E-commerce intelligent question-answering system based on access learning

Abstract

The invention discloses an e-commerce intelligent question-answering system based on access learning, and the system comprises an interaction module which is provided with a first interaction window, a second interaction window and a third interaction window, the first interaction window displays a primary answer automatically generated by a robot, the second interaction window receives an artificial customer service answer and corrects the intention of the primary answer, and the third interaction window displays the primary answer; the third interaction window outputs a final answer after the primary answer and the manual customer service answer are fused to the user terminal; the user intention analysis module is used for analyzing user problems and generating initial intention judgment; the intention auditing module is used for judging the initial intention and providing confirmation of judgment of the initial intention and a weight adjustment control; and the dynamic fusion module is used for carrying out weighted fusion on the robot output answer and the artificial customer service output answer to generate a fusion answer, outputting the fusion answer to a third interaction window, and the training module is used for training the logic of the robot through a manual input result and generating an answer intention library of the robot for user semantics through an intention manually selected in the intention auditing module.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and e-commerce, and particularly to an e-commerce intelligent question-answering system based on access learning. Background Art

[0002] Existing e-commerce intelligent question-answering systems usually rely on a pre-stored knowledge base to answer consumers' questions. This method has some limitations. Firstly, when a consumer's question exceeds the storage scope of the robot's knowledge base, the system cannot provide an accurate answer and can only transfer it to a human customer service, which not only increases the labor cost but also may lead to an extended waiting time for customers. Secondly, current robots mainly rely on a superficial analysis of the text of consumers' questions and lack the ability to deeply understand the true intentions of users. This results in the robot possibly giving an answer according to the literal meaning but actually not truly solving the consumers' problems.

[0003] In addition, existing systems perform poorly in dealing with complex or ambiguous queries. Consumers may use non-standard terms, colloquial expressions or complex queries containing multiple related questions, which all pose challenges to the intelligent question-answering system. At the same time, existing systems often lack the ability of personalization and cannot adjust the tone and content of answers according to the backgrounds, preferences and historical interactions of different users.

[0004] Another problem is that most existing systems lack an effective learning mechanism. They usually rely on regular manual updates to expand the knowledge base and cannot continuously learn and improve from daily interactions, which causes the system to be unable to adapt to new product information, market trends or changes in customer needs in a timely manner. Summary of the Invention

[0005] The purpose of this application is to provide an e-commerce intelligent question-answering system based on access learning, which has the advantages of improving the accuracy of answers, enhancing the ability to understand user intentions, realizing personalized answers, continuously learning and improving, and protecting sensitive information.

[0006] In order to solve the above technical problems, the present invention is solved by the following technical solutions: This application provides an e-commerce intelligent question-answering system based on access learning, and the technical solution is as follows: It includes: an interaction module, which is provided with a first interaction window, a second interaction window, and a third interaction window. The first interaction window displays the primary answer automatically generated by the robot in real time. The second interaction window receives the answer from the human customer service and the intention correction of the primary answer. The third interaction window outputs the final answer after fusing the primary answer and the answer from the human customer service to the user terminal; a user intention analysis module, which analyzes the user's question through a semantic understanding model and generates an initial intention judgment; an intention review module, which is configured in the second interaction window and is used to display the initial answer of the robot and the judgment of the initial intention, and provides confirmation and weight adjustment controls for the initial intention judgment; a dynamic fusion module, which performs weighted fusion on the answer output by the robot and the answer output by the human customer service based on the manually set weight to generate a fused answer output and outputs it to the third interaction window. A training module trains the logic of the robot through the manually input results, and generates an answer intention library for the robot for the user's semantics through the intention selected manually in the intention review module.

[0007] Furthermore, this application also proposes that the dynamic fusion module includes: a writing style weight library, which is configured with multi-dimensional word class feature classification, a writing style feature extraction unit, which analyzes the word usage preference, sentence structure, and emotional tendency in the user's historical conversation records in real time; a weight dynamic calculation unit, which matches the best parameter combination in the writing style weight library based on the current conversation context and user characteristics; a style conversion unit, which converts the fused answer output by the dynamic fusion module into an expression form that conforms to the user's historical writing style and outputs it to the third interaction window.

