Intelligent customer service automatic answering method and device, equipment, medium and product

Abstract

The application discloses an intelligent customer service automatic answering method and device, equipment, medium and product, and the method comprises the following steps: obtaining a question text submitted by an e-commerce customer service system chat interface; obtaining a preset question most similar in semantic composition to the question text from a knowledge base of the e-commerce customer service system as a target question, and determining an answer set corresponding to a question set to which the target question belongs as a target answer set; standard questions in the question set and preset answers in the answer set are all labeled with process labels in a label pool, and each process label in the label pool respectively represents each business link in an e-commerce order process; a classification model is used to determine a classification probability of the question text mapped to each process label, and a preset answer with a classification probability of the process label being a relative maximum value in the target answer set is determined as a target answer; and the target answer is pushed to the chat interface for display. The application can improve the accuracy of the e-commerce customer service system response.

Description

Technical Field

[0001] This application relates to the field of intelligent customer service technology, and in particular to an intelligent customer service automatic response method and its corresponding device, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] Currently, in e-commerce scenarios, due to the large number of customer inquiries, merchants typically deploy intelligent customer service robots to assist customer service representatives in answering questions. Most intelligent customer service systems on the market establish a knowledge base, which includes three parts: standard questions, similar questions, and answers. Generally, the response answer for a specific scenario is unique. To address the uniformity of responses, improvements typically involve configuring multiple answers for standard questions. When the corresponding standard question is triggered, an answer is randomly selected and returned to the customer. This ensures that different customers receive different answers, making the intelligent customer service more human-like.

[0003] In reality, customers are segmented according to business scenarios. However, due to the randomness of the answers obtained by the above methods, the responses to different customer segments may not be ideal. Therefore, in order to improve the accuracy of the matching between answers and questions, the applicant is trying to explore new ideas. Summary of the Invention

[0004] The primary objective of this application is to solve at least one of the above-mentioned problems by providing an intelligent customer service automatic response method and corresponding apparatus, computer equipment, computer-readable storage medium, and computer program product.

[0005] To achieve the various objectives of this application, the following technical solution is adopted:

[0006] A smart customer service automatic response method provided for one of the purposes of this application includes the following steps:

[0007] Retrieve the text of questions submitted through the chat interface of the e-commerce customer service system;

[0008] The system extracts the target question from the full set of preset questions in the knowledge base of the e-commerce customer service system. The target question is then selected as the preset question that is semantically most similar to the question text. The answer set that maps to the question set to which the target question belongs is determined as the target answer set. The knowledge base includes mapping relationship data between question sets and answer sets. Each question set includes multiple preset questions, including one standard question and multiple similar questions. Each answer set includes multiple preset answers. The standard question and preset answers are labeled with process tags in a tag pool. Each process tag in the tag pool represents a different business step in the e-commerce order process.

[0009] A classification model trained to convergence is used to determine the classification probability corresponding to each process tag in the tag pool that maps the question text to the tag pool. From the target answer set, a preset answer whose classification probability of the labeled process tag is the relative maximum value within the target answer set is determined as the target answer.

[0010] The target answer is pushed to the chat interface of the e-commerce customer service system for display.

[0011] In a more detailed embodiment, the process involves obtaining the most semantically similar preset question to the question text from the full set of preset questions in the knowledge base of the e-commerce customer service system, and determining the answer set mapped to the question set to which the target question belongs as the target answer set. This includes the following steps:

[0012] A text feature extraction model that has been trained to convergence is used to extract sentence vectors from the question text;

[0013] Calculate the similarity data between the sentence vector of the question text and the sentence vectors of all preset questions in the knowledge base, which are pre-extracted by the text feature extraction model;

[0014] Select the preset questions whose similarity data is higher than a preset threshold and whose similarity data is the maximum value, and identify them as target questions;

[0015] The target answer set is obtained from the knowledge base by mapping the answer set corresponding to the standard questions in the question set to which the target question belongs.

[0016] In a more detailed embodiment, a classification model trained to convergence is used to determine the classification probability corresponding to each process tag in the tag pool that maps the question text to the target answer pool. A preset answer is then determined from the target answer set whose classification probability for the labeled process tag is the maximum relative value within the target answer set, comprising the following steps:

[0017] The sentence vector of the question text is extracted using a text feature extraction model in a classification model;

[0018] The classifier preset in the classification model is used to perform classification mapping based on the sentence vector of the question text, so as to obtain the classification probability corresponding to each process tag in the tag pool to which the question text is mapped;

[0019] Based on one or more process tags carried by the preset answers in the target answer set, determine one or more classification probabilities corresponding to each preset answer;

[0020] Compare the classification probabilities of each preset answer in the target answer set, determine the preset answer with the highest classification probability, and take the preset answer with the highest classification probability as the target answer.

[0021] In a further embodiment, the training process of the classification model is pre-implemented, including the following steps:

[0022] Select a preset question from the question set in the knowledge base as a training sample, and input it into the text feature extraction model of the classification model to extract sentence vectors;

[0023] The sentence vector is classified and mapped by a classifier to obtain the classification probability of the sentence vector to each process label in the label pool, and the target process label corresponding to the highest classification probability is determined.

[0024] Using the process labels carried by the standard questions in the selected problem set as supervision labels, calculate the loss value of the target process label. If the loss value reaches the preset threshold and the convergence state is reached, terminate the training; otherwise, implement gradient update and use the next training sample to iteratively train the classification model.

[0025] In a preferred embodiment, after completing the training process of the classification model, the following steps are included:

[0026] The text feature extraction model in the classification model that has been trained to convergence is used to extract sentence vectors for each preset question in the knowledge base and store them in the knowledge base in association with the preset question.

[0027] In a specific embodiment, the e-commerce order process includes multiple different stages, each stage including one or more business links, and a process tag is set for each business link. The set of process tags corresponding to all business links constitutes the tag pool. The different stages include the collection stage, shopping cart stage, payment stage, delivery stage, and after-sales stage.

