third model to obtain the recommended words output by the third model. As shown in FIG1 , the third model can be called through the online service module. The third model is obtained by training with third training data, and the third training data includes: feature information of the fourth customer, feature information of the fourth service personnel, business scenario intentions corresponding to the historical conversations between the fourth customer and the fourth service personnel, and preset words.

For example, for service personnel with different characteristic information, the preset words also meet different characteristics. Taking the preset words for introducing products as an example, the preset words for the fourth service personnel are words for introducing the second product. The second product meets at least one of the following characteristics: the second product corresponds to the ability level of the fourth service personnel, the second product belongs to the business team of the fourth service personnel, the customer satisfaction obtained by the fourth service personnel in the history of selling the second product is greater than the first preset threshold, the second product belongs to the industry that the fourth service personnel is good at, and the number of sales history of the second product by the fourth service personnel is greater than the second preset threshold.

The data in the third training data other than the preset words correspond to the data in the target data one by one. The third training data may also include at least one of the following: historical conversation data, historical transaction products of the fourth customer, and a second keyword group (Words'). The second keyword group is obtained by extracting keywords from the historical conversation data. For example, when the third training data also includes historical conversation data, the target data also includes the first customer The target data also includes at least one preferred product of the first customer. When the third training data also includes the second keyword group, the target data also includes the first keyword group.

The explanation of each data included in the third training data can refer to the aforementioned processes 101 and 102, and the embodiments of the present application will not be repeated here. Referring to the description of the first training data of the aforementioned process 101, there are multiple combinations of fourth customers and fourth service personnel, and the third training data includes data corresponding to the multiple combinations. Taking the third training data also including: historical conversation data, historical transaction products of the fourth customer, and the second keyword group (Words′) as an example, the third training data includes multiple groups of the following data: <C-Tags′>, <S-Tags′>, <P-list′>, historical conversation data, business scenario intentions corresponding to historical conversations, Words′, and preset words.

In another example, vector retrieval can be performed in a third database (e.g., a third vector database) based on the third vector data to obtain recommended words for vector recall. The third vector data includes a vector of the target data, namely, a vector of the feature information of the first customer, a vector of the feature information of the first service personnel, and a vector of the business scenario intention corresponding to the target conversation. The third database may store vector data corresponding to the third training data, and vector retrieval is performed based on the vector data corresponding to the third training data. The vector data corresponding to the third training data includes multiple groups of the following data: vectors of <C-Tags′>, vectors of <S-Tags′>, vectors of business scenario intentions corresponding to historical conversations, and preset words.

Taking the third training data also including: historical conversation data, historical transaction products of the fourth customer, and the second keyword group (Words′) as an example, the vector data corresponding to the third training data includes multiple groups of the following data: vector of <C-Tags′>, vector of <S-Tags′>, vector of <P-list′>, vector of historical conversation data, vector of business scenario intention corresponding to historical conversations, vector of Words′, and preset words. The third vector data includes: vector of feature information of the first customer, vector of feature information of the first service staff, vector of at least one preferred product of the first customer, vector of conversation data, vector of business scenario intention corresponding to the target conversation, and vector of the first keyword group.

In another example, the two examples described above can be weighted. The target data is input into the third model to obtain the third model output result output by the third model. A vector search is performed in the third database according to the third vector data to obtain a third vector recall result. The third model output result and the third vector recall result are then weighted to obtain a recommended speech. Weighting the model output result and the vector retrieval result can effectively reduce the error of the recommended speech caused by data cold start or model error.

104. Provide a recommended script display interface, which is used to display recommended scripts.

The number of recommended scripts is at least one, and the first service staff can choose whether to click on a recommended script to reply to the first customer.

In summary, the method for recommending words of speech provided in the embodiment of the present application first determines the characteristic information of the first customer and the characteristic information of the first service personnel in the target conversation, and parses the conversation data of the target conversation to determine the business scenario intention corresponding to the target conversation, and then obtains the recommended words of speech based on the target data and provides a recommended words of speech display interface, which is used to display the recommended words of speech, and the target data includes: the characteristic information of the first customer, the characteristic information of the first service personnel and the business scenario intention corresponding to the target conversation, and the recommended words of speech are optional words for replying to the first customer, and the method combines the characteristic information of the customer and the characteristic information of the service personnel at the same time, and compared with the related technology, it can more effectively fit the personality characteristics of the service personnel and the customer to assist in the recommendation of words of speech, and comprehensively uses dynamic parameters (conversation text) and static parameters (characteristic information of the first customer and characteristic information of the first service personnel) to recommend words of speech, and compared with the related technology, it can effectively improve the accuracy and relevance of the word of speech recommendation, thereby assisting the service personnel to complete efficient services.

The following is an explanation of the process of recommending speech through the model, and takes the target data including: characteristic information of the first customer, characteristic information of the first service staff, at least one preferred product of the first customer, conversation data between the first customer and the first service staff, business scenario intention corresponding to the target conversation, and the first keyword group as an example. The process includes a model training process and a model recommendation process. Please refer to Figure 3, which is a flow chart of a model training method provided in an embodiment of the present application. The method can be applied to a speech recommendation system, such as the speech recommendation system shown in Figure 1. As shown in Figure 3, the method may include the following process:

201. Train a first initial model using first training data to obtain a first model, where the first training data includes: feature information of a second customer, feature information of a second service personnel, and historical transaction products of the second customer.

Among them, the characteristic information of the second customer can be obtained from the customer portrait library of the second customer, and the characteristic information of the second service personnel can be obtained from the service portrait library of the second service personnel. The interpretation of the characteristic information of the second customer and the characteristic information of the second service personnel can refer to the characteristic information of the first customer and the characteristic information of the first service personnel in the aforementioned process 101 respectively, and the embodiments of the present application are not elaborated here.

