User-oriented intelligent customer service method and device, equipment and medium

Abstract

The application relates to a user-oriented intelligent customer service method and device, equipment and medium in the e-commerce technical field, the method comprising: in response to a user enabling an intelligent customer service event, obtaining personal information and historical behavior data of the user, and determining a corresponding user portrait; in response to a user asking event, obtaining an asking text input by the user, determining a standard asking text matched with the asking text in a preset question and answer knowledge base, and obtaining a standard reply text of the standard asking text; applying a preset instruction template, generating a corresponding constraint instruction according to the user portrait and the standard reply text, the constraint instruction indicating that the standard reply text is re-expressed according to the user portrait to match the character style of the user; and inputting the constraint instruction into a preset personification question and answer model to generate a user-oriented reply. The application can match the character style of the user to reply, and realizes more real and personified customer service.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of e-commerce, and in particular to a user-oriented intelligent customer service method and a corresponding device, computer equipment and computer readable storage medium. BACKGROUND

[0002] With the continuous improvement of people's quality of life, more and more consumers prefer online shopping, which makes the number of users requiring customer service grow explosively. In order to provide customer service in a timely manner and efficiently reply to users, intelligent customer service has been developed.

[0003] The intelligent customer service implemented by the traditional technology usually retrieves a question and answer knowledge base according to the user's question text, determines the standard question text and the corresponding standard reply text therein, and directly replies to the user with the standard reply text. However, the standard reply text is only a professional answer to the question raised by the user, and such a reply is extremely harsh and mechanical, which cannot move and arouse the user's empathy, resulting in the user's dissatisfaction and the customer's feeling of coldness and unfriendliness.

[0004] In view of the deficiencies of the traditional technology, the present applicant has long been engaged in the research in the relevant field, and has developed a user-oriented intelligent customer service method and a corresponding device, computer equipment and computer readable storage medium to solve the problems in the e-commerce field. SUMMARY

[0005] The primary purpose of the present application is to solve at least one of the above problems and provide a user-oriented intelligent customer service method and a corresponding device, computer equipment and computer readable storage medium.

[0006] To meet the various purposes of the present application, the present application adopts the following technical solutions:

[0007] A user-oriented intelligent customer service method provided for one of the purposes of the present application comprises the following steps:

[0008] In response to a user enabling an intelligent customer service event, personal information and historical behavior data of the user are obtained, and a corresponding user portrait is determined;

[0009] In response to a user question event, a question text input by the user is obtained, a standard question text in a preset question and answer knowledge base that matches the question text is determined, and a standard reply text of the standard question text is obtained;

[0010] A preset instruction template is applied, and a corresponding constraint instruction is generated according to the user portrait and the standard reply text, the constraint instruction indicating that the standard reply text is re-expressed in a character style that fits the user according to the user portrait;

[0011] The constraint instruction is input into a preset personification question and answer model, and a user-oriented reply is generated.

[0012] In a further embodiment, determining the standard question text in the preset question and answer knowledge base that matches the question text comprises the following steps:

[0013] Determining the corresponding semantic similarity between the question text and each standard question text in the preset question and answer knowledge base using a preset text similarity model;

[0014] Filtering out the standard question text whose semantic similarity exceeds a preset threshold.

[0015] In a further embodiment, before determining the corresponding user portrait based on the user's personal information and historical behavior data, the following steps are included:

[0016] Obtaining the user portrait of the user inputting the historical question text, and the supervised reply text generated in response to the historical question text;

[0017] Determining the standard question text in the preset question and answer knowledge base that matches the historical question text, and obtaining the standard reply text of the standard question text;

[0018] Applying a preset instruction template to generate a corresponding constraint instruction as a training sample based on the user portrait and the standard reply text, and using the supervised reply text as a supervised label;

[0019] Inputting the training sample into the anthropomorphic reply model to extract deep semantic information of the training sample and generate a predicted user-oriented reply;

[0020] Using the supervised label of the training sample to determine the loss value of the predicted user-oriented reply, and when the loss value does not reach a preset threshold, implementing weight update on the anthropomorphic reply model and continuing to call other training samples for iterative training until the anthropomorphic reply model converges.

[0021] In a further embodiment, determining the corresponding user portrait comprises the following steps:

[0022] Determining the basic attribute label and the preference attribute label based on the user's personal information and historical behavior data;

[0023] Constructing the corresponding user portrait based on the user's basic attribute label and preference attribute label.

[0024] In a further embodiment, before obtaining the user portrait of the user inputting the historical question text and the supervised reply text generated in response to the historical question text, the following steps are included:

[0025] obtaining a user portrait of a user inputting a historical question text, and a historical reply text replying to the historical question text;

[0026] determining a character style of the user corresponding to the user portrait;

[0027] inputting the historical reply text into a preset style classification model to determine a corresponding reply style;

[0028] when the reply style matches the user style to achieve a fit, taking the historical reply text as a supervised reply text.

[0029] In further embodiments, before obtaining a user portrait of a user inputting a historical question text, and a historical reply text replying to the historical question text, the following steps are included:

[0030] obtaining a historical question text and its corresponding historical reply text, and constructing a text data pair;

[0031] inputting the text data pair into a preset reply matching model to determine a corresponding reply quality;

[0032] removing the historical question text and its corresponding historical reply text with a reply quality lower than a preset threshold.

[0033] In further embodiments, after responding to a user question event and obtaining a question text input by a user, the following steps are included:

[0034] in response to a matching failure event, using a preset intent classification model to determine whether the intent of the question text is clear;

[0035] when the intent is clear, sending the question text to an artificial customer service interface;

[0036] when the intent is not clear, constructing a re-input notification to be pushed to the user.

[0037] In another aspect, a user-oriented intelligent customer service device is provided for one of the purposes of the present application, comprising a customer service enabling module, a user question module, an instruction generating module, and a reply generating module. The customer service enabling module is configured to respond to a user-enabled intelligent customer service event, obtain personal information and historical behavior data of the user, and determine a corresponding user portrait. The user question module is configured to respond to a user question event, obtain a question text input by the user, determine a standard question text in a preset question and answer knowledge base that matches the question text, and obtain a standard reply text of the standard question text. The instruction generating module is configured to apply a preset instruction template, generate a corresponding constraint instruction according to the user portrait and the standard reply text, and the constraint instruction indicates that the standard reply text is re-expressed in a character style that fits the user according to the user portrait. The reply generating module is configured to input the constraint instruction into a preset personified question and answer model to generate a user-oriented reply.

[0038] In further embodiments, the user question module comprises a similarity determining submodule configured to determine a corresponding semantic similarity between the question text and each standard question text in the preset question and answer knowledge base using a preset text similarity model; and a question text screening submodule configured to screen out the standard question text whose semantic similarity exceeds a preset threshold.

