E-commerce vertical domain database construction method and related device

Abstract

Embodiments of the invention provide an e-commerce vertical domain database construction method and a related device. The method specifically comprises the steps of obtaining a user question in an e-commerce vertical domain; according to the user question and historical data of the e-commerce vertical field, calling an answer mining large model to indicate the answer mining large model to mine an answer of the user question from the historical data; wherein the user questions and answers form question and answer pairs; the question and answer pairs are used for forming an e-commerce vertical domain database; calling an evaluation large model according to the question and answer pairs to indicate the evaluation large model to identify and obtain the question and answer pairs with wrong answers, and giving out error description labels for the wrong answers; wherein the error description label is used as a cue word for training the answer mining large model, so that the question and answer pair is used as a negative sample to train the answer mining large model. The accuracy of information in the e-commerce vertical domain database can be improved to a certain extent.

Description

Technical Field

[0001] The embodiments in the present application relate to the field of artificial intelligence technology, and particularly to a method for constructing an e-commerce vertical domain database and related devices. Background Art

[0002] In the e-commerce vertical domain, users usually raise various questions regarding different products, services, or transaction processes. These questions cover aspects such as product functions, usage methods, target audiences, and after-sales services. Most traditional database construction methods rely on manual labor or simple retrieval tools to organize and maintain Q&A content, making it difficult to meet the diverse needs of users in a timely and effective manner. With the continuous increase in the types of products and services on e-commerce platforms, the cost of manually maintaining the database is high, and errors and inconsistencies are prone to occur, resulting in low accuracy of the Q&A information in the database. Summary of the Invention

[0003] In view of this, multiple embodiments of the present application are dedicated to providing a method for constructing an e-commerce vertical domain database and related devices, which can improve the accuracy of information in the e-commerce vertical domain database to a certain extent.

[0004] An embodiment of the present application provides a method for constructing an e-commerce vertical domain database, including: obtaining user questions in the e-commerce vertical domain; according to the user questions and historical data in the e-commerce vertical domain, invoking an answer mining large model to instruct the answer mining large model to mine answers to the user questions from the historical data; where user questions and answers form Q&A pairs; the Q&A pairs are used to form an e-commerce vertical domain database; according to the Q&A pairs, invoking an evaluation large model to instruct the evaluation large model to identify Q&A pairs with incorrect answers and give error description tags for the incorrect answers; where the error description tags are used as prompt words for training the answer mining large model to train the answer mining large model with the Q&A pairs as negative samples.

[0005] An embodiment of the present application also provides a construction device for an e-commerce vertical domain database, including: an acquisition module, configured to acquire user questions in the e-commerce vertical domain; a first call module, configured to call an answer mining large model according to the user questions and the historical data in the e-commerce vertical domain, so as to instruct the answer mining large model to mine the answers to the user questions from the historical data; wherein, the user questions and the answers form question-and-answer pairs; the question-and-answer pairs are used to form an e-commerce vertical domain database; a second call module, configured to call an evaluation large model according to the question-and-answer pairs, so as to instruct the evaluation large model to identify the question-and-answer pairs with wrong answers, and give wrong description tags for the wrong answers; wherein, the wrong description tags are used as prompt words for training the answer mining large model, so as to use the question-and-answer pairs as negative samples to train the answer mining large model.

[0006] An embodiment of the present application also provides a computer device, which includes a memory and a processor. At least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the method as described above.

[0007] An embodiment of the present application also provides a computer-readable storage medium, in which at least one computer program is stored, and when the at least one computer program is executed by a processor, it can implement the method as described above.

[0008] An embodiment of the present application also provides a computer program product, which is used to implement the method as described above.

[0009] In multiple embodiments provided by the present application, by acquiring user questions in the e-commerce vertical domain and calling an answer mining large model to mine corresponding answers from historical data, it is possible to quickly construct an e-commerce vertical domain database with a relatively low manual maintenance cost, and by further calling an evaluation large model to identify and use wrong description tags to mark wrong answers, and using the wrong answers as negative samples to train the answer mining large model, the accuracy of the e-commerce vertical domain database is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the application scenario of the construction method of the e-commerce vertical domain database provided by an embodiment of the present application.

[0011] Figure 2 It is a flowchart of the construction method of the e-commerce vertical domain database provided by an embodiment of the present application.

[0012] Figure 3 It is a schematic diagram of the process of rewriting user questions based on intent recognition provided by an embodiment of the present application.

[0013] Figure 4 Schematic diagram of the answer mining process provided for an embodiment of the present application.