The first model can be called a preference model according to its function. The process 201 trains the first model with successful sales as the goal. The input of the trained first model is the characteristic information of the customer and the characteristic information of the service personnel, and the output is at least one preferred product of the customer. The first training data can be regarded as a matrix of customers, sales, customer satisfaction with the product and transaction results. The first training data may include multiple groups of the following data: <C-Tags′>, <S-Tags′>, <P-list′>.

As shown in FIG1 , the customer profile library of the second customer, the service profile library of the second service personnel, and the historical transaction products of the second customer can be stored in the offline data storage submodule. The design of the model training algorithm and the model training process can be performed by the model management module. For example, the model training algorithm includes but is not limited to: collaborative filtering (CF) algorithm and deep factor decomposition machine (deep factorization machines, DeepFM) algorithms, etc. The embodiments of the present application do not limit the model training algorithm.

202. Train a second initial model using second training data to obtain a second model, where the second training data includes: historical conversation data of historical conversations between a third customer and a third service personnel and business scenario intentions corresponding to the historical conversations.

For example, a preset number of continuous dialogue texts can be obtained from the dialogue texts of the historical dialogues between the third customer and the third service personnel to obtain historical dialogue data, thereby constructing a corpus. For example, a preset number of continuous dialogue texts can be obtained from the dialogue texts between the third customer and the third service personnel through a dynamic window limitation range to obtain historical dialogue data. Acquiring dialogue texts in a rolling window manner can maximize the reuse of multiple rounds of information in a dialogue, and can continuously correct, which can improve the output accuracy and stability of the second model. As shown in FIG1 , the corpus can be located in the offline data storage submodule.

The business scenario intention corresponding to the historical conversation includes: the business scenario label and/or intention label corresponding to the historical conversation, and its explanation can refer to the business scenario intention corresponding to the target conversation in the aforementioned process 102, and the embodiment of the present application will not be elaborated here.

In the embodiment of the present application, a preset business scenario label and a preset intention label can be established, a business scenario label and/or an intention label can be selected from the preset business scenario labels and the preset intention labels, and the business scenario intention corresponding to the historical conversation is marked as the selected business scenario label and/or intention label. The process of selecting a business scenario label and/or an intention label can be performed manually through the marking module shown in FIG1, and the embodiment of the present application does not limit this.

The number of business scenario tags is one or more, and the number of intent tags is one or more. For example, business scenario tags and intent tags can be established based on business scenario presetting and tag collection and collation. Business scenario tags and intent tags can be updated according to actual conditions, and this embodiment of the application does not limit this.

The second model can be called a scenario intent classification model according to its function. The input of the trained second model is the conversation data between the customer and the service personnel, and the output is the business scenario intent corresponding to the conversation. The second training data may include multiple groups of the following data: <Text, Intent>, Text (text) represents the historical conversation data, and its specific structure is {"sentence 1 <tos> sentence 2 <tos>... sentence n"}, <tos> is a specific separator used to separate different sentences. Intent (intent) represents the business scenario intent corresponding to the historical conversation. It should be noted that the multiple groups of data included in the second training data may correspond one-to-one to the multiple groups of data included in the first training data, or may not completely correspond, and the embodiments of the present application do not limit this. The correspondence here refers to the correspondence between the data of the same combination of customers and service personnel in the second training data and the first training data. For example, taking customer A1 and service staff A2 as an example, the A1-A2 related data in the first training data corresponds to the A1-A2 related data in the second training data. The A1-A2 related data in the first training data includes: A1's feature information, A2's feature information and A2's historical transaction products. The A1-A2 related data in the second training data includes: A1 and A2's historical conversation data and the business scenario intentions corresponding to the historical conversations between A1 and A2.

In the process 202, the business scenario label and the intent label are outputted using the same model. It is understandable that two models (business scenario model and intent model) can also be trained to output the business scenario and intent respectively. The business scenario model can be obtained by training the historical conversation data between the third customer and the third service personnel and the business scenario corresponding to the historical conversation. The intent model can be obtained by training the historical conversation data between the third customer and the third service personnel and the intent corresponding to the historical conversation. The embodiment of the present application does not limit the number of models.

As shown in Figure 1, the design of the model training algorithm and the process of model training can be performed by the model management module. For example, the model training algorithm can include: Bert (bidirectional encoder representations from transformers) algorithm, and the embodiment of the present application does not limit the model training algorithm.

203. A third initial model is trained by using third training data to obtain a third model. The third training data includes: characteristic information of a fourth customer, characteristic information of a fourth service personnel, historical transaction products of the fourth customer, historical conversation data of historical conversations between the fourth customer and the fourth service personnel, business scenario intentions corresponding to the historical conversations, a second keyword group, and preset words. The second keyword group is obtained by extracting keywords from the historical conversation data.

For example, the basic corpus of the first model and the second model (including the first training data and the second training data) can be sorted and combined, that is, the two corresponding groups of data in the first training data and the second training data can be combined into one group of data. And based on all the combined data, at least one preset speech (also called a business template, which is text information customized according to business scenarios and product intentions, etc.) is enumerated, and the preset speech is annotated for each combined data to obtain the third training data of the fusion context. The fusion context is a data structure that integrates multiple information.

The third model can be called a recommendation model according to its function. The input of the trained third model is: the customer's characteristic information, the service staff's characteristic information, at least one preferred product of the customer, the conversation data between the customer and the service staff, the business scenario intention corresponding to the conversation between the customer and the service staff, and the keyword group extracted from the conversation data between the customer and the service staff. The output is the words recommended to the service staff.