[0028] An intelligent customer service automatic response device provided for one of the purposes of this application includes: a question response module, a question matching module, an answer matching module, and a question answering module. The question response module is used to acquire the question text submitted by the chat interface of an e-commerce customer service system. The question matching module is used to acquire, from the full set of preset questions in the knowledge base of the e-commerce customer service system, a preset question that is semantically most similar to the question text as a target question, and to determine the answer set mapped to the question set to which the target question belongs as a target answer set. The knowledge base includes mapping relationship data between question sets and answer sets, and each question set includes multiple preset questions, including a standard... The system includes a question and several similar questions; each answer set includes multiple preset answers. The standard questions and preset answers are labeled with process tags in a tag pool. Each process tag in the tag pool represents a business step in the e-commerce order process. The answer-hitting module is used to determine the classification probability corresponding to each process tag in the tag pool by using a classification model that has been trained to convergence. The module then selects the preset answer from the target answer set whose classification probability of the labeled process tag is the relative maximum value within the target answer set as the target answer. The question-and-response module is used to push the target answer to the chat interface of the e-commerce customer service system for display.

[0029] In a further embodiment, the question naming module includes: a vector extraction submodule, used to extract sentence vectors of the question text using a text feature extraction model trained to convergence; a similarity calculation submodule, used to calculate similarity data between the sentence vectors of the question text and the sentence vectors of all preset questions in the knowledge base, which are pre-extracted by the text feature extraction model; a target filtering submodule, used to filter out preset questions whose similarity data is higher than a preset threshold and whose similarity data is the maximum value, and determine them as target questions; and an answer selection submodule, used to obtain the answer set corresponding to the standard questions of the question set to which the target question belongs from the knowledge base as the target answer set.

[0030] In a more detailed embodiment, the answer-hitting module includes: a vector extraction submodule, used to extract sentence vectors of the question text using a text feature extraction model in the classification model; a classification mapping submodule, used to perform classification mapping based on the sentence vectors of the question text using a preset classifier in the classification model, to obtain the classification probability corresponding to each process tag in the tag pool to which the question text is mapped; and a correspondence conversion submodule, used to determine one or more classification probabilities corresponding to each preset answer based on one or more process tags carried by preset answers in the target answer set.

[0031] The comparison and selection submodule is used to compare the classification probabilities of each preset answer in the target answer set, determine the preset answer with the highest classification probability, and select the preset answer with the highest classification probability as the target answer.

[0032] In a further embodiment, the intelligent customer service automatic response device of this application further includes a structure for executing the training process of the classification model. This structure includes: a sample selection submodule, used to select a preset question from the question set in the knowledge base as a training sample, and input it into the text feature extraction model of the classification model to extract sentence vectors; a mapping prediction submodule, used to classify and map the sentence vectors through a classifier to obtain the classification probability corresponding to each process label in the label pool, and determine the target process label corresponding to the highest classification probability; and an iterative decision submodule, used to calculate the loss value of the target process label using the process label carried by the standard question in the selected question set as the supervision label. If the loss value reaches a preset threshold and convergence is achieved, the training is terminated; otherwise, gradient update is implemented, and the classification model is iteratively trained using the next training sample.

[0033] In a preferred embodiment, the intelligent customer service automatic response device of this application further includes: a vector preprocessing submodule, used to extract sentence vectors for each preset question in the knowledge base using a text feature extraction model in the classification model that has been trained to a convergent state, and to associate and store the sentence vectors with the preset question in the knowledge base.

[0034] In a specific embodiment, the e-commerce order process includes multiple different stages, each stage including one or more business links, and a process tag is set for each business link. The set of process tags corresponding to all business links constitutes the tag pool. The different stages include the collection stage, shopping cart stage, payment stage, delivery stage, and after-sales stage.

[0035] A computer device provided for one of the purposes of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the intelligent customer service automatic response method described in this application.

[0036] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the described intelligent customer service automatic response method, which, when invoked by a computer, performs the steps included in the method.

[0037] A computer program product provided for another purpose of this application includes a computer program / instructions that, when executed by a processor, implement the steps of the method described in any embodiment of this application.

[0038] Compared with existing technologies, the advantages of this application are as follows:

[0039] First, to improve the accuracy of e-commerce customer service system responses, this application pre-labels standard questions and preset answers in the question set of the knowledge base with process tags corresponding to the business links in the e-commerce order process, according to the e-commerce order process. Subsequently, when a user submits a question text, the preset question with the closest semantics is matched as the target question, and the target answer set corresponding to the standard questions associated with the target question is determined. Furthermore, a classification model is used to obtain the classification probability corresponding to each process tag in the tag pool mapped to the question text. Based on the classification probability, the preset answer with the process tag with the highest classification probability among all preset answers in the target answer set is determined as the target answer, and then the target answer is pushed to the user for display. Since the preset answers are already labeled with process tags corresponding to the business steps in the e-commerce order process, and the question text is mapped to the classification probability corresponding to each process tag through a classification model to serve as a sorting reference, the optimal selection of multiple preset answers in the target answer set yields a target answer that not only corresponds to the question text in meaning, but also, because the target answer has been pre-labeled with process tags in the e-commerce order process, and the process tag of the target answer is the one with the highest classification probability among all the process tags of the preset answers in the target answer set, a business step association can be established between the target answer and the question text based on the e-commerce order process. This makes the target answer more closely match the question text, thereby improving the user experience of the e-commerce customer service system.

[0040] Secondly, this application focuses on the mapping relationship between the question text and process tags at the semantic level, and then uses the classification probability of each process tag in the tag pool corresponding to the question text as the sorting basis for the preset answers in the target answer set. Therefore, the target answers it determines are necessarily more accurate and refined. Accuracy means that the corresponding meaning expression between the question text and the target answer is realized, and refinement means that this corresponding meaning expression is specified to the depth of different business links in the e-commerce order process, thereby comprehensively improving the intelligence level of the e-commerce customer service system.

