A method for generating dialogues for e-commerce customer service simulation training

Abstract

The present invention relates to the technical field of text data processing, and in particular to a method for generating simulated training dialogues for e-commerce customer service, comprising: clustering all complete dialogue processes of all e-commerce customer service according to the similarity of texts in a complete dialogue process between any two e-commerce customer service personnel to obtain a number of dialogue clusters; obtaining a simulated training cluster of each e-commerce customer service by the number of dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, the response time of all dialogue rounds, the number of other users served at the same time, and the user evaluation satisfaction value; selecting similar targeted business texts through the similarity between all user feedback texts and the texts in the simulated training cluster of each e-commerce customer service, generating questions for them, and conducting simulated training for each e-commerce customer service through the generated questions. The present invention improves the efficiency of targeted training for e-commerce customer service.

Description

Technical Field

[0001] The present invention relates to the technical field of text data processing, and in particular to a method for generating a simulated training dialogue for e-commerce customer service. Background Art

[0002] The e-commerce customer service simulation training dialogue generation method can not only improve the response ability of customer service personnel through reasonable data preparation, model selection and dialogue generation strategy, but also enhance their practical experience by simulating actual scenarios. Continuous optimization and feedback mechanisms will ensure the effectiveness and relevance of training, thereby improving overall customer satisfaction.

[0003] When conducting simulated training for e-commerce customer service, since different customer service staff may have their own business topics that they are proficient in and not proficient in, the generative model usually randomly extracts a topic from the question library and asks questions based on the topic. Some customer service staff may be more familiar with a certain business topic. If the customer service staff happens to extract questions related to this business, it is impossible to train the customer service staff on the business that they are not familiar with, and targeted training cannot be provided to the customer service staff at this time. Summary of the invention

[0004] The present invention provides an e-commerce customer service simulation training dialogue generation method to solve the existing problems.

[0005] The e-commerce customer service simulation training dialogue generation method of the present invention adopts the following technical solutions:

[0006] An embodiment of the present invention provides a method for generating a simulated training dialogue for e-commerce customer service, the method comprising the following steps:

[0007] Obtain the text of the complete conversation between each e-commerce customer service and the user; obtain the number of conversation rounds in a complete conversation with each e-commerce customer service, the response time of each conversation round, and the total number of users served at the same time; obtain the content of all questions raised by users on the platform, as well as the content of user complaints received by the platform, and record a question raised by a user or a user complaint as a user feedback text;

[0008] The complete conversation process between each e-commerce customer service and a user purchasing a product is recorded as a complete conversation process of each e-commerce customer service. According to the topic overlap in the text of a complete conversation process between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's words, all the complete conversation processes of all e-commerce customer service staff are clustered to obtain several conversation clusters;

[0009] The business topic performance of each e-commerce customer service in each cluster is obtained through the number of dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, the response time of all dialogue rounds, and the number of other users served in each dialogue round; the business topic familiarity of each e-commerce customer service in each cluster is obtained through the business topic performance of each e-commerce customer service in each cluster and the user evaluation satisfaction value of a complete dialogue process of each e-commerce customer service; a cluster with the smallest business topic familiarity of each e-commerce customer service in all clusters is selected and recorded as the simulation training cluster of each e-commerce customer service, and the first simulation training question set of each e-commerce customer service is generated through the text in the simulation training cluster of each e-commerce customer service;

[0010] Through the similarity between all user feedback texts and the texts in the simulation training cluster of each e-commerce customer service, targeted business texts similar to the texts in the simulation training cluster of each e-commerce customer service are selected, and the similar targeted business texts are used to generate the second simulation training question set for each e-commerce customer service; each e-commerce customer service is simulated trained through the first simulation training question set and the second simulation training question set of each e-commerce customer service.

[0011] Furthermore, according to the topic overlap in the text of a complete conversation between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's speech, all complete conversation processes of all e-commerce customer service staff are clustered to obtain a number of conversation clusters, including the following specific steps:

[0012] For the text of a complete conversation with each e-commerce customer service representative, use Jieba word segmentation to obtain all the words in the complete conversation, and then use the keyword extraction algorithm TF-IDF to obtain all the keywords from all the words. All the keywords in the complete conversation are grouped into a set, which is recorded as a keyword set for each e-commerce customer service representative.

[0013] Obtain the intersection and union of keywords between any two keyword sets, and obtain the topic overlap between any two keyword sets by the difference in the number of keywords in the intersection and union of keywords;

[0014] Use the Word2vec algorithm to obtain the word vector of each word in the text corresponding to the e-commerce customer service during a complete conversation between any two e-commerce customer service staff.

