An item attribute question answering method and system

Abstract

The application discloses an article attribute question and answer method and system, and relates to the technical field of computers.A specific implementation of the method comprises: receiving operation information of a user on a live broadcast room, and determining a target article corresponding to the operation information; wherein the operation information comprises an operation mode and input questions related to attributes of the target article; obtaining candidate knowledge information corresponding to the target article from an article attribute question and answer knowledge base, and calculating the similarity between the questions and each candidate knowledge information; determining target knowledge information most matched with the questions based on the similarity, the questions and current background information of the live broadcast room; generating answers to the questions based on the target knowledge information, and displaying the answers in the live broadcast room.The implementation not only improves the semantic understanding and scene adaptation capability of the article attribute question and answer system, but also enhances the accuracy and context relevance of the answers, thereby effectively improving the user experience and live broadcast interaction efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for answering questions about the attributes of items. Background Technology

[0002] With the rapid development of e-commerce live streaming, product attribute Q&A systems have become a key tool for enhancing the user shopping experience. In live streaming rooms, the types of products are becoming increasingly diverse (products are just one type of item; for example, "toothbrush" and "mobile phone" may appear in the same live streaming room), and consumers' demand for real-time understanding of product details is constantly growing. Traditional product recommendation and information acquisition methods are no longer sufficient to meet this demand for high real-time and strong interactivity.

[0003] Existing e-commerce digital human live streaming room item attribute Q&A systems mainly rely on deep learning models or rule-based methods. Furthermore, the construction of the item attribute Q&A knowledge base currently depends primarily on manual configuration by operations personnel. This results in poor real-time performance, high maintenance costs, incomplete coverage, and insufficient intelligence in the item attribute Q&A system, thereby affecting the user experience. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and system for answering questions about item attributes, which can at least solve the problems of poor real-time performance, high maintenance costs, incomplete coverage and insufficient intelligence level in existing item attribute question-and-answer systems.

[0005] To achieve the above objectives, according to one aspect of the present invention, a question-and-answer method for item attributes is provided, comprising:

[0006] Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item;

[0007] From the item attribute question and answer knowledge base, obtain candidate knowledge information corresponding to the target item, and calculate the similarity between the question and each candidate knowledge information;

[0008] Based on the similarity, the question, and the current background information of the live stream, the target knowledge information that best matches the question is determined.

[0009] Based on the target knowledge information, an answer to the question is generated and displayed in the live broadcast room.

[0010] To achieve the above objectives, according to one aspect of the present invention, an item attribute question-answering system is provided, including a knowledge mining agent and an item attribute question-answering agent:

[0011] A knowledge mining agent is used to build a knowledge base for question-and-answer questions about item attributes.

[0012] An intelligent agent for answering questions about item attributes, used for:

[0013] Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item;

[0014] From the item attribute question and answer knowledge base, obtain candidate knowledge information corresponding to the target item, and calculate the similarity between the question and each candidate knowledge information;

[0015] Based on the similarity, the question, and the current background information of the live stream, the target knowledge information that best matches the question is determined.

[0016] Based on the target knowledge information, an answer to the question is generated and displayed in the live broadcast room.

[0017] To achieve the above objectives, according to another aspect of the present invention, an electronic device for answering questions about item attributes is provided.

[0018] The electronic device of this invention includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement any of the above-described item attribute question-and-answer methods.

[0019] To achieve the above objectives, according to another aspect of the present invention, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the above-described item attribute question-and-answer methods.

[0020] To achieve the above objectives, according to another aspect of the present invention, a computing program product is provided. One such computing program product includes a computer program that, when executed by a processor, implements the item attribute question-and-answer method provided in the present invention.

[0021] According to the solution provided by the present invention, one embodiment of the invention has the following advantages or beneficial effects: by combining user operation information in the live broadcast room to accurately identify target items, and based on the similarity matching between the question and candidate knowledge information in the knowledge base, and then integrating the real-time background information of the live broadcast room to make a final judgment, the complexity problem of item attribute question and answer in live broadcast scenarios is effectively solved. This solution, through multi-dimensional information fusion and large-model reasoning, not only improves the accuracy of question and answer matching, but also generates answers that help users make shopping decisions in complex environments, thereby solving the problems of "poor real-time performance, high maintenance costs, incomplete coverage, and insufficient intelligence" in existing technologies, and significantly improving the efficiency of live broadcast interaction and user experience.

