Training method of reply key point determination model, verbal skill reply method and related products

Abstract

The invention discloses a training method of a reply key point determination model, a verbal skill reply method and a related product, and the method comprises the steps: obtaining instruction information used for model training, and the instruction information is used for instructing a to-be-trained model to output historical reply key points corresponding to historical dialogue data based on the historical dialogue data, the historical reply key points are to-be-replied keyword information analyzed from historical dialogue data; constructing a training data set, wherein the training data set comprises historical dialogue data and historical reply key points; the to-be-trained model is trained based on the instruction information and the training data set, a reply key point determination model is obtained after training is finished, and the reply key point determination model is used for outputting reply key points corresponding to the to-be-processed dialogue data based on the to-be-processed dialogue data. Therefore, the customer service reply efficiency and response speed can be remarkably improved, and the customer satisfaction degree is further improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a training method for a reply key point determination model, a speech reply method, and related products. Background Art

[0002] In today's digital age, companies need to establish strong customer service teams to achieve deep interaction and continuous connection with customers. High-quality customer service is not only an important reflection of the corporate image, but also a key factor in improving customer satisfaction and loyalty.

[0003] In practical applications, for example, customer service systems in auto insurance marketing scenarios require one-on-one communication with customers. However, when faced with a large number of customer inquiries, limited staffing often leads to inefficient service and extended response times. This is especially true when complex questions are beyond the customer's knowledge base, requiring additional time to research and confirm information. This further increases response time and reduces customer satisfaction. Summary of the Invention

[0004] The embodiments of the present application provide a training method for a reply key point determination model, a speech reply method and related products, which effectively reduce the reply response time and improve customer service reply efficiency and satisfaction.

[0005] In a first aspect, an embodiment of the present application provides a method for training a reply key point determination model, comprising:

[0006] Obtain instruction information for model training, wherein the instruction information is used to instruct the to-be-trained model to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, wherein the historical reply key points are keyword information to be replied to analyzed from the historical conversation data;

[0007] Constructing a training data set, wherein the training data set includes the historical conversation data and the historical reply key points;

[0008] The model to be trained is trained based on the instruction information and the training data set, and a reply key point determination model is obtained after the training is completed. The reply key point determination model is used to output reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

[0009] Optionally, the instruction information includes role information, target task information and output format information;

[0010] The role information is used to set the role type of the model to be trained when executing the target task information, the target task information is used to instruct the model to be trained to determine the historical response points from the historical conversation data, and the output format information is used to set the output format of the historical response points.

[0011] Optionally, constructing a training data set includes:

[0012] Obtaining the historical conversation data;

[0013] Performing semantic analysis on the historical conversation data to obtain the key points of the historical responses;

[0014] The training data set is constructed based on the historical conversation data and the historical reply key points.

[0015] In a second aspect, an embodiment of the present application provides a method for replying to a conversation, including:

[0016] Get the conversation data to be processed;

[0017] Processing the dialogue data to be processed based on a reply key point determination model to obtain reply key points corresponding to the dialogue data to be processed, wherein the reply key point determination model is obtained based on any implementation step of the training method of the reply key point determination model;

[0018] Determining a reply script for the conversation data to be processed based on the reply key points;

[0019] Output the reply words.

[0020] Optionally, determining a reply wording for the conversation data to be processed based on the reply key points includes:

[0021] Determining a first reply script and / or a second reply script based on the reply key points, wherein the second reply script includes a slot to be filled;

[0022] The outputting of the reply speech includes:

[0023] Output the first reply script; and / or,

[0024] Call a filling tool to fill the to-be-filled slot in the second reply speech to obtain a third reply speech, and output the third reply speech.

[0025] Optionally, before determining a reply wording for the conversation data to be processed based on the reply key points, the method further includes:

[0026] Constructing a reply speech knowledge base, wherein the reply speech knowledge base includes a first reply speech and a second reply speech;

[0027] The determining of the first reply words and / or the second reply words based on the reply key points includes:

[0028] Based on the reply key points, the first reply speech and / or the second reply speech are retrieved from the reply speech knowledge base.