[0014] Figure 5 Schematic diagram of the Q&A pair extraction process for picture data provided for an embodiment of the present application.

[0015] Figure 6 Visualization display interface for evaluating the output results of a large model provided for an embodiment of the present application.

[0016] Figure 7 Schematic diagram of the answer mining accuracy rate statistical results provided for an embodiment of the present application.

[0017] Figure 8 Schematic diagram of the modules of the device for constructing an e-commerce vertical domain database provided for an embodiment of the present application.

[0018] Figure 9 Schematic diagram of a computer device provided for an embodiment of the present application. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0020] In the description of the embodiments of the present application, it should be understood that the terms "first" and "second" are only used for descriptive purposes, and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of the described features. In the description of the embodiments of the present application, "a plurality" means two or more, unless otherwise specifically defined.

[0021] In the related art, in the e-commerce vertical domain, the common database construction methods usually rely on manual collation or simple retrieval tools to maintain Q&A information. Specifically, it is necessary to manually select typical questions from various questions raised by users, and manually write answers based on the historical data of products or services to establish Q&A content. However, when faced with the increasing types of products and services, this method often has difficulty in responding to the diverse question needs of users in a timely manner, and may also include incorrect answers in the database due to understanding deviations or omissions during the manual writing process, thereby reducing the accuracy of the e-commerce vertical domain database. With the development of e-commerce platforms, the workload of manual maintenance has been increasing day by day, and the required labor cost has also been continuously rising.

[0022] With the rapid development of natural language processing technology, large models have gradually been widely applied in question-and-answer scenarios. Large models are usually pre-trained with extremely large-scale data and possess strong question-understanding and answer-generation capabilities. Therefore, in some scenarios of building question-and-answer pairs in e-commerce vertical domains, large models can directly be used to extract answers from historical data and automatically generate question-and-answer pairs.

[0023] However, although large models have shown good potential in building e-commerce vertical domain databases, there are still problems with insufficient accuracy in actual applications. For example, when generating answers or marking incorrect answers, large models may make mistakes due to insufficient context inference. Once incorrect answers are incorporated into the database, it will directly affect the accuracy of subsequent user queries, and the efficiency of manually identifying or deleting these incorrect answers is relatively low.

[0024] In summary, there is still a problem of relatively low construction accuracy of e-commerce vertical domain databases in the existing technology.

[0025] In multiple embodiments provided in this application, the method for building an e-commerce vertical domain database can be applied to a device for building an e-commerce vertical domain database. The device for building an e-commerce vertical domain database can be an electronic device with certain computing power and network access capabilities. This electronic device can be a desktop computer, a laptop computer, a tablet computer, a smart phone, or a server. The server can also be a distributed server, including multiple processors, memories, network communication modules, etc., which cooperate to achieve various functions. Or, the server can also be a server cluster formed by several servers, with higher computing and data processing capabilities. With the development of science and technology, the server can also be realized by new forms of technical means, such as a new type of "server" based on quantum computing. Of course, in some embodiments, the device for building an e-commerce vertical domain database can also be a program module running in an electronic device.

[0026] Please refer to Figure 1 . An application scenario example of a method for building an e-commerce vertical domain database is provided in an embodiment of this application. The method for building an e-commerce vertical domain database is applied to a device for building an e-commerce vertical domain database to improve the accuracy of the e-commerce vertical domain database, thereby improving the reply accuracy of the question-and-answer scenario.

[0027] The e-commerce platform includes functions such as product display, online shopping, and customer service. Users can browse and purchase products on the e-commerce platform through the client or website. When a user triggers the customer service function for a product "XXX Guipi Pills" expecting the customer service to answer questions about this product, the client or website displays a Q&A interface for the user. In the Q&A interface, the user sends the question "Can the elderly take this product?" The customer service replies "Yes, dear. We see that you have submitted an order. Thank you for your trust and support. We will arrange the shipment for you as soon as possible!". Among them, the content of the reply can be edited by human customer service, intelligent customer service, or professionals, and the content of this reply can be obtained based on the e-commerce vertical domain database.