[0039] In further embodiments, the user question module comprises a first data obtaining submodule configured to obtain a user portrait of a user who inputs a historical question text and a supervised reply text generated in response to the historical question text; determine a standard question text in a preset question and answer knowledge base that matches the historical question text, and obtain a standard reply text of the standard question text; a sample supervision constructing submodule configured to apply a preset instruction template, generate a corresponding constraint instruction as a training sample according to the user portrait and the standard reply text, and use the supervised reply text as a supervision label; a reply generating submodule configured to input the training sample into a personified reply model, extract deep semantic information of the training sample, and generate a predicted user-oriented reply; and an iterative training submodule configured to use the supervision label of the training sample to determine a loss value of the predicted user-oriented reply, update weights of the personified reply model when the loss value does not reach a preset threshold, and continue to call other training samples for iterative training until the personified reply model converges.

[0040] In further embodiments, the customer service enabling module comprises a label determining submodule configured to determine basic attribute labels and preference attribute labels according to the personal information and historical behavior data of the user; and a portrait constructing submodule configured to construct a corresponding user portrait according to the basic attribute labels and the preference attribute labels of the user.

[0041] In a further embodiment, the data acquisition submodule is preceded by a second data acquisition submodule configured to acquire a user portrait of a user inputting the historical question text and historical reply text replying to the historical question text, a character style determination submodule configured to determine a character style of the user corresponding to the user portrait, a reply style determination submodule configured to input the historical reply text into a preset style classification model to determine a corresponding reply style, and a supervision determination submodule configured to, when the reply style matches the user style to achieve a fit, take the historical reply text as a supervision reply text.

[0042] In a further embodiment, the second data acquisition submodule is preceded by a third data acquisition submodule configured to acquire historical question text and corresponding historical reply text to construct a text data pair, a quality determination submodule configured to input the text data pair into a preset reply matching model to determine a corresponding reply quality, and a text removal submodule configured to remove historical question text and corresponding historical reply text with a reply quality lower than a preset threshold.

[0043] In a further embodiment, the user question module is followed by an intent determination submodule configured to, in response to a matching failure event, determine whether the intent of the question text is clear using a preset intent classification model, an intent clear submodule configured to, when the intent is clear, send the question text to an artificial customer service interface, and an intent unclear submodule configured to, when the intent is unclear, construct a re-input notification to push to the user.

[0044] In another aspect, a computer device is provided to adapt to one of the purposes of the present application, comprising a central processing unit and a memory, the central processing unit being configured to invoke a computer program stored in the memory to execute the steps of the user-oriented intelligent customer service method described in the present application.

[0045] In another aspect, a computer readable storage medium is provided to adapt to another purpose of the present application, which stores a computer program implemented according to the user-oriented intelligent customer service method in the form of computer readable instructions, and the computer program is invoked and run by a computer to execute the steps included in the method.

[0046] The technical solutions of the present application have multiple advantages, including but not limited to the following aspects:

[0047] The application responds to the user to enable the intelligent customer service event, obtains the personal information and historical behavior data of the user, determines the corresponding user portrait, responds to the user question event, obtains the question text input by the user, determines the standard question text in the preset question and answer knowledge base matched with the question text, obtains the standard reply text of the standard question text, applies the preset instruction template, generates the corresponding constraint instruction according to the user portrait and the standard reply text, the constraint instruction indicates that the standard reply text is re-expressed according to the user portrait to match the user's character style, the constraint instruction is input into the preset personification question and answer model, and a reply to the user is generated. It can be seen that the user portrait of the user is understood, the expression of the standard reply text is improved to match the user's character style on the basis of maintaining the professional answering ability of the standard reply text, the intelligent customer service is more realistic, the user experiences the feeling of communicating with a real person who understands himself, the user satisfaction is greatly improved, and the user stickiness is increased. BRIEF DESCRIPTION OF DRAWINGS

[0048] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0049] Figure 1 A flowchart of a typical embodiment of the user-oriented intelligent customer service method of the present application;

[0050] Figure 2 A flowchart of screening out the standard question text similar to the question text in the question and answer knowledge base in the embodiment of the present application;

[0051] Figure 3 A flowchart of the training process of the personification reply model in the embodiment of the present application;

[0052] Figure 4 A flowchart of constructing a user portrait in the embodiment of the present application;

[0053] Figure 5 A flowchart of determining the supervision reply text in the embodiment of the present application;

[0054] Figure 6 A flowchart of filtering the historical question text and the corresponding historical reply text according to the reply quality in the embodiment of the present application;

[0055] Figure 7 A flowchart of whether the corresponding processing is performed for the question text input by the user when the question text input by the user fails to match in the embodiment of the present application;

[0056] Figure 8 A principle block diagram of the user-oriented intelligent customer service device of the present application;

[0057] Figure 9 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0058] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0059] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0060] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0061] Those skilled in the art will understand that, as used herein, the terms "client," "terminal," and "terminal device" include both devices that are solely wireless signal receivers and devices that have both receiving and transmitting hardware that can communicate bi-directionally over a bi-directional communication link. Such devices can include cellular or other communication devices with single-line or multiple-line displays, or no display, Personal Communications Service (PCS) devices that can combine a voice and / or data processor, a PDA that can include a radio frequency receiver and a pager, Internet and / or Intranet access, a Web browser, a calendar, and / or a GPS receiver, a conventional laptop and / or palmtop computer and / or other devices that have a radio frequency receiver. As used herein, the terms "client," "terminal," and "terminal device" can be portable, transportable, mounted in a vehicle (aeronautical, maritime, and / or land), or adapted and / or configured for local and / or distributed operation on Earth and / or any other location in space. As used herein, the terms "client," "terminal," and "terminal device" can also be a communication terminal, an Internet terminal, a music / video playing terminal, such as a PDA, a Mobile Internet Device (MID), and / or a mobile phone with music / video playing function, a smart television, a set-top box, and / or the like.

[0062] As used herein, the terms "server," "client," "service node," and the like refer to hardware that has the equivalent capability of a personal computer, i.e., an electronic device having a central processing unit (including an arithmetic unit and a controller), a memory, an input device, and an output device, and the like necessary components disclosed by the Von Neumann principle. A computer program is stored in the memory, the central processing unit loads the program stored in the external memory into the memory and runs it, executes the instructions in the program, and interacts with the input and output devices, thereby completing a specific function.

[0063] It should be noted that the concept of "server" in the present application can also be extended to the case of a server cluster. According to the principle of network deployment understood by those skilled in the art, the servers should be logically divided, and in physical space, these servers can be independent of each other but can be called through an interface, or can be integrated into a physical computer or a computer cluster. Those skilled in the art should understand this variation and should not be restricted by the implementation of the network deployment of the present application.

[0064] One or more technical features of the present application, unless explicitly specified, can be deployed on a server for implementation and accessed by a client remotely calling an online service interface provided by the server, or can be directly deployed and run on a client for implementation.

[0065] The neural network model referred to or possibly referred to in the present application, unless explicitly specified, can be deployed on a remote server and remotely called by a client, or can be deployed on a client with sufficient device capability for direct calling. In some embodiments, when it is run on a client, its corresponding intelligence can be obtained through transfer learning, so as to reduce the requirement for client hardware running resources and avoid excessive occupation of client hardware running resources.

[0066] The various data involved in the present application, unless explicitly specified, can be remotely stored on a server or stored on a local terminal device, as long as it is suitable for being called by the technical solutions of the present application.