[0029] Optionally, there are multiple reply points, and determining a reply wording for the conversation data to be processed based on the reply points includes:

[0030] Determining a plurality of intermediate reply words based on a plurality of reply key points, wherein the plurality of reply key points correspond to the plurality of intermediate reply words in a one-to-one manner;

[0031] The multiple intermediate reply dialogues are concatenated to obtain the reply dialogue.

[0032] In a third aspect, an embodiment of the present application provides a training device for a reply key point determination model, comprising:

[0033] An instruction information acquisition module, configured to acquire instruction information for model training, wherein the instruction information is configured to instruct the model to be trained to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, wherein the historical reply key points are keyword information to be replied to analyzed from the historical conversation data;

[0034] A data set construction module, configured to construct a training data set, wherein the training data set includes the historical conversation data and the historical reply key points;

[0035] A training module is used to train the model to be trained based on the instruction information and the training data set, and obtain a reply key point determination model after the training is completed. The reply key point determination model is used to output the reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

[0036] In a fourth aspect, an embodiment of the present application provides a speech reply device, comprising:

[0037] A conversation data acquisition module, used to acquire conversation data to be processed;

[0038] a conversation data processing module, configured to process the conversation data to be processed based on a reply key point determination model to obtain reply key points corresponding to the conversation data to be processed, wherein the reply key point determination model is obtained based on any implementation step of the training method for the reply key point determination model described above;

[0039] A reply word determination module, configured to determine a reply word for the conversation data to be processed based on the reply key points;

[0040] An output module is used to output the reply words.

[0041] In a fifth aspect, an embodiment of the present application provides an electronic device, characterized in that the device includes: a processor, a memory, and a system bus;

[0042] The processor and the memory are connected via the system bus;

[0043] The memory is used to store a program, which includes instructions. When the instructions are executed by the processor, the processor executes any implementation step of the training method for determining the reply key points model, or any implementation step of the above-mentioned speech reply method.

[0044] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0045] In an embodiment of the present application, instruction information for model training is first obtained. The instruction information is used to instruct the model to be trained to output historical reply key points corresponding to the historical conversation data based on historical conversation data. The historical reply key points are keyword information of the reply to be answered analyzed from the historical conversation data. A training dataset is then constructed, which includes the historical conversation data and historical reply key points. Finally, the model to be trained is trained based on the instruction information and the training dataset. After training, a reply key point determination model is obtained. The reply key point determination model outputs reply key points corresponding to the pending conversation data based on the pending conversation data. Thus, by using the instruction information and the training dataset as input for training, a reply key point determination model is obtained. The reply key point determination model is then used to extract key information from the pending conversation data to obtain reply key points, thereby quickly identifying customer intent and concerns. In this way, taking a customer service scenario as an example, the reply key points can be used to respond to the pending conversation content, significantly improving customer service response efficiency and response speed, thereby increasing customer satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a method for training a response key point determination model provided in an embodiment of the present application;

[0047] Figure 2 A flowchart of a speech reply method provided in an embodiment of the present application;

[0048] Figure 3 A flowchart of another speech reply method provided in an embodiment of the present application;

[0049] Figure 4 A schematic diagram of the structure of a training device for a response key point determination model provided in an embodiment of the present application;

[0050] Figure 5 A schematic structural diagram of a speech reply device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] As mentioned earlier, when companies build strong customer service teams, they aim to achieve deep engagement and continuous connections with customers. High-quality customer service not only impacts a company's image but is also key to improving customer satisfaction and loyalty. However, in practice, taking auto insurance marketing as an example, customer service personnel typically need to communicate with customers one-on-one. Faced with a large number of inquiries, limited human resources often lead to inefficient service and extended response times. This is especially true when dealing with complex issues that exceed customer knowledge. Customer service personnel must spend additional time searching and confirming information, further lengthening the response cycle and impacting the customer experience.