[0028] To improve the richness and accuracy of the e-commerce vertical domain database, the construction device of the e-commerce vertical domain database extracts the question with the intention of product consultation "Can the elderly take this product?" and rewrites the question "Can the elderly take this product?" based on the product "XXX Guipi Pills", thus obtaining the user question "Can the elderly take this Guipi Pills?". Further, the construction device of the e-commerce vertical domain database can call the answer mining large model according to the customer service reply "Yes, dear. We see that you have submitted an order. Thank you for your trust and support. We will arrange the shipment for you as soon as possible!", the content of the product details page of "XXX Guipi Pills", the picture data with product information, and the Q&A pairs corresponding to the near-synonym user questions semantically similar to the user question "Can the elderly take this Guipi Pills?" to instruct the answer mining large model to mine the answer to the user question from them. For example, "Yes, dear".

[0029] Thus, the user question "Can the elderly take this Guipi Pills?" and the answer "Yes, dear" are formed into a Q&A pair. Executing the above solution for a large number of user questions can obtain a large number of Q&A pairs, and these Q&A pairs can form an e-commerce vertical domain database with richness and accuracy for human customer service, intelligent customer service, or professionals to call and refer to when answering user questions later.

[0030] To ensure the accuracy of the e-commerce vertical domain database, the construction device of the e-commerce vertical domain database calls the evaluation large model to identify the Q&A pairs with wrong answers, and gives error description labels for the wrong answers. For example, the knowledge does not match the product, the knowledge answer is wrong, the knowledge answer is off-topic, the knowledge content is incomplete, or it is not in the reference materials. For example, the Q&A pair with a wrong answer includes the user question "How many drops should be added when the baby's milk intake is 120 ml?" and the answer "15 ml / bottle, and one bottle can be used 75 times" for the product "XXX Lactase Infant Drops XXX". The error description label corresponding to this Q&A pair is that the knowledge answer is off-topic. This error description label, as a prompt word, can use this Q&A pair as a negative sample to train the answer mining large model, thereby improving the answer mining quality of the answer mining large model, and further improving the accuracy of the e-commerce vertical domain database.

[0031] In this scenario example, by obtaining user questions in the e-commerce vertical domain and calling the answer mining large model to mine corresponding answers from historical data, it is possible to quickly build an e-commerce vertical domain database with a relatively low manual maintenance cost. And by further calling the evaluation large model to identify and mark wrong answers with error description labels, and using the wrong answers as negative samples to train the answer mining large model, the accuracy of the e-commerce vertical domain database is improved.

[0032] Please refer to Figure 2 One embodiment of the present application provides a method for constructing an e-commerce vertical domain database. The method for constructing the e-commerce vertical domain database may include the following steps.

[0033] Step S110: Obtain user questions in the e-commerce vertical domain.

[0034] Step S120: According to the user question and the historical data of the e-commerce vertical domain, call the answer mining large model to instruct the answer mining large model to mine the answer to the user question from the historical data; wherein, the user question and the answer form a Q&A pair; the Q&A pair is used to form the e-commerce vertical domain database.

[0035] Step S130: According to the Q&A pair, call the evaluation large model to instruct the evaluation large model to identify the Q&A pairs with wrong answers, and give error description labels for the wrong answers; wherein, the error description label is used as a prompt word for training the answer mining large model, and the Q&A pair is used as a negative sample to train the answer mining large model.

[0036] In this embodiment, in order to explore the correspondence between user questions and answers and thus construct an e-commerce vertical domain database, the construction device of the e-commerce vertical domain database starts to execute the database construction task by obtaining user questions in the e-commerce vertical domain. Among them, the e-commerce vertical domain refers to a business model in e-commerce that focuses on a specific industry or niche market and provides highly specialized products or services to meet the needs of specific user groups. In the actual application process, this embodiment can be applied to various e-commerce vertical domains to form corresponding e-commerce vertical domain databases. User questions can include multiple question types, such as product function consultation, usage method query, applicable population assessment, after-sales service request, etc. The historical data of the e-commerce vertical domain can include stored product information, user feedback, transaction records, and historical customer service conversation data, etc.

[0037] During the database construction process, the construction device can call the answer mining large model to mine the answers to user questions from the historical data. The answer mining large model can be an artificial intelligence model pre-trained based on a large-scale language model and fine-tuned in combination with domain knowledge, and has the ability to reason and generate high-quality answers from unstructured or semi-structured text data. By calling this answer mining model, the e-commerce vertical domain database can automatically extract highly relevant answers that meet business requirements from historical data and form question-and-answer pairs. Among them, each question-and-answer pair consists of a user question and its corresponding answer and serves as the basic data unit of the database.