The third training data may include multiple groups of the following data: <Tags′>, <P-list′>, Cw′, Words′, <Answers>. Each element in a group of data may be separated by <eos>. <Tags′> represents the characteristic information of the fourth customer and the characteristic information of the fourth service personnel, and the characteristic information may be separated by the "##" symbol. <P-list′> may refer to the explanation of the first training data, and the "##" symbol may be used to separate the products. Cw′ is a composite data structure, which includes the conversation window information (i.e., historical conversation data and the speaker of each sentence) and the business scenario intention corresponding to the historical conversation. Words′ represents the second keyword group, <Answers> represents the preset words, and when there are multiple preset words, the "##" symbol can be used to separate each preset word.

It should be noted that the multiple groups of data in the first training data, the second training data, and the third training data may correspond one to one or may not correspond completely, and the embodiments of the present application do not limit this. The correspondence here refers to the data correspondence between the same combination of customers and service personnel in the first training data, the second training data, and the third training data. For example, taking customer A1 and service personnel A2 as an example, the A1-A2 related data in the first training data, the A1-A2 related data in the second training data, and the A1-A2 related data in the third training data correspond. The A1-A2 related data in the first training data and the A1-A2 related data in the second training data can refer to the relevant explanations of the aforementioned process 202, and the A1-A2 related data in the third training data includes: feature information of A1, feature information of A2, historical transaction products of A2, historical conversation data of historical conversations between A1 and A2, business scenario intentions corresponding to historical conversations between A1 and A2, and preset words corresponding to A1 and A2.

As shown in Figure 1, the design of the model training algorithm and the model training process can be performed by the model management module. For example, the third initial model can be a large model that can be fine-tuned (such as a large language model), and the large model can be fine-tuned through the third training data to obtain a third model, and the third model is a recommended large model. The embodiment of the present application does not limit the model training process.

For example, please refer to Figure 4, which is a schematic diagram of a model training process provided in an embodiment of the present application, and Figure 4 shows the training process of the preference model (i.e., the first model) M1, the scene intent classification model (i.e., the second model) M2, and the large language model (i.e., the third model) M3. As shown in Figure 4, the feature information of the second customer and the feature information of the second service staff are constructed, and the preference model M1 is obtained by training the feature information of the second customer, the feature information of the second service staff, and the historical transaction products of the second customer. The scene intent allocation model M2 is obtained by training the historical conversation data of the historical conversation between the third customer and the third service staff and the business scene intent corresponding to the historical conversation. Afterwards, a fusion context is constructed based on the preference model M1 and the scene intent allocation model M2, and the large language model M3 is fine-tuned using the data of the constructed fusion context.

With reference to the foregoing description, the embodiment of the present application can combine the model output with the vector retrieval when performing speech recommendation. By way of example, as shown in FIG4 , the vector data can be stored in a database (e.g., a vector database), and then the vector retrieval can be performed in the database to implement speech recommendation. The vector data may include multiple groups of the following data: [V1′, V2′, V3′, V4′], <Answers>. V1′ represents the vector of <Tags′>, V2′ represents the vector of <P-list′>, V3′ represents the vector of Cw′, and V4′ represents the vector of Words′. By way of example, the data in the third training data can be converted into vectors through the word vector capability of the unified language model.

The following describes the process of recommending speech through the above three models. Please refer to Figure 5, which is a schematic diagram of a speech recommendation process provided in an embodiment of the present application. As shown in Figure 5, the characteristic information <C-Tags> of the first customer and the characteristic information <S-Tags> of the first service personnel are determined. Then call the preference model M1 to input <C-Tags> and <S-Tags> into the preference model M1, and obtain at least one preferred product <P-list> of the first customer output by the preference model M1. Then construct the context information as the static parameter Context-s (<C-Tags>, <S-Tags>, <P-list>).

The scenario intention allocation model M2 is called to input the conversation data between the first customer and the first service personnel into the scenario intention allocation model M2, and the business scenario intention corresponding to the target conversation output by the scenario intention allocation model M2 is obtained. Keywords are extracted from the conversation data between the first customer and the first service personnel to obtain a first keyword group.

Then, a fusion context Context-d (Cw, Words) is constructed, where Cw includes the conversation data between the first customer and the first service personnel and the business scenario intention corresponding to the target conversation, and Words represents the first keyword group.

Then continue to build the fusion context, fuse Context-s and Context-d to get Context (<Tags>, <P-list>, Cw, Words), and call the large language model M3 to input the Context object into the large language model M3 to get the model recall result output by the large language model M3. And get the recommended words based on the model recall result.

For example, [V1, V2, V3, V4] can also be obtained based on the aforementioned Context-s and Context-d, where V1 represents the vector of <Tags>, V2 represents the vector of <P-list>, V3 represents the vector of Cw, and V4 represents the vector of Words. Then, as shown in FIG5 , vector retrieval is performed based on [V1, V2, V3, V4] to obtain the vector recall result. The model recall result and the vector recall result are weighted to obtain the recommended words. The recommended words can be displayed in the form of a recommended words list.

For example, a weighted comparison can be made between the model recall result and the vector recall result, and the recommended words can be output according to the preset rules. The preset rules can be relevance priority or similarity priority. Relevance priority means that the model recall result is ranked first in the recommended words list, and the vector recall result is ranked last. Similarity priority means that the vector recall result is ranked first in the recommended words list, and the model recall result is ranked last. In the embodiment of the present application, the preset rules that take effect can be selected according to the rule configuration parameters.

The process of vector retrieval may specifically include: comparing [V1, V2, V3, V4] with [V1′, V2′, V3′, V4′] stored in the database to obtain the most approximate [V1′, V2′, V3′, V4′], and calling the <Answers> corresponding to the most approximate [V1′, V2′, V3′, V4′]. Back.