[0015] According to the similarity between the word vectors in the corresponding texts of any two e-commerce customer service representatives during a complete conversation, the consistency of the words used in a complete conversation between any two e-commerce customer service representatives is obtained;

[0016] The product of the topic overlap between any two keyword sets and the consistency of the words in a complete conversation between any two e-commerce customer service staff is recorded as the similarity of the conversation in a complete conversation between any two e-commerce customer service staff; sentence vectors are constructed based on all the keywords in the keyword set corresponding to a complete conversation between each e-commerce customer service staff; the similarity of the conversation in a complete conversation between any two e-commerce customer service staff is taken as the cosine similarity between the sentence vectors in the complete conversation between any two e-commerce customer service staff, and all complete conversation processes of all e-commerce customer service staff are clustered using the ISODATA clustering algorithm based on the cosine similarity between the sentence vectors in the complete conversation between any two e-commerce customer service staff to obtain several conversation clusters.

[0017] Furthermore, the topic overlap between any two keyword sets is obtained by the difference in the number of keywords in the intersection and union of the keywords, and the specific steps include the following:

[0018] The ratio between the number of keywords in the intersection and the number of keywords in the union is recorded as the topic overlap between any two keyword sets.

[0019] Furthermore, according to the similarity between the word vectors in the corresponding texts of the e-commerce customer service during a complete conversation between any two e-commerce customer service staff, the degree of consistency of the words in a complete conversation between any two e-commerce customer service staff is obtained, and the specific calculation method included is as follows:

[0020] ;

[0021] In the formula, Indicates the first e-commerce customer service in a complete conversation The word vector of each word, Indicates the second e-commerce customer service in a complete conversation The word vector of each word, Representing word vectors and word vectors The cosine similarity between It indicates the number of all words in the text corresponding to the second e-commerce customer service during a complete conversation. Indicates the first e-commerce customer service in a complete conversation The maximum value of the cosine similarity between the word vector of the word and the word vector of all words in a complete conversation with the second e-commerce customer service, It indicates the number of all words in the text corresponding to the first e-commerce customer service during a complete conversation. It indicates the consistency of the words used in a complete conversation between any two e-commerce customer service representatives.

[0022] Furthermore, the business topic performance of each e-commerce customer service in each cluster is obtained by the number of dialogue rounds, the response time of all dialogue rounds, and the number of other users served in each dialogue round during a complete dialogue process of each e-commerce customer service in each cluster, and the specific steps include the following:

[0023] The time conversation performance of each e-commerce customer service in each cluster is obtained according to the response time of all conversation rounds in a complete conversation process of each e-commerce customer service in each cluster, and the number of other users served in each conversation round;

[0024] Calculate the mean of all conversation rounds in all complete conversations of each e-commerce customer service in each cluster, and negatively correlate the mean of the number of conversation rounds of each e-commerce customer service in each cluster to obtain the conversation round performance of each e-commerce customer service in each cluster;

[0025] The product of the time conversation expression and the conversation turn expression of each e-commerce customer service in each cluster is recorded as the business topic expression of each e-commerce customer service in each cluster.

[0026] Furthermore, according to the response time of all dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, and the number of other users served in each dialogue round, the time dialogue performance of each e-commerce customer service in each cluster is obtained, and the specific calculation method is as follows:

[0027] ;

[0028] In the formula, Indicates the number of complete conversations of each e-commerce customer service in each cluster. The average response time of all dialogue rounds in a complete dialogue process, Indicates the number of complete conversations of each e-commerce customer service in each cluster. The total number of users served during the complete conversation, represents the total number of all complete conversations of each e-commerce customer service in each cluster, represents the linear normalization function, represents an exponential function with a natural constant as base, Indicates the temporal conversation performance of each e-commerce customer service in each cluster.

[0029] Furthermore, the business topic familiarity of each e-commerce customer service in each cluster is obtained through the business topic expression of each e-commerce customer service in each cluster and the user evaluation satisfaction value of a complete conversation process of each e-commerce customer service, including the following specific steps:

[0030] The average of the user satisfaction values ​​of all complete conversations with each e-commerce customer service in each cluster is recorded as the service satisfaction of each e-commerce customer service in each cluster;

[0031] The product of the business topic expression of each e-commerce customer service in each cluster and the service satisfaction of each e-commerce customer service in each cluster is recorded as the business topic familiarity of each e-commerce customer service in each cluster.

