E-commerce customer service question and answer method and device, equipment and medium

Abstract

The application relates to an e-commerce customer service question and answer method and device, equipment and medium in the technical field of e-commerce information processing. The method comprises the following steps: in response to a customer service enabling event, obtaining a question text; performing semantic matching on the question text and the text content of each preset instruction document with different structures to determine the matching score corresponding to each text content; determining the correlation score between each preset instruction document and the question file according to the matching score; screening at least one preset instruction document with a correlation score meeting a preset condition as a candidate document; and determining a reply text according to the at least one candidate document by using a large language model. The application can accurately reply to a question.

Description

Technical Field

[0001] This application relates to the field of e-commerce information processing technology, and in particular to an e-commerce customer service question-and-answer method and its corresponding apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] With the rapid development of e-commerce, the number of merchants and buyers on e-commerce platforms has increased dramatically, leading to a surge in user inquiries. To address these numerous questions, intelligent customer service systems have been developed. These systems can identify user questions and provide appropriate responses.

[0003] In traditional technologies, to achieve intelligent customer service, a search engine is typically used to extract all or part of the user's question text as keywords. This search engine then retrieves the relevant documentation from multiple available resources to answer the user's question. However, this approach has several drawbacks. First, the documentation must contain the keywords to be found. Documents that are semantically related to the question but lack the keywords cannot be retrieved, leading to inaccurate or incomplete responses. Second, it is easily affected by spelling errors, synonyms, polysemous words, and grammatical variations, which can also result in inaccurate or incomplete responses.

[0004] Given the shortcomings of traditional technologies, the applicant has long been engaged in research in related fields and has therefore explored alternative approaches to solve industry problems in the e-commerce sector. Summary of the Invention

[0005] The primary objective of this application is to address at least one of the aforementioned problems by providing an e-commerce customer service question-and-answer method and corresponding apparatus, computer equipment, and computer-readable storage medium.

[0006] To achieve the various objectives of this application, the following technical solution is adopted:

[0007] An e-commerce customer service question-and-answer method provided for one of the purposes of this application includes the following steps:

[0008] Respond to customer service activation event and obtain the query text;

[0009] The question text is semantically matched with multiple text contents of different structures in each preset explanatory document to determine the matching score corresponding to each text content;

[0010] Based on the matching score, a relevant score is determined between each preset instruction document and the question file;

[0011] At least one of the preset explanatory documents that meet the preset conditions for the relevant scores is selected as candidate documents, and the response text is determined based on the at least one candidate document using a large language model.

[0012] In a further embodiment, semantic matching is performed between the question text and multiple text contents with different structures in each preset instruction document to determine the matching score corresponding to each text content, including the following steps:

[0013] A preset text vectorization model is used to determine the question semantic vector based on the semantics of the question text, and the text semantic vector corresponding to each text content is determined based on the semantics of multiple text contents with different structures in each preset explanatory document.

[0014] The similarity between the question semantic vector and each of the text semantic vectors is determined and used as the matching score corresponding to each text content.

[0015] In a further embodiment, the relevant score between each preset instruction document and the question file is determined based on the matching score, including:

[0016] The matching scores corresponding to multiple text contents of each preset instruction document are weighted and summed to obtain the relevant score for each preset instruction document.

[0017] In a further embodiment, the following steps are included before responding to the customer service activation event:

[0018] The document title of the preset instruction document is used as the text content of the first structure;

[0019] A pre-defined summary generation model is used to generate a document summary based on the original text of the pre-defined explanatory document, which serves as the text content of the second structure.

[0020] The original text of the preset instruction document is segmented to obtain each document fragment as the text content of the third structure.

[0021] In a further embodiment, the original text of the preset instruction document is segmented to obtain various document fragments, including the following steps:

[0022] A sliding window of a preset window size is used to slide and extract words from the original text of the preset instruction document with a preset step size, so as to obtain the document segment to be verified corresponding to each word extraction;

[0023] The relevance between each document segment to be verified is determined based on the semantics of each document segment to be verified;

[0024] When the relevance meets the preset conditions, the document fragment to be verified is confirmed to pass the verification and is recognized as a document fragment.

[0025] In response to a verification failure event, adjust the current sliding window size and / or step size, and then iterate the above process until the document fragment to be verified obtained by sliding the original text with the sliding window at the step size passes the verification.

[0026] In a further embodiment, the original text of the preset instruction document is segmented to obtain various document fragments, including the following steps:

[0027] The original text of the preset instruction document is segmented at the sentence level to obtain the corresponding sentence sequence;

[0028] Based on the semantics of each pair of adjacent clauses in the sentence sequence, a text semantic model is used to determine whether there is a sentence relationship between each pair of adjacent clauses. Each pair of adjacent clauses that do not have a sentence relationship is then segmented to determine the corresponding document fragments.

[0029] In a further embodiment, after determining the response text based on the at least one candidate document using a large language model, the following steps are included:

[0030] The compliance and authenticity of the reply text are verified using a preset reply verification model, and the verification pass rate of the reply text is determined.

[0031] When the verification pass rate reaches the preset standard, the reply text will be pushed to the user.

[0032] On the other hand, an e-commerce customer service Q&A device provided to meet one of the purposes of this application includes an event response module, a score determination module, a rating determination module, and a response determination module. The event response module is used to respond to a customer service activation event and obtain the question text. The score determination module is used to semantically match the question text with multiple text contents of different structures in each preset instruction document to determine the matching score corresponding to each text content. The rating determination module is used to determine the relevant rating between each preset instruction document and the question document based on the matching score. The response determination module is used to filter at least one preset instruction document whose relevant rating meets preset conditions as a candidate document, and use a large language model to determine the response text based on the at least one candidate document.