In the embodiment of the present application, the aforementioned data can be converted into vectors by word embedding technology. For example, <Tags> can be converted into V1, <P-list> can be converted into V2, Cw can be converted into V3, and Words can be converted into V4 by word embedding technology.

The following is an explanation of the method provided in the embodiment of the present application with a specific example. Please refer to FIG6 , which is a business flow chart of a speech recommendation provided in the embodiment of the present application. FIG6 further explains the aforementioned process in detail.

First, process 1 is executed: establishing a dialogue window between the first customer and the first service personnel. Then process 2 is executed: understanding the business scenario intention. This process 2 can be executed by the scenario intention classification model M2. The scenario intention classification model M2 in FIG6 can include multiple sub-models to perform various recognition processes.

As shown in Figure 6, process 2 specifically includes the following processes: First, keyword extraction is performed, and the input of the keyword extraction process is N rounds of conversations in the conversation window, and the output is a keyword group (such as bold words in the conversation window). Then, sales scenario recognition is performed. The input of the sales scenario recognition process is N rounds of conversations and keyword groups, and the output is a sales scenario (including promotion/renewal/transfer/non-cloud services/others). And objection scenario recognition is performed. The input of the objection scenario recognition process is N rounds of conversations and keyword groups, and the output is an objection scenario (including price/friendly business/complaint). And product intent recognition is performed. The input of the product intent recognition process is N rounds of conversations and keyword groups, and the output is the entities and concerns of the first customer's intention.

The aforementioned process 2 divides the business scenario intention recognition of process 102 into three processes: sales scenario recognition, objection scenario recognition, and product intention recognition. When the process is executed by the scenario intention classification model M2, the scenario intention classification model M2 may include a sales scenario recognition sub-model, an objection scenario recognition sub-model, and a product intention recognition sub-model, which are respectively used to execute the three processes. The sales scenario recognition sub-model is obtained by training the historical dialogue data of the historical dialogue between the third customer and the third service personnel, the second keyword group, and the sales scenario label corresponding to the historical dialogue. The objection scenario recognition sub-model is obtained by training the historical dialogue data of the historical dialogue between the third customer and the third service personnel, the second keyword group, and the objection scenario label corresponding to the historical dialogue. The product intention recognition sub-model is obtained by training the historical dialogue data of the historical dialogue between the third customer and the third service personnel, the second keyword group, and the product intention label corresponding to the historical dialogue. It can be understood that there can be other division combinations, such as combining sales scenario recognition and objection scenario recognition into one scenario recognition process, or merging these three processes into one recognition process, which is not limited in the embodiments of the present application.

Then, process 3 is executed. This process 3 can be executed by the large language model M3, or by fusing the large language model M3 and the vector retrieval process. In Figure 6, the large language model M3 can include multiple sub-models to perform various recommendation processes. Figure 6 divides process 3 into two parts: 3.1 recommending sales scripts and 3.2 recommending products/solutions. In 3.1, the preset sales scripts are first retrieved. The input of the process of retrieving the preset sales scripts is the characteristic information of the first customer, the characteristic information of the first service personnel, the previously output sales scenarios, objection scenarios, and keyword groups, and the output is M sales scripts selected from the preset sales scripts. Then, relevance sorting is performed. The input of the relevance sorting process is the previously output M sales scripts, sales scenarios, and objection scenarios, and the output is the top X sales scripts among the M sales scripts.

In 3.2, first, a list of products/solutions is obtained. The input of the process of obtaining the list of products/solutions is the characteristic information of the first customer, the characteristic information of the first service personnel, and the entities and concerns of the first customer's intentions previously output, and the output is Y related products. Then, relevance sorting is performed. The input of the relevance sorting process is the concerns of the first customer's intentions previously output, the Y related products, and the characteristic information of the first customer, and the output is the first Z related products among the Y related products.

For the Z related products outputted in 3.2, 3.3 may be continued to be performed to recommend marketing activities. The process 3.3 includes associating the Z related products with the marketing activities in the activity library and determining the timing of the promotion activities.

Then, process 4 is executed. In process 4, the output contents of 3.1 or 3.2 and 3.3 are assembled to obtain the final list of words displayed on the platform. In process 4, manual strategies (rules) can be integrated to assemble the results, and manual intervention can be made on the results, such as setting some results not to be displayed. As shown in Figure 6, the last paragraph of the first service staff is the words selected by the first service staff from the displayed list of words.

The aforementioned process 3 divides the speech recommendation of process 103 into two processes: sales speech recommendation and product/solution recommendation. When the process is executed by the large language model M3, the large language model M3 may include a sales speech recommendation sub-model and a product recommendation sub-model, which are respectively used to execute the two processes. The sales speech recommendation sub-model is obtained through the characteristic information of the fourth customer, the characteristic information of the fourth service personnel, the business scenario label corresponding to the historical conversation, the second keyword group, and the preset sales speech training. The product recommendation sub-model is obtained through the characteristic information of the fourth customer, the characteristic information of the fourth service personnel, the intention label corresponding to the historical conversation between the fourth customer and the fourth service personnel, and the preset product training. It can be understood that there can be other division combinations, such as combining the sales speech recommendation and the product/solution recommendation into one recommendation process, which is not limited in the embodiments of the present application.

When both the aforementioned processes 2 and 3 are implemented through the model, as shown in FIG6 , process 5: manual labeling process needs to be executed during the model training process. In process 5, dialogue window data is selected from the dialogue corpus, and then keyword groups are labeled for the dialogue window data, so as to obtain a model through training the dialogue window data and the labeled keyword groups, and the model is used to implement the keyword extraction process in the aforementioned process 2.