[0032] Furthermore, the step of generating the first set of simulated training questions for each e-commerce customer service representative through the text in the simulated training cluster of each e-commerce customer service representative includes the following specific steps:

[0033] The generative model is used to generate questions for the simulated training clusters of each e-commerce customer service, and all the generated questions are grouped into a set, which is recorded as the first simulated training question set of each e-commerce customer service.

[0034] Furthermore, the method of selecting targeted business texts similar to the texts in the simulated training cluster of each e-commerce customer service through the similarity between all user feedback texts and the texts in the simulated training cluster of each e-commerce customer service includes the following specific steps:

[0035] The text vector of each feedback text is obtained through the bag-of-words model, and then the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service is obtained;

[0036] Calculate the cosine similarity between the text vectors of all feedback texts and the text vectors corresponding to the central text in the simulated training cluster of each e-commerce customer service, and normalize the cosine similarity. The normalized result is recorded as the approximation between the text vector of each feedback text and the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service. The text vectors with an approximation greater than the preset threshold are recorded as the approximation between the text vectors of each feedback text and the text vectors corresponding to the central text in the simulated training cluster of each e-commerce customer service. The feedback text is recorded as the targeted business text of each e-commerce customer service.

[0037] Furthermore, the method of generating a second set of simulated training questions for each e-commerce customer service representative using similar targeted business texts includes the following specific steps:

[0038] The targeted business text of each e-commerce customer service is used to generate questions through a generative model, and all the generated questions are grouped into a set, which is recorded as the second simulation training question set for each e-commerce customer service.

[0039] The beneficial effects of the technical solution of the present invention are as follows: according to the topic overlap in the text of a complete conversation between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's words, all complete conversation processes of all e-commerce customer service staff are clustered to obtain a number of conversation clusters, thereby improving the accuracy of similar conversation analysis; through the number of conversation rounds in a complete conversation process of each e-commerce customer service staff in each cluster, the response time of all conversation rounds, and the number of other users served in each conversation round, the business topic expression of each e-commerce customer service staff in each cluster is obtained, thereby improving the accuracy of clear analysis of business topic content; through the business topic expression of each e-commerce customer service staff in each cluster and the user evaluation satisfaction value of a complete conversation process of each e-commerce customer service staff, The familiarity of business topics of each e-commerce customer service in each cluster is obtained, thereby improving the accuracy of familiarity with business topics; the first set of simulated training questions for each e-commerce customer service is generated through the text in the simulated training cluster of each e-commerce customer service; through the similarity between all user feedback texts and the text in the simulated training cluster of each e-commerce customer service, targeted business texts similar to the texts in the simulated training cluster of each e-commerce customer service are selected to improve the accuracy of targeted analysis of real-time problems, and the second set of simulated training questions for each e-commerce customer service is generated from the similar targeted business texts; each e-commerce customer service is simulated trained through the first set of simulated training questions and the second set of simulated training questions of each e-commerce customer service, thereby improving the efficiency of targeted training for e-commerce customers. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 This is a flowchart of the steps of the method for generating a simulated training dialogue for e-commerce customer service of the present invention;

[0042] Figure 2 A flowchart generated for a simulated training conversation for e-commerce customer service. DETAILED DESCRIPTION

[0043] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the method for generating a simulated training dialogue for e-commerce customer service proposed by the present invention, its specific implementation method, structure, features and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0044] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0045] The specific scheme of the method for generating a simulated training dialogue for e-commerce customer service provided by the present invention is described in detail below with reference to the accompanying drawings.

[0046] See also Figure 1 , which shows a flowchart of a method for generating a simulated training dialogue for e-commerce customer service provided by an embodiment of the present invention, the method comprising the following steps:

[0047] Step S001: Collect the text of the complete conversation between each e-commerce customer service and the user, as well as the number of conversation rounds in a complete conversation process, the response time of each conversation round and the total number of users served simultaneously; then collect the content of questions raised by all users on the platform, as well as the content of user complaints received by the platform.

[0048] It should be noted that when conducting simulated conversation training for e-commerce customer service, it may not be possible to accurately locate business topics that the e-commerce customer service is not familiar with, resulting in simulated training under conventional random topic extraction. Not only can it fail to improve the proficiency of the e-commerce customer service, but it also wastes a certain amount of time. Therefore, the text of each e-commerce customer service's conversation with the user is analyzed to analyze the business topics that each e-commerce customer service is not familiar with, so as to conduct targeted training. Therefore, it is first necessary to collect the text of the e-commerce customer service during the conversation.