[0038] In this embodiment, the question-and-answer pairs are used to form the e-commerce vertical domain database. However, since the answer mining large model may have problems such as inaccurate information, inconsistency with product or service descriptions, redundant or missing expressions during the reasoning process, directly storing the unvalidated question-and-answer pairs may affect the quality of the database. To improve the accuracy of the database, the construction device further calls the evaluation large model to evaluate the quality of the generated question-and-answer pairs. The evaluation large model can be a text quality discrimination model fine-tuned through supervision, and its role is to analyze the correctness, consistency, and integrity of the question-and-answer pairs and identify the wrong answers among them.

[0039] For the question-and-answer pairs identified as wrong, the evaluation large model further generates error description labels to specifically describe the error types. The error description labels can include categories such as "incomplete answer", "answer inconsistent with product description", "wrong answer", or "answer irrelevant to user question" to identify different types of errors. These error description labels can not only be used to filter low-quality data to prevent it from being incorporated into the e-commerce vertical domain database, but also serve as auxiliary information for training data, enabling the answer mining large model to further optimize the answer generation strategy in subsequent training.

[0040] In this embodiment, the error description tag can be used as a training prompt to construct a negative sample training set for training and optimizing the answer mining large model. Specifically, the construction device can input the Q&A pairs with error description tags as negative samples into the answer mining large model to adjust its parameters, so that it can avoid similar incorrect inferences in subsequent inference processes and improve the quality of answer generation. Through the above optimization strategy, the e-commerce vertical domain database can improve the accuracy and reliability of Q&A pairs during continuous training and iteration.

[0041] In this embodiment, by calling the answer mining large model to mine the answers to user questions from historical data and combining with the evaluation large model for quality screening, the database construction process can reduce manual intervention and improve data processing efficiency. At the same time, using the error description tag to construct a negative sample training set enables the answer mining large model to be continuously optimized in subsequent training processes, improving the accuracy of Q&A pairs, and ultimately realizing the automated construction of a high-quality e-commerce vertical domain database.

[0042] In some embodiments, the construction device of the e-commerce vertical domain database can obtain the original user questions with the intention of commodity consultation from the historical data of the e-commerce vertical domain; rewrite the original user questions based on the commodity information corresponding to the original user questions to obtain user questions; wherein, at least part of the commodity information corresponding to the original user questions is included in the user questions.

[0043] In this embodiment, since each user question in the historical data may correspond to multiple intentions, for example, commodity consultation intention, chatting intention, logistics consultation intention, etc., in order to improve the accuracy of information in the e-commerce vertical domain database, the construction device can extract the original user questions with the intention of commodity consultation from the historical data based on the intention recognition model. The intention recognition model is an artificial intelligence model pre-trained based on a large-scale language model, used to identify the intention corresponding to the user question, and accordingly extract some of the user questions specifically for constructing the e-commerce vertical domain database.

[0044] After identifying the intention of the original user question, the original user questions can be clustered to obtain multiple question sets corresponding to different intentions respectively, and the question set corresponding to the commodity consultation intention can be obtained from them and applied to the e-commerce vertical domain database construction process.

[0045] On this basis, since there may be problems such as incomplete information in the original user questions, the construction device can trigger the text normalization model to rewrite the original user questions based on the product information corresponding to these original user questions, so as to normalize the original user questions and obtain user questions with complete information. Among them, the text normalization model is an artificial intelligence model obtained through supervised training for rewriting user questions in a specified text format. The rewriting here includes, but is not limited to, the following operations: adding, deleting, and replacing. Each user question obtained in this way has completeness and can be used in the subsequent database construction process. Among them, the product information corresponding to the original user questions can include, but is not limited to, product names, product specifications, product parameters, product descriptions, etc.

[0046] For details, please refer to Figure 3 , such as Figure 3 As shown, the intent recognition model can recognize that the intent corresponding to the original user question "Can the elderly eat this product?" is the product consultation intent, and also recognize that the intent corresponding to the original user question "Okay" is the chatting intent. On this basis, the text normalization model can be triggered to rewrite the original user question "Can the elderly eat this product?" corresponding to the product consultation intent into a user question "Can the elderly eat this Guipi Pills?" that includes some / all of the content in the product information. Among them, the rewritten user question follows a preset rule, and the preset rule is at least used to limit the fields that the user question needs to include, such as the product name. For different e-commerce vertical fields, the corresponding preset rules may be different. In addition, it should be clear that the original user questions can come from the same or different Q&A interfaces. In the same Q&A interface, different original user questions can correspond to the same or different product information.