[0008] Furthermore, this application also proposes that when the interaction module detects that the human customer service fails to complete the response within the specified time or the human customer service is offline, the system automatically starts the primary answer of the robot as a temporary answer.

[0009] Furthermore, this application also proposes that it further includes a supplementary training module. The supplementary training module includes: a temporary record library, which is used to store the temporary answers of the robot; a retrieval unit, when there is no user consultation from the human customer, retrieves the temporary answers in the temporary record library and sends them to the second interaction window, and obtains the answers from the human customer service through the second interaction window and inputs them into the training module.

[0010] Furthermore, this application also proposes that the supplementary training module further includes a background training unit, which generates simulated service samples based on historical question-and-answer data and inputs the historical training samples into the training module.

[0011] Furthermore, the present application also proposes that the supplementary training module further includes: an automatic error correction unit that generates error type tags by comparing the differences between the robot's autonomous answers and the human customer service answers in the temporary record library; an active learning unit that automatically generates reinforcement training samples when it detects that the cumulative number of the same error type tags exceeds a threshold; and an incremental training unit that automatically slices the Q&A records marked by the human customer service according to the time stamp and batches them into the training module.

[0012] Furthermore, the present application also proposes that the system further includes a new customer service training module, including: a historical Q&A knowledge base for storing historical Q&A data; a new customer service training unit that generates a training case set including typical scenarios according to the historical Q&A library, and an evaluation unit that tests the new customer service's intention judgment ability through simulated conversations.

[0013] Furthermore, the present application also proposes that the evaluation unit includes: an intention deviation detection sub-unit that compares the semantic differences between the customer service answer and the standard answer; and an intelligent tutoring sub-unit that dynamically adjusts the question difficulty level according to the semantic difference results.

[0014] Furthermore, the present application also proposes that the dynamic fusion module further includes a sensitive word detection unit for retrieving and marking sensitive words in the answers output by the human customer service, and the style conversion unit that replaces the sensitive words in the answers output by the human customer service with synonyms.

[0015] As can be seen from the above, an e-commerce intelligent Q&A system based on access learning provided by the present application includes an interaction module, a user intention analysis module, an intention review module, a dynamic fusion module, and a training module. The system automatically generates a primary answer through the robot, and the human customer service corrects the intention and answers, and dynamically fuses the two to finally output an answer that meets the user's needs. The system analyzes the user's question through a semantic understanding model, improving the accuracy of the answer; through intention review and dynamic fusion, enhancing the ability to understand the user's intention; through extraction of writing style features and style conversion, realizing personalized answers; through the training module and the supplementary training module, realizing the continuous learning and improvement of the system; through sensitive word detection and replacement, protecting sensitive information. Therefore, the present application has the advantages of improving the accuracy of answers, enhancing the ability to understand the user's intention, realizing personalized answers, continuous learning and improvement, and protecting sensitive information. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a system structure diagram of an e-commerce intelligent Q&A system based on access learning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0018] Embodiment: Existing e-commerce intelligent question answering systems generally answer based on the content stored in the knowledge base. When a consumer's question exceeds the storage scope of the robot's knowledge base, it can only be answered manually. Moreover, current robots can only perform text analysis on consumers' questions and cannot recognize the true intention behind the user's words, resulting in the robot's inability to truly solve consumers' problems based on the word meaning.