[0033] In a further embodiment, the score determination module includes: a text vectorization submodule, used to determine a question semantic vector based on the semantics of the question text using a preset text vectorization model, and to determine the text semantic vector corresponding to each text content based on the semantics of multiple text contents with different structures in each preset explanatory document; and a similarity determination submodule, used to determine the similarity between the question semantic vector and each text semantic vector, as a matching score corresponding to each text content.

[0034] In a further embodiment, the scoring determination module includes: a scoring calculation submodule, used to perform a weighted summation of the matching scores corresponding to multiple text contents of each preset instruction document to obtain the relevant score corresponding to each preset instruction document.

[0035] In a further embodiment, before the event response module, there are: a first structure submodule, used to obtain the document title of a preset explanatory document as the text content of the first structure; a second structure submodule, used to generate a document summary based on the original text of the preset explanatory document using a preset summary generation model as the text content of the second structure; and a third structure submodule, used to segment the original text of the preset explanatory document to obtain each document fragment as the text content of the third structure.

[0036] In a further embodiment, the third structural submodule includes: a sliding word extraction submodule, used to extract words from the original text of the preset explanatory document using a sliding window of a preset window size with a preset step size, to obtain the document fragment to be verified corresponding to each word extraction; a relevance determination submodule, used to determine the relevance between each document fragment to be verified based on the semantics of each document fragment to be verified; a fragment verification submodule, used to confirm that the document fragment to be verified passes the verification when the relevance meets a preset condition, and is thus recognized as a document fragment; and a verification failure submodule, used to respond to a verification failure event, adjust the window size and / or step size of the current sliding window, and iterate the above process until the document fragment to be verified obtained by the sliding window extracting words from the original text with the specified step size passes the verification.

[0037] In a further embodiment, the third structural submodule includes: a sentence segmentation module, used to perform sentence-level segmentation on the original text of the preset explanatory document to obtain a corresponding sentence sequence; and a segmentation module, used to sequentially determine whether there is a sentence relationship between each two adjacent sentences based on the semantics corresponding to each two adjacent sentences in the sentence sequence, and to segment each two adjacent sentences that do not have a sentence relationship into segments to determine the corresponding document segments.

[0038] In a further embodiment, after the response determination module, there is a post-verification submodule, used to verify the compliance and authenticity of the response text using a preset response verification model, and determine the verification pass rate of the response text; and a user push submodule, used to push the response text to the user when the verification pass rate reaches a preset standard.

[0039] On another front, a computer device provided for one of the purposes of this application includes a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the e-commerce customer service question-and-answer method described in this application.

[0040] In another aspect, a computer-readable storage medium is provided to suit another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the e-commerce customer service question-and-answer method, which, when invoked by a computer, performs the steps included in the method.

[0041] The technical solution of this application has many advantages, including but not limited to the following aspects:

[0042] This application semantically matches the user's query text provided by customer service with multiple text contents of different structures in each preset instruction document, determining a matching score for each text content. Based on the matching scores of the multiple text contents of the preset instruction documents, a correlation score is determined between each preset instruction document and the query text. At least one preset instruction document whose correlation score meets preset conditions is selected as a candidate document. A large language model is then used to determine the response text based on the at least one candidate document. Therefore, the correlation score derived from the matching scores between multiple text contents of the preset instruction documents and the query text considers the relevance between each text content representing the preset instruction document and the query text, ensuring the accuracy and reliability of the correlation score. This makes the candidate documents selected based on the correlation score highly relevant to the query text, and thus, the response text determined based on the candidate documents accurately replies to the user. Attached Figure Description

[0043] 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:

[0044] Figure 1 This is a flowchart illustrating a typical embodiment of the e-commerce customer service question-and-answer method of this application;

[0045] Figure 2 This is a flowchart illustrating the process of determining the matching score of each text content based on the vectorized representation of the semantics corresponding to multiple text contents with different structures in each preset specification document.

[0046] Figure 3 This is a flowchart illustrating the process of determining the text content of each structure in an embodiment of this application;

[0047] Figure 4 This is a flowchart illustrating the process of determining a verified document fragment in one embodiment of this application.

[0048] Figure 5 This is a schematic diagram of the process for determining a document fragment in another embodiment of this application;

[0049] Figure 6 This is a schematic diagram illustrating the process of verifying the response text and pushing the verified response text to the user in an embodiment of this application;

[0050] Figure 7 This is a schematic diagram of the e-commerce customer service Q&A device of this application;

[0051] Figure 8 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0052] 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.

[0053] 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.

[0054] 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.

[0055] 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.

[0056] 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.

[0057] It should be noted that the concept of "server" used in this application can also be extended to apply to server clusters. Based on network deployment principles as understood by those skilled in the art, the servers should be logically divided; physically, these servers can be independent yet 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 described in this application.

[0058] 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.

[0059] 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.

[0060] 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.

[0061] 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.

[0062] 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.

[0063] The e-commerce customer service question-and-answer 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 exemplary application scenario of this application, it can be deployed on the server of an e-commerce platform, thereby allowing 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 runs.