[0047] In some embodiments, after the construction device obtains the original user questions with the product consultation intent from the historical data in the e-commerce vertical field, it can also traverse the e-commerce vertical field database based on semantic similarity. If an existing user question with a similarity greater than the preset value to the original user question is found through traversal, the original user question is deleted. If no existing user question with a similarity greater than the preset value to the original user question is found through traversal, the subsequent steps are continued. This can avoid repeated processing of the same or highly similar questions, thereby avoiding wasting computing resources.

[0048] In some embodiments, the historical data includes a communication data set formed by the communication data between the user account and the merchant account; the construction device of the e-commerce vertical field database can preferentially use the communication data set to call the answer mining large model to obtain the answer to the user question.

[0049] In this embodiment, the user account and the merchant account can communicate on the Q&A interface. The user raises questions about the product through the user account, and the customer service replies to the questions raised by the user through the merchant account. The communication results between the two can be stored as communication data for use as the input to the answer mining large model. The communication results include any number of user questions and any number of customer service replies.

[0050] Among them, the communication data collected from different Q&A interfaces or Q&A periods can form a communication data set. Since there may be similar questions and similar replies among different communication data in the communication data set, after the communication data set is input into the answer mining large model, the answer mining large model can analyze and obtain the answers corresponding to the user questions by combining the natural language understanding of the communication data and the relevance and similarity between the communication data. Since this answer is generated using the natural language understanding ability of the large model and the similarity between the communication data, this answer has a certain degree of correctness and reference significance.

[0051] In some embodiments, the above-mentioned communication data set refers to a set composed of communication data related to user questions obtained through screening, which can save computing power.

[0052] In some embodiments, the historical data includes product information for introducing the product; the construction device of the e-commerce vertical domain database can construct a prompt instruction according to the communication data set and the product information; among them, the product information serves as an information supplement prompt word for the communication data set.

[0053] In this embodiment, in order to improve the mining quality of the answer mining large model and thus improve the matching degree between user questions and answers, the construction device can construct a prompt instruction by combining the communication data set and the product information. Among them, the prompt instruction, as the input to the answer mining large model, can prompt the answer mining large model to mine the answers corresponding to the user questions from the communication data set and the product information. The product information includes but is not limited to the description on the product detail page, the product detail page pictures, the advertising page slogans, and the user comments on the product. The Q&A pairs mined based on the communication data set and the product information can have higher accuracy.

[0054] In some embodiments, in order to improve the output accuracy of the answer mining large model, the construction device of the e-commerce vertical domain database inputs the thought chain prompting instruction for prompting the answer mining large model to display the logical reasoning process into the answer mining large model, so as to trigger the answer mining large model to output the entire logical reasoning process including the answer corresponding to the user's question, thereby counteracting model hallucinations and improving the large model's ability to summarize and output accuracy. For example, the thought chain prompting instruction is expressed as: You need to tell me which piece of knowledge you used as the source, and then extract, summarize, and give a reasonable explanation from it. Optionally, the entire logical reasoning process may include information such as the answer corresponding to the user's question, the reference information source, and the reason for generating the answer.

[0055] In some embodiments, the construction device of the e-commerce vertical domain database may also input a scoring instruction for instructing the answer mining large model to output the quality score of the answer it mines into the answer mining large model. The scoring instruction may include, but is not limited to, scoring criteria, existing question-and-answer pairs and their scores.

[0056] In some embodiments, the historical data may also include external knowledge, and the sources of external knowledge may be books, technical manuals, web pages, etc.

[0057] As an example, specific reference can be made to Figure 4 , such as Figure 4 shown, for the "user question: Can this ventilator improve snoring? For example, how long does it take to stop snoring after using it?", it is rewritten to obtain the "rewritten user question: Can the XX anti-snoring bilevel ventilator improve snoring? For example, how long does it take to stop snoring after using it?".

[0058] Input into the answer mining large model but not limited to the following content: communication data related to the user's question {Customer service: No, dear. User: Does it only stop snoring when wearing it? Customer service: The machine is for auxiliary use. That is, when you use the machine, the machine helps you breathe to prevent snoring. But if you don't use it, the machine can't help you. Yes, dear...}, external knowledge "The ventilator can treat snoring, but it can't cure it. It is mainly used for obstructive sleep apnea syndrome. Ventilator treatment can't eradicate the disease, it can only control symptoms and improve manifestations. Because after using the ventilator for obstructive sleep apnea, the manifestations of sleep apnea can be eliminated, hypoventilation can be eliminated, and snoring can be eliminated.", "Product information: XX anti-snoring bilevel ventilator".