[0019] To overcome the limitations of the prior art, the present invention proposes an e-commerce intelligent question answering system based on access learning. The system includes an interaction module, a user intention analysis module, an intention review module, a dynamic fusion module, and a training module. The interaction module is provided with a first interaction window, a second interaction window, and a third interaction window. The first interaction window displays in real time the primary answer automatically generated by the robot. The second interaction window receives the answer from the human customer service and the intention correction of the primary answer. The third interaction window outputs the final answer after fusing the primary answer and the answer from the human customer service to the user terminal. The user intention analysis module analyzes the user's question through a semantic understanding model and generates an initial intention judgment. The intention review module is configured in the second interaction window to display the initial answer of the robot and the judgment of the initial intention, and provides confirmation and weight adjustment controls for the initial intention judgment. The dynamic fusion module performs weighted fusion on the answer output by the robot and the answer output by the human customer service based on the manually set weight to generate a fused answer output and outputs it to the third interaction window. The training module trains the logic of the robot through the manually input results and generates an answer intention library for the robot to respond to the user's semantics through the intention selected manually in the intention review module.

[0020] The e-commerce intelligent question answering system of the present invention solves the problem that the robot in the prior art cannot accurately understand the true intention of the user. Through the user intention analysis module, the system can analyze the user's question and generate an initial intention judgment. The intention review module provides confirmation and weight adjustment controls for the initial intention judgment, ensuring the accuracy of the robot's answer. The dynamic fusion module generates a final answer that better meets the user's needs by weighted fusing the answers of the robot and the human customer service. The training module continuously optimizes the answer ability of the robot through the manually input results and the manual selection in the intention review module.

[0021] The interaction module is provided with three interaction windows. The first interaction window is used to display in real time the primary answer automatically generated by the robot to ensure that users can quickly obtain preliminary information. The second interaction window receives the answer from the human customer service and the intention correction of the primary answer, providing an interface for manual intervention. The third interaction window outputs the fused final answer to the user terminal to improve the user experience.

[0022] The user intention analysis module analyzes the user's question through a semantic understanding model and generates an initial intention judgment. The introduction of this module enables the system to more accurately understand the user's true needs, rather than just performing a superficial analysis of words. The intention review module displays the robot's initial answer and the judgment of the initial intention, and provides confirmation and weight adjustment controls to ensure the accuracy and reliability of the answer.

[0023] The dynamic fusion module performs weighted fusion on the answers of the robot and the human customer service based on the manually set weights. This module ensures the quality of the final answer and can also be dynamically adjusted according to actual needs. The training module continuously optimizes the robot's answer ability through the manually input results and the manual selection in the intention review module, building a system that continuously learns and evolves.

[0024] Compared with the prior art, the e-commerce intelligent question answering system of the present invention has significant advantages. Through the user intention analysis module and the intention review module, the system can more accurately understand the user's needs and provide higher-quality answers. The dynamic fusion module ensures the accuracy and reliability of the final answer by performing weighted fusion on the answers of the robot and the human customer service. The training module improves the overall performance of the system through continuous optimization.

[0025] Through the implementation of the present invention, the e-commerce intelligent question answering system can better understand the user's needs and provide higher-quality services. The user intention analysis module analyzes the user's question and generates an initial intention judgment to ensure that the system can accurately understand the user's needs. The intention review module provides confirmation and weight adjustment controls to ensure the accuracy of the answer. The dynamic fusion module generates a final answer that better meets the user's needs by performing weighted fusion on the answers of the robot and the human customer service. The training module improves the overall performance of the system through continuous optimization. Thus, the present invention effectively solves the problem that the robot in the prior art cannot accurately understand the user's true intention and improves the service quality and user experience of the e-commerce intelligent question answering system.

[0026] Furthermore, the present application also proposes that the dynamic fusion module includes: The dynamic fusion module analyzes the word usage preferences, sentence structures, and emotional tendencies in the user's historical conversation records in real time through a writing style weight library. The weight dynamic calculation unit matches the best parameter combination in the writing style weight library based on the current conversation context and user characteristics. The style conversion unit converts the fused answer output by the dynamic fusion module into an expression form that conforms to the user's historical writing style and outputs it to the third interaction window.