[0064] Please see Figure 1 The e-commerce customer service Q&A method of this application, in its typical embodiment, includes the following steps:

[0065] Step S1100: Respond to the customer service activation event and obtain the query text;

[0066] Users of e-commerce platforms can activate customer service through various entry points, such as the "Contact Customer Service" control in the bottom navigation bar, the "Contact Customer Service" control in the menu bar, the "Contact Customer Service" control in the floating window or bubble, or operations on the help center page. This allows them to provide a question to customer service, triggering the generation of a customer service activation request, which carries the user's identifier and the question text. The question text can be edited and entered by the user based on their question, or it can be specified by the user from a preset question text. The preset question text can be pre-set by those skilled in the art as needed, for example, pre-set questions that users frequently ask to human customer service. The user identifier is used to uniquely identify a single user and whether the user belongs to a merchant or a buyer. Those skilled in the art can set it as needed, for example, a user ID.

[0067] The e-commerce platform's server receives the customer service activation request, responds to the customer service activation event, and obtains the question text and user identifier carried in the customer service activation request.

[0068] Step S1200: Semantically match the question text with multiple text contents of different structures in each preset explanatory document to determine the matching score corresponding to each text content;

[0069] For merchants who open online stores on e-commerce platforms, the pre-set instruction documents are typically pre-configured by the e-commerce platform's operations personnel and / or product managers. The content of these different pre-set instruction documents includes introductions to the e-commerce platform, instructions for opening an online store on the platform, instructions for configuring and / or managing the online store, introductions to marketing and promotion, instructions for marketing and promotion, introductions to live streaming, instructions for live streaming, customer management, etc. Thus, customer service is provided to merchants based on these pre-set instruction documents.

[0070] For buyers who purchase or wish to purchase goods on e-commerce platforms, the pre-set instruction documents are typically pre-set by merchants on the e-commerce platform for buyers, or provided with a template by the platform's operations staff and / or product managers, which merchants can then edit and modify according to their own needs. The content of these pre-set instruction documents includes product purchase instructions, refund instructions, logistics and delivery instructions, and product-related operational instructions. Thus, customer service is provided to buyers using these pre-set instruction documents.

[0071] The preset instruction document comprises several text contents with different structures, namely, a document title, a document summary, and various document fragments. The document title is a concise description of the core theme or key points of the corresponding preset instruction document, typically less than 30 characters. The document summary is a more detailed description of the core theme or key points of the corresponding preset instruction document, typically less than 100 characters. Each document fragment is a portion of the text content within the corresponding preset instruction document; all document fragments belonging to the same preset instruction document constitute the original text of the corresponding preset instruction document. Those skilled in the art can flexibly adapt the disclosed information to obtain the document title, document summary, and various document fragments of the preset instruction document.

[0072] The user identifier is identified. When the user identifier belongs to a merchant user, it indicates that the user who activated customer service is a merchant user. Therefore, various preset instruction documents pre-set for merchant users are obtained. The question text is semantically matched with the document title, document summary, and various document fragments of each preset instruction document. Specifically, taking a single preset instruction document as an example, a preset text vectorization model is used to determine the question semantic vector that vectorizes the semantics of the prompt text, and to determine the text semantic vectors that vectorize the semantics of the document title, document summary, and various document fragments of the preset instruction document. Furthermore, the vector distance between the question semantic vector and each text semantic vector is determined. The vector distance represents the semantic similarity between the question text and the document title, document summary, and various document fragments. Thus, the vector distance between the question text and the document title, document summary, and various document fragments of each preset instruction document can be determined. The vector distance of the document title is used as the matching score of the document title, the vector distance of the document summary is used as the matching score of the document summary, and the maximum value of the vector distances corresponding to each document fragment is used as the matching score of the document fragment.

[0073] When the user identifier belongs to a buyer user, it indicates that the user who activated customer service is a buyer user. Therefore, various preset instruction documents pre-set for the buyer user are obtained. The question text is semantically matched with the document title, document summary, and various document fragments of each preset instruction document. Specifically, taking a single preset instruction document as an example, a preset text vectorization model is used to determine the question semantic vector representing the semantics of the prompt text, and to determine the text semantic vectors corresponding to the semantics of the document title, document summary, and various document fragments of the preset instruction document. Furthermore, the vector distance between the question semantic vector and each text semantic vector is determined, and the vector distance represents the semantic similarity between the question text and the document title, document summary, and various document fragments. Thus, the vector distance between the question text and the document title, document summary, and various document fragments of each preset instruction document can be determined. The vector distance of the document title is used as the matching score of the document title, the vector distance of the document summary is used as the matching score of the document summary, and the maximum value of the vector distances corresponding to each document fragment is used as the matching score of the document fragment.

[0074] The text vectorization model is pre-trained to convergence, learning to extract deep semantic information from the input text and outputting vectorized feature vectors representing the semantics of the text. Specific models can include Text Transformer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc., and those skilled in the art can choose one to implement as needed. Since the training process of these models is known in the art, it will not be described in detail.

[0075] The vector distance can be determined using any available vector distance algorithm, including but not limited to: cosine similarity algorithm, vector dot product algorithm, Manhattan distance, Euclidean distance algorithm, Pearson correlation coefficient, etc.

[0076] Step S1300: Determine the relevant score between each preset instruction document and the question file based on the matching score;

[0077] Furthermore, to reasonably synthesize the matching scores corresponding to the text content of each structure of the preset instruction document, thereby deriving a correlation score between the preset instruction document and the question text, weights are preset for the text content of each structure. The matching scores corresponding to the text content of each structure of the preset instruction document are multiplied by their respective weights and then summed to obtain the correlation score between the preset instruction document and the question text. An example formula is provided below:

[0078] Score=a*doc_score+b*doc_title_score+c*doc_summary_score

[0079] Wherein: Score is the relevant score, doc_score is the matching score of the document fragment, doc_title_score is the matching score of the document title, doc_summary_score is the matching score of the document summary, and a+b+c=1. a, b, and c can be preset as needed by those skilled in the art based on the disclosure herein.