[0022] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0023] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein:

[0024] Figure 1 This is a schematic diagram of the main process for generating a knowledge base according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the system architecture of a knowledge mining agent;

[0026] Figure 3 This is a flowchart illustrating an item attribute question-and-answer method according to an embodiment of the present invention;

[0027] Figure 4(a) is a schematic diagram of the process of a user opening the live shopping bag and clicking on the question option for a specific item in the live room;

[0028] Figure 4(b) is a schematic diagram of the system architecture of the item attribute question-and-answer agent;

[0029] Figure 5 This is a flowchart illustrating an optional question-and-answer method for item attributes according to an embodiment of the present invention;

[0030] Figure 6 This is a schematic diagram of the system architecture of the user simulator Agent;

[0031] Figure 7 This is a flowchart illustrating another optional question-and-answer method for item attributes according to an embodiment of the present invention;

[0032] Figure 8(a) is a schematic diagram of the system architecture for automatically evaluating agents;

[0033] Figure 8(b) is a schematic diagram of the overall process of the knowledge base iteration agent;

[0034] Figure 9(a) is a schematic diagram of the question-and-answer process for item attributes according to an embodiment of the present invention;

[0035] Figure 9(b) is a schematic diagram of the question-and-answer architecture for item attributes according to an embodiment of the present invention;

[0036] Figure 10 This is a schematic diagram of the main modules of an item attribute question-and-answer system according to an embodiment of the present invention;

[0037] Figure 11 This is an exemplary system architecture diagram in which embodiments of the present invention can be applied;

[0038] Figure 12 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention, such as a mobile device or server. Detailed Implementation

[0039] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0040] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0041] Where there is no conflict, the embodiments and features in the embodiments of this invention can be combined with each other. The acquisition, transmission, storage, use, and processing of data in the technical solutions of this invention comply with the relevant provisions of national laws and regulations, are used for legal and reasonable purposes, and are not shared, disclosed, or sold outside of these legal uses, and are subject to supervision and management by regulatory authorities.

[0042] Regarding user information, necessary measures should be taken to prevent unauthorized access to such personal information data, ensure that personnel authorized to access such data comply with relevant laws and regulations, and safeguard the security of user personal information. Once this user personal information data is no longer needed, risks should be minimized by restricting or even prohibiting data collection and / or deleting the data. Where applicable, including in certain relevant applications, user privacy should be protected through data de-identification, such as by removing specific identifiers (e.g., date of birth), controlling the amount or specificity of stored data (e.g., collecting location data at the city level rather than the specific address level), controlling how data is stored, and / or other de-identification methods.

[0043] In e-commerce live streaming scenarios, item attribute Q&A systems aim to improve user purchase decision efficiency and satisfaction by analyzing factors such as "user comments," "candidate items in the live stream shopping bag and their various attributes," and "background information of the live stream." However, existing technologies have the following drawbacks:

[0044] First, the construction of the item attribute Q&A knowledge base is highly dependent on manual operation, which requires a large investment of manpower and is inefficient.

[0045] Secondly, in item attribute question-answering systems, traditional deep learning models can provide the top item attribute that best matches the user's question and a fixed response script through "recall and ranking". However, traditional deep learning models are mostly small-scale models with limited semantic understanding capabilities. When dealing with complex and varied user questions in e-commerce live streaming scenarios, especially in multi-turn dialogues and situations requiring high contextual coherence, they are inadequate and prone to problems such as misunderstandings or stiff responses.

[0046] Third, in the early stages of the item attribute question and answer system, due to the limited accumulation of online dialogue data, how to simulate user behavior in a cold start scenario to generate more high-quality dialogue data in order to quickly iterate the online model is an urgent problem to be solved.

[0047] Fourth, current evaluation methods mainly rely on manual annotation, which is time-consuming and labor-intensive. How to conduct large-scale item attribute question-and-answer dialogue evaluation quickly and at low cost, thereby efficiently iterating online item attribute question-and-answer system models and knowledge bases, is also a current technological challenge.

[0048] Fifth, building a knowledge base is only the initial stage. Its subsequent continuous iteration and optimization are equally crucial. It is necessary to ensure the accuracy, completeness, and unambiguity of the content in the knowledge base in order to support the long-term stable operation of the item attribute question and answer system.

[0049] To address the five types of problems mentioned above, this solution proposes a multi-agent collaborative architecture for an item attribute question-answering system, comprising: a knowledge mining agent, an item attribute question-answering agent (such as a customer service robot), a user simulator agent, an automatic evaluation agent, and a knowledge base iteration agent. By integrating large-scale model capabilities with multi-agent technology, the item attribute question-answering system can not only quickly respond to users' personalized inquiries about item attributes but also provide accurate question-answering services in complex and diverse multi-category item environments. Simultaneously, a closed-loop collaboration mechanism is formed among the agents, enabling continuous updates, optimizations, and automatic iterations of the knowledge base and model. This significantly improves the system's adaptability and operational efficiency, ultimately constructing an item attribute question-answering system with self-evolving capabilities.