[0027] The writing style weight library is a database for multi-dimensional classification of word class features, used to store and classify users' word usage preferences, sentence structures, and emotional tendencies. The writing style feature extraction unit extracts these features from the user's historical conversation records and updates the writing style weight library in real time. The weight dynamic calculation unit matches the best parameter combination from the writing style weight library according to the context of the current conversation and user characteristics to ensure that the writing style of the answer is consistent with the user's historical writing style. The style conversion unit applies these parameters to the answer content when generating the final answer to make it conform to the user's expression habits and emotional tendencies.

[0028] In this way, the dynamic fusion module can generate answers that are more in line with the user's habits and emotions, improving the user experience and satisfaction. This method not only solves the problem of lack of personalization in the existing robot answers but also enables the answers to be more in line with the user's true intentions and needs.

[0029] Furthermore, this application also proposes that when the interaction module detects that the human customer service fails to complete the response within the specified time or the human customer service is offline, the system automatically activates the primary answer of the robot as a temporary answer.

[0030] In this application, the interaction module determines whether to activate the primary answer of the robot by detecting the online status and response time of the human customer service. When it detects that the human customer service fails to complete the response within the specified time or the human customer service is offline, the system automatically generates and outputs the primary answer of the robot as a temporary answer. This can ensure that the user can still get a timely answer in the case where the human customer service cannot respond in time, improving the user experience.

[0031] Specifically, the interaction module can set a predetermined response time threshold. When the human customer service fails to complete the response within this time, the system automatically triggers the robot to generate a primary answer. This primary answer is generated by the robot according to the pre-set rules and the content of the knowledge base and is immediately output to the user as a temporary answer. In this way, it can effectively avoid the dissatisfaction of users caused by waiting too long and also maintain the continuity and availability of the system in the case where the human customer service is offline.

[0032] Thus, this application realizes the function of automatically activating the primary answer of the robot as a temporary answer when the human customer service cannot respond in time by adding a detection and response mechanism in the interaction module. Compared with the existing technology, this application can better ensure that the user can get a timely answer in any case, improving the intelligence of the system and the user satisfaction.

[0033] Furthermore, the present application also proposes that the supplementary training module includes a temporary record library for storing the temporary answers of the robot, and a retrieval unit. When there is no user consultation by the human customer service, the retrieval unit retrieves the temporary answers in the temporary record library and sends them to the second interaction window, and obtains the answers from the human customer service through the second interaction window and inputs them into the training module.

[0034] The supplementary training module of the present application stores the temporary answers of the robot through the temporary record library, and when there is no user consultation by the human customer service, retrieves these temporary answers and sends them to the second interaction window, so as to obtain the answers from the human customer service and input them into the training module. This method ensures that the robot can still provide temporary answers when the human customer service is offline or fails to respond in time, and through the correction and supplementation of the human customer service subsequently, further improves the quality and accuracy of the robot's answers.

[0035] The temporary record library can be implemented using a relational database or a NoSQL database. The specific selection can be determined according to the system requirements and data volume. The retrieval unit can be implemented through a scheduled task or an event trigger mechanism. When it is detected that there is no user consultation by the human customer service, it automatically retrieves the temporary answers from the temporary record library and sends them to the second interaction window. After obtaining the answers from the human customer service through the second interaction window, the system inputs these answers into the training module to further optimize the answer strategy of the robot.

[0036] Through the above technical means, the present application effectively solves the problem that the robot in the prior art cannot provide effective answers when the human customer service is offline or fails to respond in time, and improves the overall response efficiency and user satisfaction of the intelligent question-and-answer system.

[0037] Furthermore, the present application also proposes that the supplementary training module further includes a background training unit, which generates simulated service samples according to historical question-and-answer data and inputs the historical training samples into the training module.

[0038] This solution uses the background training unit in the supplementary training module to generate simulated service samples using historical question-and-answer data and inputs these samples into the training module for training. In this way, the answer ability of the robot can be continuously optimized and improved, enabling it to more accurately understand the user's intention and provide more effective answers. The introduction of the background training unit enables the system to automatically perform self-optimization and improvement without manual intervention, significantly improving the system's autonomous learning ability and intelligent level.