[0059] Thus, it is indicated that the answer mining large model mines the answer corresponding to the "rewritten user question: Can the XX anti-snoring bilevel ventilator improve snoring? For example, how long does it take to stop snoring after using it?" based on data such as communication data, external knowledge, and product information, and outputs it based on the chain-of-thought prompting instruction and the scoring instruction. "Answer: The machine is for auxiliary use and cannot cure the problem. That is, when you use the machine, the machine helps you breathe to prevent snoring. But if you don't use it, the machine can't help you." "Reason: The machine is for auxiliary use. That is, when you use the machine, the machine helps you breathe to prevent snoring. But if you don't use it, the machine can't help you. Yes, dear. The ventilator can treat snoring, but it can't cure it." "Quality score: 5".

[0060] Furthermore, the answer output by the answer mining large model and the corresponding user question can be stored as a question-and-answer pair, so as to obtain an e-commerce vertical domain database composed of question-and-answer pairs. The reason output by the answer mining large model can not only assist the answer mining large model to output more accurate answers to combat model hallucinations, but also be used as a basis for answer review. The quality score output by the answer mining large model can also be used as a basis for answer review, and can also be used as a basis for model optimization.

[0061] In some embodiments, the answer, reason, and quality score output by the answer mining large model can be submitted to the answer evaluation process, so that relevant personnel or the evaluation large model can make an online / offline evaluation of the answer with reference to the reason and quality score. The evaluation result can be expressed as "correct" or "wrong". In addition, an error description label can be generated for the answer with an evaluation result of "wrong" based on the reference reason and / or quality score to indicate the error type to which the "wrong" answer belongs.

[0062] In some embodiments, the device for constructing the e-commerce vertical domain database can obtain a synonymous user question that is semantically similar to the user question; among them, the question-and-answer pair including the synonymous user question is used as a synonymous question-and-answer pair; and a prompting instruction is jointly constructed according to the user question, the historical data of the e-commerce vertical domain, and the synonymous question-and-answer pair to call the answer mining large model.

[0063] In this embodiment, the near-synonym user questions semantically similar to the user question can be retrieved by index based on a semantic recognition model. The Q&A pairs corresponding to the near-synonym user questions contain the corresponding answers. Since the generation basis / reason for the answers corresponding to different near-synonym user questions may come from different types of customer service, such as physician customer service, general customer service, intelligent customer service, etc. There may be differences in quality, conflicts, etc. among the answers corresponding to the near-synonym user questions. Based on the statistics and analysis of the answers corresponding to most near-synonym user questions, the correct answer corresponding to the user question can be determined. Thus, the answers corresponding to multiple near-synonym user questions can be used as the basis for obtaining the correct answer. Therefore, the construction device can jointly construct a prompt instruction according to the user question, the historical data in the e-commerce vertical domain, and the near-synonym Q&A pairs, and input the prompt instruction into the answer mining large model to instruct the answer mining large model to output the answer to the user question. The answer generated based on this method has higher accuracy.

[0064] In some embodiments, the historical data includes picture data with product information; the construction device of the e-commerce vertical domain database can call a multi-modal large model according to the picture data to instruct the multi-modal large model to generate multiple Q&A pairs based on the product information in the picture data.

[0065] In this embodiment, the multi-modal large model is an artificial intelligence model pre-trained based on a large-scale language model for analyzing multi-type data inputs. To improve the accuracy and abundance of the Q&A pairs in the e-commerce vertical domain database and solve the problem of insufficient accuracy in the structured processing of picture data containing product information, the construction device can input the picture data into the multi-modal large model, so that the multi-modal large model can perform structured extraction on the product information contained in the picture data to generate multiple Q&A pairs. Among them, the picture data can be pictures included in the product details page, pictures in the advertising page, and web pictures related to the product obtained by searching.

[0066] As an example, reference can be made to Figure 5 , such as Figure 5As shown, taking the picture data as the input of the multimodal large model can instruct the multimodal large model to generate multiple question-and-answer pairs based on the product information in the picture data, specifically including: Question 1: How to select the appropriate nasal mask size? Answer 1: Please refer to the nasal mask specification comparison table, measure the nasal bridge width (W), and select the appropriate size. Question 2: How to select the appropriate nose and mouth mask size? Answer 2: Please refer to the nose and mouth mask specification comparison table, measure the height from the nasal root to the lower lip edge (H), and select the appropriate size. Question 3: Can an inappropriate nasal mask or nose and mouth mask be returned or exchanged? Answer 3: If the nasal mask or nose and mouth mask is inappropriate, do not unpack it. It cannot be returned after unpacking. Please consult the online customer service for replacement service. Writing multiple question-and-answer pairs into the e-commerce vertical domain database can improve the data abundance of the e-commerce vertical domain database.