[0067] Those skilled in the art should know that the various methods of the present application, although based on the same concept and described to present commonality between them, are independently executable unless otherwise specified. Similarly, for each embodiment disclosed in the present application, it is based on the same inventive concept, so the same concept is understood to be equivalent, and although the concept is expressed differently, it is only a suitable transformation for convenience.

[0068] Unless it is explicitly stated that the various embodiments disclosed in the present application are mutually exclusive, the related technical features involved in each embodiment can be cross-combined to flexibly construct new embodiments, as long as such combination does not deviate from the spirit of the present application and can meet the needs of the prior art or solve some aspects of the deficiencies in the prior art. For this variation, those skilled in the art should know.

[0069] The user-oriented intelligent customer service method of the present application can be programmed as a computer program product, deployed in a client or a server for running and implementation, for example, in the exemplary application scenario of the present application, it can be deployed and implemented in the server of an e-commerce platform, thereby the interface opened after the running of the computer program product can be accessed, the process of the computer program product can be interacted with through the graphical user interface to execute the method.

[0070] Please refer to Figure 1 In a typical embodiment of the user-oriented intelligent customer service method of the present application, the following steps are included:

[0071] Step S1100, in response to the user enabling the intelligent customer service event, obtaining the personal information and historical behavior data of the user, and determining the corresponding user portrait;

[0072] The user of the e-commerce platform can trigger the intelligent customer service by the customer service entrance on the page of the online store or other pages of the e-commerce platform, such as the touch "contact customer service" bottom navigation bar control, or the user can trigger the intelligent customer service by the customer service entrance on the current page, such as the "contact customer service" bubble control in the lower right corner of the page, when the user is browsing goods, adding goods to the shopping cart, settling accounts, paying, and other e-commerce business links, and the corresponding intelligent customer service request is sent to the server.

[0073] The server of the e-commerce platform responds to the user enabling the intelligent customer service event by receiving the intelligent customer service request, connects the user to the corresponding online store or the official intelligent customer service of the e-commerce platform, and provides the intelligent customer service service of automatically replying to the user's questions.

[0074] In order for the intelligent customer service to be based on as much as possible understanding of the user, so as to reply to the user, in an embodiment, according to the unique identifier of the user, such as ID, the user database is accessed to obtain the personal information and historical behavior data of the user, the personal information can be the relevant information input by the user when registering on the e-commerce platform, including any one or more of gender, delivery address, education level, age / birth date, occupation, etc., the historical behavior data can be the behavior record of the user in the past on the e-commerce platform, including any one or more of the behaviors of liking, forwarding, collecting, adding to the shopping cart, purchasing, browsing, etc. Based on the personal information and historical behavior data, the corresponding user portrait can be determined, and the specific implementation is further disclosed by the subsequent embodiments, and this step will not be described here, and the exemplary examples are as follows: age: 28, education level: master, gender: female, occupation: finance, location: Guangzhou Haizhu District, Guangdong Province, China, brand preference: Chanel, Gucci, purchase cycle: three times a month, consumption level: high consumption level, product type preference: lipstick, bag, cosmetics.

[0075] Step S1200, in response to the user asking event, obtaining the user input question text, determining the standard question text in the preset Q&A knowledge base matched with the question text, obtaining the standard reply text of the standard question text;

[0076] The Q&A knowledge base is constructed in advance, covering a wide range of commodity fields and service categories, such as commodity inquiry, service guidance, order tracking, after-sales maintenance, etc., including standard question texts and corresponding standard reply texts of various aspects of e-commerce platforms. The standard question text has the characteristics of concise and brief, accurately indicating the intention, and expression without emotion. The standard reply text has the characteristics of concise and brief, accurate and professional reply, and expression without emotion. For ease of understanding, exemplary examples are as follows: standard question text: What colors can this product be selected? Standard reply text: This product has three colors of red, blue and green to choose from. The color may vary slightly due to batch differences, please refer to the actual product.

[0077] The Q&A knowledge base can be directly constructed by the operation personnel of the e-commerce platform, which is suitable for the entire e-commerce platform, or it can be constructed by the operation personnel of the e-commerce platform providing a basic Q&A template, and the merchant user of the online store on the operation platform constructs the Q&A suitable for his own online store according to the basic Q&A template combined with his own needs.

[0078] Generally, the e-commerce platform provides a Q&A window for communication between the user and the intelligent customer service. The user can input the question text to the Q&A window to trigger the generation of a question request carrying the question text sent to the intelligent customer service.

[0079] The intelligent customer service receives the question request, responds to the user question event, and obtains the question text. When the intelligent customer service is an intelligent customer service of an online store, a standard question text similar in semantic to the question text is matched from a question and answer database preset for the online store. To ensure the accuracy of the matching, in the embodiments, a text semantic model pre-trained to convergence is used to determine a text feature vector that vectorizes the semantic of the question text, and a text feature vector that vectorizes the semantic corresponding to each standard question text in the question and answer database. The vector distance between the question text and the text feature vector corresponding to each standard question text is determined, which accurately represents the semantic similarity between the question text and each standard question text. Accordingly, the standard question text matching the question text can be screened out, and a standard reply text replying to the standard question text is obtained from the question and answer database. When the intelligent customer service is an intelligent customer service of an e-commerce platform, a standard question text similar in semantic to the question text is matched from a question and answer database preset for the e-commerce platform, and a standard reply text replying to the standard question text is obtained from the question and answer database.

[0080] The text semantic model is pre-trained to convergence, learns to extract deep semantic information of the text input thereto, and outputs a text feature vector that vectorizes the semantic of the text. The specific type can be TextTransfomer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc., and a person skilled in the art can select one as needed. Since the training process of these models is known in the art, it is not described in detail.

[0081] Step S1300, applying a preset instruction template, generating a corresponding constraint instruction according to the user portrait and the standard reply text, the constraint instruction instructing to re-express the standard reply text according to the user portrait to fit the character style of the user;

[0082] The preset instruction template includes an instruction sentence, a user portrait field, and a standard reply text field, and a person skilled in the art can flexibly modify and implement according to the disclosure. For example, the instruction template is: re-express the standard reply text according to the user portrait to fit the character style of the user, the user portrait is, and the standard reply text is.

[0083] The instruction template is applied, and the user portrait and the standard reply text are filled into the corresponding positions of the template to generate a corresponding constraint instruction. A person skilled in the art should know that the constraint instruction is an instruction for limiting or regulating the artificial intelligence system to control its behavior or performance.

[0084] Step S1400, input the constraint instruction into a preset human-like question and answer model to generate a user-oriented reply.

[0085] The human-like question and answer model is trained in advance to convergence, learns to understand the semantics of the constraint instruction, and has the ability to generate a user-oriented reply according to the corresponding instruction. The specific type can be T5, GPT, BERT, Encoder-Decoder model, Transformer model, etc. Those skilled in the art can choose one as needed. The training process of the human-like question and answer model is further disclosed in the subsequent embodiments. This step is temporarily omitted.