[0041] Furthermore, the implementation of the technical solution in this application can eliminate a large amount of manpower in large customer service scenarios such as e-commerce platforms, thereby saving corresponding implementation costs and achieving economies of scale. Attached Figure Description

[0042] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0043] Figure 1This is a flowchart illustrating a typical embodiment of the intelligent customer service automatic response method of this application;

[0044] Figure 2 This is a flowchart illustrating the process of matching the target answer set based on the question text in an embodiment of this application;

[0045] Figure 3 This is a flowchart illustrating the process of determining the classification probability of each process label based on the question text and further determining the target answer from the target answer set in this embodiment of the application.

[0046] Figure 4 This is a flowchart illustrating the training process of the classification model in this application.

[0047] Figure 5 This is a schematic block diagram of the intelligent customer service automatic response device of this application;

[0048] Figure 6 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0049] 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.

[0050] 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.

[0051] 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.

[0052] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDA (Personal Digital Assistant) that may include a radio frequency receiver, pager, internet / intranet access, web browser, notepad, calendar, and / or GPS (Global Positioning System) receiver; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0053] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, execute the instructions in the program, and interact with the input and output devices to complete specific functions.

[0054] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0055] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0056] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0057] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0058] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0059] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0060] The intelligent customer service automatic response method of this application can be programmed into a computer program product and deployed on a client or server to run. For example, in the e-commerce platform application scenario of this application, it is generally deployed on a server. In this way, the method can be executed by human-computer interaction with the process of the computer program product through a graphical user interface by accessing the interface opened after the computer program product is running.

[0061] One exemplary application scenario of this application is an application in an e-commerce platform based on independent websites. Each independent website is a merchant instance of the e-commerce platform, with its own independent access domain name, and its actual owner is responsible for publishing and updating products.

[0062] Each merchant instance on an independent website can be configured with an e-commerce customer service system provided by the e-commerce platform to introduce an intelligent customer service robot. This system provides consultation services to relevant consumers. Consumers enter the corresponding chat interface of the merchant instance and input their questions as the question text. The e-commerce platform's customer service system receives this question text and performs semantic matching against a question set pre-configured in the knowledge base for that independent website. It identifies a preset question that is semantically most similar to the question text, determines the standard question within that question set, and then calls a pre-stored answer set mapped to that standard question. Using the relevant technical means of this application, it determines one of the preset answers as the target answer and outputs it to the chat interface to respond to the consumer's question and meet their consultation needs.

[0063] During the process of a consumer user chatting with the intelligent customer service robot, the user is usually allowed to introduce a human customer service representative. When a human customer service representative is connected, the e-commerce customer service system will establish a dialogue channel between the user asking the question and the human customer service representative of the independent website. The two parties will continue to have a human dialogue. Thus, the user asking the question enters the question text, and the human customer service representative replies with the answer text, generating chat data in turn.

[0064] Chat logs generated from the chat interface, including the text of questions asked by users and the text of answers from human customer service representatives or automated responses from chatbots, can be archived and stored in a database carrying the speaker's characteristic information. This data can then be used for data mining to expand the knowledge base.

[0065] Please see Figure 1 The intelligent customer service automatic response method of this application, in its typical embodiment, includes the following steps:

[0066] Step S1100: Obtain the question text submitted by the e-commerce customer service system chat interface:

[0067] When a user asks a question to the intelligent customer service robot in the chat interface after entering the e-commerce customer service system, they enter the corresponding text (the question text) in the input box of the chat interface and then confirm and submit. In response to the user's submission event, the client device where the chat interface is located will submit the question text to the backend server where the e-commerce customer service system is deployed. The server, based on the business logic implemented in the program, performs general formatting processing on the question text as needed, including removing extra spaces, removing emoji elements, etc., to obtain the final question text.

[0068] Step S1200: Obtain the preset question that is semantically most similar to the question text from the full set of preset questions in the knowledge base of the e-commerce customer service system as the target question, and determine the answer set mapped to the question set to which the target question belongs as the target answer set; wherein, the knowledge base includes mapping relationship data between question sets and answer sets, each question set includes multiple preset questions, including one standard question and multiple similar questions; each answer set includes multiple preset answers, and the standard question and preset answers are labeled with process tags in the tag pool, and each process tag in the tag pool represents a corresponding business link in the e-commerce order process:

[0069] The e-commerce customer service system of this application is configured with a knowledge base, which is a collection of information on the correspondence between questions and answers. The knowledge base includes multiple question sets and multiple answer sets. Each question set includes multiple preset questions. Within the same question set, there is one preset question as the standard question, and the remaining preset questions are similar questions to the standard question. Similar questions within the same question set are generally consistent or similar in meaning to their standard question, differing only slightly in expression and tone. Each standard question in each question set is mapped to an answer set, thus establishing a corresponding mapping between the question set and the answer set. Each answer set contains multiple preset answers, which are used to answer the preset questions in the question set mapped to that answer set.

[0070] Building upon this knowledge base structure, this application further introduces process tags to annotate standard questions and preset answers. To this end, a tag pool is prepared. This tag pool is essentially a pre-planned tag system containing multiple process tags corresponding to different business stages of the e-commerce order process. Accordingly, each process tag represents a business stage of the e-commerce order process. The e-commerce order process can be divided into multiple different stages according to the order processing procedure. For example, it includes the "favorites" stage (representing the e-commerce product being added to the user's favorites), the "shopping cart" stage (representing the e-commerce product being added to the user's shopping cart), the payment stage (representing the e-commerce product being ordered and paid for), the shipping stage (representing the e-commerce product entering the logistics stage), and the after-sales stage (representing the e-commerce product being delivered to the user). Such stages can be included as many times as needed by the e-commerce customer service system. Furthermore, for each stage, multiple specific business stages are subdivided. For example, the payment stage can be further subdivided into paid business stages and unpaid business stages; similarly, the after-sales stage can be further subdivided into refund business stages and repair business stages. In this way, various corresponding process labels can be flexibly customized according to different stages, so that the process labels correspond to the business links of the e-commerce order process.