[0049] It should be further explained that in order to provide targeted training to users, the familiarity of e-commerce customer service with business topics can be analyzed from the performance of the e-commerce customer service in answering users. That is, the more familiar the e-commerce customer service is with business topics, the faster the response time to users will be. Since e-commerce customer service may reply to multiple users within a period of time, the grayscale response time will be slower. Therefore, the familiarity with business topics can be analyzed through the response of e-commerce customer service. Since each e-commerce customer service will evaluate the user based on the response after answering the user's question, this indicator can also be used to analyze the familiarity of e-commerce customer service with business topics.

[0050] It should be further explained that, because some business topics are time-sensitive, it is also necessary to collect the content of questions and complaints raised by users on the platform to ensure the timeliness of e-commerce customer service training.

[0051] Specifically, the background collects the text of the complete conversation between each e-commerce customer service representative and the user in the past year; collects the user evaluation satisfaction value of a complete conversation process of each e-commerce customer service representative, the response time of a conversation round in a complete conversation process of each e-commerce customer service representative, and the total number of users served at the same time.

[0052] Collect more The content of all questions raised by users on the platform within the month, as well as the content of user complaints received by the platform, will record a piece of content of a question raised by a user or a user complaint as a user feedback text. Among them, the time of obtaining the feedback text is not specifically limited. In this embodiment, , implementers can adjust according to specific circumstances.

[0053] Step S002: Based on the topic overlap in the text of a complete conversation between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's speech, all complete conversation processes of all e-commerce customer service staff are clustered to obtain a number of conversation clusters.

[0054] It should be noted that e-commerce customer service can simulate actual user dialogue scenarios and communicate with the generative model. This practice mode can improve the service quality of e-commerce customer service, enhance its adaptability, and better serve users. Usually, the generative model randomly extracts a topic from the question library and asks questions based on the topic. If the e-commerce customer service happens to extract questions related to business topics that he is familiar with, it will cause the e-commerce customer service to be unable to train on business topics that he is not familiar with.

[0055] It should be further explained that, in practice, when e-commerce customer service and users are communicating, users ask questions continuously, which prompts the e-commerce customer service to answer the questions asked, and the record of the conversation between users and customer service forms a conversation; each conversation may include topics such as after-sales issues, product issues, and usage issues. When the keywords contained in the questions asked by users in two conversations are the same, it means that the two conversations may be discussing the same topic. In addition, when the customer service answers, the rhetoric in the same field is similar; therefore, the similarity of any two conversations is determined by the overlap of user questions in the two topics and the consistency of the customer service answers.

[0056] Specifically, the complete conversation process between each e-commerce customer service and a user purchasing a product is recorded as a complete conversation process of each e-commerce customer service. For the text of a complete conversation process of each e-commerce customer service, Jieba word segmentation is used to obtain all words in the complete conversation process, and then the keyword extraction algorithm TF-IDF is used to obtain all keywords from all words, and all keywords in the complete conversation process are combined into a set, which is recorded as a keyword set of each e-commerce customer service. Among them, Jieba word segmentation and TF-IDF are both well-known technologies and will not be described in detail here.

[0057] Get the intersection and union of keywords between any two keyword sets, and analyze the number of keywords in the intersection and union of keywords to get the topic overlap between any two keyword sets. The specific formula is:

[0058] ;

[0059] In the formula, represents the number of keywords in the intersection of keywords between any two keyword sets, represents the number of keywords in the union of keywords between any two keyword sets, Represents the topic overlap between any two keyword sets.

[0060] Among them, the larger the ratio of the number of keywords in the intersection of keywords to the number of keywords in the union, that is, the closer it is to 1, the higher the topic overlap between any two keyword sets; conversely, the smaller the ratio, the lower the topic overlap between any two keyword sets.

[0061] It should be noted that when the same issue is discussed during a complete conversation between any two e-commerce customer service staff, the e-commerce customer service staff will use consistent words when answering the user. Therefore, the similarity between the words in the two conversations is very high. Therefore, the similarity between the two conversations is analyzed by the similarity between the word vectors of all words in the conversation between any two e-commerce customer service staff.

[0062] Specifically, the word vector of each word in the text corresponding to the e-commerce customer service is obtained through the Word2vec algorithm during a complete conversation between any two e-commerce customer service staff. Among them, the Word2vec algorithm is a well-known technology and will not be described in detail here.