[0080] Step S1400: Select at least one of the preset description documents that meet the preset conditions for the relevant scores as candidate documents, and use a large language model to determine the response text based on the at least one candidate document.

[0081] The large language model described is suitable for text processing in the field of NLP. It is pre-trained to convergence using an extremely large corpus, acquires the ability to generate human language, and possesses accurate text semantic understanding and logical reasoning capabilities. The selected large language models include OPT, Chinchilla, PaLM, LLaMA, Alpaca, Vicuna, GPT3, GPT3.5, GPT4, and ChatGPT.

[0082] In one embodiment, a first instruction statement is preset to instruct the large language model to determine the candidate document that best matches the question text from a given plurality of candidate documents. Those skilled in the art can edit the first instruction statement as disclosed herein. As an example, the first instruction statement is: "Determine the candidate document that best matches the given question text from the following plurality of candidate documents, question text: xxxx".

[0083] Based on the relevant scores, from highest to lowest, the scores corresponding to each preset instruction document are sorted, and the top N preset instruction documents are selected as candidate documents. Alternatively, multiple preset instruction documents whose relevant scores exceed a preset threshold are selected as candidate documents. These candidate documents are then concatenated with the first instruction statement to form the input text of the large language model. This input text is then input into the large language model to determine the candidate document that best matches the question text as the response text. The N and preset threshold can be set by those skilled in the art according to business needs and the input limit of the large language model, ensuring that the number of characters in the input text is lower than the input limit.

[0084] In another embodiment, a second instruction statement is preset to instruct the large language model to generate a response text to the question text based on a plurality of given candidate documents. Those skilled in the art can edit the second instruction statement as disclosed herein. As an example, the second instruction statement is: "Generate a response text to the given question text based on the following plurality of given candidate documents, question text: xxxx".

[0085] Based on the relevant scores in descending order, the scores corresponding to each preset instruction document are sorted, and the top N preset instruction documents are selected as candidate documents. Alternatively, multiple preset instruction documents whose relevant scores exceed a preset threshold are selected as candidate documents. The multiple candidate documents are concatenated with the first instruction statement to form the input text of the large language model. The input text is then input into the large language model to generate a response text that replies to the prompt text. The N and preset threshold can be set by those skilled in the art according to business needs and the input limit of the large language model, ensuring that the number of characters in the input text is lower than the input limit.

[0086] As can be seen from the typical embodiments of this application, the technical solution of this application has many advantages, including but not limited to the following aspects:

[0087] This application semantically matches the user's query text provided by customer service with multiple text contents of different structures in each preset instruction document, determining a matching score for each text content. Based on the matching scores of the multiple text contents of the preset instruction documents, a correlation score is determined between each preset instruction document and the query text. At least one preset instruction document whose correlation score meets preset conditions is selected as a candidate document. A large language model is then used to determine the response text based on the at least one candidate document. Therefore, the correlation score derived from the matching scores between multiple text contents of the preset instruction documents and the query text considers the relevance between each text content representing the preset instruction document and the query text, ensuring the accuracy and reliability of the correlation score. This makes the candidate documents selected based on the correlation score highly relevant to the query text, and thus, the response text determined based on the candidate documents accurately replies to the user.

[0088] Please see Figure 2 In a further embodiment, step S1200, semantically matching the question text with multiple text contents of different structures in each preset explanatory document to determine the matching score corresponding to each text content, includes the following steps:

[0089] Step S1210: Using a preset text vectorization model, determine the question semantic vector based on the semantics of the question text, and determine the text semantic vector corresponding to each text content based on the semantics of multiple text contents with different structures in each preset explanatory document.

[0090] The preset instruction document comprises several text contents with different structures, namely, a document title, a document summary, and various document fragments. The document title is a concise description of the core theme or key points of the corresponding preset instruction document, typically less than 30 characters. The document summary is a more detailed description of the core theme or key points of the corresponding preset instruction document, typically less than 100 characters. Each document fragment is a portion of the text content within the corresponding preset instruction document; all document fragments belonging to the same preset instruction document constitute the original text of the corresponding preset instruction document. Those skilled in the art can flexibly adapt the disclosed information to obtain the document title, document summary, and various document fragments of the preset instruction document.

[0091] The text vectorization model is pre-trained to convergence, learning to extract deep semantic information from the input text and outputting vectorized feature vectors representing the semantics of the text. Specific models can include Text Transformer, RoBERTa, XLM-RoBERTa, MPNet, BERT, etc., and those skilled in the art can choose one to implement as needed. Since the training process of these models is known in the art, it will not be described in detail.

[0092] The deep semantic information of the question text is extracted using a text vectorization model. Based on the deep semantic information, the corresponding question semantic vector is determined. The deep semantic information of the document title, document summary, and content of each preset explanatory document is also extracted. Based on the deep semantic information, the text semantic vectors corresponding to the document title, document summary, and content of each document are determined.

[0093] Step S1220: Determine the similarity between the question semantic vector and each of the text semantic vectors, and use it as the matching score corresponding to each text content.

[0094] For each of the preset explanatory documents, the similarity is calculated using a vector distance algorithm between the question semantic vector and the corresponding text semantic vector. It can be understood that the similarity represents the degree of similarity between the semantics of the question text and the semantics of the corresponding text content. Therefore, the similarity of the document title is used as the matching score of the document title, the similarity of the document summary is used as the matching score of the document summary, and the maximum value of the similarity among the document fragments is used as the matching score of the document fragment.