[0050] See Figure 1 The diagram shows a main flowchart of a knowledge base generation method provided by an embodiment of the present invention, which includes the following steps:

[0051] S101: Obtain item information related to the items in the live broadcast room, and extract key features from the item information;

[0052] S102: Based on the key features, predict the questions that users will ask about the attributes of the item during the live broadcast, and determine the item attributes corresponding to the predicted questions;

[0053] S103: Generate an answer based on the item information, the predicted question, and the item attributes corresponding to the predicted question;

[0054] S104: Based on the predicted question, the generated answer, the item information, and the determined item attributes, generate knowledge information and store it in the item attribute question and answer knowledge base, and at the same time establish the association between the knowledge information and the item identifier of the item.

[0055] This implementation describes how to construct an item attribute question-and-answer knowledge base based on a "knowledge mining agent," thereby addressing the problem that the construction of existing item attribute question-and-answer knowledge bases is highly dependent on manual operation, requiring a large investment of manpower and resulting in low efficiency. Traditional knowledge base construction methods are limited by labor and time costs, making it difficult to quickly respond to the massive and ever-changing item information demands in e-commerce scenarios. The "knowledge mining agent," however, leverages the advantages of large-scale models in semantic understanding, reasoning, and reflection to automatically extract and generate high-quality, diverse, and coherent item attribute question-and-answer pairs (QAs) from unstructured, multi-source, and heterogeneous item information. This effectively alleviates the reliance on manual resources and improves the efficiency and quality of knowledge base construction.

[0056] In applications with high real-time requirements, such as e-commerce live streaming, there is a vast amount of product information. This information is diverse in form and format, constituting a multi-source, heterogeneous source of product information. The "Knowledge Mining Agent" leverages the powerful semantic understanding and summarization capabilities of large-scale models to integrate this multi-source, heterogeneous, "unstructured" information into a "structured" text knowledge base. (See [link to relevant documentation]). Figure 2 As shown. This structured text knowledge base not only makes knowledge information clearer and easier to use, but also improves the readability and retrieval of knowledge information. It also provides high-quality information support for subsequent item attribute question-and-answer agents, enabling them to provide more accurate and efficient answers when faced with user questions about item attributes.

[0057] In the constructed item attribute question-and-answer knowledge base, each piece of knowledge consists of four parts: question (optional), answer (optional), item title (required), and item attribute (required). For item attributes with configured question-and-answer pairs, all four parts can be presented; for item attributes without configured question-and-answer pairs, at least the item title and item attribute should be included. Based on this, the "knowledge mining agent" can further expand the questions and answers through automated means, thereby enriching the content of the entire knowledge information.

[0058] This solution, the "Knowledge Mining Agent," employs a four-stage prompting scheme: "CoT + few-shot + Self-Refine" (four-stage prompting for new knowledge mining). Through layer-by-layer filtering, high-quality QA pairs are obtained and ultimately stored in the item attribute question-and-answer knowledge base. CoT (Chain of Thought) guides the model's logical reasoning, few-shot provides a small number of examples to enhance the generation effect, and Self-Refine continuously improves the quality of the generated results through an iterative feedback mechanism. This approach organically combines multiple existing technologies to form a data flywheel-like system architecture. As data accumulates and iterates, the overall performance and output quality of the system will continuously improve.

[0059] Furthermore, the "knowledge mining agent" can be further divided into two specific implementation schemes: "new knowledge mining" and "existing knowledge expansion" (see below). Figure 5 (As described in the image). The core difference between the two is that "existing knowledge expansion" incorporates "existing QA pairs in the knowledge base under the current attribute dimension" as reference information into the input. This allows for better consistency and complementarity when generating new question-answer pairs, avoiding content duplication and improving knowledge coverage. Through this mechanism, the system can achieve continuous evolution and improvement of the knowledge base while ensuring knowledge diversity.

[0060] The specific execution process is described here:

[0061] Phase one is the question generation phase, implemented using a few-shot strategy. This phase can be executed by a live-streaming question generation expert, such as a Large Language Model (LLM), which possesses information integration capabilities, predictive analysis skills, and language expression abilities. By integrating multi-source heterogeneous product information, key features are extracted from these sources, and based on this, questions that users might ask about product attributes during the live stream are predicted. Generating these questions not only helps the live streamer prepare corresponding answers in advance but also effectively enhances the audience's interactive experience, improves live-streaming interaction efficiency and information delivery quality, thereby further optimizing communication quality and conversion rates in e-commerce live-streaming scenarios.