[0052] Based on this, in order to solve the above problems, an embodiment of the present application provides a training method for a reply key point determination model, first obtaining instruction information for model training, wherein the instruction information is used to instruct the model to be trained to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, and the historical reply key points are the keyword information to be replied analyzed from the historical conversation data, and then a training data set is constructed, the training data set includes historical conversation data and historical reply key points, and finally, the model to be trained is trained based on the instruction information and the training data set, and a reply key point determination model is obtained after the training is completed, and the reply key point determination model outputs the reply key points corresponding to the conversation data to be processed based on the conversation data to be processed.

[0053] As can be seen, by training the response key point determination model using instruction information and a training dataset as input, the response key point determination model is then used to extract key information from the pending conversation data to obtain response key points, thereby quickly identifying customer intent and focus. In this way, for example, in customer service scenarios, these response key points can be used to answer the pending conversation content, significantly improving customer service response efficiency and speed, thereby increasing customer satisfaction.

[0054] It should be noted that the embodiments of the present application may not limit the execution subject of the training method of the reply key point determination model. For example, the training method of the reply key point determination model of the embodiments of the present application can be applied to information processing devices such as servers or terminal devices, and accordingly, the large language model can be installed in the information processing device. Among them, the server can be a stand-alone server, a cluster server or a cloud server. The terminal device can be an electronic device such as a smart phone, a computer, a personal digital assistant (PDA), a tablet computer, etc.

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0056] Figure 1 A flowchart of a training method for a reply key point determination model provided in an embodiment of the present application. Figure 1 As shown, the training method of the reply key point determination model may include the following steps S101 to S103.

[0057] S101: Obtain instruction information for model training, where the instruction information is used to instruct the model to be trained to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, where the historical reply key points are keyword information to be replied analyzed from the historical conversation data.

[0058] Instruction information refers to information used to instruct or assist the model in making decisions. In the customer service response scenario involved in the embodiments of this application, the above instruction information can be used to instruct the trained model to determine and output the historical response points corresponding to the historical conversation data based on the historical conversation data.

[0059] Among them, the key points of historical replies are the keyword information to be replied to extracted through fine-grained semantic analysis of historical conversation data. Fine-grained semantic analysis includes word segmentation, entity recognition, intent classification and slot extraction of historical conversation data. By deeply understanding the contextual semantics, the important information of the historical conversation data can be accurately extracted, thereby accurately identifying the specific needs of customers and forming accurate and effective reply points.

[0060] It should be noted that the instruction information includes three parts: role information, target task information and output format information. Among them, the role information is used to set the role type of the model to be trained when executing the target task information. The target task information is used to instruct the model to be trained to determine the key points of historical responses from the historical dialogue data. The output format information is used to set the output format of the key points of historical responses.

[0061] In specific implementation, using the auto insurance customer service marketing scenario as an example, the role information can be set as a professional auto insurance salesperson with excellent and comprehensive auto insurance sales knowledge and practical experience. The target task information is based on historical conversation data. The keyword information in this historical conversation data is analyzed to obtain the historical response points corresponding to the historical conversation data. The output format can require that up to three historical response points be extracted for each historical conversation data, with "#" used as a separator between multiple historical response points, for example, "Historical response point 1#Historical response point 2#Historical response point 3".

[0062] S102: Construct a training dataset, which includes historical conversation data and historical response points.

[0063] In the process of building a training dataset, we can obtain historical conversation data and perform semantic analysis on it to obtain the key points of historical responses. Then, we can build a training dataset based on the historical conversation data and the key points of historical responses.

[0064] In the above process, the embodiment of the present application does not specifically limit the process of semantic analysis, and can be implemented using any existing or future semantic analysis algorithm.

[0065] S103: The training model is trained based on the instruction information and the training data set. After the training, a reply key point determination model is obtained. The reply key point determination model is used to output reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

[0066] After configuring the role information, target task information, and output format information included in the instruction information, and constructing the training dataset, the response determination model undergoes supervised training using the instruction information and training dataset as input and the response key points corresponding to the pending conversation data as output. Specifically, during training, the response determination model learns the role information and target task information set in the instruction information, understands and extracts key information from the pending conversation data, and gradually optimizes its parameters, ultimately obtaining a response key point determination model capable of automatically extracting response key points.