[0071] Of course, different business links in the e-commerce order process can also correspond to different stages, that is, each stage can be regarded as a business link. Therefore, the depth at which process tags correspond to the e-commerce order process can be flexibly set by those skilled in the art as needed. The more specific the corresponding business link, the more refined the node in the e-commerce order process corresponding to the process tag. Therefore, by setting the level of refinement of the business links corresponding to the process tags, the accuracy of the standard questions or preset answers labeled by the process tags can be enhanced. Thus, the role of process tags in this application is not simply to classify the preset questions and answers they label, but to map the specific subdivided business links of the e-commerce order process. This plays a crucial role in guiding users to provide correct answers to their questions based on different business links.

[0072] Each standard question in each of the aforementioned question sets can be labeled with a corresponding process tag based on the actual relevance of its textual content. Similarly, each preset answer in each of the aforementioned answer sets can also be labeled with a corresponding process tag based on the actual relevance of its textual content. When labeling process tags, a standard question can be labeled with multiple process tags, and similarly, a preset answer can also be labeled with multiple process tags. Therefore, preset answers can be invoked based on process tags.

[0073] Based on the knowledge base provided in this application, the question text submitted by the user can be further semantically matched with all the preset questions in the knowledge base to determine the preset question that best matches the question text semantically, that is, the preset question that is most similar in semantics.

[0074] When performing semantic matching, a convolutional neural network (CNN) model can be used to first extract the deep semantic information of the question text. Then, based on the data distance between the deep semantic information of the question text and the deep semantic information of each preset question in the knowledge base, the preset question that is closest in data distance to the question text is determined. The determined preset question is the target question that best matches the question text. Specific embodiments of this matching process will be further provided in subsequent embodiments of this application, and will not be discussed here. The CNN model can be a CNN or RNN-based CNN model, including but not limited to TextCNN, LSTM, Transformer, BERT, Albert, Electra, etc., which are suitable basic models for learning text representations to obtain corresponding sentence vectors. Preferably, the CNN model can also be built as a Siamese network to improve the computational efficiency of similarity matching. Those skilled in the art can select a suitable basic model from the prior art to extract sentence vectors representing the deep semantic information of the question text based on the principles disclosed in this application.

[0075] Once the target question is identified, the question set containing it is also determined, allowing us to identify the standard question within that set. Based on the mapping relationship between the answer set and the question set, the answer set corresponding to the standard question can be determined, and this answer set can serve as the target answer set. As mentioned earlier, this target answer set contains multiple preset answers, each of which is labeled with one or more process tags from the aforementioned tag pool.

[0076] Step S1300: Using a classification model trained to convergence, determine the classification probability corresponding to each process label in the label pool that maps the question text to the target answer set. Then, determine the preset answer from the target answer set whose classification probability for the labeled process label is the maximum relative value within the target answer set as the target answer.

[0077] To implement this application, a classification model is prepared. After the classification model is trained to a convergent state, it is used in this method, mainly to determine the classification probability of the query text to each process label in the label pool.

[0078] The classification model mainly includes a text feature extraction model and a classifier. The text feature extraction model can be the same model used in step S1200 to extract sentence vectors for the question text and the preset question, so that after the classification model is trained to convergence, the text feature extraction model can be used to extract sentence vectors for the question text and the preset question. Alternatively, it can be a different, independently configured model. Those skilled in the art can flexibly implement it based on the disclosure herein. The classifier is used to perform classification mapping based on the sentence vectors obtained by the text feature extraction model, so as to calculate the classification probabilities of the sentence vector mapped to multiple classification labels in the classification space. Therefore, the classifier can be a multi-classifier constructed using the softmax function, whose obtained classification probabilities are normalized to the (0, 1) numerical space, and the sum of the classification probabilities of all classification labels is 1.

[0079] The classification model is pre-trained to convergence before being put into use. A subsequent embodiment will exemplarily reveal the steps involved in this training process in detail, which will not be discussed here. After training, the classification model learns the ability to classify and map a given question text to various classification labels in a preset classification space, that is, to various process labels in the label pool, thereby obtaining the classification probability corresponding to each process label.

[0080] Therefore, each process tag in the tag pool has a corresponding classification probability for the question text. Obviously, this classification probability itself serves as a sorting index. Moreover, the preset answers in the target answer set are associated with process tags, and the corresponding classification probability can be determined through these process tags. Therefore, this classification probability can be used to compare the degree of matching of all preset answers in the target answer set.

[0081] Specifically, the mapping relationship data between each process tag and its classification probability determined from the question text can be regarded as a table. Then, the corresponding classification probability is determined by querying this table based on the process tags carried by the preset answers in the target answer set. If a preset answer contains multiple process tags, the classification probabilities corresponding to each process tag can be determined separately. The process tag with the highest classification probability in this preset answer is taken as the target classification probability for comparison with other preset answers. Thus, each preset answer in the target answer set obtains a target classification probability and a target classification probability. The preset answers in the target answer set can then be sorted and compared based on this target classification probability to determine the preset answer with the highest target classification probability. This target answer is the preset answer in the target answer set whose classification probability of the labeled process tag is the relative maximum value within the target answer set. Theoretically, this target answer is the appropriate answer that corresponds to the question contained in the question text in meaning and corresponds to the business process associated with the question text in terms of business process.

[0082] Step S1400: Push the target answer to the e-commerce customer service system chat interface for display.

[0083] After obtaining the target answer, the server can push it to the user's terminal device that submitted the question text. After receiving the target answer, the user's terminal device displays it in the chat interface of its e-commerce customer service system, thus providing the user with a more accurate target answer.