In process 5, sales scene labels can also be annotated for the dialogue window data, so as to obtain a model through training the dialogue window data and the annotated sales scene labels. The model is used to implement the sales scene recognition process in the aforementioned process 2.

In process 5, objection scene labels can also be annotated for the dialogue window data, so as to obtain a model through training the dialogue window data and the annotated objection scene labels, and the model is used to implement the objection scene identification process in the above-mentioned process 2.

In process 5, product intent labels can also be annotated for the dialogue window data, so as to obtain a model through training the dialogue window data and the annotated product intent labels. The model is used to implement the product intent recognition process in the aforementioned process 2.

In the embodiment of the present application, feedback data of the recommendation results can be obtained. As shown in FIG6 , transaction data and click data based on the recommendation words can be obtained to obtain product solution feedback samples, based on which the manually labeled results can be modified, and the first model to the third model can be retrained to achieve model update.

It should be noted that the input data of each model shown in FIG6 is for illustrative purposes only, and there may be other combinations. For details, please refer to the above description, and the embodiments of the present application will not be elaborated here.

In summary, the method for recommending words of speech provided in the embodiment of the present application trains a first initial model with first training data to obtain a first model, the first training data includes: characteristic information of the second customer, characteristic information of the second service personnel, and historical transaction products of the second customer, trains a second initial model with second training data to obtain a second model, the second training data includes: historical conversation data of historical conversations between a third customer and a third service personnel, and business scenario intentions corresponding to the historical conversations, trains a third initial model with third training data to obtain a third model, the third training data includes: characteristic information of a fourth customer, characteristic information of a fourth service personnel, historical transaction products of the fourth customer, historical conversation data of historical conversations between the fourth customer and the fourth service personnel, business scenario intentions corresponding to the historical conversations, a second keyword group, and preset words of speech, the second keyword group is obtained by extracting keywords from the historical conversation data, and the third initial model is obtained by training the third initial model with third training data. The characteristic information of a customer and the characteristic information of the first service staff are input into the first model to obtain at least one preferred product of the first customer output by the first model, the conversation data between the first customer and the first service staff is input into the second model to obtain the business scenario intention corresponding to the target conversation output by the second model, and then the target data is input into the third model to obtain the recommended speech based on the model recall result output by the third model. The target data includes: the characteristic information of the first customer, the characteristic information of the first service staff, at least one preferred product of the first customer, the conversation data and the business scenario intention corresponding to the target conversation. The method trains and fine-tunes the model through the data structure of the fused context, and the obtained large model can recommend speech according to the data structure of the fused context. At the same time, combined with the customer characteristic information and the characteristic information of the service staff, compared with the related technology, it can more effectively fit the personality characteristics of the service staff and the customer to assist in the recommendation of speech.

In addition, the combined use of dynamic parameters (conversation text) and static parameters (feature information of the first customer and feature information of the first service personnel) for speech recommendation can effectively improve the accuracy and relevance of speech recommendation compared to related technologies, thereby assisting service personnel to provide efficient services. For example, in the sales process over the phone or other channels that can directly or indirectly generate text data, accurate recommendations for sales speech and products can be achieved, assisting sales personnel to complete efficient sales. In addition, the first model, the second model, and the third model can achieve self-closed-loop sustainable iterative updates, thereby further achieving the real-time and accuracy of speech recommendations.

The sequence of the methods provided in the embodiments of the present application can be adjusted appropriately, and the process can be increased or decreased accordingly according to the situation. Any technician familiar with the technical field can easily think of a method of change within the technical scope disclosed in this application, which should be included in the protection scope of this application, and the embodiments of the present application do not limit this.

The embodiment of the present application also provides a speech recommendation device, as shown in FIG7 , in the case of dividing each functional module according to each function, the speech recommendation device 300 may include: a processing module 301, a recommendation module 302 and a display module 303. Exemplarily, the speech recommendation device may be a speech recommendation device, or a chip therein or other combined devices, components, etc. having the functions of the above-mentioned speech recommendation device. The functions of each module of the device are as follows:

A processing module is used to determine the characteristic information of the first customer and the characteristic information of the first service staff in the target dialogue, and the target dialogue is the dialogue between the first customer and the first service staff; the processing module is also used to parse the dialogue data of the target dialogue to determine the business scenario intention corresponding to the target dialogue, and the dialogue data includes the dialogue context of the target dialogue; a recommendation module is used to obtain recommended words based on the target data, and the target data includes: the characteristic information of the first customer, the characteristic information of the first service staff and the business scenario intention corresponding to the target dialogue, and the recommended words are optional words for replying to the first customer; a display module is used to provide a recommended words display interface, and the recommended words display interface is used to display the recommended words.

In combination with the above solution, the characteristic information of the first service personnel includes at least one of the following: the ability level of the first service personnel, the business team of the first service personnel, the sales status of the first service personnel, and the sales history of the first service personnel.

In combination with the above scheme, the recommended script is the script for introducing the target product; the target product meets at least one of the following characteristics: the target product corresponds to the ability level of the first service personnel, the target product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the target product is greater than a first preset threshold, the target product belongs to the industry that the first service personnel is good at, and the number of historical sales orders of the target product by the first service personnel is greater than a second preset threshold.

In combination with the above solution, the processing module is further used to determine at least one preferred product of the first customer based on the characteristic information of the first customer and the characteristic information of the first service personnel, and the target data also includes at least one preferred product.

In combination with the above scheme, the processing module is specifically used to input the characteristic information of the first customer and the characteristic information of the first service personnel into the first model to obtain at least one preferred product output by the first model; wherein the first model is obtained by training the first initial model with the first training data, and the first training data includes: the characteristic information of the second customer, the characteristic information of the second service personnel and the historical transaction products of the second customer.