[0067] It's important to note that in real-world conversations, customers often include multiple intents in a single sentence, such as inquiring about the vehicle's license plate, event details, and service procedures. This response key point determination model extracts multiple historical response key points to accurately deconstruct and respond to these multiple intents, avoiding overlooking customer needs. Furthermore, the response key point confirmation model performs semantic analysis based on historical conversation data and provides predefined historical response key points. This reduces the computational complexity of the response key point determination model, alleviating the optimization and iteration requirements for the model and improving its practicality.

[0068] Based on the relevant content of steps S101 to S103 described above, it can be seen that in this embodiment of the present application, instruction information for model training is first obtained, wherein the instruction information is used to instruct the trained model to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, and the historical reply key points are the keyword information of the reply to be replied analyzed from the historical conversation data. Then, a training dataset is constructed, which includes the historical conversation data and the historical reply key points. Finally, the trained model is trained based on the instruction information and the training dataset. After the training, a reply key point determination model is obtained. The reply key point determination model outputs the reply key points corresponding to the processed conversation data based on the processed conversation data. It can be seen that by training the reply key point determination model using the instruction information and the training dataset as input, the reply key point determination model is obtained, and the reply key point determination model is used to extract key information from the processed conversation data to obtain reply key points, thereby quickly identifying customer intent and concerns. In this way, taking the customer service scenario as an example, the reply key points can be used to respond to the content of the processed conversation, thereby significantly improving the efficiency and response speed of customer service responses, thereby increasing customer satisfaction.

[0069] Furthermore, based on the training method for the reply key point determination model provided in the above embodiment, the embodiment of the present application correspondingly provides a speech reply method. The speech reply method is described below in conjunction with the embodiments and drawings.

[0070] Figure 2 A flowchart of a speech reply method provided in an embodiment of the present application, combined with Figure 2 As shown, the speech reply method may include the following steps S201-S204.

[0071] S201: Acquire conversation data to be processed.

[0072] Unprocessed conversation data refers to the conversation content between customer service and customers, which usually includes multiple rounds of conversations, specifically covering the questions raised by the customer, the customer service's responses, and related contextual content. This unprocessed conversation data can fully reflect the customer's needs and intentions, and provides key support for the subsequent extraction of reply key points from the unprocessed conversation data by the reply key point determination model.

[0073] S202: Processing the dialogue data to be processed based on the reply key point determination model to obtain reply key points corresponding to the dialogue data to be processed, wherein the reply key point determination model is obtained based on any implementation step of the training method of the reply key point determination model.

[0074] The reply key points determination model can process the conversation data to be processed, thereby obtaining the reply key points corresponding to the conversation data to be processed. It should be noted that when the reply key points determination model fails to accurately identify the core content of the conversation data to be processed during the processing process, or the generated reply key points deviate from the actual context, the reply key points determination model can be fine-tuned to improve the reply key points determination model's ability to understand the semantics of the conversation data to be processed, and the accuracy of the reply key points corresponding to the conversation data to be processed.

[0075] S203: Based on the reply key points, determine the reply words for the conversation data to be processed.

[0076] Before determining the model based on the reply points to process the conversation data to be processed and obtaining the reply words for the conversation data to be processed, it is first necessary to build a reply words knowledge base, and the reply words knowledge base includes the first reply words and the second reply words. Then, based on the reply points, the first reply words and / or the second reply words can be retrieved from the reply words knowledge base.

[0077] It should be noted that the first response is a static, fixed-format line. For example, in the auto insurance marketing scenario, the first response might be, "Hello, thank you for your interest in our auto insurance products. If you have any questions, please feel free to contact us. We will be happy to assist you."

[0078] Because the second reply includes slots to be filled and can be dynamically populated based on specific information, the second reply is considered dynamic. For example, in a car insurance marketing scenario, the second reply might be, "Hello, we've received your request for a quote for [car insurance type] for [license plate number]. We'll get back to you within [processing time]. Please keep your line open." "[license plate number]," "[car insurance type]," and "[processing time]" represent the slots to be filled.