[0062] The generated questions must be strictly based on the provided item information to ensure their relevance and accuracy. Item information includes, but is not limited to, the item's livestream script, item title (including item details), item OCR (Optical Character Recognition) information, and current item attribute descriptions. The output format is uniformly: [{"Q":""}, ..., {"Q":""}].

[0063] As an optimized implementation method, "extracting key features from item information sources" specifically includes: analyzing the four provided inputs: 1) extracting the overall style of the live stream and the key item attributes from the item live stream script; 2) identifying information related to the item transaction from the item title, such as the item's core selling points and benefits (e.g., discount information); 3) obtaining detailed technical specifications and features from the item's OCR information, such as identifying specific specifications and item features using OCR technology on item screenshots; and 4) extracting the item's unique advantages (e.g., unique attributes) and details that users may care about from the item's current attribute description. Based on one or more of the above information, and preferably more, a series of questions that users may ask are predicted and listed.

[0064] Assume the input includes the following four elements: 1) Item livestream script: A livestream script introducing a new smartphone, emphasizing its high-speed processor and long battery life. 2) Item title: XX brand's new generation smartphone. 3) Item OCR information: Including processor model, memory size, screen resolution, etc. 4) Item current attribute description: Emphasizing camera functionality. The output question is: [{"Q": "What new features does this phone's camera have?"}, {"Q": "What is the pixel count of this phone's camera?"},...].

[0065] Phase two is the issue verification phase, implemented using a CoT + few-shot + Self-refine strategy. This phase can be executed by live-stream issue generation and verification experts, such as large-scale modeling (LLM), who possess keen insight, rigorous analytical skills, comparison capabilities, logical thinking, and attention to detail. Their core task is to systematically evaluate and verify the issues generated in Phase one, ensuring that the generated issues accurately reflect the current attributes of the items and maintain a high degree of consistency with the provided input, thereby guaranteeing the relevance, accuracy, and completeness of the issues. It should be noted that the evaluation process must be strictly based on the item information provided by the user and must not introduce any external knowledge or subjective judgment.

[0066] Specifically, each generated question is analyzed one by one. First, it is determined whether the question is closely related to the item's attributes. Second, the question is compared with the input content to verify whether the question truly and accurately reflects the item's characteristics and advantages. Finally, based on the evaluation results, a result indicating whether the question meets the requirements is given, along with specific analytical reasons. If the evaluation result is negative, the question is further modified to better meet the requirements for replacement. The output format should be consistent as follows: {"Evaluation Result": "Yes / No", "Reason": "Specific Analysis", "Modified Output": "If the evaluation result is negative, a more suitable question is provided"}.

[0067] Using the input from Phase 1 as input, let's assume the output question is {"Q": "How waterproof is this phone?"}. The evaluation information includes: {"Evaluation Result": "No", "Reason": "The generated question is inconsistent with the current attribute description of the item. The current attribute description emphasizes the camera function, while the question focuses on waterproof performance, which is irrelevant to the input content.", "Modified Output": "What new features does this phone's camera have?"}.

[0068] Phase three is the answer generation phase, implemented using a few-shot strategy. This phase can be executed by live-stream answer generation experts, such as large-scale modeling (LLM), who possess information integration capabilities, creative writing skills, and language expression abilities. Their core task is to generate accurate, relevant, and engaging answers based on the item information and the generated question. These answers will be directly applied in the live-stream scenario to enhance the audience's interactive experience and improve the effectiveness of item information delivery. It should be noted that the generated answers must be created based on the item information, ensuring that the content is authentic, accurate, and consistent with reality. The output format is: {"Question": "", "Answer": ""}.

[0069] Specifically, the answer generation process requires first thoroughly reading and understanding the item information. Then, based on the predicted questions, the specific item attributes that the audience is interested in are identified. Building on this, the answer is generated by combining the item information with the predicted questions, ensuring it accurately conveys the item information while being engaging and appealing. This ensures the answer is easy to understand, logically clear, and effectively stimulates audience interest and trust. Furthermore, since the core purpose of the item is marketing, the content planning for the live stream must closely revolve around the item's attributes.