[0086] In an embodiment, the human-like question and answer model uses GPT3.5, takes the constraint instruction as an input sequence of the human-like question and answer model, performs word segmentation on the input sequence to obtain a corresponding segmented sequence, then inputs the segmented sequence into the encoding end of the Transformer model, encodes each word in the segmented sequence by stacking multiple layers of multi-head self-attention layers and fully connected layers. Specifically, for each word, the multi-head attention layer performs multi-head attention calculation on it, thereby performing self-attention weighting on different dimensions of the word to obtain a corresponding weighted vector representation. After passing through the fully connected layer, the encoding vector representation corresponding to the word is obtained. Accordingly, the deep semantic information of the constraint text is extracted, and the encoding representation corresponding to each word in the segmented sequence is input into the decoding section of the Transformer model. The encoding representation corresponding to each word is decoded to generate a corresponding user-oriented reply. Specifically, during the decoding process, the generation probability of each generated word is calculated according to the generated word, the current word position, and the encoding representation corresponding to the word, and the word with the highest generation probability is selected. The user-oriented reply can be obtained by concatenating the generated words in sequence.

[0087] According to the typical embodiments of the present application, it can be known that the technical solutions of the present application have many advantages, including but not limited to the following aspects:

[0088] The present application responds to the user to enable the intelligent customer service event, obtains the personal information and historical behavior data of the user, determines the corresponding user portrait, responds to the user question event, obtains the question text input by the user, determines the standard question text in the preset question and answer knowledge base matched with the question text, obtains the standard reply text of the standard question text, applies the preset instruction template, generates the corresponding constraint instruction according to the user portrait and the standard reply text, the constraint instruction indicates that the standard reply text is re-expressed according to the user portrait to match the user's character style, the constraint instruction is input into the preset personification question and answer model, and the reply to the user is generated. It can be seen that the user portrait of the user is understood, the professional answering ability of the standard reply text is maintained, the expression of the standard reply text is improved to match the character style of the user, and then the reply is made, so that the intelligent customer service is more realistic, the user experiences the feeling of communicating with a real person who understands himself, the user satisfaction is greatly improved, and the user stickiness is increased.

[0089] Please refer to Figure 2 In further embodiments, step S1200, determining the standard question text in the preset question and answer knowledge base matched with the question text, comprises the following steps:

[0090] Step S1210, determining the corresponding semantic similarity between the question text and each standard question text in the preset question and answer knowledge base by using a preset text similarity model;

[0091] The text similarity model can be a double-tower model, which includes two processing branches, each processing branch includes the same network structure, and each processing branch includes a text feature representation layer. Then, the outputs of the two processing branches are respectively connected to a linear layer for semantic similarity calculation. The text feature extraction layer can use Text Transfomer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc. Those skilled in the art can choose one of them as needed. The linear layer can use MLP (Multi-Layer Perceptron), FC (Fully Connected Layer), etc. Those skilled in the art can choose one of them as needed. Since the training process of RoBERTa+CRF, Text Transfomer, RoBERTa, XLM-RoBERTa, MPNet, and BERT is known in the art, it will not be described in detail.

[0092] The question text and each standard question text in the preset question and answer knowledge base form a text data pair, the text data pair is taken as an input of the text similarity model, a processing branch extracts deep semantic features of the question text in the text data pair to obtain a corresponding vectorized feature representation, meanwhile, another processing branch extracts deep semantic features of the standard question text in the text data pair to obtain a corresponding vectorized feature representation, then, a linear layer calculates a vector distance between the feature representations corresponding to the question text and the standard question text as a semantic similarity, the vector distance calculation can be implemented by using any available vector distance algorithm, including but not limited to any one of the following: cosine similarity algorithm, vector dot product algorithm, Manhattan distance, Euclidean distance algorithm, Pearson correlation coefficient, etc.

[0093] Step S1220, screening out the standard question text with the semantic similarity exceeding the preset threshold.

[0094] The standard question text with the semantic similarity exceeding the preset threshold in the question and answer knowledge base is screened out, the standard question text is semantically related to the question text, and the preset threshold can be set by a person skilled in the art as needed.

[0095] In the embodiment, on the one hand, the text similarity model is used to determine the standard question text semantically related to the question text in the question and answer knowledge base, which is efficient and can ensure the accuracy of the semantic similarity, so as to accurately represent the semantic similarity between two texts, on the other hand, by matching the standard question text semantically similar to the question text, the expression of the question text can be standardized, so that the corresponding standard reply text is determined according to the standard question text, laying a solid foundation for the professionalism and accuracy of subsequent replies.

[0096] Please refer to Figure 3 In further embodiments, before step S1100, obtaining personal information and historical behavior data of the user to determine a corresponding user portrait, the following steps are included:

[0097] Step S1010, obtaining a user portrait of a user inputting a historical question text, and a supervised reply text generated by replying to the historical question text;

[0098] The historical question text is input by the user to the artificial customer service in the past, and the historical reply text is input by the artificial customer service to the user in the past, the corresponding question text and reply text can be stored for calling when the artificial customer service provides customer service to the user.

[0099] Often in the strategy of selling goods, the user portrait of the corresponding consumer group is positioned for the sold goods, so that the artificial customer service of the e-commerce platform can reply to the user who asks questions according to the user portrait. Accordingly, the historical reply text with the professional, accurate, complete answer to all questions in the question is taken as the supervision reply text, which is the historical reply text input by the artificial customer service to reply to the historical question text input by the user. The specific implementation can be to filter out the historical reply text by the operator of the e-commerce platform according to the disclosure, or to filter out the historical reply text by artificial intelligence algorithm. The specific implementation is further disclosed in the subsequent embodiments, and this step is temporarily omitted.

[0100] The personal information and historical behavior data of the user are obtained, and the corresponding user portrait is determined according to steps S1110-S1120.

[0101] Step S1020, determining the standard question text in the preset question and answer knowledge base matched with the historical question text, and obtaining the standard reply text of the standard question text;

[0102] The historical question text and each standard question text in the preset question and answer knowledge base form a text data pair, and the text data pair is taken as the input of the text similarity model. The deep semantic features of the historical question text in the text data pair are extracted by one processing branch to obtain the corresponding vectorized feature representation. At the same time, the deep semantic features of the standard question text in the text data pair are extracted by another processing branch to obtain the corresponding vectorized feature representation. Then, the vector distance between the feature representations corresponding to the historical question text and the standard question text is calculated by a linear layer as the semantic similarity. Accordingly, the standard question text in the question and answer knowledge base with the semantic similarity exceeding the preset threshold is screened out, and the standard reply text of the standard question text is obtained. The preset threshold can be set by a person skilled in the art as needed.

[0103] Step S1030, applying a preset instruction template to generate a corresponding constraint instruction as a training sample according to the user portrait and the standard reply text, and taking the supervision reply text as a supervision label;

[0104] The preset instruction template includes an instruction sentence, a user portrait field, and a standard reply text field, which can be flexibly implemented by a person skilled in the art according to the disclosure. For example, the instruction template is: according to the user portrait, the standard reply text is re-expressed to fit the user's character style. The user portrait is, and the standard reply text is.

[0105] The user portrait and the standard reply text are filled into the corresponding positions of the template according to the instruction template to generate a corresponding constraint instruction as the training sample. Those skilled in the art should know that the constraint instruction is an instruction for limiting or regulating the artificial intelligence system to control its behavior or performance.