[0063] Through the similarity between the word vectors of any two e-commerce customer service staff in a complete conversation, the consistency of the words in a complete conversation between any two e-commerce customer service staff is obtained, which is specifically expressed by the formula:

[0064] ;

[0065] In the formula, Indicates the first e-commerce customer service in a complete conversation The word vector of each word, Indicates the second e-commerce customer service in a complete conversation The word vector of each word, Representing word vectors and word vectors The cosine similarity between It indicates the number of all words in the text corresponding to the second e-commerce customer service during a complete conversation. Indicates the first e-commerce customer service in a complete conversation The maximum value of the cosine similarity between the word vector of the word and the word vector of all words in a complete conversation with the second e-commerce customer service, It indicates the number of all words in the text corresponding to the first e-commerce customer service during a complete conversation. It indicates the consistency of the words used in a complete conversation between any two e-commerce customer service representatives.

[0066] Among them, the higher the cosine similarity between the word vectors of any two e-commerce customer service staff during a complete conversation, the greater the consistency of the words used during a complete conversation between any two e-commerce customer service staff.

[0067] Through the consistency of words in a complete conversation between any two e-commerce customer service staff and the topic overlap between any two keyword sets, the similarity of the conversation between any two e-commerce customer service staff is obtained, which is specifically expressed by the formula:

[0068] ;

[0069] In the formula, represents the topic overlap between any two keyword sets, It indicates the consistency of words used by any two e-commerce customer service representatives during a complete conversation. Indicates the similarity of the conversation between any two e-commerce customer service staff during the complete conversation.

[0070] Among them, when there is a complete conversation between any two e-commerce customer service staff, the higher the topic overlap and the consistency of the words, the higher the similarity of the conversation between any two e-commerce customer service staff. Conversely, the lower the similarity of the conversation between any two e-commerce customer service staff.

[0071] According to all the keywords in the keyword set corresponding to a complete conversation process of each e-commerce customer service, a sentence vector is constructed; the similarity of the conversations between any two e-commerce customer service complete conversations is used as the cosine similarity between the sentence vectors in the complete conversation process of any two e-commerce customer service, and all the complete conversation processes of all e-commerce customer service are clustered by the ISODATA clustering algorithm according to the cosine similarity between the sentence vectors in the complete conversation process of any two e-commerce customer service, and several conversation clusters are obtained; wherein the ISODATA clustering algorithm is a well-known technology and will not be described in detail here. Among them, each cluster corresponds to a business topic.

[0072] At this point, several conversation clusters are obtained.

[0073] Step S003: Obtain the business topic performance of each e-commerce customer service in each cluster through the number of dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, the response time of all dialogue rounds, and the number of other users served in each dialogue round; obtain the business topic familiarity of each e-commerce customer service in each cluster through the business topic performance of each e-commerce customer service in each cluster and the user evaluation satisfaction value of a complete dialogue process of each e-commerce customer service; select a cluster with the smallest business topic familiarity of each e-commerce customer service in all clusters, record it as the simulation training cluster of each e-commerce customer service, and generate the first simulation training question set for each e-commerce customer service through the text in the simulation training cluster of each e-commerce customer service.

[0074] It should be noted that when the e-commerce customer service is more familiar with the business topic, the e-commerce customer service is more proficient in the business topic and can provide users with better solutions. Therefore, the e-commerce customer service's familiarity with each business topic is calculated based on the user satisfaction when answering the business topic and the customer service's performance when answering, and then targeted business topic simulation dialogue training is carried out for each e-commerce customer service based on the familiarity.

[0075] Specifically, according to the user evaluation satisfaction value of all complete conversations of each e-commerce customer service in each cluster, the service satisfaction of each e-commerce customer service in each cluster is obtained, which is specifically expressed by the formula:

[0076] ;

[0077] In the formula, Indicates the number of complete conversations of each e-commerce customer service in each cluster. User satisfaction value of the complete conversation process, represents the total number of all complete conversations of each e-commerce customer service in each cluster, Represents the service satisfaction of each e-commerce customer service in each cluster.

[0078] It should be noted that when e-commerce customer service is very familiar with business topics, when users ask questions, they can quickly locate relevant knowledge in their minds and give answers quickly. Therefore, the shorter the response time of e-commerce customer service, the more familiar the e-commerce customer service is with the business topic. In addition, skilled e-commerce customer service can solve user problems in fewer conversation rounds. They can accurately understand user needs and give effective solutions. Therefore, the performance characteristics of e-commerce customer service when answering business topics are quantified through the response time of e-commerce customer service and the number of conversation rounds with users.