[0070] Using the input from Phase 1 as input, assuming the question is "What new features does this phone's camera have?", the output answer would be: {"Question": "What new features does this phone's camera have?", "Answer": "This phone's camera is equipped with the latest XX technology, supports high-definition shooting up to XX pixels, and also has multiple functions such as night mode, AI scene recognition, and professional mode, allowing you to easily take high-quality photos whether it's day or night."}

[0071] Phase four is the answer verification phase, also implemented using the CoT + few-shot + Self-refine strategy. This phase can be performed by live answer generation verification experts, such as large-scale model LLM, who possess keen insight, rigorous analytical skills, comparison capabilities, logical thinking, and attention to detail. Their core task is to systematically evaluate and verify the answers generated in Phase three. Through this process, it is ensured that the generated answers accurately reflect the item attributes in content and maintain a high degree of consistency with the provided input information and predicted questions, thereby guaranteeing the accuracy, relevance, and marketing effectiveness of the answers. It should be noted that the entire evaluation process must be strictly based on the item information provided by the user and must not introduce any external knowledge or subjective judgment.

[0072] Specifically, each generated answer is analyzed one by one. First, it is determined whether the answer is closely related to the item's attributes. Second, the correspondence between the answer and the input information and question is compared to verify whether it truly and comprehensively reflects the item's characteristics, advantages, and related information. Finally, based on the evaluation results, a result indicating whether the answer meets the requirements is given, along with specific reasons. If the evaluation result is negative, the answer is further modified to better meet the requirements for replacement. The output format is: {"Evaluation Result": "Yes / No", "Reason": "Specific Analysis", "Modified Output": "If the evaluation result is negative, a more suitable answer is provided"}.

[0073] Using the input from Phase 1 as input, let's assume the output question and answer are: {"Question": "What new features does this phone's camera have?", "Answer": "This phone's camera is waterproof and can be used for underwater photography."}. The evaluation information includes: {"Evaluation Result": "No", "Reason": "The generated answer is inconsistent with the item's current attribute description. The description emphasizes features such as high-resolution pixels, night mode, AI scene recognition, and professional mode, while the answer focuses on waterproofing, which is irrelevant to the input.", "Modified Output": "This phone's camera is equipped with the latest XX technology, supporting high-resolution shooting up to XX pixels. It also features night mode, AI scene recognition, and professional mode, allowing you to easily take high-quality photos day or night."}.

[0074] All four stages described above can be accomplished using a Large Language Model (LLM). The system can flexibly choose how to use the model during implementation: a single large model can be used throughout all stages, or separate large models can be configured for different stages to achieve specialized processing; alternatively, some stages can share a single model, while other stages are handled by different models. Which model executes each stage does not affect the overall workflow's functionality; this solution does not impose restrictions on this, and adjustments can be made based on computing resources, performance requirements, or business scenarios during actual deployment.

[0075] After generating questions and answers and performing verification and optimization, this solution can generate knowledge information based on the predicted questions, generated answers, item titles in the item information, and item attributes corresponding to the predicted questions, and store this information in an item attribute question-and-answer knowledge base. For different items, such as XX and YY15, associations can be established between their item identifiers and different knowledge information, as shown in the following example:

[0076] 1. Item title: XX || Item attribute name: Color (QA pair not configured);

[0077] 2. Item Title: XX || Item Attribute Name: Price (QA pair not configured);

[0078] 3. Question: xx || Answer: xx || Item Title: XX || Item Attribute Name: 5G Network;

[0079] 4. Question: xx || Answer: xx || Item Title: YY || Item Attribute Name: Fast Charge.

[0080] Taking XX as an example, associations can be established with knowledge information 1, 2, and 3 based on its item identifier. These knowledge information are generated based on different item attributes. For color and price, no QA pairs are configured, so knowledge information is generated only based on the item title and item attributes. However, for the 5G network attribute, a QA pair is configured, and knowledge information is generated based on the question, answer, item title, and item attributes. For YY15, the association with the fourth knowledge information is established only based on its item identifier, indicating that the knowledge information corresponding to this item under the current attribute configuration is relatively limited. This demonstrates that different items can generate or be associated with different structured knowledge information based on their own attribute characteristics, reflecting the flexibility and adaptability of the knowledge mining agent when handling multiple categories of items.

[0081] The method provided in the above embodiments can efficiently generate question-and-answer pairs that match the attributes of the items by automatically extracting key features from the item information and predicting the questions that users may ask, and construct structured knowledge information. This not only significantly improves the efficiency and coverage of knowledge base construction, but also reduces the cost of manual annotation. At the same time, the generated knowledge content is close to the real user needs, which helps to improve the accuracy and practicality of the subsequent question-and-answer system.