[0106] In step S1040, the training sample is input into the anthropomorphic reply model to extract deep semantic information of the training sample and generate a predicted user-oriented reply.

[0107] In an embodiment, the anthropomorphic question and answer model adopts GPT3.5, takes the constraint instruction as an input sequence of the anthropomorphic question and answer model, performs word segmentation on the input sequence to obtain a corresponding segmented sequence, then inputs the segmented sequence into an encoding end of a Transformer model, encodes each segmented word in the segmented sequence by stacking multiple layers of multi-head self-attention layers and fully connected layers. Specifically, for each segmented word, multi-head attention calculation is performed on it through the multi-head attention layer to perform self-attention weighting on different dimensions of the segmented word and obtain a corresponding weighted vector representation. After passing through the fully connected layer, the encoding vector representation corresponding to the segmented word is obtained. Accordingly, the encoding representation corresponding to each segmented word in the segmented sequence is input into a decoding section of the Transformer model to decode the encoding representation corresponding to each segmented word and generate a predicted user-oriented reply. Specifically, in the decoding process, the generation probability of each generated word is calculated according to the generated word, the current word position, and the encoding representation corresponding to the segmented word, and the word with the highest generation probability is selected. The user-oriented reply can be obtained by concatenating the generated words in sequence.

[0108] In step S1050, the loss value of the predicted user-oriented reply is determined by using the supervision label of the training sample. When the loss value does not reach a preset threshold, the weight of the anthropomorphic reply model is updated, and the iterative training of other training samples is continued until the anthropomorphic reply model converges.

[0109] The preset cross-entropy loss function is called, which can be set flexibly by those skilled in the art according to prior knowledge or experimental experience, and a cross-entropy loss value of the predicted user-oriented reply is calculated based on the supervised label of the training sample; when the cross-entropy loss value reaches a preset threshold, it indicates that the anthropomorphic reply model has been trained to a convergent state, so that the model training can be terminated; when the cross-entropy loss value does not reach the preset threshold, it indicates that the model has not converged, so the model is updated according to the cross-entropy loss value, usually by back propagation to correct the weight parameters of each link of the model to make the model further approach convergence, then the other training samples are continuously called to implement iterative training on the model until the model is trained to a convergent state, and the preset threshold can be set as needed by those skilled in the art.

[0110] In this embodiment, the anthropomorphic reply model is supervised and trained until convergence, so that the anthropomorphic reply model learns to understand the semantics of the constraint instructions and generates user-oriented replies according to the corresponding instructions, which can ensure that the standard reply text is re-expressed in the character style of the corresponding user according to the user portrait, and user-oriented replies are generated.

[0111] Please refer to Figure 4 In further embodiments, the step S1100 of determining the corresponding user portrait comprises the following steps:

[0112] The step S1110 comprises the following steps:

[0113] In one embodiment, the personal information and historical behavior data of the user are obtained from a user database according to the unique identifier of the user, such as an ID. The personal information can be the relevant information input by the user when registering on the e-commerce platform, including any one or more of gender, delivery address, education level, age / birth date, occupation, etc. The historical behavior data can be the behavior records of the user on the e-commerce platform in the past, including any one or more of likes, forwards, collections, adding to shopping cart, purchases, browsing, etc.

[0114] The age is calculated according to the birth date in the personal information, and the administrative division part of the text is determined as the place of residence according to the delivery address in the personal information, such as Guangzhou Haizhu District, Guangdong Province, China. The age, place of residence, and gender, occupation, and education level in the personal information are taken as basic attribute labels.

[0115] According to any one of the brands and product types of the commodities involved in the likes, forwards, collections, shopping cart additions, purchases, and browsing in the historical behavior data as brand preferences and product type preferences, such as Chanel, Gucci, lipstick, bags, and cosmetics, according to the purchase time of the commodities in the historical behavior data, the purchase cycle is determined, such as three times a month, according to the purchase data of the commodities in the historical behavior data, the total payment and the number of purchased commodities are determined, and the consumption level is determined according to the total payment and the number of purchased commodities. In an embodiment, a preset total payment and a number of purchased commodities corresponding to a stepwise growth threshold value, and a consumption level corresponding to each step of the mapping association are determined. When the total payment and the number of purchased commodities reach the threshold value of the corresponding step, the corresponding consumption level is determined, for example, the consumption level: low consumption level [first step: 10,000-100,000 yuan, 100-500 pieces of goods], medium consumption level [second step: 100,000-500,000 yuan, 500-5000 pieces], and high consumption level [third step: more than 500,000 yuan, more than 5000 pieces]. Those skilled in the art can flexibly set the total payment and the number of purchased commodities corresponding to the stepwise growth threshold value according to the disclosure. The brand preferences, product type preferences, consumption levels, and purchase cycles are used as preference attribute labels.

[0116] Step S1120, constructing a corresponding user portrait according to the basic attribute labels and the preference attribute labels of the user.

[0117] The basic attribute labels and the preference attribute labels of the user are associated with corresponding identifiers, i.e., age, education level, gender, occupation, location, brand preference, purchase cycle, consumption level, and product type preference, to construct a corresponding user portrait.

[0118] In this embodiment, the basic attribute labels and the preference attribute labels are determined through the personal information and the historical behavior data of the user, and the user portrait is constructed accordingly, which can depict rich user characteristics, so that the subsequent personification reply model can fully understand the user according to the user portrait.

[0119] Please refer to Figure 5 In further embodiments, before step S1010, obtaining the user portrait of the user inputting the historical question text, and the supervised reply text generated by replying to the historical question text, the following steps are included:

[0120] Step S1000, obtaining the user portrait of the user inputting the historical question text, and the historical reply text replying to the historical question text.

[0121] The historical question text is input by the user in the past to the artificial customer service, and the historical reply text is input by the artificial customer service in the past to reply to the user, so that the corresponding question text and reply text can be stored for calling when the artificial customer service provides customer service to the user.

[0122] Step S1001, determining the character style of the user corresponding to the user portrait;

[0123] A character style set collecting multiple different character styles can be constructed in advance, and a character style in the character style set matched with the user portrait of the user is matched. The matching can be realized by human or by a deep learning model suitable for a multi-classification task trained to convergence in advance. The user portrait is taken as the model input, and is mapped to the corresponding matched character style in the character style set.

[0124] The character styles in the character style set can be flexibly set by those skilled in the art according to business needs. For example, the character styles in the character style set include serious, youthful, fashionable, brandism, qualityism, pragmatism, pursuit of enjoyment, and simplicity.

[0125] Step S1002, inputting the historical reply text into a preset style classification model to determine the corresponding reply style;

[0126] The style classification model is trained to convergence in advance, and acquires the ability to determine the reply style of the input historical reply text. The structure of the model can be a text feature extraction layer followed by a classifier. The text feature extraction layer is suitable for deep semantic task of text extraction, and Text Transfomer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc. can be used. Those skilled in the art can realize one of them as needed. The classifier is used for multi-classification task, and LightGBM, XGBoost, CatBoost, AdaBoost, GBDT, Random Forest, etc. can be used. Those skilled in the art can realize one of them as needed.