[0095] The vector distance algorithm includes, but is not limited to, any of the following: cosine similarity algorithm, vector dot product algorithm, Manhattan distance, Euclidean distance algorithm, Pearson correlation coefficient, etc.

[0096] In this embodiment, the question semantic vector of the question text and the text semantic vectors corresponding to multiple text contents of each preset explanatory document are determined by the text vectorization model. The similarity between the question semantic vector and each text semantic vector is determined as the matching score of each text content, which ensures the accuracy and reliability of the matching score and the high efficiency of the process.

[0097] Please see Figure 3 In a further embodiment, before step S1100, responding to the customer service activation event, the following steps are included:

[0098] Step S1000: Obtain the document title of the preset instruction document as the text content of the first structure;

[0099] The document title is a concise description of the core theme or key points of the document, and usually has fewer than 30 characters.

[0100] The document title can be generated by using a large language model with the original text of a preset explanatory document as input, based on the core theme or key points in the original text. Alternatively, the document title can be determined manually. Those skilled in the art can flexibly adapt the method to achieve this.

[0101] Step S1010: Using a preset summary generation model, generate a document summary as the text content of the second structure based on the original text of the preset explanatory document;

[0102] The document summary is a more detailed description of the core theme or key points of the corresponding pre-set explanatory document, usually within 100 words.

[0103] The proposed summary generation model is a large language model suitable for text processing in the field of NLP. It is pre-trained to convergence using an extremely large corpus, acquiring the ability to generate human language and possessing accurate text semantic understanding and logical reasoning capabilities. The proposed summary generation model includes OPT, Chinchilla, PaLM, LLaMA, Alpaca, Vicuna, GPT3, GPT3.5, GPT4, and ChatGPT, among others.

[0104] In one embodiment, the summary generation model is implemented using GPT-4. The original text of the preset explanatory document is used as the input to the summary generation model. The input text is segmented into words to obtain a segmented sequence composed of all the words in the input text. This sequence is then input to the encoding end of the model. By stacking multiple multi-head attention layers and fully connected layers, the segmented sequence is encoded. Specifically, when the segmented sequence passes through the multi-head attention layer, multi-head attention calculation is performed on it to apply self-attention weights to the segmented sequence in different dimensions, thereby obtaining corresponding weighted vector representations. After passing through the fully connected layer, the encoded vector representation of the segmented sequence is obtained. Further, the encoded vector representation is input to the decoding end of the model for decoding. Specifically, during the decoding process, the generation probability of each word is calculated based on the generated words, the current word position, and the corresponding encoded vector representation of the word. A greedy search strategy is used to generate a document summary based on the generation probability of each word.

[0105] Step S1020: Segment the original text of the preset instruction document to obtain each document fragment as the text content of the third structure.

[0106] It is understood that the question text provided by the user is usually related to part of the text content in the original text of the preset instruction document. Therefore, the original text of the preset instruction document is segmented. Specifically, a sliding window of preset window size is used to slide and extract words from the original text with a preset step size to obtain the document segment corresponding to each word extraction. The window size and step size can be set as needed by those skilled in the art based on the number of words in the original text. For example, the window size can be 100 words and the step size can be 50 words.

[0107] In this embodiment, the text content of multiple structures of the preset instruction document is disclosed, so that the text content of each structure can represent the preset instruction document from multiple perspectives, thereby representing the correlation between the preset instruction document and the question text from multiple perspectives.

[0108] Please see Figure 4 In a further embodiment, step S1020, segmenting the original text of the preset instruction document to obtain various document fragments, includes the following steps:

[0109] Step S1021: Using a sliding window of a preset window size, the original text of the preset instruction document is slid-by-sliding to extract words with a preset step size, so as to obtain the document segment to be verified corresponding to each word extraction.

[0110] The window size and step size can be set as needed by those skilled in the art based on the number of characters in the original text. For example, the window size can be 100 characters and the step size can be 50 characters.

[0111] In one embodiment, a sliding window with a window size of 100 characters is used to slide and extract words from the original text of the preset document in a preset step size of 50 characters. The text content of the original text covered by the sliding window for each time is obtained as the document segment to be verified. It can be understood that since the window size is larger than the step size of 50 characters, there are 50 repeated characters of text content among the various document segments to be verified.

[0112] Step S1022: Determine the relevance between each document segment to be verified based on the semantics of each document segment to be verified;

[0113] The deep semantic information of each document segment to be verified is extracted using the text vectorization model. Based on the deep semantic information, the text feature vector corresponding to each document segment to be verified is determined. Furthermore, the vector distance algorithm is used to calculate the vector distance between every two text feature vectors. It can be understood that the vector distance represents the semantic similarity between the corresponding two document segments to be verified, so the vector distance is used as the similarity.

[0114] The vector distance algorithm includes, but is not limited to, any of the following: cosine similarity algorithm, vector dot product algorithm, Manhattan distance, Euclidean distance algorithm, Pearson correlation coefficient, etc.

[0115] Step S1023: When the relevance meets the preset conditions, confirm that the document fragment to be verified passes the verification and is used as a document fragment;

[0116] When all the aforementioned relevance scores are below a preset threshold, it indicates that the original text has been reasonably segmented by the sliding word extraction method. Consequently, the resulting document segments to be verified exhibit certain semantic differences. Therefore, all document segments to be verified are confirmed to have passed verification and are thus recognized as document segments. The preset threshold can be set as needed by those skilled in the art based on the information disclosed herein.