[0039] Specifically, the background training unit generates simulated service samples based on historical Q&A data. These samples can include various typical user questions and corresponding standard answers. By inputting these simulated service samples into the training module, the system can continuously adjust and optimize its answering model during the training process, thereby improving its performance in actual applications. Thus, the system can better handle various complex and changing user needs and improve user satisfaction.

[0040] Compared with the prior art, the solution of the present application realizes the automated and intelligent training of the system by introducing a background training unit. This can not only reduce the need for manual intervention but also improve the accuracy and effectiveness of the robot's answers by continuously optimizing the training samples, solving the problem in the prior art that the robot cannot accurately understand the user's intention and significantly enhancing the overall performance of the e-commerce intelligent Q&A system.

[0041] Furthermore, the present application also proposes that the supplementary training module further includes an automatic error correction unit that generates error type labels by comparing the differences between the robot's autonomous answers and the human customer service answers in the temporary record library, and an active learning unit that automatically generates enhanced training samples when it detects that the same error type labels exceed a threshold, and an incremental training unit that automatically slices the Q&A records annotated by the human customer service according to timestamps and inputs them into the training module in batches.

[0042] The present application aims to improve the accuracy and intelligence level of the robot's answers by adding an automatic error correction unit, an active learning unit, and an incremental training unit. The automatic error correction unit generates error type labels by comparing the differences between the robot's answers and the human customer service answers, thereby identifying common errors in the robot's answers. The active learning unit automatically generates enhanced training samples when it detects that the same error type labels exceed a preset threshold and inputs them into the training module for further training of the robot. The incremental training unit slices the Q&A records according to timestamps and inputs them into the training module in batches to continuously improve the robot's answering ability.

[0043] The automatic error correction unit can analyze the differences between the robot's answers and the human customer service answers through natural language processing technology and generate specific error type labels. For example, the error type labels can include semantic errors, grammar errors, sentiment tendency errors, etc. The active learning unit can automatically generate enhanced training samples through machine learning algorithms. When it detects that a specific error type label exceeds a preset threshold, the system will automatically generate a new set of training samples and input them into the training module for reinforcement learning. The incremental training unit can slice the Q&A records by timestamps and input them into the training module in batches to ensure the timely update of the training data and the gradual improvement of the robot's answering ability.

[0044] By introducing an automatic error correction unit, an active learning unit, and an incremental training unit, this application realizes the dynamic monitoring and continuous improvement of the quality of the robot's answers. Compared with the prior art, this application can more effectively identify and correct errors in the robot's answers, and continuously optimize the robot's answering ability through an active learning mechanism, thereby improving the overall performance of the intelligent Q&A system and user satisfaction.

[0045] Furthermore, this application also proposes a new customer service training module, including a historical Q&A knowledge base for storing historical Q&A data; a new customer service training unit for generating a training case set containing typical scenarios based on the historical Q&A library; and an evaluation unit for testing the new customer service's intention judgment ability through simulated conversations.

[0046] The technical solution of the new customer service training module aims to improve the response ability and user intention judgment ability of new customer service through a systematic training and evaluation mechanism. The historical Q&A knowledge base stores a large amount of historical Q&A data, which serves as an important resource for the training of new customer service. The new customer service training unit generates a training case set containing typical scenarios based on these data. These case sets can cover common and complex user questions, helping new customer service quickly master response strategies and skills. The evaluation unit tests the new customer service's intention judgment ability through simulated conversations to ensure that the new customer service can accurately understand the true intention of the user and provide appropriate answers.

[0047] The historical Q&A knowledge base in the new customer service training module can be updated and maintained regularly to ensure the timeliness and accuracy of its content. The new customer service training unit can adopt various forms of training methods, such as online courses, scenario simulations, etc., to further improve the training effect. The evaluation unit can score and provide feedback on the simulated conversations of the new customer service to help the new customer service continuously improve and enhance their service level.