[0127] In an embodiment, the historical reply text is taken as a training sample, the reply style corresponding to the historical reply text is artificially labeled as a supervised label of the training sample, the training sample is input into the style classification model, deep semantic information of the training sample is extracted by a text feature extraction layer in the style classification model to obtain a corresponding vectorized text semantic feature representation, the text semantic feature representation is input into a classifier to be mapped to a plurality of preset categories, each category corresponds to a reply style, a classification probability corresponding to each category is obtained, one or more categories with a classification probability exceeding a preset threshold are determined, a corresponding predicted reply style is output, a preset cross-entropy loss function is called, and a cross-entropy loss value of the predicted reply style is calculated based on the supervised label of the training sample; when the cross-entropy loss value reaches the preset threshold, it is indicated that the style classification model has been trained to a convergent state, so that the model training can be terminated; when the cross-entropy loss value does not reach the preset threshold, it is indicated that the model has not converged, so that the model is updated according to the cross-entropy loss value, and the weight parameters of each link of the model are usually corrected through back propagation to make the model further approach the convergent state, then, other training samples are continuously called to implement iterative training on the model until the model is trained to the convergent state.

[0128] The preset threshold and the cross-entropy loss function can be set as needed by a person skilled in the art. The reply styles respectively represented by the plurality of categories can be flexibly set as needed by a person skilled in the art, and exemplary examples include humorous, relaxed, substantial, concise, warm, sincere, frank, considerate, interesting, practical, elegant, and artistic.

[0129] In step S1003, when the reply style matches the user style to achieve a fit, the historical reply text is taken as a supervised reply text.

[0130] The reply styles respectively represented by the plurality of categories and the user styles of the characters in the character style set can be artificially matched according to whether the reply style fits the user style, a mapping relationship between the user style and the reply style that are matched can be a one-to-many mapping relationship or a one-to-one association relationship, and a mapping relationship table can be constructed by summarizing and structuring.

[0131] The mapping relationship table is searched to determine whether a mapping relationship exists between the reply style of the historical reply text and the user style corresponding to the user portrait, when the mapping relationship exists, it is determined that the reply style matches the user style, and the reply style can fit the user style, and accordingly, the historical reply text is taken as a supervised reply text.

[0132] In this embodiment, the process of determining the supervised reply text is disclosed, which can ensure that the expression of the supervised reply text can match the character style of the user. Accordingly, the expression of the supervised reply text as a supervised label can match the character style of the user, and the expression of the reply to the user generated by the personified question and answer model can match the character style of the user.

[0133] Please refer to Figure 6 In further embodiments, before the step S1000, acquiring the user portrait of the user inputting the historical question text, and the historical reply text replying to the historical question text, the following steps are included:

[0134] Step S2000, acquiring the historical question text and the corresponding historical reply text, and constructing a text data pair;

[0135] The historical question text is input by the user to the artificial customer service in the past, and the historical reply text is input by the artificial customer service to the user in the past. The corresponding question text and reply text can be stored for calling when the artificial customer service provides customer service to the user.

[0136] Step S2001, inputting the text data pair into a preset reply matching model to determine the corresponding reply quality;

[0137] The reply matching model is trained to convergence in advance, and learns the ability to determine the reply quality of the historical reply text in the input text data pair replying to the historical question text in the text data pair. The model structure can be a double-tower model, which includes two processing branches, each processing branch contains the same network structure, and each processing branch includes a text feature representation layer. Then, the outputs of the two processing branches are connected to a classifier for binary classification, and the probability of belonging to an effective reply is classified as the reply quality. The binary classification includes effective reply and ineffective reply. The text feature extraction layer can use Text Transfomer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc. Those skilled in the art can choose one of them as needed. The classifier used for binary classification task can use LightGBM, XGBoost, CatBoost, AdaBoost, GBDT, Random Forest, etc. Those skilled in the art can choose one of them as needed.

[0138] In an embodiment, the historical question text and its corresponding historical reply text are constructed into a text data pair as a training sample, and it is judged whether the historical reply text in the text data pair is effective in replying to the historical question text in the text data pair, and a supervision label of the training sample is labeled accordingly. In an embodiment, it is judged whether the historical reply text in the training sample is effective in replying to the historical question text in the training sample, and the reply is professional, accurate, and complete in answering all questions in the question. If yes, the supervision label of the training sample is labeled as 1, otherwise, the supervision label of the training sample is labeled as 0.

[0139] The training sample is input into the reply matching model. A processing branch extracts deep semantic features of the historical question text in the text data pair to obtain a corresponding vectorized feature representation. Meanwhile, another processing branch extracts deep semantic features of the historical reply text in the text data pair to obtain a corresponding vectorized feature representation. Then, the feature representations corresponding to the historical question text and the historical reply text are input into a classifier for binary classification, which is mapped to a preset two-class, including a class representing an effective reply and a class representing an ineffective reply, to obtain a classification probability of the predicted class representing an effective reply, i.e., a probability of belonging to an effective reply. A preset cross-entropy loss function is called to calculate a cross-entropy loss value of the predicted probability of belonging to an effective reply based on the supervision label of the training sample. When the cross-entropy loss value reaches a preset threshold, it indicates that the reply matching model has been trained to a convergent state, so that the model training can be terminated. When the cross-entropy loss value does not reach the preset threshold, it indicates that the model has not converged, and then the model is updated according to the cross-entropy loss value, usually by back propagation to correct the weight parameters of each link of the model to make the model further approach convergence. Then, other training samples are continuously called to iteratively train the model until the model is trained to a convergent state. The cross-entropy loss function can be set as needed by those skilled in the art.

[0140] In step S2002, the historical question text and its corresponding historical reply text with a reply quality lower than a preset threshold are removed.

[0141] The preset threshold is used to determine whether the reply quality is too low, and can be flexibly set by those skilled in the art according to the disclosure.

[0142] It can be understood that when the reply quality is lower than the preset threshold, it means that the corresponding historical reply text is not professional, accurate, and complete in replying to the corresponding historical question text, which belongs to an ineffective reply.

[0143] In this embodiment, the historical reply text is determined by using the reply matching model to determine the reply quality of the historical reply text to the corresponding historical question text, thereby filtering out the historical question text and its corresponding historical reply text with substandard reply quality. The remaining historical reply text can ensure that it replies to the corresponding historical question text, and the reply is professional, accurate and complete to a certain extent.

[0144] Please refer to Figure 7 In further embodiments, after step S1200, in response to the user question event, obtaining the question text input by the user, the following steps are included:

[0145] Step S1201, in response to a matching failure event, using a preset intent classification model to determine whether the intent of the question text is clear;

[0146] The intent classification model is pre-trained to convergence and learns the ability to determine the probability of the intent of the input question text being clear. The structure of the model can be a text feature extraction layer followed by a classifier. The text feature extraction layer is suitable for deep semantic task of text extraction, and Text Transfomer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc. can be used. A person skilled in the art can choose one of them as needed. The classifier is used for binary classification task, and LightGBM, XGBoost, CatBoost, AdaBoost, GBDT, Random Forest, etc. can be used. A person skilled in the art can choose one of them as needed.