[0117] Step S1024: In response to the verification failure event, adjust the window size and / or step size of the current sliding window and iterate the above process until the document fragment to be verified obtained by sliding the original text with the step size passes the verification.

[0118] When the similarity is greater than or equal to the preset threshold, it indicates that the sliding word extraction of the original text has resulted in incomplete and unreasonable segmentation. Consequently, the resulting document segments to be verified exhibit semantic similarities. All document segments to be verified fail the verification, triggering a verification failure event and responding accordingly. The window size and / or step size of the current sliding window are adjusted. Those skilled in the art can adjust these as needed; an example is adjusting the current step size by 40 characters and / or adjusting the current window size by 10 characters. Then, steps S1021-1022 are iterated until the adjusted sliding window or the current sliding window re-performs sliding word extraction with the adjusted step size or the current step size, resulting in multiple document segments to be verified. The relevance between each document segment is then determined. If all relevances are lower than the preset threshold, all document segments to be verified pass the verification.

[0119] In this embodiment, a sliding window of a preset window size is used to extract words from a preset explanatory document, resulting in a document segment to be verified for each extracted word. When the document segment passes verification, it is confirmed as a document segment. When the document segment fails verification, the current window size and / or step size are adjusted and the above process is iterated until the corresponding document segment passes verification. This allows the window size and step size that can reasonably segment the original text to be determined through experiments, and ensures that there are certain semantic differences between the corresponding document segments.

[0120] Please see Figure 5 In a further embodiment, step S1020, segmenting the original text of the preset instruction document to obtain various document fragments, includes the following steps:

[0121] Step S10200: Perform sentence-level segmentation on the original text of the preset instruction document to obtain the corresponding sentence sequence;

[0122] Sentence-level segmentation is performed based on line breaks and / or periods in the original text to divide the original text into sentences, thereby obtaining all the sentences in the original text. Based on the order of each sentence in the original text, all the sentences are used to form a sentence sequence.

[0123] Step S10201: Based on the semantics of each pair of adjacent clauses in the sentence sequence, use a text semantic model to determine whether there is a sentence relationship between each pair of adjacent clauses. Cut each pair of adjacent clauses that do not have a sentence relationship into segments and determine the corresponding document fragments.

[0124] The text semantic model is pre-trained to convergence, acquiring the ability to determine the confidence level of a sentence-to-sentence relationship between two input sentences. Since the training task is NSP (Next Sentence Prediction), which is well-known in the field, the training process will not be detailed. The network architecture of the text semantic model includes a text encoding layer and a classifier. The text encoding layer can use a model suitable for extracting text features in the NLP field; BERT is recommended, but other models such as Transformer Encoder, RoBERTa, XLM-RoBERTa, MPNet, BiLSTM, and GPT can also be used. The classifier is suitable for binary classification tasks and can be an MLP (Feedforward Neural Network) or a FC (Fully Connected) layer.

[0125] It is understandable that the sentence sequence of the original text usually revolves around the same core theme or key point, and there is a sentence-upper relationship between each two adjacent clauses that describe the same content. That is, the clause that comes first is the upper clause, and the clause that comes later is the lower clause.

[0126] Each pair of adjacent clauses in the sentence sequence constitutes a single input text. Each input text is used as input to a text semantic model. Taking a single input text as an example, the text encoding layer in the text semantic model extracts the deep semantic information corresponding to each clause in the input text, determining the corresponding text semantic vector. The classifier of the text semantic model maps the text semantic vector to preset binary categories, determining the classification probability of mapping to the positive category. The binary categories include a positive category representing a relationship between two clauses in the input text, and a negative category representing a relationship between two clauses in the input text. It is easy to understand that the classification probability represents the confidence level that two sentences in the corresponding input text have a sentence-line relationship. The higher the classification probability, the higher the confidence level. Accordingly, every two adjacent sentences in the sentence sequence whose classification probability is lower than a preset threshold can be identified sequentially. These are two sentences that do not have a sentence-line relationship. Each pair of adjacent sentences is then segmented, so that adjacent sentences with a sentence-line relationship form a text segment, and adjacent sentences without a sentence-line relationship are divided into different text segments, thereby determining each document segment in the sentence sequence.

[0127] In this embodiment, the original text of the preset instructional document is divided into sentence-level segments to obtain a sentence sequence. Then, a text semantic model is used to determine whether there is a sentence-upper relationship between each two adjacent sentences in the sentence sequence. Based on this, each document segment in the sentence sequence is divided, which can accurately and efficiently segment the original text based on the semantic relationship between each sentence in the original text.

[0128] Please see Figure 6In a further embodiment, after step S1400, which uses a large language model to determine the response text based on the at least one candidate document, the following steps are included:

[0129] Step S1500: Verify the compliance and authenticity of the reply text using a preset reply verification model, and determine the verification pass rate of the reply text;

[0130] The response verification model is pre-trained to convergence, acquiring the ability to determine the verification pass rate of the input text. The network architecture of the response verification model includes a text encoding layer and a classifier. The text encoding layer can use a model suitable for extracting text features in the NLP field, with BERT being recommended, but other models such as Transformer Encoder, RoBERTa, XLM-RoBERTa, MPNet, BiLSTM, and GPT can also be used. The classifier is suitable for binary classification tasks and can be an MLP (feedforward neural network) or a FC (fully connected layer). The response text is input into the response verification model, where the text encoding layer extracts the deep semantic information of the response text, determining the text semantic vector. The classifier then maps the text semantic vector to preset binary categories, determining the classification probability of the positive category as the verification pass rate. The binary categories include a positive category representing compliant and truthful input text, and a negative category representing non-compliant and / or untruthful input text. It is easy to understand that the pass rate represents the confidence level of the corresponding input text's compliance and authenticity; the higher the pass rate, the higher the confidence level.