In combination with the above scheme, the processing module is specifically used to input the conversation data into the second model to obtain the business scenario intention corresponding to the target conversation output by the second model; wherein the second model is obtained by training the second initial model with second training data, and the second training data includes: historical conversation data of historical conversations between the third customer and the third service personnel and the business scenario intentions corresponding to the historical conversations.

In combination with the above scheme, the recommendation module is specifically used to: input the target data into the third model to obtain the model recall result output by the third model; and obtain the recommended words based on the model recall result; wherein the third model is obtained by training the third initial model with the third training data, and the third training data includes: the characteristic information of the fourth customer, the characteristic information of the fourth service personnel, the business scenario intentions corresponding to the historical conversations between the fourth customer and the fourth service personnel, and the preset words.

In combination with the above solution, please refer to Figure 8, which is a block diagram of another speech recommendation device provided in an embodiment of the present application. Based on Figure 7, the device also includes: a storage module 304.

A storage module is used to store first vector data and preset words in a vector database, wherein the first vector data includes: a vector of characteristic information of a fourth customer, a vector of characteristic information of a fourth service staff, and a vector of business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff.

In combination with the above-mentioned solution recommendation module, it is also used to perform vector retrieval based on the second vector data to obtain a vector recall result, and the second vector data includes a vector of the target data; the recommendation module is specifically used to: weight the model recall result and the vector recall result to obtain a recommended speech.

In combination with the above solution, the processing module is also used to obtain a preset number of continuous dialogue texts from the dialogue text of the target dialogue to obtain dialogue data.

The cloud management platform provided in this application includes a processing module, a recommendation module, a display module, and a storage module.

Among them, the processing module, the recommendation module, the display module, and the storage module can all be implemented by software, or can be implemented by hardware. Exemplarily, the implementation of the processing module is introduced below by taking the processing module as an example. Similarly, the implementation of the recommendation module, the display module, and the storage module can refer to the implementation of the processing module.

As an example of a software functional unit, a processing module may include code running on a computing instance. Among them, the computing instance may include at least one of a physical host (computing device), a virtual machine, and a container. Furthermore, the above-mentioned computing instance may be one or more. For example, a processing module may include code running on multiple hosts/virtual machines/containers. It should be noted that the multiple hosts/virtual machines/containers used to run the code may be distributed in the same region or in different regions. Furthermore, the multiple hosts/virtual machines/containers used to run the code may be distributed in the same availability zone (AZ) or in different AZs, each AZ including one data center or multiple data centers with close geographical locations. Among them, usually a region may include multiple AZs.

Similarly, multiple hosts/virtual machines/containers used to run the code can be distributed in the same virtual private cloud (VPC) or in multiple VPCs. Usually, a VPC is set up in a region. For cross-region communication between two VPCs in the same region and between VPCs in different regions, a communication gateway needs to be set up in each VPC to achieve interconnection between VPCs through the communication gateway.

As an example of a hardware functional unit, a processing module may include at least one computing device, such as a server, etc. Alternatively, the processing module may also be a device implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD). The PLD may be a complex programmable logical device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL) or any combination thereof.

The multiple computing devices included in the processing module can be distributed in the same region or in different regions. The multiple computing devices included in the processing module can be distributed in the same AZ or in different AZs. Similarly, the multiple computing devices included in the processing module can be distributed in the same VPC or in multiple VPCs. The multiple computing devices can be any combination of computing devices such as servers, ASICs, PLDs, CPLDs, FPGAs, and GALs.

It should be noted that, in other embodiments, the processing module can be used to execute any step in the speech recommendation method, the recommendation module can be used to execute any step in the speech recommendation method, the display module can be used to execute any step in the speech recommendation method, and the storage module can be used to execute any step in the speech recommendation method. Used to execute any step in the speech recommendation method. The steps that the processing module, recommendation module, display module, and storage module are responsible for implementing can be specified as needed, and the processing module, recommendation module, display module, and storage module respectively implement different steps in any method of claims 1 to 10 to realize all the functions of the cloud management platform.

The present application also provides a speech recommendation system, including a cloud management platform and infrastructure.

The cloud management platform and the infrastructure can be implemented by software or hardware. As an example, the implementation of the cloud management platform is introduced below. Similarly, the implementation of the infrastructure can refer to the implementation of the cloud management platform.

As an example of a software functional unit, a cloud management platform may include code running on a computing instance. The computing instance may be at least one of a physical host (computing device), a virtual machine, a container, and other computing devices. Furthermore, the computing device may be one or more. For example, the cloud management platform may include code running on multiple hosts/virtual machines/containers. It should be noted that the multiple hosts/virtual machines/containers used to run the application may be distributed in the same region or in different regions. The multiple hosts/virtual machines/containers used to run the code may be distributed in the same AZ or in different AZs, each AZ including one data center or multiple data centers with close geographical locations. Typically, a region may include multiple AZs.

Similarly, multiple hosts/virtual machines/containers used to run the code can be distributed in the same VPC or in multiple VPCs. Usually, a VPC is set up in a region. For cross-region communication between two VPCs in the same region and between VPCs in different regions, a communication gateway must be set up in each VPC to achieve interconnection between VPCs through the communication gateway.

As an example of a hardware functional unit, the cloud management platform may include at least one computing device, such as a server, etc. Alternatively, the cloud management platform may also be a device implemented using ASIC or PLD, etc. The PLD may be implemented using CPLD, FPGA, GAL or any combination thereof.