[0079] Specifically, according to the response time of all dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, and the number of other users served in each dialogue round, the time dialogue performance of each e-commerce customer service in each cluster is obtained; the specific formula is:

[0080] ;

[0081] In the formula, Indicates the number of complete conversations of each e-commerce customer service in each cluster. The average response time of all dialogue rounds in a complete dialogue process, Indicates the number of complete conversations of each e-commerce customer service in each cluster. The total number of users served during the complete conversation, represents the total number of all complete conversations of each e-commerce customer service in each cluster, represents the linear normalization function, represents an exponential function with a natural constant as base, Indicates the temporal conversation performance of each e-commerce customer service in each cluster.

[0082] Among them, during all the complete conversations of each e-commerce customer service in each cluster, the shorter the response time of all conversation rounds, the higher the proficiency of the e-commerce customer service in the business topic corresponding to the cluster; conversely, the lower the proficiency of the e-commerce customer service in the business topic corresponding to the cluster. When the e-commerce customer service has to handle a lot of user information at the same time, the response time of the e-commerce customer service for this conversation reflects the less authenticity of the e-commerce customer service's familiarity with the business. Therefore, the reciprocal of the number of other user conversations that the e-commerce customer service has to handle while handling the current conversation is used as the weight of the response time of this conversation. The weight is expressed by express.

[0083] It should be further explained that the fewer rounds of conversation the e-commerce customer service has during a complete conversation, the more proficient the e-commerce customer service is in business topics, and the faster the customer service can solve user problems.

[0084] Specifically, according to all the complete conversations of each e-commerce customer service in each cluster and the number of all conversation rounds in each complete conversation, the conversation round performance of each e-commerce customer service in each cluster is obtained; the specific formula is:

[0085] ;

[0086] In the formula, Indicates the number of complete conversations of each e-commerce customer service in each cluster. The number of all dialogue rounds during a complete dialogue; represents the total number of all complete conversations of each e-commerce customer service in each cluster, represents an exponential function with a natural constant as base, Represents the conversation turn performance of each e-commerce customer service in each cluster.

[0087] Among them, the fewer the number of conversation rounds in a complete conversation process, the higher the proficiency of the e-commerce customer service in the business topics corresponding to the cluster; conversely, the lower the proficiency of the e-commerce customer service in the business topics corresponding to the cluster.

[0088] According to the time conversation performance and conversation turn performance of each e-commerce customer service in each cluster, the business topic performance of each e-commerce customer service in each cluster is obtained, which is expressed by the formula:

[0089] ;

[0090] In the formula, Indicates the time conversation performance of each e-commerce customer service in each cluster, represents the dialogue turn performance of each e-commerce customer service in each cluster, Indicates the business topic expression of each e-commerce customer service in each cluster.

[0091] According to the service satisfaction and business topic performance of each e-commerce customer service in each cluster, the business topic familiarity of each e-commerce customer service in each cluster is obtained, which is expressed by the formula:

[0092] ;

[0093] In the formula, represents the service satisfaction of each e-commerce customer service in each cluster, Indicates the business topic performance of each e-commerce customer service in each cluster, Indicates the familiarity of each e-commerce customer service representative in each cluster with respect to the business topic.

[0094] Among them, when the service satisfaction of each e-commerce customer service in each cluster is greater and the business topic expression is higher, it means that each e-commerce customer service in each cluster is more familiar with the business topic, that is, the familiarity is greater; otherwise, the familiarity is smaller.

[0095] Select a cluster with the smallest familiarity of business topics for each e-commerce customer service from all clusters, and record it as the simulation training cluster for each e-commerce customer service; generate questions for each e-commerce customer service simulation training cluster through a generative model, and form all generated questions into a set, which is recorded as the first simulation training question set for each e-commerce customer service. The generative model is a well-known technology and will not be described in detail here.

[0096] Step S004: through the similarity between all user feedback texts and the texts in the simulation training cluster of each e-commerce customer service, select the targeted business texts similar to the texts in the simulation training cluster of each e-commerce customer service, and generate the second simulation training question set for each e-commerce customer service with the similar targeted business texts; conduct simulation training for each e-commerce customer service through the first simulation training question set and the second simulation training question set of each e-commerce customer service.

[0097] It should be noted that when conducting simulation training for e-commerce customer service, the virtual questions of the generative model are generally fixed questions (the content of the questions remains unchanged). The virtual questions in the fixed mode have the following disadvantages. First, the virtual questions of fixed questions cannot test the customer service's ability to respond. Second, they cannot target the user's pain points. Due to the rapid development of e-commerce, the questions raised by users are time-sensitive. After a certain period of time, the question is outdated, and it will be meaningless for the e-commerce customer service to conduct training on this question.

[0098] Specifically, the text vector of each feedback text is obtained through the bag-of-words model, and then the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service is obtained; wherein, the bag-of-words model is a well-known technology and will not be described in detail here.