[0131] To train the response verification model, multiple response texts generated by the model can be pre-collected. Compliant and genuine response texts are manually selected as training samples belonging to the positive category, while non-compliant and / or inauthentic response texts are selected as training samples belonging to the negative category. The corresponding verification pass rate is labeled based on whether the training samples are compliant and authentic, serving as a supervision label. Those skilled in the art can flexibly modify the labeling; for example, the supervision label for training samples belonging to the positive category represents a verification pass rate of 1, and the supervision label for training samples belonging to the negative category represents a verification pass rate of 0. Each training sample is associated with its supervision label to form a training set. A single training sample and its supervision label are obtained from the training set and input into the response verification model. The text encoding layer in the model extracts the deep semantic information of the training sample, determining the text semantic vector. The classifier in the model maps the text semantic vector to a preset binary category, determining the classification probability mapped to the positive category as the predicted verification pass rate. The system calls a preset cross-entropy loss function or mean squared error loss function, which can be flexibly set by those skilled in the art based on prior knowledge or experimental experience. It calculates the loss value of the predicted validation pass rate based on the supervision labels of the training samples. When the loss value reaches a preset threshold, it indicates that the validation model has been trained to a convergent state, and model training can be terminated. When the loss value does not reach the preset threshold, it indicates that the validation model has not converged. Therefore, gradient updates are performed on the model based on the loss value. Typically, backpropagation is used to correct the weight parameters of each stage of the model to make the model further approach convergence. Then, other training samples are called to iteratively train the model until it is trained to a convergent state.

[0132] Step S1510: When the verification pass rate reaches the preset standard, the reply text is pushed to the user.

[0133] When the pass rate exceeds a preset threshold, it means that the reply text has passed the verification, i.e., the reply text is compliant and authentic, and the reply text is pushed to the user; when the pass rate is less than or equal to the preset threshold, it means that the reply text has failed the verification, i.e., the reply text is non-compliant and / or unauthentic, and the reply text is prohibited from being pushed to the user. Instead, the user can be directly replied with a statement such as "Sorry, unable to reply" or "Reply failed".

[0134] In this embodiment, a response verification model is used to verify the response text, thereby ensuring that the user receives a compliant and authentic response text that passes the verification, avoiding the user from being misled by receiving non-compliant and / or authentic response text, and protecting the user experience.

[0135] Please see Figure 7This application provides an e-commerce customer service Q&A device, which is a functional embodiment of the e-commerce customer service Q&A method of this application. On another note, this e-commerce customer service Q&A device, also provided to fulfill one of the purposes of this application, includes an event response module 1100, a score determination module 1200, a rating determination module 1300, and a response determination module 1400. The event response module 1100 is used to respond to a customer service activation event and obtain the question text. The score determination module 1200 is used to perform semantic matching between the question text and multiple text contents with different structures of each preset instruction document to determine the matching score corresponding to each text content. The rating determination module 1300 is used to determine the relevant rating between each preset instruction document and the question document based on the matching score. The response determination module 1400 is used to filter at least one preset instruction document whose relevant rating meets preset conditions as a candidate document, and use a large language model to determine the response text based on the at least one candidate document.

[0136] In a further embodiment, the score determination module 1200 includes: a text vectorization submodule, used to determine a question semantic vector based on the semantics of the question text using a preset text vectorization model, and to determine the text semantic vector corresponding to each text content based on the semantics of multiple text contents with different structures in each preset explanatory document; and a similarity determination submodule, used to determine the similarity between the question semantic vector and each text semantic vector, as a matching score corresponding to each text content.

[0137] In a further embodiment, the scoring determination module 1300 includes: a scoring calculation submodule, used to perform a weighted summation of the matching scores corresponding to multiple text contents of each preset instruction document to obtain the relevant score corresponding to each preset instruction document.

[0138] In a further embodiment, before the event response module 1100, there are: a first structure submodule, used to obtain the document title of a preset instructional document as the text content of the first structure; a second structure submodule, used to generate a document summary based on the original text of the preset instructional document using a preset summary generation model as the text content of the second structure; and a third structure submodule, used to segment the original text of the preset instructional document to obtain each document fragment as the text content of the third structure.

[0139] In a further embodiment, the third structural submodule includes: a sliding word extraction submodule, used to extract words from the original text of the preset explanatory document using a sliding window of a preset window size with a preset step size, to obtain the document fragment to be verified corresponding to each word extraction; a relevance determination submodule, used to determine the relevance between each document fragment to be verified based on the semantics of each document fragment to be verified; a fragment verification submodule, used to confirm that the document fragment to be verified passes the verification when the relevance meets a preset condition, and is thus recognized as a document fragment; and a verification failure submodule, used to respond to a verification failure event, adjust the window size and / or step size of the current sliding window, and iterate the above process until the document fragment to be verified obtained by the sliding window extracting words from the original text with the specified step size passes the verification.