[0079] As can be seen, by searching the reply script knowledge base based on key response points, the content of the reply script can be controlled, avoiding inappropriate or non-standard expressions. Furthermore, when applying it to new marketing scenarios, simply adjust or expand the corresponding reply script knowledge base to quickly adapt and deploy it into the reply key response determination model, thus achieving efficient online launch and flexible iteration, fully meeting the business needs of fast-paced marketing scenarios.

[0080] S204: Output reply words.

[0081] As mentioned earlier, the first reply is a static reply with a fixed format, while the second reply is a dynamic reply that includes slots to be filled. Therefore, if the determined reply includes the first reply, the reply key determination model can output the retrieved first reply. If the determined reply includes the second reply, the reply key determination module can first call a filling tool to fill the slots to be filled in the second reply, obtain a third reply, and then output the third reply.

[0082] Taking the auto insurance marketing scenario as an example, if the second reply is "Hello, we have received your request for a quote on [auto insurance type] for [license plate number]. We will provide you with feedback within [processing time]. Please keep your phone open." The fill-in tool is called to fill the slot to be filled in with "Beijing A88888," "[Auto Insurance Type]" with "Compulsory Traffic Insurance," and "[Processing Time]" with "Three Days." The fill-in tool is then called to fill in the slot to be filled in the second reply. The resulting third reply is "Hello, we have received your request for a quote on compulsory traffic insurance for Beijing A88888. We will provide you with feedback within three days. Please keep your phone open." Therefore, after the fill-in tool is called to fill in the slots to be filled in, the response key determination model can output the third reply.

[0083] In addition, when there are multiple reply points, multiple intermediate reply phrases can be determined based on the multiple reply points, and the multiple reply points correspond one-to-one to the multiple intermediate reply phrases. The multiple intermediate reply phrases can be spliced ​​together to obtain the reply phrases, where the multiple intermediate reply phrases can include one or more first reply phrases, and / or, one or more second reply phrases.

[0084] Based on the relevant contents of steps S201-S204 described above, it can be seen that in the embodiment of the present application, the conversation data to be processed is first obtained, and then the conversation data to be processed is processed based on the reply key point determination model to obtain the reply key points corresponding to the conversation data to be processed. The reply key point determination model is obtained based on the above-mentioned training method. Then, based on the reply key points, the reply words for the conversation data to be processed are determined, and finally, the reply words are output. It can be seen that by inputting the conversation data to be processed into the reply key point determination model for processing, thereby obtaining the reply key points of the conversation data to be processed, the key information of the conversation data to be processed can be quickly determined. In this way, the reply words for the conversation data to be processed can be determined based on the reply key points, thereby significantly improving the efficiency and response speed of customer service responses, thereby improving customer satisfaction.

[0085] Furthermore, for easier understanding, it is also possible to combine specific application logic and attached Figure 3Let me introduce an example of a method of replying with dialogue.

[0086] Figure 3 A flowchart of another method of replying with dialogue provided in an embodiment of the present application, combined with a specific example, first loads a historical conversation (i.e., the conversation data to be processed in the above embodiment), the content of which is customer: "Hello", the customer service responds: "Hello, I am a car insurance customer service staff, I see that your car insurance is about to expire. Our company is currently carrying out an activity to give back to old customers, and provides exclusive benefits and discounts for renewing old customers. I can send you the detailed activity information via SMS, is it convenient for you to learn about it?" Customer: "That's my vehicle, please tell me the license plate number."

[0087] Next, the above-mentioned dialogue data to be processed is input into the large model (i.e., the reply key point determination model in the above-mentioned embodiment) for analysis and processing, and then two reply key points of "Reply to expired vehicle license plate#Introduction to activity benefits" are extracted. Subsequently, for reply key point 1 "Reply to expired vehicle license plate", a search is performed from the reply script knowledge base to obtain the second reply script "Your car with the last license plate (car_license) is about to expire", and for reply key point 2 "Introduction to activity benefits", a search is performed from the reply script knowledge base to obtain the first reply script "For old customers, today I will give you a special 1 yuan car wash coupon. I will send the collection method via SMS, please pay attention to it."