[0099] Calculate the cosine similarity between the text vectors of all feedback texts and the text vectors corresponding to the central text in the simulated training cluster of each e-commerce customer service, and normalize the cosine similarity. The normalized result is recorded as the approximation between the text vector of each feedback text and the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service. The text vectors with an approximation greater than the preset threshold are recorded as the approximation between the text vectors of each feedback text and the text vectors corresponding to the central text in the simulated training cluster of each e-commerce customer service. The feedback text is recorded as the targeted business text of each e-commerce customer service. In this embodiment, the preset threshold , in which no Make specific restrictions.

[0100] Generate questions for each e-commerce customer service's targeted business text through a generative model, and group all generated questions into a set, which is recorded as the second simulation training question set for each e-commerce customer service; simulate training is performed on each e-commerce customer service through the first simulation training question set and the second simulation training question set of each e-commerce customer service. The flowchart generated by the e-commerce customer service simulation training dialogue is as follows: Figure 2 shown.

[0101] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating a simulated training dialogue for e-commerce customer service, characterized in that: The method comprises the following steps: Obtain the text of the complete conversation between each e-commerce customer service and the user; obtain the number of conversation rounds in a complete conversation with each e-commerce customer service, the response time of each conversation round, and the total number of users served at the same time; obtain the content of all questions raised by users on the platform, as well as the content of user complaints received by the platform, and record a question raised by a user or a user complaint as a user feedback text; The complete conversation process between each e-commerce customer service and a user purchasing a product is recorded as a complete conversation process of each e-commerce customer service. According to the topic overlap in the text of a complete conversation process between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's words, all the complete conversation processes of all e-commerce customer service staff are clustered to obtain several conversation clusters; The business topic performance of each e-commerce customer service in each cluster is obtained through the number of dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, the response time of all dialogue rounds, and the number of other users served in each dialogue round; the business topic familiarity of each e-commerce customer service in each cluster is obtained through the business topic performance of each e-commerce customer service in each cluster and the user evaluation satisfaction value of a complete dialogue process of each e-commerce customer service; a cluster with the smallest business topic familiarity of each e-commerce customer service in all clusters is selected and recorded as the simulation training cluster of each e-commerce customer service, and the first simulation training question set of each e-commerce customer service is generated through the text in the simulation training cluster of each e-commerce customer service; Through the similarity between all user feedback texts and the texts in the simulated training cluster of each e-commerce customer service, select targeted business texts similar to the texts in the simulated training cluster of each e-commerce customer service, and generate the second simulated training question set for each e-commerce customer service with the similar targeted business texts; conduct simulated training for each e-commerce customer service through the first simulated training question set and the second simulated training question set of each e-commerce customer service; The method for obtaining the business topic performance is as follows: according to the response time of all dialogue rounds in a complete dialogue process of each e-commerce customer service in each cluster, and the number of other users served in each dialogue round, the time dialogue performance of each e-commerce customer service in each cluster is obtained; Calculate the mean of all conversation rounds in all complete conversations of each e-commerce customer service in each cluster, and negatively correlate the mean of the number of conversation rounds of each e-commerce customer service in each cluster to obtain the conversation round performance of each e-commerce customer service in each cluster; The product of the time conversation performance and the conversation turn performance of each e-commerce customer service in each cluster is recorded as the business topic performance of each e-commerce customer service in each cluster; The method for obtaining the business topic familiarity is: taking the average of the user satisfaction values ​​of all complete conversations of each e-commerce customer service in each cluster as the service satisfaction of each e-commerce customer service in each cluster; The product of the business topic expression of each e-commerce customer service in each cluster and the service satisfaction of each e-commerce customer service in each cluster is recorded as the business topic familiarity of each e-commerce customer service in each cluster.