[0147] When there is no standard question text matching the question text from the question and answer knowledge base, a corresponding matching failure event occurs. The intent classification model is used to determine the probability that the intent of the question text is clear. Further, when the probability is greater than a preset threshold, it is determined that the intent of the question text is clear, otherwise, it is determined that the intent of the question text is not clear. The preset threshold can be set by a person skilled in the art as needed.

[0148] In an embodiment, historical question texts are obtained as training samples, and it is artificially judged whether the intent of the historical question text is clear, i.e., whether the question is clearly expressed. The supervision label of the training sample is labeled accordingly. The acquisition of the historical question text can be implemented according to step S2000. In an embodiment, the intent of the historical question text of the training sample is artificially judged to be clear, and the supervision label of the training sample is labeled as 1, otherwise, the supervision label of the training sample is labeled as 0.

[0149] The training sample is input into the intent classification model, deep semantic information of the training sample is extracted by a text feature extraction layer in the intent classification model, a corresponding vectorized text semantic feature representation is obtained, the text semantic feature representation is input into a classifier, and is mapped to two preset categories, the two categories include a category representing an explicit intent and a category representing an implicit intent, a classification probability of the predicted category representing the explicit intent is obtained, a preset cross-entropy loss function is called, and a cross-entropy loss value of the classification probability of the predicted category representing the explicit intent is calculated based on a supervision label of the training sample; when the cross-entropy loss value reaches a preset threshold value, it is indicated that the intent classification model has been trained to a convergent state, so that the model training can be terminated; when the cross-entropy loss value does not reach the preset threshold value, it is indicated that the model has not converged, and then the model is updated in gradient according to the cross-entropy loss value, and the weight parameters of each link of the model are usually corrected through back propagation to make the model further approach the convergent state, then other training samples are continuously called to implement iterative training on the model until the model is trained to the convergent state. The cross-entropy loss function can be set as required by a person skilled in the art.

[0150] Step S1202, when the intent is explicit, the question text is sent to the artificial customer service interface;

[0151] When the intent is explicit, it is indicated that the question text cannot be replied by the intelligent customer service, and the question text is sent to the artificial customer service interface, and a reply is performed by the artificial customer service to ensure the reply quality.

[0152] Step S1203, when the intent is implicit, a re-input notification is constructed and pushed to the user.

[0153] When the intent is implicit, it is indicated that the question text does not clearly express the doubt, and accordingly, a re-input notification is constructed and pushed to the user, and the content of the re-input notification can be "the intent of the input question text is implicit, and the doubt cannot be clearly understood, please re-input".

[0154] In the embodiment, it is disclosed that in response to a matching failure event, whether the intent of the question text is explicit or not is determined according to the intent classification model, and a corresponding solution is provided, so that the robustness and reliability of the entire technical solution are effectively ensured.

[0155] Please refer to Figure 8The user-oriented intelligent customer service device provided for one of the purposes of the application is a functional embodiment of the user-oriented intelligent customer service method of the application. The device comprises a customer service enabling module 1100, a user question module 1200, an instruction generating module 1300, and a reply generating module 1400. The customer service enabling module 1100 is configured to respond to a user-enabled intelligent customer service event, obtain personal information and historical behavior data of the user, and determine a corresponding user portrait. The user question module 1200 is configured to respond to a user question event, obtain a question text input by the user, determine a standard question text in a preset question and answer knowledge base that matches the question text, and obtain a standard reply text of the standard question text. The instruction generating module 1300 is configured to apply a preset instruction template, generate a corresponding constraint instruction according to the user portrait and the standard reply text, and the constraint instruction indicates that the standard reply text is re-expressed in a character style that fits the user according to the user portrait. The reply generating module 1400 is configured to input the constraint instruction into a preset personification question and answer model to generate a user-oriented reply.

[0156] In a further embodiment, the user question module 1200 comprises a similarity determining submodule configured to determine a corresponding semantic similarity between the question text and each standard question text in the preset question and answer knowledge base using a preset text similarity model; and a question text screening submodule configured to screen out the standard question text whose semantic similarity exceeds a preset threshold.

[0157] In a further embodiment, before the user question module 1200, a first data acquisition submodule is configured to acquire a user portrait of a user who inputs a historical question text and a supervised reply text generated in response to the historical question text; determine a standard question text in a preset question and answer knowledge base that matches the historical question text, and obtain a standard reply text of the standard question text; a sample supervision construction submodule is configured to apply a preset instruction template, generate a corresponding constraint instruction as a training sample according to the user portrait and the standard reply text, and use the supervised reply text as a supervision label; a reply generating submodule is configured to input the training sample into a personification reply model, extract deep semantic information of the training sample, and generate a predicted user-oriented reply; and an iterative training submodule is configured to use the supervision label of the training sample to determine a loss value of the predicted user-oriented reply, update the weight of the personification reply model when the loss value does not reach a preset threshold, and continue to call other training samples for iterative training until the personification reply model converges.

[0158] In a further embodiment, the customer service enabling module 1100 comprises a label determining submodule for determining basic attribute labels and preference attribute labels according to personal information and historical behavior data of the user; and a portrait constructing submodule for constructing a corresponding user portrait according to the basic attribute labels and the preference attribute labels of the user.

[0159] In a further embodiment, before the data obtaining submodule, the customer service enabling module 1100 comprises a second data obtaining submodule for obtaining a user portrait of a user inputting a historical question text and a historical reply text replying to the historical question text; a character style determining submodule for determining a character style of the user corresponding to the user portrait; a reply style determining submodule for inputting the historical reply text into a preset style classification model to determine a corresponding reply style; and a supervision determining submodule for determining the historical reply text as a supervision reply text when the reply style matches the user style.

[0160] In a further embodiment, before the second data obtaining submodule, the customer service enabling module 1100 comprises a third data obtaining submodule for obtaining a historical question text and a corresponding historical reply text to construct a text data pair; a quality determining submodule for inputting the text data pair into a preset reply matching model to determine a corresponding reply quality; and a text removing submodule for removing the historical question text and the corresponding historical reply text with a reply quality lower than a preset threshold.

[0161] In a further embodiment, after the user question module 1200, the customer service enabling module 1100 comprises an intent determining submodule for determining whether an intent of the question text is clear by using a preset intent classification model in response to a matching failure event; an intent clear submodule for sending the question text to an artificial customer service interface when the intent is clear; and an intent unclear submodule for constructing a re-input notification to push to the user when the intent is unclear.

[0162] To solve the above technical problems, the embodiments of the present application further provide a computer device. Figure 9As shown, the internal structure diagram of the computer device is shown. The computer device includes a processor, a computer readable storage medium, a memory and a network interface connected by a system bus. Among them, the computer readable storage medium of the computer device stores an operating system, a database and computer readable instructions, the database can store control information sequence, and the computer readable instructions are executed by the processor to enable the processor to implement a user-oriented intelligent customer service method. The processor of the computer device is used to provide computing and control capability to support the operation of the entire computer device. The memory of the computer device can store computer readable instructions, and the computer readable instructions are executed by the processor to enable the processor to execute the user-oriented intelligent customer service method of the present application. The network interface of the computer device is used to connect and communicate with the terminal. Those skilled in the art can understand, Figure 9 The structure shown in the above embodiment is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0163] The processor in the embodiment is used to execute the specific functions of each module and its sub-modules in the above embodiment, and the memory stores the program codes and various data required for executing the above modules or sub-modules. The network interface is used for data transmission between the user terminal or the server. The memory in the embodiment stores the program codes and data required for executing all modules / sub-modules in the user-oriented intelligent customer service device of the present application, and the server can call the program codes and data of the server to execute the functions of all sub-modules. Figure 8 The memory in the embodiment stores the program codes and data required for executing all modules / sub-modules in the user-oriented intelligent customer service device of the present application, and the server can call the program codes and data of the server to execute the functions of all sub-modules.