[0067] Among them, the above question-and-answer pairs written into the e-commerce vertical domain database can be used as the basis for answer mining during the subsequent answer mining process to improve the answer mining quality.

[0068] In some embodiments, the error description label includes at least one of the following: knowledge does not match the product, knowledge answer is incorrect, knowledge answer is off-topic, knowledge content is incomplete, or not in the reference materials.

[0069] In this embodiment, the evaluation large model can identify the question-and-answer pairs with incorrect answers in the question-and-answer pairs. In order to use these question-and-answer pairs to optimize the answer mining large model, it can be instructed to mark the error types of the question-and-answer pairs with incorrect answers to obtain the error description labels corresponding to the question-and-answer pairs of each incorrect answer. For example, it can refer to Figure 6 , Figure 6 shows a visual display interface for the output results of the evaluation large model. The knowledge names (such as user questions, question IDs, etc.) corresponding to each question-and-answer pair containing incorrect answers are displayed under the "knowledge name" field. Correspondingly, the production time of the question-and-answer pair, the detection party, the status indicating whether the inspection has been completed, the recognition result indicating correct or incorrect, and the error description label indicating the error type can also be displayed. Based on Figure 6 the content shown, it is possible to clarify each question-and-answer pair containing incorrect answers and the corresponding information and error description labels. The error description label combined with the corresponding question-and-answer pair containing incorrect answers can form negative samples for further optimizing and strengthening the learning of the answer mining large model.

[0070] Furthermore, based on the recognition results of the question-and-answer pairs with incorrect answers and the manual review results / evaluation large model recognition results of the question-and-answer pairs without incorrect answers, it can be statistically obtained as Figure 7The statistical results of answer mining accuracy shown. Specifically, the statistical results of answer mining accuracy include the following fields: the task name for characterizing the execution of the answer evaluation task for a certain commodity or a certain category of commodities (such as mouth and nose masks), the task time for indicating the execution of this task, the number of answer samples based on which the answer evaluation task is performed, the evaluation progress for all samples in this task, and the accuracy comprehensively characterized by each evaluation result for all samples in this task. This accuracy can represent the ratio of the answers evaluated as correct in this task to the number of samples. Based on this statistical result, the answer mining large model is further optimized to improve the output quality of the answer mining large model.

[0071] Please refer to Figure 8 。This application embodiment also provides a device for constructing an e-commerce vertical domain database. The device for constructing an e-commerce vertical domain database may include: An acquisition module, configured to acquire user questions in the e-commerce vertical domain; A first call module, configured to call an answer mining large model according to the user questions and the historical data in the e-commerce vertical domain, so as to instruct the answer mining large model to mine the answers to the user questions from the historical data; wherein, the user questions and the answers form question-and-answer pairs; the question-and-answer pairs are used to form an e-commerce vertical domain database; A second call module, configured to call an evaluation large model according to the question-and-answer pairs, so as to instruct the evaluation large model to identify the question-and-answer pairs with wrong answers, and give wrong description labels for the wrong answers; wherein, the wrong description labels are used as prompt words for training the answer mining large model, so as to use the question-and-answer pairs as negative samples to train the answer mining large model.

[0072] In this embodiment, the specific functions and effects realized by the device for constructing an e-commerce vertical domain database can be explained by referring to other embodiments of this application, and will not be elaborated here.

[0073] Please refer to Figure 9 。This application embodiment also provides a computer device, which includes: a memory and a processor. At least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the method as described above.

[0074] This application embodiment also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor implements the method as described above.

[0075] This application embodiment also provides a computer program product containing instructions. When the computer program product is executed by a processor, the method as described above is implemented.

[0076] The user information or user account information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, etc.) involved in multiple embodiments of the present application are all information and data that have been authorized by the user or fully authorized by all parties. Moreover, the collection, use, and processing of the relevant data need to comply with the relevant laws and regulations and standards of the relevant countries and regions, and corresponding operation entrances are provided for users to choose to authorize or reject.

[0077] It can be understood that the specific examples in this article are only to help those skilled in the art better understand the embodiments of the present application, rather than limiting the scope of the present invention.

[0078] It can be understood that in various embodiments of the present application, the magnitudes of the sequence numbers of the various processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0079] It can be understood that the various embodiments described in the present application can be implemented alone or in combination, and the embodiments of the present application do not limit this.