The multiple computing devices included in the cloud management platform can be distributed in the same region or in different regions. The multiple computing devices included in the cloud management platform can be distributed in the same AZ or in different AZs. Similarly, the multiple computing devices included in the speech recommendation device can be distributed in the same VPC or in multiple VPCs. The multiple computing devices can be any combination of computing devices such as servers, ASICs, PLDs, CPLDs, FPGAs, and GALs.

The present application also provides a computing device 400. As shown in FIG9 , the computing device 400 includes: a bus 402, a processor 404, a memory 406, and a communication interface 408. The processor 404, the memory 406, and the communication interface 408 communicate with each other through the bus 402. The computing device 400 can be a server or a terminal device. It should be understood that the present application does not limit the number of processors and memories in the computing device 400.

The bus 402 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, FIG. 9 is represented by only one line, but does not mean that there is only one bus or one type of bus. The bus 402 may include a path for transmitting information between various components of the computing device 400 (e.g., the memory 406, the processor 404, and the communication interface 408).

Processor 404 may include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

The memory 406 may include a volatile memory (volatile memory), such as a random access memory (RAM). The processor 404 may also include a non-volatile memory (non-volatile memory), such as a read-only memory (ROM), a flash memory, a hard disk drive (HDD), or a solid state drive (SSD).

The memory 406 stores executable program codes, and the processor 404 executes the executable program codes to respectively implement the functions of the aforementioned processing module, recommendation module, display module, and storage module, thereby implementing any one of the speech recommendation methods of claims 1 to 10. That is, the memory 406 stores instructions for executing any one of the speech recommendation methods of claims 1 to 10.

The communication interface 408 uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the computing device 400 and other devices or a communication network.

The embodiment of the present application also provides a computing device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device such as a desktop computer, a laptop computer, or a smart phone.

As shown in Fig. 10, the computing device cluster includes at least one computing device 400. The memory 406 in one or more computing devices 400 in the computing device cluster may store the same instructions for executing any one of the speech recommendation methods of claims 1 to 10.

In some possible implementations, the memory 406 of one or more computing devices 400 in the computing device cluster may also respectively store partial instructions for executing any one of the speech recommendation methods of claims 1 to 10. In other words, the combination of one or more computing devices 400 may jointly execute instructions for executing any one of the speech recommendation methods of claims 1 to 10.

It should be noted that the memory 406 in different computing devices 400 in the computing device cluster can store different instructions, which are respectively used to execute part of the functions of the cloud management platform. That is, the instructions stored in the memory 406 in different computing devices 400 can implement the functions of one or more modules among the processing module, the recommendation module, the display module and the storage module.

In some possible implementations, one or more computing devices in a computing device cluster may be connected via a network. The network may be a wide area network or a local area network, etc. FIG. 11 shows a possible implementation. As shown in FIG. 11 , two computing devices 400A and 400B are connected via a network. Specifically, the network is connected via a communication interface in each computing device. In this type of possible implementation, the memory 406 in the computing device 400A stores instructions for executing the functions of the processing module and the recommendation module. At the same time, the memory 406 in the computing device 400B stores instructions for executing the functions of the display module and the storage module.

The connection method between the computing device clusters shown in Figure 11 can be that considering that any of the speech recommendation methods of claims 1 to 10 provided in the present application requires speech recommendation, it is considered to entrust the functions implemented by the display module and the storage module to the computing device 400B for execution.

It should be understood that the functions of the computing device 400A shown in FIG11 may also be completed by multiple computing devices 400. Similarly, the functions of the computing device 400B may also be completed by multiple computing devices 400.

The embodiment of the present application also provides a computer program product including instructions. The computer program product may be a software or program product including instructions that can be run on a computing device or stored in any available medium. When the computer program product is run on at least one computing device, the at least one computing device executes any one of the speech recommendation methods of claims 1 to 10.

The embodiment of the present application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that can be stored by a computing device or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk). The computer-readable storage medium includes instructions that instruct the computing device to execute any one of the speech recommendation methods of claims 1 to 10.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims (23)