[0088] Then, the to-be-filled slot in the second reply script "Your car with the license plate number (car_license) is about to expire" is filled, and the third reply script "Your car with the license plate number 88888" is obtained.

[0089] Finally, the third reply and the first reply are spliced ​​and integrated to get the final reply: "Your car with license plate number ending in 88888 is about to expire. As an old customer, I will give you a 1 yuan car wash coupon today. I will send the collection instructions via SMS. Please check it out."

[0090] Further, Figure 4 A structural diagram of a training device for a reply key point determination model provided in an embodiment of the present application. Figure 4 As shown, the training device 400 for the reply key point determination model provided in the embodiment of the present application may include:

[0091] Instruction information acquisition module 401 is used to obtain instruction information for model training, wherein the instruction information is used to instruct the to-be-trained model to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, wherein the historical reply key points are keyword information to be replied to analyzed from the historical conversation data;

[0092] A data set construction module 402 is used to construct a training data set, wherein the training data set includes the historical conversation data and the historical reply key points;

[0093] The training module 403 is used to train the model to be trained based on the instruction information and the training data set. After the training, a reply key point determination model is obtained. The reply key point determination model is used to output the reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

[0094] Optionally, the data set construction module 402 is specifically configured to:

[0095] Obtaining the historical conversation data;

[0096] Performing semantic analysis on the historical conversation data to obtain the key points of the historical responses;

[0097] The training data set is constructed based on the historical conversation data and the historical reply key points.

[0098] Figure 5 This is a structural diagram of a speech reply device provided in an embodiment of the present application. Figure 5 As shown, the speech reply device 500 provided in the embodiment of the present application may include:

[0099] A conversation data acquisition module 501 is used to acquire conversation data to be processed;

[0100] a conversation data processing module 502 configured to process the conversation data to be processed based on a reply key point determination model to obtain reply key points corresponding to the conversation data to be processed, wherein the reply key point determination model is obtained based on any implementation step of the training method of the reply key point determination model;

[0101] A reply word determination module 503 is configured to obtain a reply word for the conversation data to be processed based on the reply key points;

[0102] The output module 504 is used to output the reply words.

[0103] Optionally, the reply speech determination module 503 may include:

[0104] A speech word determination submodule determines a first reply speech word and / or a second reply speech word based on the reply key points, wherein the second reply speech word includes a slot to be filled;

[0105] The output module 504 is specifically configured to:

[0106] Output the first reply script; and / or,

[0107] Call a filling tool to fill the to-be-filled slot in the second reply speech to obtain a third reply speech, and output the third reply speech.

[0108] Optionally, the speech reply device 500 further includes:

[0109] Constructing a knowledge base module to construct a reply speech knowledge base, wherein the reply speech knowledge base includes a first reply speech and a second reply speech;

[0110] The speech technique determination submodule is specifically used to:

[0111] Based on the reply key points, the first reply speech and / or the second reply speech are retrieved from the reply speech knowledge base.

[0112] Optionally, the reply speech determination module 503 is specifically configured to:

[0113] Determining a plurality of intermediate reply words based on a plurality of reply key points, wherein the plurality of reply key points correspond to the plurality of intermediate reply words in a one-to-one manner;

[0114] The multiple intermediate reply dialogues are concatenated to obtain the reply dialogue.

[0115] Furthermore, an embodiment of the present application also provides an electronic device, comprising: a processor, a memory, and a system bus;

[0116] The processor and the memory are connected via the system bus;

[0117] The memory is used to store one or more programs, which include instructions. When executed by the processor, the instructions enable the processor to perform any implementation step of the training method for determining the reply key points model, or any implementation step of the above-mentioned speech reply method.

[0118] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, which is used to store a computer program. When the computer program is executed by a terminal device, it implements any implementation step of the training method for determining the reply key points model, or any implementation step of the method for determining the speech reply.

[0119] It can be seen from the description of the above implementation methods that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a storage medium, such as ROM / RAM, a disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in each embodiment of the present application or certain parts of the embodiments. It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other.