[0080] Unless otherwise specified, all technical and scientific terms used in the embodiments of the present application have the same meaning as commonly understood by those skilled in the technical field of the present application. The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the scope of the present application. The term "and / or" used in the present application includes any and all combinations of one or more of the related listed items. The singular forms "a", "above", and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0081] It can be understood that the processor in the implementation manner of the present application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method implementation manner can be completed by the integrated logic circuit in the hardware of the processor or instructions in the form of software. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute each method, step, and logic block diagram disclosed in the implementation manner of the present application. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in combination with the implementation manner of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or can be executed and completed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0082] It can be understood that the memory in the implementation manner of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0083] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0084] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0085] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the couplings or direct couplings or communication connections shown or discussed with each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0087] In addition, in each embodiment of the present application, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0088] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0089] The above is only the specific embodiment of the present application, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application and should be covered by the protection scope of the present application. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for constructing an e-commerce vertical database, characterized in that: include: Obtain user questions in the e-commerce vertical field; Based on the user question and the historical data of the e-commerce vertical field, the answer mining big model is called to instruct the answer mining big model to mine the answer to the user question from the historical data; wherein the user question and the answer form a question-answer pair; and the question-answer pair is used to form an e-commerce vertical field database; The evaluation big model is called according to the question-answer pair to instruct the evaluation big model to identify the question-answer pair with the wrong answer and give an error description label for the wrong answer; wherein the error description label is used as a prompt word for training the answer mining big model, so as to train the answer mining big model by taking the question-answer pair as a negative sample.

2. The method according to claim 1, characterized in that The steps to obtain user questions in the e-commerce vertical include: Obtain original user questions with product inquiry intent from historical data in the e-commerce vertical field; The original user question is rewritten based on the product information corresponding to the original user question to obtain a user question; wherein the user question includes at least part of the product information corresponding to the original user question.

3. The method according to claim 1, characterized in that The historical data includes a communication data set formed by communication data between the user account and the merchant account; According to the user question and the historical data of the e-commerce vertical field, the answer mining big model is called to instruct the answer mining big model to mine the answer to the user question from the historical data, including: The communication data set is preferentially used to call the answer mining model to obtain the answer to the user's question.

4. The method according to claim 3, characterized in that The historical data includes product information for introducing products; The communication data set is preferentially used to call the answer mining model to obtain the answer to the user's question, including: A prompt instruction is constructed according to the communication data set and the product information; wherein the product information serves as an information supplement prompt word for the communication data set.

5. The method according to claim 1, characterized in that According to the user question and the historical data of the e-commerce vertical field, the answer mining big model is called to instruct the answer mining big model to mine the answer to the user question from the historical data, including: Acquire a user question that is semantically similar to the user question; wherein the question-answer pair including the user question is used as the similar question-answer pair; The answer mining big model is called by jointly constructing a prompt instruction based on the user question, the historical data of the e-commerce vertical field and the similar question and answer pairs.

6. The method according to claim 1, characterized in that The historical data includes picture data with product information; The method further comprises: The multimodal large model is called according to the image data to instruct the multimodal large model to generate a plurality of question-answer pairs based on the commodity information of the image data.

7. The method according to claim 1, characterized in that The error description label includes at least one of the following: the knowledge does not match the product, the knowledge answer is wrong, the knowledge answer is irrelevant, the knowledge content is incomplete, or it is not in the reference material.

8. A device for constructing an e-commerce vertical database, characterized in that: include: The acquisition module is used to obtain user questions in the e-commerce vertical field; A first calling module is used to call the answer mining big model according to the user question and the historical data of the e-commerce vertical field, so as to instruct the answer mining big model to mine the answer to the user question from the historical data; wherein the user question and the answer form a question-answer pair; and the question-answer pair is used to form an e-commerce vertical field database; The second calling module is used to call the evaluation big model according to the question-answer pair, so as to instruct the evaluation big model to identify the question-answer pair with the wrong answer, and to give an error description label for the wrong answer; wherein the error description label is used as a prompt word for training the answer mining big model, so as to train the answer mining big model by taking the question-answer pair as a negative sample.

9. A computer device, characterized in that: The computer device includes a memory and a processor, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one computer program, and when the at least one computer program is executed by a processor, the method according to any one of claims 1 to 7 can be implemented.

11. A computer program product, characterized in that The computer program product is used to implement the method according to any one of claims 1 to 7.

Images