A method for recommending a speech technique, characterized in that the method comprises: Determining characteristic information of a first customer and characteristic information of a first service staff in a target conversation, wherein the target conversation is a conversation between the first customer and the first service staff; Parsing the conversation data of the target conversation to determine the business scenario intention corresponding to the target conversation, the conversation data including the conversation context of the target conversation; obtaining a recommended speech based on target data, the target data including: characteristic information of the first customer, characteristic information of the first service personnel, and business scenario intention corresponding to the target dialogue, the recommended speech being an optional speech for replying to the first customer; A recommended speech display interface is provided, wherein the recommended speech display interface is used to display the recommended speech. The method according to claim 1 is characterized in that the characteristic information of the first service personnel includes at least one of the following: the ability level of the first service personnel, the business team of the first service personnel, the sales status of the first service personnel, and the sales history of the first service personnel. The method according to claim 2, characterized in that the recommendation words are words that introduce the target product; The target product satisfies at least one of the following characteristics: the target product corresponds to the ability level of the first service personnel, the target product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the target product is greater than a first preset threshold, the target product belongs to the industry that the first service personnel is good at, and the number of sales history of the target product by the first service personnel is greater than a second preset threshold. The method according to any one of claims 1 to 3, characterized in that the method further comprises: Based on the feature information of the first customer and the feature information of the first service personnel, at least one preferred product of the first customer is determined, and the target data also includes the at least one preferred product. The method according to claim 4, characterized in that the determining, based on the characteristic information of the first customer and the characteristic information of the first service personnel, at least one preferred product of the first customer comprises: The characteristic information of the first customer and the characteristic information of the first service staff are input into a first model to obtain the at least one preferred product output by the first model, wherein the first model is obtained by training a first initial model with first training data, and the first training data includes: the characteristic information of the second customer, the characteristic information of the second service staff and the historical transaction products of the second customer. The method according to any one of claims 1 to 5, characterized in that parsing the dialog data of the target dialog to determine the business scenario intention corresponding to the target dialog comprises: The conversation data is input into a second model to obtain the business scenario intention corresponding to the target conversation output by the second model, wherein the second model is obtained by training a second initial model with second training data, and the second training data includes: historical conversation data of historical conversations between a third customer and a third service staff and the business scenario intention corresponding to the historical conversations. The method according to any one of claims 1 to 6, characterized in that obtaining the recommended words based on the target data comprises: Inputting the target data into a third model to obtain a model recall result output by the third model, wherein the third model is obtained by training a third initial model with third training data, and the third training data includes: feature information of a fourth customer, feature information of a fourth service staff, business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff, and preset speech; Based on the model recall result, the recommended words are obtained. The method according to claim 7, characterized in that the method further comprises: The first vector data and the preset words are stored in the vector database, wherein the first vector data includes: a vector of feature information of the fourth customer, a vector of feature information of the fourth service staff, and a vector of business scenario intention corresponding to the historical conversation between the fourth customer and the fourth service staff. The method according to claim 8, characterized in that the method further comprises: Performing vector retrieval according to second vector data to obtain a vector recall result, wherein the second vector data includes a vector of the target data; The step of obtaining the recommended words based on the model recall result includes: The model recall result and the vector recall result are weighted to obtain the recommended words. The method according to any one of claims 1 to 9, characterized in that the method further comprises: A preset number of continuous dialogue texts are obtained from the dialogue text of the target dialogue to obtain the dialogue data. A speech technique recommendation device, characterized in that the device comprises: a processing module, configured to determine characteristic information of a first customer and characteristic information of a first service staff in a target conversation, wherein the target conversation is a conversation between the first customer and the first service staff; The processing module is further used to parse the dialogue data of the target dialogue to determine the business scenario intention corresponding to the target dialogue, wherein the dialogue data includes the dialogue context of the target dialogue; A recommendation module, configured to obtain a recommended speech based on target data, wherein the target data includes: characteristic information of the first customer, characteristic information of the first service personnel, and business scenario intention corresponding to the target dialogue, and the recommended speech is an optional speech for replying to the first customer; The display module is used to provide a recommended speech display interface, and the recommended speech display interface is used to display the recommended speech. The device according to claim 11 is characterized in that the characteristic information of the first service personnel includes at least one of the following: the ability level of the first service personnel, the business team of the first service personnel, the sales status of the first service personnel, and the sales history of the first service personnel. The device according to claim 12, characterized in that the recommendation words are words that introduce the target product; The target product satisfies at least one of the following characteristics: the target product corresponds to the ability level of the first service personnel, the target product belongs to the business team of the first service personnel, the customer satisfaction obtained by the first service personnel in the history of selling the target product is greater than a first preset threshold, the target product belongs to the industry that the first service personnel is good at, and the number of sales history of the target product by the first service personnel is greater than a second preset threshold. The device according to any one of claims 11 to 13 is characterized in that the processing module is also used to determine at least one preferred product of the first customer based on the characteristic information of the first customer and the characteristic information of the first service personnel, and the target data also includes the at least one preferred product. The device according to claim 14, characterized in that the processing module is specifically used to input the characteristic information of the first customer and the characteristic information of the first service personnel into the first model to obtain the at least one preferred product output by the first model; The first model is obtained by training the first initial model with first training data, and the first training data includes: characteristic information of the second customer, characteristic information of the second service personnel, and historical transaction products of the second customer. The device according to any one of claims 11 to 15, characterized in that the processing module is specifically used to input the dialogue data into a second model to obtain the business scenario intention corresponding to the target dialogue output by the second model; The second model is obtained by training the second initial model with second training data, and the second training data includes: historical conversation data of historical conversations between a third customer and a third service personnel and business scenario intentions corresponding to the historical conversations. The device according to any one of claims 11 to 16 is characterized in that the recommendation module is specifically used to: input the target data into a third model to obtain a model recall result output by the third model; and obtain the recommended words based on the model recall result; Among them, the third model is obtained by training the third initial model through third training data, and the third training data includes: characteristic information of the fourth customer, characteristic information of the fourth service personnel, business scenario intentions corresponding to the historical conversation between the fourth customer and the fourth service personnel, and preset words. The device according to claim 17, characterized in that the device further comprises: A storage module is used to store first vector data and the preset words in a vector database, wherein the first vector data includes: a vector of feature information of the fourth customer, a vector of feature information of the fourth service staff, and a vector of business scenario intentions corresponding to historical conversations between the fourth customer and the fourth service staff. The device according to claim 18, characterized in that the recommendation module is further used to perform vector retrieval based on second vector data to obtain a vector recall result, wherein the second vector data includes the vector of the target data; The recommendation module is specifically used to: perform weighted processing on the model recall result and the vector recall result to obtain the recommended words. The device according to any one of claims 11 to 19 is characterized in that the processing module is also used to obtain a preset number of continuous dialogue texts from the dialogue text of the target dialogue to obtain the dialogue data. A computing device cluster, characterized in that it comprises at least one computing device, each of which comprises a processor and a memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device, so that the computing device cluster executes the method according to any one of claims 1 to 10. A computer program product comprising instructions, wherein when the instructions are executed by a computing device cluster, the computing device cluster executes the method according to any one of claims 1 to 10. A computer-readable storage medium, characterized in that it includes computer program instructions. When the computer program instructions are executed by a computing device cluster, the computing device cluster executes the method as described in any one of claims 1 to 10.