[0164] The present application also provides a storage medium storing computer readable instructions, which are executed by one or more processors to enable the one or more processors to execute the steps of the user-oriented intelligent customer service method of any embodiment of the present application.

[0165] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiments of the method can be included. Among them, the storage medium can be a computer readable storage medium such as a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM).

[0166] In summary, the present application can reply according to the character style of the user to realize more real and personalized customer service.

[0167] Those skilled in the art can understand that the steps, measures, and schemes in various operations, methods, and processes discussed in the present application can be alternated, changed, combined, or deleted. Further, other steps, measures, and schemes in various operations, methods, and processes discussed in the present application can also be alternated, changed, rearranged, decomposed, combined, or deleted. Further, steps, measures, and schemes in various operations, methods, and processes in the prior art can also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0168] The above only describes some embodiments of the present application. It should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A user-oriented intelligent customer service method, characterized in that, Includes the following steps: In response to a user's activation of intelligent customer service, the system obtains the user's personal information and historical behavior data to determine the corresponding user profile. In response to a user's question event, the system obtains the question text input by the user, determines the standard question text that matches the question text in a preset question and answer knowledge base, and obtains the standard response text of the standard question text. Using a preset instruction template, corresponding constraint instructions are generated based on the user profile and the standard reply text. The constraint instructions instruct the standard reply text to be rephrased in a way that matches the user's personality style based on the user profile. The constraint instructions are input into a preset anthropomorphic question-and-answer model to generate user-oriented responses; Before obtaining users' personal information and historical behavior data to determine the corresponding user profile, the process includes: Obtain user profiles of users who input historical question texts, and historical reply texts of those users; determine the user's personality style corresponding to the user profile; input the historical reply texts into a preset style classification model to determine the corresponding reply style; when the reply style matches the user style and achieves a fit, use the historical reply texts as supervisory reply texts. Obtain user profiles of users who input historical question texts, as well as supervised response texts generated from responding to the historical question texts; Determine the standard question text that matches the historical question text in the preset question-and-answer knowledge base, and obtain the standard response text of the standard question text; Using a preset instruction template, corresponding constraint instructions are generated based on the user profile and the standard response text as training samples, and the supervised response text is used as a supervision label; The training samples are input into the anthropomorphic response model to extract the deep semantic information of the training samples and generate predicted user-oriented responses. Using the supervision labels of the training samples, the loss value of the predicted user-oriented response is determined. When the loss value does not reach a preset threshold, the weights of the anthropomorphic response model are updated, and other training samples are called to carry out iterative training until the anthropomorphic response model converges.

2. The user-oriented intelligent customer service method according to claim 1, characterized in that, Determining the standard question text that matches the question text in the preset question-and-answer knowledge base includes the following steps: A preset text similarity model is used to determine the semantic similarity between the question text and each standard question text in the preset question-answering knowledge base; The standard question texts with semantic similarity exceeding a preset threshold are selected.

3. The user-oriented intelligent customer service method according to claim 1, characterized in that, Determining the appropriate user profile includes the following steps: Based on users' personal information and historical behavior data, basic attribute tags and preference attribute tags are determined; Based on the user's basic attribute tags and preference attribute tags, a corresponding user profile is constructed.

4. The user-oriented intelligent customer service method according to claim 1, characterized in that, Before obtaining the user profile of the user who entered the historical question text, and the historical reply text of the reply to the historical question text, the following steps are included: Obtain the historical question texts and their corresponding historical reply texts, and construct text data pairs; The text data is input into a preset response matching model to determine the corresponding response quality; Remove historical question texts and their corresponding historical reply texts whose reply quality is below a preset threshold.

5. The user-oriented intelligent customer service method according to claim 1, characterized in that, After responding to a user's question event and obtaining the text of the question entered by the user, the process includes the following steps: In response to a matching failure event, a preset intent classification model is used to determine whether the intent of the question text is clear; When the intent is clear, the question text will be sent to the human customer service interface; When the intent is unclear, a re-entry notification is generated and pushed to the user.

6. A user-oriented intelligent customer service device, characterized in that, include: The customer service activation module is used to respond to user activation of intelligent customer service events, obtain user personal information and historical behavior data, and determine the corresponding user profile; The user question module is used to respond to user question events, obtain the question text input by the user, determine the standard question text that matches the question text in the preset question and answer knowledge base, and obtain the standard reply text of the standard question text. The instruction generation module is used to apply a preset instruction template to generate corresponding constraint instructions based on the user profile and the standard reply text. The constraint instructions instruct the standard reply text to be rephrased in a way that matches the user's personality style based on the user profile. The response generation module is used to input the constraint instructions into a preset anthropomorphic question-and-answer model to generate user-facing responses; Before obtaining users' personal information and historical behavior data to determine the corresponding user profile, the process includes: Obtain user profiles of users who input historical question texts, and historical reply texts of those users; determine the user's personality style corresponding to the user profile; input the historical reply texts into a preset style classification model to determine the corresponding reply style; when the reply style matches the user style and achieves a fit, use the historical reply texts as supervisory reply texts. Obtain user profiles of users who input historical question texts, as well as supervised response texts generated from responding to the historical question texts; Determine the standard question text that matches the historical question text in the preset question-and-answer knowledge base, and obtain the standard response text of the standard question text; Using a preset instruction template, corresponding constraint instructions are generated based on the user profile and the standard response text as training samples, and the supervised response text is used as a supervision label; The training samples are input into the anthropomorphic response model to extract the deep semantic information of the training samples and generate predicted user-oriented responses. Using the supervision labels of the training samples, the loss value of the predicted user-oriented response is determined. When the loss value does not reach a preset threshold, the weights of the anthropomorphic response model are updated, and other training samples are called to carry out iterative training until the anthropomorphic response model converges.

7. The user-oriented intelligent customer service device according to claim 6, characterized in that, The user question module includes: a similarity determination submodule, used to determine the semantic similarity between the question text and each standard question text in the preset question-and-answer knowledge base using a preset text similarity model; and a question text filtering submodule, used to filter out standard question texts whose semantic similarity exceeds a preset threshold.

8. The user-oriented intelligent customer service device according to claim 6, characterized in that, The customer service activation module includes: a tag determination submodule, used to determine basic attribute tags and preference attribute tags based on the user's personal information and historical behavior data; and a profile construction submodule, used to construct a corresponding user profile based on the user's basic attribute tags and preference attribute tags.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 5, which, when invoked by a computer, executes the steps included in the corresponding method.

Images