[0084] Based on the disclosure of the above typical embodiments and their modified embodiments, it can be understood that the technical solution of this application has many positive effects, including but not limited to the following aspects:

[0085] First, to improve the accuracy of e-commerce customer service system responses, this application pre-labels standard questions and preset answers in the question set of the knowledge base with process tags corresponding to the business links in the e-commerce order process, according to the e-commerce order process. Subsequently, when a user submits a question text, the preset question with the closest semantics is matched as the target question, and the target answer set corresponding to the standard questions associated with the target question is determined. Furthermore, a classification model is used to obtain the classification probability corresponding to each process tag in the tag pool mapped to the question text. Based on the classification probability, the preset answer with the process tag with the highest classification probability among all preset answers in the target answer set is determined as the target answer, and then the target answer is pushed to the user for display. Since the preset answers are already labeled with process tags corresponding to the business steps in the e-commerce order process, and the question text is mapped to the classification probability corresponding to each process tag through a classification model to serve as a sorting reference, the optimal selection of multiple preset answers in the target answer set yields a target answer that not only corresponds to the question text in meaning, but also, because the target answer has been pre-labeled with process tags in the e-commerce order process, and the process tag of the target answer is the one with the highest classification probability among all the process tags of the preset answers in the target answer set, a business step association can be established between the target answer and the question text based on the e-commerce order process. This makes the target answer more closely match the question text, thereby improving the user experience of the e-commerce customer service system.

[0086] Secondly, this application focuses on the mapping relationship between the question text and process tags at the semantic level, and then uses the classification probability of each process tag in the tag pool corresponding to the question text as the sorting basis for the preset answers in the target answer set. Therefore, the target answers it determines are necessarily more accurate and refined. Accuracy means that the corresponding meaning expression between the question text and the target answer is realized, and refinement means that this corresponding meaning expression is specified to the depth of different business links in the e-commerce order process, thereby comprehensively improving the intelligence level of the e-commerce customer service system.

[0087] Furthermore, the implementation of the technical solution in this application can eliminate a large amount of manpower in large customer service scenarios such as e-commerce platforms, thereby saving corresponding implementation costs and achieving economies of scale.

[0088] Please see Figure 2 In a more detailed embodiment, step S1200, obtaining the preset question that is semantically most similar to the question text from the full set of preset questions in the knowledge base of the e-commerce customer service system as the target question, and determining the answer set mapped to the question set to which the target question belongs as the target answer set, includes the following steps:

[0089] Step S1210: Using a text feature extraction model that has been trained to convergence, extract the sentence vector of the question text:

[0090] A text feature extraction model trained to convergence is used to extract sentence vectors from the question text. This text feature extraction model can be a convolutional neural network model based on CNN or RNN, including but not limited to TextCNN, LSTM, Transformer, BERT, Albert, Electra, etc., which is a basic model suitable for learning representations of text to obtain corresponding sentence vectors. Those skilled in the art can train it themselves using various training methods to enable it to learn representations of text information and obtain sentence vectors representing the deep semantic information of the text.

[0091] Step S1220: Calculate the similarity data between the sentence vector of the question text and the sentence vectors of all preset questions in the knowledge base, which are pre-extracted by the text feature extraction model:

[0092] In order to achieve similarity matching between the question text and all the preset questions in the knowledge base, the text feature extraction model that has been trained to convergence is used in advance to extract the sentence vector of each preset question in the knowledge base, and these sentence vectors are associated with their corresponding preset questions and stored in the knowledge base for easy retrieval in this step.

[0093] Furthermore, a data distance algorithm, including any one of the cosine similarity algorithm, Euclidean distance algorithm, Pearson correlation coefficient algorithm, or Jaccard algorithm, is used to calculate the similarity data between the sentence vector of the question text and the sentence vector of each preset question in the knowledge base. This similarity data is represented by a higher value indicating greater similarity and a lower value indicating less similarity. This yields a similarity data sequence, which can be sorted as needed for operational convenience. Alternatively, as an alternative, a Siamese network can be used to simultaneously match the question text with each preset question in real time to determine the similarity data between them.

[0094] Step S1230: Select preset problems whose similarity data is higher than a preset threshold and whose similarity data is the maximum value, and determine them as target problems:

[0095] A preset threshold, which can be an empirical or experimental threshold, is used to determine whether the determined similarity data meets the most basic similarity requirements. Based on this, the similarity data corresponding to each preset question in the knowledge base can be sorted to determine the maximum similarity data and its corresponding preset question. Then, the similarity data of the preset question is compared with the preset threshold. If the similarity data of the preset question is higher than the preset threshold, the preset question is determined to be the target question; otherwise, it can be determined that there is no matching preset question in the knowledge base, and an entry point for redirecting to human customer service is pushed to the user so that the user can communicate with a human customer service representative through this entry point.

[0096] Step S1240: Obtain the answer set corresponding to the standard questions of the question set to which the target question belongs from the knowledge base as the target answer set:

[0097] Once the target problem is identified, the set of problems to which it belongs can be determined, that is, the set of problems containing the target problem. After determining this set of problems, the standard problems within it can be identified. After determining the standard problems, the set of answers that corresponds to them can be determined, and this set of answers is the target answer set.

[0098] This embodiment uses a specific similarity matching method to determine the corresponding target answer set for the question text, which facilitates the subsequent pre-selection of preset answers as target answers from the target answer set, thereby achieving the screening of matching preset questions and narrowing the data range for subsequent determination of target answers.

[0099] Please see Figure 3 In a more detailed embodiment, step S1300, which involves using a classification model trained to convergence to determine the classification probability corresponding to each process tag in the tag pool to which the question text is mapped, and determining the preset answer from the target answer set whose classification probability of the labeled process tag is the relative maximum value within the target answer set, includes the following steps:

[0100] Step S1310: Extract the sentence vector of the question text using the text feature extraction model in the classification model:

[0101] In this embodiment, the classification model includes a text feature extraction model and a classifier, which are pre-trained to convergence before use. The text feature extraction model is adapted to extract sentence vectors from the question text, and then the classifier performs classification mapping on these sentence vectors. Accordingly, the question text is input into the classification model, which extracts its sentence vectors. Similarly, the text feature extraction model can be a convolutional neural network model based on CNN or RNN, including but not limited to TextCNN, LSTM, Transformer, BERT, Albert, Electra, etc., which are basic models suitable for learning text representations to obtain corresponding sentence vectors.