[0082] See Figure 3 The diagram illustrates a flowchart of an item attribute question-and-answer method according to an embodiment of the present invention, including the following steps:

[0083] S301: Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item;

[0084] S302: Obtain candidate knowledge information corresponding to the target item from the item attribute question and answer knowledge base, and calculate the similarity between the question and each candidate knowledge information;

[0085] S303: Based on the similarity, the question, and the current background information of the live stream, determine the target knowledge information that best matches the question;

[0086] S304: Generate an answer to the question based on the target knowledge information, and display the answer in the live broadcast room.

[0087] This implementation describes how to intelligently answer user questions about item attributes in e-commerce scenarios such as live streaming, based on an "item attribute question-answering agent." Addressing the limitations of traditional small deep learning models in practical applications, such as "limited semantic understanding capabilities" and "prone to misunderstandings or awkward responses," this solution utilizes a large model approach. Leveraging its powerful semantic understanding and natural language dialogue capabilities, it can quickly locate the top-ranked knowledge information most relevant to the user's question and generate fluent, natural, and targeted answers based on context, thereby significantly improving the user's interactive experience and satisfaction.

[0088] For step S301, unlike the item attribute Q&A in a shopping guide scenario, a live-stream shopping bag typically contains hundreds of items, each with dozens or even hundreds of attributes. User questions are often vague or do not explicitly point to a specific item, making it difficult to accurately match the specific attributes of a particular item. This solution designs a rule-based method for identifying item SKUs (StockKeeping Units) in a live-stream scenario to receive and parse user operation information to determine the target item the user is currently asking about; wherein, the operation information includes, but is not limited to, questions related to the attributes of the target item. This method supports the following three identification methods:

[0089] Method 1: As shown in Figure 4(a), when a user opens the live shopping bag in the live room and clicks on a specific item (i.e., the first item), the icon area of ​​that specific item will display a "question option." After clicking this option, the user can enter a specific question. At this time, the rule-based item identification method will, upon receiving the user's question, clearly identify the specific item clicked by the user as the target item, achieving accurate item positioning and effectively narrowing the scope of the knowledge base recall candidate.

[0090] Method 2: When a user enters a question in the live stream chat box (as shown in Figure 4(a)), and the question contains explicit item name information (such as "XX"), keyword extraction and precise matching can be performed using rule-based methods. The rule-based item identification method uses the item in the live stream that corresponds to the item name information (i.e., the second item) as the target item, thereby improving the accuracy and response efficiency of item retrieval.

[0091] Method 3: When the user does not actively select an item, and the input information does not contain a specific item name, or contains a specific item name but there is no item in the live stream corresponding to that name, the rule-based item identification method uses all items in the live stream as the target item and performs a full knowledge base retrieval. Although this method has a large candidate range, combined with the large-model semantic capabilities of the item attribute question-answering agent, it can still filter out the most relevant information for response from a wide range.

[0092] In summary, the three methods described above are applicable to different user behavior scenarios. Methods one and two achieve accurate item identification through rule-based means, which helps improve the response accuracy and efficiency of the item attribute question-and-answer system. Method three, on the other hand, can still ensure question-and-answer coverage through a full search even when there is no clear item reference.

[0093] For step S302, after identifying the target item and clarifying the recall scope of the item attribute question-and-answer knowledge base, the "Item Attribute Question-and-Answer Agent" obtains candidate knowledge information corresponding to the target item. Using vector recall technology, it calculates the text vector similarity between the user's question and each candidate knowledge piece, selecting the top 50 to 70 candidate knowledge pieces with the highest similarity. Specifically, the system will filter the corresponding candidate knowledge information from the knowledge base based on the target item's item identifier (e.g., SKU) for matching calculation. Since the target item has already been determined in the item identification stage as either a specific item or all items in the live stream, the recall process here can be divided into "partial recall" (for a specific item) or "full recall" (for all items in the live stream).

[0094] Referring to Figure 4(b), when the user inputs the question "Does XX support 5G networks?", the system identifies the target item as "XX" through rule recognition and then only retrieves knowledge information associated with "XX" from the knowledge base as candidate knowledge information. However, if the user inputs "Does XY support 5G networks?", since "XY" is not a standard item name, the rule recognition fails to match a specific item SKU. In this case, the system includes all items displayed in the live stream in the target scope for full retrieval.

[0095] For step S303, in the previous steps, vector recall technology has been used to filter out the top 50 to 70 candidate knowledge information with the highest similarity to the user's question. The core of this step is how to accurately identify the unique target knowledge information that best matches the user's question from these candidate knowledge information. To achieve this goal, the "Item Attribute Question Answering Agent" further introduces the current background information of the live broadcast room as contextual input, including real-time scene elements such as the information of the item being introduced, the pace of the live broadcast content, and the item display status.