[0048] By introducing the new customer service training module, this application can effectively solve the problem of insufficient training of new customer service existing in the prior art. Compared with traditional training methods, the solution of this application is more systematic and comprehensive, which can help new customer service adapt to the working environment faster, improve service quality and user satisfaction. Therefore, the new customer service training module has significant advantages in improving the comprehensive ability and service level of new customer service.

[0049] Furthermore, this application also proposes that the evaluation unit includes an intention deviation detection subunit and an intelligent tutoring subunit.

[0050] The intention deviation detection subunit detects the deviation existing in the customer service's answer by comparing the semantic differences between the customer service's answer and the standard answer. The intelligent tutoring subunit then dynamically adjusts the difficulty level of the questions according to the semantic difference results of the intention deviation detection subunit to better train new customer service.

[0051] The role of the intention deviation detection subunit is to ensure that new customer service representatives can accurately understand the true intentions of users and provide correct answers. By comparing the customer service answers with the standard answers, deficiencies in the new customer service representatives' understanding of user intentions can be identified, enabling targeted improvement. The intelligent tutoring subunit then dynamically adjusts the difficulty of training questions based on these differences, enabling new customer service representatives to gradually improve the accuracy of their answers and their ability to judge intentions.

[0052] Thus, through the cooperation of the intention deviation detection subunit and the intelligent tutoring subunit, the training effect of new customer service representatives can be effectively improved, ensuring that they can accurately judge user intentions and provide high-quality services in actual work. Compared with the prior art, this application solves the problem of the inability to accurately judge user intentions in existing e-commerce intelligent question-and-answer systems, improving the pertinence and effectiveness of customer service training.

[0053] Furthermore, this application also proposes that the dynamic fusion module further includes a sensitive word detection unit for retrieving and marking sensitive words in the answers output by human customer service representatives, and a style conversion unit for replacing the sensitive words in the answers output by human customer service representatives with synonyms.

[0054] The sensitive word detection unit can automatically identify sensitive words in the answers of human customer service representatives and mark them. When the style conversion unit receives the marked answer, it will replace the marked sensitive words with appropriate synonyms according to a preset synonym library, thereby generating an answer that meets the user's needs and does not contain sensitive words. Specifically, the sensitive word detection unit can adopt natural language processing-based technologies to identify sensitive words by segmenting and matching the answer content, and the style conversion unit can perform the replacement operation of sensitive words by searching for matching items in the synonym library.

[0055] The combined use of the sensitive word detection unit and the style conversion unit can effectively prevent the appearance of sensitive words in the answers, enhancing the security of the system and the user experience. Through these technical means, this application can not only provide answers that conform to user habits but also ensure the security and compliance of the answer content. Thus, compared with the prior art, the solution of this application can better identify and handle sensitive words when processing user inquiries, avoiding potential risks caused by sensitive words and improving the intelligence and security of the system.

[0056] The technical solution of this application further improves the intelligence and security of the system by adding a sensitive word detection unit and a style conversion unit to the dynamic fusion module. The sensitive word detection unit can retrieve and mark sensitive words in the answers output by human customer service representatives to ensure the security and compliance of the answer content, and the style conversion unit replaces the marked sensitive words with synonyms, making the final output answer conform to the user's writing style preference and avoiding the appearance of sensitive content.

[0057] For example, the sensitive word detection unit can perform real-time analysis and comparison on the answers of the human customer service through a pre-set sensitive word library. Once a sensitive word is detected, the system will automatically mark it and transmit the marked information to the style conversion unit. The style conversion unit will then select appropriate synonyms for replacement based on the context and the user's historical word usage preferences, so as to generate an answer that conforms to the user's habits and is safe.

[0058] Thus, the technical solution proposed in this application not only solves the problem in the prior art that robots cannot recognize the true intentions of users, but also enhances the security and user experience of the system through the sensitive word detection and replacement functions. Compared with the prior art, the solution of this application has significantly improved in terms of intelligence and security.