[0102] Step S1320: Using a pre-set classifier in the classification model, classify and map the question text according to its sentence vector to obtain the classification probability corresponding to each process label in the label pool.

[0103] The sentence vector of the question text is fed into a classifier via a fully connected layer. The classifier calculates the classification probability corresponding to each process label in the label pool based on the sentence vector of the question text. As mentioned earlier, this classifier can be a multi-classifier. As previously stated, the set of classification probabilities obtained by the classifier can be regarded as a lookup table, which contains the mapping relationship between process labels and their classification probabilities.

[0104] Step S1330: Based on one or more process tags carried by the preset answers in the target answer set, determine one or more classification probabilities corresponding to each preset answer:

[0105] The pre-determined target answer set includes multiple preset answers, each carrying one or more process tags. These process tags can be classified by querying the aforementioned set of classification probabilities. For each preset answer within the target answer set, regardless of the number of process tags it has, only the process tag with the highest classification probability is retained, and this probability is used as the target classification probability.

[0106] Step S1340: Compare the classification probabilities corresponding to each preset answer in the target answer set, determine the preset answer with the highest classification probability, and take the preset answer with the highest classification probability as the target answer:

[0107] The classification probabilities of each preset answer in the target answer set are compared. With the target classification probability of each preset answer already selected, the target classification probabilities of each preset answer are compared. Based on this, the preset answers in the target answer set are sorted from largest to smallest according to their target classification probabilities. The preset answer with the highest ranking can be determined as the target answer.

[0108] In this embodiment, the classification probability obtained by classifying the sentence vectors of the question text is used as the sorting criterion to sort the preset answers in the target answer set. Finally, the preset answer corresponding to the process label with the highest classification probability is determined. Accordingly, the determined preset answer is semantically matched with a preset question in the question set that maps to the answer set. Therefore, it has a semantic correspondence with the question text. Furthermore, since it is the preset answer corresponding to the process label with the highest classification probability obtained by classifying the sentence vectors of the question text, it has a business link correspondence with the question text. It can be seen that the target answer determined for the question text is more accurate and refined. Taking the question corresponding to e-commerce orders as an example, the answer text can be provided more closely to the business links implied or explicitly stated in the question text, which can significantly improve the user experience of the e-commerce customer service system.

[0109] Please see Figure 4 In a further embodiment, to enable the classification model to learn the ability to extract sentence vectors from the question text and perform classification mapping based on those sentence vectors, the training process of the classification model is pre-implemented, including the following steps:

[0110] Step S2100: Select a preset question from the question set in the knowledge base as a training sample, and input it into the text feature extraction model of the classification model to extract sentence vectors:

[0111] As mentioned earlier, the question set in the knowledge base establishes a mapping relationship with the process labels in the label pool through the pre-labeled standard questions. Therefore, the preset questions in the question set also have a mapping relationship with the process labels of the standard questions. Accordingly, a preset question and its corresponding process label labeled with a standard question can be used to construct the training data for training the classification model. Based on this principle, multiple pairs of such training data can be organized to form a training dataset for training the classification model.

[0112] When training the classification model, a preset question from the training data is selected as a training sample and input into the text feature extraction model of the classification model to extract deep semantic information and obtain its corresponding sentence vector.

[0113] Step S2200: Classify and map the sentence vector using a classifier to obtain the classification probability of the sentence vector to each process label in the label pool, and determine the target process label corresponding to the highest classification probability:

[0114] The sentence vector of the training sample is further mapped to the classification space corresponding to the classifier through a fully connected layer. This classification space is the classification space defined by the label pool. Therefore, the classification space contains classification labels that correspond one-to-one with all process labels in the label pool. Thus, the classifier can calculate the classification probability corresponding to each process label based on the sentence vector, and then determine the process label corresponding to the highest classification probability as the target process label.

[0115] Step S2300: Using the process labels carried by the standard questions in the selected question set as supervision labels, calculate the loss value of the target process label. If the loss value reaches a preset threshold and convergence is achieved, terminate training; otherwise, implement gradient update and use the next training sample to iteratively train the classification model.

[0116] Since the training data containing the training sample already includes a process label for the preset question, this process label can be directly used as the supervision label for the classification model. Under the cross-entropy loss function, the loss value of the target process label relative to the supervision label is calculated. Then, it is determined whether this loss value reaches the preset threshold required for training the classification model. If the preset threshold is reached, it indicates that the classification model has been trained to a convergent state, and training can be terminated, allowing the model to be used normally. Otherwise, if the preset threshold has not been reached, it indicates that the classification model has not yet been trained to a convergent state. Therefore, gradient updates can be performed on the classification model based on the loss value, and backpropagation can be used to correct the weight parameters of each stage of the classification model, so that the loss function of the classification model further approaches convergence. Then, the training samples from the next training data are called from step S2100 for iterative training, continuously training the classification model until it finally reaches a convergent state.

[0117] The text feature extraction model in the classification model trained in this embodiment can be used in the technical solution of this application to extract sentence vectors for the preset questions of the knowledge base and the question text submitted by the user. Therefore, by training the classification model, the text feature extraction model required for step S1200 and its specific steps and step S1300 can be obtained. Compared with training two different models separately, it can save implementation costs.

[0118] This embodiment trains the classification model using data from the knowledge base, avoiding the need to construct a separate training dataset, thus saving training costs. Ultimately, it enables the classification model to more accurately obtain the sentence vector of the question text and the corresponding classification probability set, guiding the effective sorting of preset answers within the target answer set to determine the target answer that matches the question text.