[0096] The user's question, candidate knowledge information, and the current background information of the live stream are input into the large language model (LLM) of the "Item Attribute Question Answering Agent." Leveraging the powerful semantic understanding, context fusion, and multi-dimensional reasoning capabilities of the large model, 50 to 70 candidate knowledge information items are deeply analyzed and compared. This analysis comprehensively considers multiple dimensions such as semantic similarity, scenario relevance, and logical consistency between the question and the knowledge, thereby inferring the target knowledge information that best matches the user's intent and the current live stream context. This process not only relies on keyword matching at the text level but also emphasizes the understanding of the user's question intent and the integration of this understanding with the actual live stream scenario. This ensures that the final selected knowledge information is not only accurate in content but also possesses natural interactivity and business applicability, improving the intelligence level of the question answering system and the user experience.

[0097] For step S304, in conjunction with Figure 1 and Figure 2 As shown in the structure, each piece of knowledge information in the item attribute question-and-answer knowledge base can be composed of different forms: First, it only contains "item title" and "item attribute", which is suitable for item attributes that have not yet been configured with question-and-answer pairs; second, based on the configuration by the operators, it further includes complete question-and-answer pairs; third, question-and-answer pairs automatically generated by the "item attribute knowledge base agent" (see later for details). Figure 5 (Description). This knowledge information comprehensively covers four components: "question," "answer," "item title," and "item attributes." These three forms together constitute a structured, multi-layered item attribute question-and-answer knowledge base, providing a rich and accurate information foundation for vector retrieval and answer generation.

[0098] After identifying the target knowledge information, the "Item Attribute Q&A Agent" can obtain a unique identifier (such as 1, 2, 3, 4, etc.) for the target knowledge information, which is used for subsequent answer generation and knowledge tracking. Most importantly, the "Item Attribute Q&A Agent" will generate a response to the user's question based on the target knowledge information and display it to the host or directly in the live broadcast interactive interface.

[0099] Based on the completeness of the knowledge information, there are two ways to process the answer:

[0100] 1. The target knowledge information contains the answer: At this point, the knowledge structure is complete, including "question," "answer," "item title," and "item attributes." The system will optimize and refine the answer within the target knowledge information while maintaining semantic integrity, making it more conversational, natural, and engaging, thereby enhancing the user experience.

[0101] 2. No answer configured in the target knowledge information: At this time, the knowledge structure only contains "item title" and "item attributes," which is an incomplete question-and-answer structure. In this case, the system will automatically generate an accurate, natural, and context-appropriate answer based on the "item title" and "item attributes" in the target knowledge information, combined with the user's question and the current background information of the live broadcast room (such as the content of the host's explanation, the item display status, etc.), using a large-scale model (LLM), to ensure the completeness of information delivery and the smoothness of interaction.

[0102] Through the aforementioned mechanism, the system can flexibly respond to knowledge information of different structures in the knowledge base, ensuring the quality of answers while improving the intelligence level and real-time response capability of the question-and-answer system in live streaming scenarios. As an optimized implementation method, in the process of generating answers to user questions, this solution can further obtain core benefit information related to the target item, such as promotional information, product highlights, limited-time offers, and other marketing content, and generate persuasive answers with guiding and conversion intentions based on this information to enhance interactive effects and promote user decision-making.

[0103] As an optional implementation, this solution generates a "clarifying counter-question" or provides a "fallback general answer" when the large model fails to match suitable target knowledge information. For example, if a user's question is vague or unclear, the system can automatically generate a "clarifying counter-question," further confirming the user's needs through questioning, such as "Do you mean whether this phone supports 5G?". Alternatively, it provides a "fallback general answer" to ensure the continuity of Q&A and user experience during the live stream, such as a general answer like, "The host is providing you with more information about this item; please wait."

[0104] By introducing incentive information to generate guiding responses and combining a question-and-answer mechanism with a fallback answer strategy, this solution effectively improves the robustness, adaptability, and user-friendliness of the product attribute question-and-answer system, making it particularly suitable for high-concurrency, fast-paced e-commerce live streaming scenarios.