Claims

1. An e-commerce intelligent question-answering system based on access learning, comprising: An interaction module, which is provided with a first interaction window, a second interaction window, and a third interaction window. The first interaction window displays the primary answer automatically generated by the robot in real time. The second interaction window receives the answer from the human customer service and the intention correction of the primary answer. The third interaction window outputs the final answer after fusing the primary answer and the answer from the human customer service to the user terminal. A user intention analysis module, which analyzes the user's question through a semantic understanding model and generates an initial intention judgment. An intention review module, which is configured in the second interaction window, is used to display the initial answer of the robot and the judgment of the initial intention, and provides confirmation and weight adjustment controls for the initial intention judgment. A dynamic fusion module, which performs weighted fusion on the answer output by the robot and the answer output by the human customer service based on the manually set weight to generate a fused answer output. A training module, which trains the logic of the robot through the manually input results, and generates an answer intention library for the robot to respond to the user's semantics through the intention selected manually in the intention review module.

2. The e-commerce intelligent question-answering system based on access learning according to claim 1, characterized in that: The dynamic fusion module includes: A writing style weight library, which is configured with multi-dimensional word class feature classifications. A writing style feature extraction unit, which analyzes the word usage preference, sentence structure, and emotional tendency in the user's historical conversation records in real time. A weight dynamic calculation unit, which matches the best parameter combination in the writing style weight library based on the current conversation context and user characteristics. A style conversion unit, which converts the fused answer output by the dynamic fusion module into an expression form that conforms to the user's historical writing style and outputs it to the third interaction window.

3. The e-commerce intelligent question-answering system based on access learning according to claim 1, characterized in that: When the interaction module detects that the human customer service fails to complete the response within the specified time or the human customer service is offline, the system automatically starts the primary answer of the robot as a temporary answer.

4. The e-commerce intelligent question answering system based on access learning according to claim 3, wherein: It further includes a supplementary training module, and the supplementary training module includes: A temporary record library, which is used to store the temporary answers of the robot. A retrieval unit, when there is no user consultation from the human customer, retrieves the temporary answers in the temporary record library and sends them to the second interaction window. Obtains the answer from the human customer service through the second interaction window and inputs it into the training module.

5. The e-commerce intelligent question answering system based on access learning according to claim 4, characterized in that: The supplementary training module further includes a background training unit, and the background training unit generates simulated service samples according to the historical Q&A data and inputs the historical training samples into the training module.

6. Referring to claim 4, characterized in that: The supplementary training module further includes: An automatic error correction unit, which generates error type labels by comparing the differences between the robot's autonomous answers and the answers from the human customer service in the temporary record library. An active learning unit, when it detects that the cumulative number of the same error type labels exceeds the threshold, automatically generates enhanced training samples. An incremental training unit, which automatically slices the Q&A records marked by the human customer service according to the time stamp and inputs them into the training module in batches.

7. An e-commerce intelligent question answering system based on access learning according to claim 1, characterized in that: The system further includes a new customer service training module, including: A historical Q&A knowledge base, which is used to store historical Q&A data. A new customer service training unit, which generates a training case set containing typical scenarios according to the historical Q&A library. An evaluation unit, which tests the new customer service's intention judgment ability through simulated conversations.

8. An e-commerce intelligent question-answering system based on access learning according to claim 7, characterized in that: The evaluation unit includes: An intention deviation detection sub-unit, which compares the semantic differences between the customer service's answer and the standard answer. An intelligent tutoring sub-unit, which dynamically adjusts the difficulty level of the questions according to the semantic difference results.

9. The e-commerce intelligent question-answering system based on access learning according to claim 1, wherein: The dynamic fusion module further includes: A sensitive word detection unit for retrieving and marking sensitive words in the answers output by the artificial customer service; The style conversion unit replaces the sensitive words in the answers output by the artificial customer service with synonyms.

Images