[0119] In a preferred embodiment, after completing the training process of the classification model, the following steps are included:

[0120] Step S2400: Using the text feature extraction model in the classification model that has been trained to convergence, extract sentence vectors for each preset question in the knowledge base and store them in the knowledge base in association with the preset question.

[0121] After the classification model is trained to convergence, the text feature extraction model within it can be used to extract sentence vectors for each preset question in the knowledge base. The extracted sentence vectors are then associated with their preset questions and stored in the knowledge base. This allows for direct referencing of the sentence vectors from the preset questions, improving the data access speed of the e-commerce customer service system and thus increasing the response speed to query texts. Therefore, there is no need to train a new text feature extraction model separately for extracting sentence vectors from the preset questions, which naturally saves on the deployment costs of the technical solution presented in this application.

[0122] Please see Figure 5 This invention provides an intelligent customer service automatic response device to meet one of the purposes of this application. It is a functional embodiment of the intelligent customer service automatic response method of this application. The device includes: a question response module 1100, a question matching module 1200, an answer matching module 1300, and a question response module 1400. The question response module 1100 is used to acquire the question text submitted by the chat interface of the e-commerce customer service system. The question matching module 1200 is used to acquire the preset question that is semantically most similar to the question text from the full set of preset questions in the knowledge base of the e-commerce customer service system as the target question, and determine the answer set mapped to the question set to which the target question belongs as the target answer set. The knowledge base includes mapping relationship data between question sets and answer sets. Each question set includes multiple preset questions, including one standard question and multiple similar questions; each answer set includes multiple preset answers, where the standard question and preset answers are labeled with process tags in a tag pool, and each process tag in the tag pool represents a business step in the e-commerce order process; the answer-hitting module 1300 is used to determine the classification probability corresponding to each process tag in the tag pool by using a classification model trained to convergence, and to determine the preset answer from the target answer set whose classification probability of the labeled process tag is the relative maximum value within the target answer set as the target answer; the question-response module 1400 is used to push the target answer to the e-commerce customer service system chat interface for display.

[0123] In a further embodiment, the question naming module includes: a vector extraction submodule, used to extract sentence vectors of the question text using a text feature extraction model trained to convergence; a similarity calculation submodule, used to calculate similarity data between the sentence vectors of the question text and the sentence vectors of all preset questions in the knowledge base, which are pre-extracted by the text feature extraction model; a target filtering submodule, used to filter out preset questions whose similarity data is higher than a preset threshold and whose similarity data is the maximum value, and determine them as target questions; and an answer selection submodule, used to obtain the answer set corresponding to the standard questions of the question set to which the target question belongs from the knowledge base as the target answer set.

[0124] In a more detailed embodiment, the answer-hitting module 1300 includes: a vector extraction submodule, used to extract sentence vectors of the question text using a text feature extraction model in a classification model; a classification mapping submodule, used to perform classification mapping based on the sentence vectors of the question text using a preset classifier in the classification model, to obtain the classification probability corresponding to each process tag in the tag pool to which the question text is mapped; and a correspondence conversion submodule, used to determine one or more classification probabilities corresponding to each preset answer based on one or more process tags carried by preset answers in the target answer set.

[0125] The comparison and selection submodule is used to compare the classification probabilities of each preset answer in the target answer set, determine the preset answer with the highest classification probability, and select the preset answer with the highest classification probability as the target answer.

[0126] In a further embodiment, the intelligent customer service automatic response device of this application further includes a structure for executing the training process of the classification model. This structure includes: a sample selection submodule, used to select a preset question from the question set in the knowledge base as a training sample, and input it into the text feature extraction model of the classification model to extract sentence vectors; a mapping prediction submodule, used to classify and map the sentence vectors through a classifier to obtain the classification probability corresponding to each process label in the label pool, and determine the target process label corresponding to the highest classification probability; and an iterative decision submodule, used to calculate the loss value of the target process label using the process label carried by the standard question in the selected question set as the supervision label. If the loss value reaches a preset threshold and convergence is achieved, the training is terminated; otherwise, gradient update is implemented, and the classification model is iteratively trained using the next training sample.

[0127] In a preferred embodiment, the intelligent customer service automatic response device of this application further includes: a vector preprocessing submodule, used to extract sentence vectors for each preset question in the knowledge base using a text feature extraction model in the classification model that has been trained to a convergent state, and to associate and store the sentence vectors with the preset question in the knowledge base.

[0128] In a specific embodiment, the e-commerce order process includes multiple different stages, each stage including one or more business links, and a process tag is set for each business link. The set of process tags corresponding to all business links constitutes the tag pool. The different stages include the collection stage, shopping cart stage, payment stage, delivery stage, and after-sales stage.

[0129] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 6 The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store control information sequences. When the computer-readable instructions are executed by the processor, the processor can implement an intelligent customer service automatic response method. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the intelligent customer service automatic response method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0130] In this embodiment, the processor is used to execute... Figure 5 The system defines the specific functions of each module and its sub-modules. The memory stores the program code and various data required to execute these modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the intelligent customer service automatic response device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0131] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent customer service automatic response method of any embodiment of this application.

[0132] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0133] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0134] In summary, this application can comprehensively improve the intelligence level of e-commerce customer service systems, determine answers to user questions that are more closely matched to the specific business links in the e-commerce order process, and improve the accuracy of e-commerce customer service system responses.