[0120] As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0121] It should also be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0122] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for a reply key point determination model, characterized in that: include: Obtain instruction information for model training, wherein the instruction information is used to instruct the to-be-trained model to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, wherein the historical reply key points are keyword information to be replied to analyzed from the historical conversation data; Constructing a training data set, wherein the training data set includes the historical conversation data and the historical reply key points; The model to be trained is trained based on the instruction information and the training data set, and a reply key point determination model is obtained after the training is completed. The reply key point determination model is used to output reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

2. The training method for the reply key point determination model according to claim 1, characterized in that: The instruction information includes role information, target task information and output format information; The role information is used to set the role type of the model to be trained when executing the target task information, the target task information is used to instruct the model to be trained to determine the historical response points from the historical conversation data, and the output format information is used to set the output format of the historical response points.

3. The training method for the reply key point determination model according to claim 1, characterized in that: The constructing of the training data set includes: Obtaining the historical conversation data; Performing semantic analysis on the historical conversation data to obtain the key points of the historical responses; The training data set is constructed based on the historical conversation data and the historical reply key points.

4. A speech reply method, characterized in that: include: Get the conversation data to be processed; Processing the conversation data to be processed based on a reply key point determination model to obtain reply key points corresponding to the conversation data to be processed, wherein the reply key point determination model is obtained based on the training method of the reply key point determination model according to any one of claims 1 to 3; Determining a reply script for the conversation data to be processed based on the reply key points; Output the reply words.

5. The speech reply method according to claim 4, characterized in that: The step of determining a reply wording for the conversation data to be processed based on the reply key points includes: Determining a first reply script and / or a second reply script based on the reply key points, wherein the second reply script includes a slot to be filled; The outputting of the reply speech includes: Output the first reply script; and / or, Call a filling tool to fill the to-be-filled slot in the second reply speech to obtain a third reply speech, and output the third reply speech.

6. The speech reply method according to claim 5, characterized in that: Before determining a reply wording for the conversation data to be processed based on the reply key points, the method further includes: Building a reply speech knowledge base, the reply speech knowledge base including the first reply speech and the second reply speech; The determining of the first reply words and / or the second reply words based on the reply key points includes: Based on the reply key points, the first reply speech and / or the second reply speech are retrieved from the reply speech knowledge base.

7. The speech reply method according to claim 4, characterized in that: There are multiple reply points, and determining a reply script for the conversation data to be processed based on the reply points includes: Determining a plurality of intermediate reply words based on a plurality of reply key points, wherein the plurality of reply key points correspond to the plurality of intermediate reply words in a one-to-one manner; The multiple intermediate reply dialogues are concatenated to obtain the reply dialogue.

8. A training device for a reply key point determination model, characterized in that: include: An instruction information acquisition module, configured to acquire instruction information for model training, wherein the instruction information is configured to instruct the model to be trained to output historical reply key points corresponding to the historical conversation data based on the historical conversation data, wherein the historical reply key points are keyword information to be replied to analyzed from the historical conversation data; A data set construction module, configured to construct a training data set, wherein the training data set includes the historical conversation data and the historical reply key points; A training module is used to train the model to be trained based on the instruction information and the training data set, and obtain a reply key point determination model after the training is completed. The reply key point determination model is used to output the reply key points corresponding to the dialogue data to be processed based on the dialogue data to be processed.

9. A speech reply device, characterized in that: include: A conversation data acquisition module, used to acquire conversation data to be processed; a conversation data processing module, configured to process the conversation data to be processed based on a reply key point determination model to obtain reply key points corresponding to the conversation data to be processed, wherein the reply key point determination model is obtained based on the training method for the reply key point determination model according to any one of claims 1 to 3; A reply word determination module, configured to determine a reply word for the conversation data to be processed based on the reply key points; An output module is used to output the reply words.

10. An electronic device, characterized in that: The device includes: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store a program, which includes instructions. When the instructions are executed by the processor, the processor executes the steps of the training method for determining the reply key points model described in any one of claims 1 to 3, or the steps of the verbal reply method described in any one of claims 4 to 7.

Images