[0140] In a further embodiment, the third structural submodule includes: a sentence segmentation module, used to perform sentence-level segmentation on the original text of the preset explanatory document to obtain a corresponding sentence sequence; and a segmentation module, used to sequentially determine whether there is a sentence relationship between each two adjacent sentences based on the semantics corresponding to each two adjacent sentences in the sentence sequence, and to segment each two adjacent sentences that do not have a sentence relationship into segments to determine the corresponding document segments.

[0141] In a further embodiment, after the response determination module 1400, there are: a post-verification submodule, used to verify the compliance and authenticity of the response text using a preset response verification model, and determine the verification pass rate of the response text; and a user push submodule, used to push the response text to the user when the verification pass rate reaches a preset standard.

[0142] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 8 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 a sequence of control information. When the computer-readable instructions are executed by the processor, the processor can implement an e-commerce customer service question-and-answer 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 e-commerce customer service question-and-answer 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 8The 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.

[0143] In this embodiment, the processor is used to execute... Figure 7 The system contains 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 e-commerce customer service Q&A device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0144] 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 e-commerce customer service question-and-answer method of any embodiment of this application.

[0145] 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.

[0146] In conclusion, this application accurately answers questions and ensures that the responses are compliant and truthful.

[0147] 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.

[0148] 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. A question-and-answer method for e-commerce customer service, characterized in that, Includes the following steps: Respond to the customer service activation event and retrieve the query text; The question text is semantically matched with multiple text contents of different structures in each preset explanatory document to determine the matching score corresponding to each text content; Based on the matching score, a relevant score is determined between each preset instruction document and the question text; At least one of the preset explanatory documents that meet the preset conditions for the relevant scores is selected as candidate documents, and the response text is determined based on the at least one candidate document using a large language model; Before the customer service activation event is initiated, the process includes: obtaining the title of a preset instructional document as the text content of the first structure; generating a document summary based on the original text of the preset instructional document using a preset summary generation model as the text content of the second structure; and segmenting the original text of the preset instructional document to obtain each document fragment as the text content of the third structure.

2. The e-commerce customer service Q&A method according to claim 1, characterized in that, The process involves semantically matching the question text with multiple text contents of different structures in each preset instruction document to determine the matching score corresponding to each text content, including the following steps: A preset text vectorization model is used to determine the question semantic vector based on the semantics of the question text, and the text semantic vector corresponding to each text content is determined based on the semantics of multiple text contents with different structures in each preset explanatory document. The similarity between the question semantic vector and each of the text semantic vectors is determined and used as the matching score corresponding to each text content.

3. The e-commerce customer service Q&A method according to claim 1, characterized in that, Based on the matching score, a correlation score is determined between each preset explanatory document and the question text, including: The matching scores corresponding to multiple text contents of each preset instruction document are weighted and summed to obtain the relevant score for each preset instruction document.

4. The e-commerce customer service question-and-answer method according to claim 1, characterized in that, The original text of the preset instruction document is segmented to obtain individual document fragments, including the following steps: A sliding window of a preset window size is used to slide and extract words from the original text of the preset instruction document with a preset step size, so as to obtain the document segment to be verified corresponding to each word extraction. The relevance between each document segment to be verified is determined based on the semantics of each document segment to be verified; When the relevance meets the preset conditions, the document fragment to be verified is confirmed to pass the verification and is recognized as a document fragment. In response to a verification failure event, adjust the current sliding window size and / or step size, and then iterate the above process until the document fragment to be verified obtained by sliding the original text with the sliding window at the step size passes the verification.

5. The e-commerce customer service Q&A method according to claim 1, characterized in that, The original text of the preset instruction document is segmented to obtain individual document fragments, including the following steps: The original text of the preset instruction document is segmented at the sentence level to obtain the corresponding sentence sequence; Based on the semantics of each pair of adjacent clauses in the sentence sequence, a text semantic model is used to determine whether there is a sentence relationship between each pair of adjacent clauses. Each pair of adjacent clauses that do not have a sentence relationship is then segmented to determine the corresponding document fragments.

6. The e-commerce customer service question-and-answer method according to claim 1, characterized in that, After determining the response text based on the at least one candidate document using a large language model, the process includes the following steps: The compliance and authenticity of the reply text are verified using a preset reply verification model, and the verification pass rate of the reply text is determined. When the verification pass rate reaches the preset standard, the reply text will be pushed to the user.

7. An e-commerce customer service Q&A device, characterized in that, include: The event response module is used to respond to customer service activation events and obtain the question text; The scoring determination module is used to semantically match the question text with multiple text contents of different structures in each preset explanatory document, and determine the matching score corresponding to each text content. The scoring determination module is used to determine the relevant score between each preset instruction document and the question text based on the matching score; The response confirmation module is used to filter at least one of the preset description documents that meet the preset conditions for the relevant rating as candidate documents, and to determine the response text based on the at least one candidate document using a large language model; Before the customer service activation event is initiated, the process includes: obtaining the title of a preset instructional document as the text content of the first structure; generating a document summary based on the original text of the preset instructional document using a preset summary generation model as the text content of the second structure; and segmenting the original text of the preset instructional document to obtain each document fragment as the text content of the third structure.

8. The e-commerce customer service Q&A device according to claim 7, characterized in that, The scoring determination module includes: a text vectorization submodule, used to determine the question semantic vector based on the semantics of the question text using a preset text vectorization model, and to determine the text semantic vector corresponding to each text content based on the semantics of multiple text contents with different structures in each preset explanatory document; and a similarity determination submodule, used to determine the similarity between the question semantic vector and each text semantic vector, as the matching score corresponding to each text content.

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 6.

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 6, which, when invoked by a computer, executes the steps included in the corresponding method.

Images