2. According to claim 1, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: According to the topic overlap in the text of a complete conversation between any two e-commerce customer service staff and the consistency of the e-commerce customer service staff's speech, all the complete conversation processes of all e-commerce customer service staff are clustered to obtain several conversation clusters, including the following specific steps: For the text of a complete conversation with each e-commerce customer service representative, use Jieba word segmentation to obtain all the words in the complete conversation, and then use the keyword extraction algorithm TF-IDF to obtain all the keywords from all the words. All the keywords in the complete conversation are grouped into a set, which is recorded as a keyword set for each e-commerce customer service representative. Obtain the intersection and union of keywords between any two keyword sets, and obtain the topic overlap between any two keyword sets by the difference in the number of keywords in the intersection and union of keywords; Use the Word2vec algorithm to obtain the word vector of each word in the text corresponding to the e-commerce customer service during a complete conversation between any two e-commerce customer service staff. According to the similarity between the word vectors in the corresponding texts of any two e-commerce customer service representatives during a complete conversation, the consistency of the words used in a complete conversation between any two e-commerce customer service representatives is obtained; The product of the topic overlap between any two keyword sets and the consistency of the words in a complete conversation between any two e-commerce customer service staff is recorded as the similarity of the conversation in a complete conversation between any two e-commerce customer service staff; sentence vectors are constructed based on all the keywords in the keyword set corresponding to a complete conversation between each e-commerce customer service staff; the similarity of the conversation in a complete conversation between any two e-commerce customer service staff is taken as the cosine similarity between the sentence vectors in the complete conversation between any two e-commerce customer service staff, and all complete conversation processes of all e-commerce customer service staff are clustered using the ISODATA clustering algorithm based on the cosine similarity between the sentence vectors in the complete conversation between any two e-commerce customer service staff to obtain several conversation clusters.

3. According to claim 2, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: The topic overlap between any two keyword sets is obtained by using the difference in the number of keywords in the intersection and union of the keywords, and the specific steps include the following: The ratio between the number of keywords in the intersection and the number of keywords in the union is recorded as the topic overlap between any two keyword sets.

4. According to claim 2, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: According to the similarity between the word vectors in the corresponding texts of any two e-commerce customer service during a complete conversation, the degree of consistency of the words in a complete conversation between any two e-commerce customer service is obtained, and the specific calculation method is as follows: Where C1 i C2 represents the word vector of the i-th word in a complete conversation with the first e-commerce customer service. j represents the word vector of the jth word in a complete conversation with the second e-commerce customer service, cos(C1 i ,C2 j ) represents the word vector C1 i And word vector C2 j The cosine similarity between them, n2 represents the number of all words in the text corresponding to the second e-commerce customer service in a complete conversation, max j∈n2 [cos(C1 i ,C2 j )] represents the maximum value of the cosine similarity between the word vector of the ith word in a complete conversation with the first e-commerce customer service and the word vector of all words in a complete conversation with the second e-commerce customer service. n1 represents the number of all words in the corresponding text of the first e-commerce customer service in a complete conversation. Z represents the degree of consistency in the words used in a complete conversation between any two e-commerce customer service staff.

5. According to claim 1, the method for generating a simulated training dialogue for e-commerce customer service is characterized in that: The time conversation performance of each e-commerce customer service in each cluster is obtained according to the response time of all conversation rounds in a complete conversation process of each e-commerce customer service in each cluster, and the number of other users served in each conversation round. The specific calculation method is as follows: In the formula, represents the mean response time of all dialogue rounds in the rth complete dialogue process of all complete dialogue processes of each e-commerce customer service in each cluster, q r represents the total number of users served in the rth complete conversation process among all the complete conversation processes of each e-commerce customer service in each cluster, m represents the total number of all complete conversation processes of each e-commerce customer service in each cluster, norm() represents the linear normalization function, exp() represents the exponential function with natural constant as the base, and T represents the time conversation performance of each e-commerce customer service in each cluster.

6. According to claim 1, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: The specific steps of generating the first set of simulated training questions for each e-commerce customer service by using the text in the simulated training cluster of each e-commerce customer service include: The generative model is used to generate questions for the simulated training clusters of each e-commerce customer service, and all the generated questions are grouped into a set, which is recorded as the first simulated training question set of each e-commerce customer service.

7. According to claim 1, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: The specific steps of selecting targeted business texts similar to the texts in the simulated training cluster of each e-commerce customer service by comparing the similarities between all user feedback texts and the texts in the simulated training cluster of each e-commerce customer service include the following: The text vector of each feedback text is obtained through the bag-of-words model, and then the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service is obtained; The cosine similarity between the text vectors of all feedback texts and the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service is calculated, and the cosine similarity is normalized. The normalized result is recorded as the approximation between the text vector of each feedback text and the text vector corresponding to the central text in the simulated training cluster of each e-commerce customer service. The feedback text with a similarity greater than the preset threshold ε is recorded as the targeted business text of each e-commerce customer service.

8. According to claim 1, a method for generating a simulated training dialogue for e-commerce customer service, characterized in that: The specific steps of generating a second set of simulated training questions for each e-commerce customer service representative using similar targeted business texts are as follows: The targeted business text of each e-commerce customer service is used to generate questions through a generative model, and all the generated questions are grouped into a set, which is recorded as the second simulation training question set for each e-commerce customer service.

Images