[0135] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0136] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An intelligent customer service automatic answering method, characterized by, The method comprises the following steps: Obtaining a question text submitted by an e-commerce customer service system chat interface, wherein the question text is a corresponding text input by a question user from an input box of the chat interface; Obtaining a preset question with the most similar semantics as the question text from all preset questions in a knowledge base of the e-commerce customer service system as a target question, and determining an answer set corresponding to a question set to which the target question belongs as a target answer set; wherein the knowledge base comprises mapping relationship data between the question set and the answer set, each question set comprises a plurality of preset questions including a standard question and a plurality of similar questions, and each answer set comprises a plurality of preset answers, wherein the standard question and the preset answers are labeled with process labels in a label pool, and each process label in the label pool represents a corresponding business link in an e-commerce order process; the e-commerce order process comprises a plurality of stages, each stage comprises one or more business links, and a process label corresponding to each business link is set; and a set of process labels corresponding to all business links constitutes the label pool; Using a trained classification model to determine a classification probability corresponding to each process label in the label pool to which the question text is mapped, and determining a preset answer with a classification probability of the labeled process label as a target answer in a range of the target answer set, comprising: using a text feature extraction model in the classification model to extract a sentence vector of the question text; using a preset classifier in the classification model to classify and map the sentence vector of the question text to obtain a classification probability corresponding to each process label in the label pool to which the question text is mapped; determining one or more classification probabilities corresponding to each preset answer in the target answer set according to one or more process labels carried by the preset answer; comparing the classification probabilities corresponding to each preset answer in the target answer set to determine a maximum classification probability, and determining a preset answer with the maximum classification probability as a target answer; Pushing the target answer to the e-commerce customer service system chat interface for display.

2. The intelligent customer service automatic answering method of claim 1, wherein, Obtaining a preset question with the most similar semantics as the question text from all preset questions in a knowledge base of the e-commerce customer service system as a target question, and determining an answer set corresponding to a question set to which the target question belongs as a target answer set, comprising the following steps: Using a trained text feature extraction model to extract a sentence vector of the question text; Calculating similarity data between the sentence vector of the question text and a sentence vector of all preset questions in the knowledge base pre-extracted by the text feature extraction model; Selecting a preset question with a similarity data higher than a preset threshold and a maximum similarity data as a target question; Obtaining an answer set corresponding to a standard question of a question set to which the target question belongs from the knowledge base as a target answer set. 3.The intelligent customer service automatic answering method of claim 1, wherein, The training process of the classification model is pre-implemented, comprising the following steps: Selecting a preset question from the question set in the knowledge base as a training sample, and inputting the training sample into a text feature extraction model of the classification model to extract a sentence vector; The sentence vector is classified and mapped by the classifier to obtain classification probabilities of the sentence vector mapped to each process label in the label pool, and a target process label corresponding to the maximum classification probability is determined; The loss value of the target process label is calculated by taking the process label carried by the standard question in the selected question set as the supervised label, and if the loss value reaches a preset threshold and reaches a convergence state, the training is terminated; Otherwise, gradient updating is performed, and the next training sample is used to implement iterative training of the classification model.

4. The intelligent customer service automatic answering method according to claim 3, characterized in that, After completing the training process of the classification model, the following steps are included: The text feature extraction model in the trained classification model is used to extract sentence vectors for each preset question in the knowledge base and store them in association with the preset question in the knowledge base.

5. The intelligent customer service automatic answering method according to any one of claims 1 to 4, characterized in that, The e-commerce order process includes multiple different stages, each stage including one or more business links, and a process label is set for each business link. The set of process labels corresponding to all business links constitutes the label pool. The different stages include a collection stage, a shopping cart stage, a payment stage, a delivery stage, and an after-sales stage.

6. An intelligent customer service automatic answering apparatus, characterized by comprising: It includes: A question response module is configured to obtain a question text submitted by an e-commerce customer service system chat interface, wherein the question text is a corresponding text input by a question user from an input box of the chat interface; A question hitting module is configured to obtain a preset question that is most similar in semantics to the question text from the full amount of preset questions in the knowledge base of the e-commerce customer service system as a target question, and determine the answer set corresponding to the question set to which the target question belongs as a target answer set. The knowledge base includes mapping relationship data between question sets and answer sets. Each question set includes multiple preset questions, including a standard question and multiple similar questions. Each answer set includes multiple preset answers. The standard question and the preset answers are labeled with process labels in the label pool, and each process label in the label pool corresponds to a business link in the e-commerce order process. The e-commerce order process includes multiple different stages, each stage including one or more business links, and a process label is set for each business link. The set of process labels corresponding to all business links constitutes the label pool. The answer hitting module is configured to determine a classification probability corresponding to each process label in the label pool to which the question text is mapped by using a classification model trained to a convergent state, and determine a preset answer in which the classification probability of the labeled process label is a relatively maximum value in the target answer set as a target answer from the target answer set, including: extracting a sentence vector of the question text by using a text feature extraction model in the classification model; obtaining a classification probability corresponding to each process label in the label pool to which the question text is mapped by using a preset classifier in the classification model according to the sentence vector of the question text; determining one or more classification probabilities corresponding to each preset answer according to one or more process labels carried by the preset answer in the target answer set; comparing the classification probabilities corresponding to each preset answer in the target answer set to determine a maximum classification probability, and determining the preset answer with the maximum classification probability as the target answer; The question and answer module is configured to push the target answer to a chat interface of the e-commerce customer service system for display.

7. The intelligent customer service automatic answering device according to claim 6, characterized in that, The question hitting module includes: The vector extraction submodule is configured to extract a sentence vector of the question text by using a text feature extraction model trained to a convergent state; The similarity calculation submodule is configured to calculate similarity data between the sentence vector of the question text and a sentence vector of all preset questions in the knowledge base pre-extracted by the text feature extraction model; The target screening submodule is configured to screen a preset question in which the similarity data is higher than a preset threshold and the similarity data is a maximum value, and determine the preset question as a target question; The answer selection submodule is configured to obtain an answer set corresponding to a standard question of a question set to which the target question belongs from the knowledge base as a target answer set.

8. A computer device comprising a central processing unit and a memory, characterized in that The central processing unit is configured to call and run a computer program stored in the memory to perform the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer program is stored in the form of computer readable instructions, and the computer program is called and run by a computer to perform the steps included in the corresponding method.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method according to any one of claims 1 to 5.

Images