[0105] The entire operation described above can be performed by the "Item Attribute Q&A Agent," an item attribute Q&A expert. This expert, like a large-scale LLM model, possesses keen insight, rigorous analytical skills, information comparison capabilities, logical thinking, and attention to detail. In a live streaming scenario, when a user asks a question about item attributes, the system first identifies the target item corresponding to the question based on rule-based methods, determining whether the user's question refers to a specific item SKU, multiple items in the live stream, or all items. Based on this, the core task of the item attribute Q&A expert is to quickly filter the most relevant knowledge information from candidate knowledge information associated with the target item, and generate a positive and accurate answer based on this knowledge information. Simultaneously, it combines the item's core benefits to generate guiding statements to enhance interaction and conversion rates.

[0106] The specific execution process is as follows:

[0107] 1. Understanding Input Information: The system needs to understand the content of the questions raised by the user and obtain information about the items currently being streamed, as an important reference to help determine the context.

[0108] 2. Knowledge Matching Analysis: Browse all knowledge information in the knowledge base that is associated with the target item, analyze the semantic relevance of each item to the user's question, identify the most matching unique candidate knowledge information, and determine its unique knowledge number.

[0109] 3. Answer Generation Strategy: If an answer already exists for the candidate knowledge information, then based on that answer, a positive answer to the user's question is generated while maintaining semantic integrity; simultaneously, guiding statements are provided in conjunction with the "benefit points". If no answer exists for the candidate knowledge information, then based on its "item title" and "item attributes", a positive answer to the user's question is generated; simultaneously, guiding statements are provided in conjunction with the "benefit points".

[0110] 4. Output Results: The final output includes the knowledge ID of the most matching knowledge information, as well as the generated answer content, which includes both positive responses and guiding statements.

[0111] Throughout the process, all evaluation and generation actions must be strictly based on user questions, candidate knowledge information, and current live-stream item information, without introducing any external knowledge or subjective judgment, to ensure the authenticity, accuracy, and business suitability of the answers.

[0112] Example 1: It can identify a specific item, but the scope of the knowledge base recall is relatively small.

[0113] Suppose a user asks: Does XX support 5G networks?

[0114] Items currently being streamed (for reference): OnePlus Ace 3 Pro 12GB+256GB Titanium Mirror Silver, Snapdragon 8 flagship chip (3rd generation) || Price: 2699.00 yuan || Features: 6100mAh Glacier Battery.

[0115] Identify the candidate knowledge information corresponding to "XX":

[0116] 1. Question: Does this XX support 5G? || Answer: XXX runs smoothly || Item Attributes: 5G network || Benefit: Instant discount of 1000 yuan;

[0117] 2. Question: Does XX support dual SIM dual standby? || Answer: XXX || Item Attribute: Dual SIM Dual Standby || Benefit: Instant discount of 1000 yuan;

[0118] ...

[0119] The model matches the knowledge information "1" that best matches the user's question. Based on this knowledge information, a response is generated, along with guiding statements based on the user's benefit. Assuming the output format is {"Knowledge ID":"ID", "Reply":"Answer"}, the final answer displayed in the live stream might be: {"Knowledge ID": "1", "Reply": "XX supports 5G networks, allowing you to enjoy high-speed network connectivity and a smooth online experience. Order now and get a 1000 yuan discount! Don't miss out!"}. However, if no item fully meets the user's needs, or if the current information makes a judgment impossible, then 'Cannot be judged' is entered in 'Knowledge ID', and a clarification or a general fallback answer is entered in 'Reply'.

[0120] Example 2: If a specific item cannot be identified, the knowledge base recall scope is the entire knowledge base.

[0121] Suppose a user's question is "Does XY support 5G networks?", and since the item "XY" does not exist in the live stream, the candidate knowledge information is the knowledge information in the knowledge base corresponding to all items in the live stream:

[0122] 1. Question: Does this XX support 5G? || Answer: XXX runs smoothly || Item Title: XX || Item Attributes: 5G network || Benefit: Instant discount of 1000 yuan;

[0123] 2. Question: Does XX support dual SIM dual standby? || Answer: XXX || Item Title: XX || Item Attribute: Dual SIM Dual Standby || Benefit: Instant discount of 1000 yuan;

[0124] ...

[0125] N-1. Question: Does this YY (Yangtze) support fast charging? || Answer: XXX || Item Title: YY || Item Attribute: Fast Charging || Benefit: 10% discount;

[0126] Question: Does this YY15 support 5G? || Answer: XXX || Item Title: YY15 || Item Attributes: 5G Network || Benefit: 10% discount.

[0127] Assuming the current live stream item is the same as in Example 1, the output answer might be {"Knowledge ID": "1", "Reply": "XX supports 5G network, allowing you to enjoy high-speed network connection and smooth online experience. Order now and get a discount of 1000 yuan, don't miss it~"}.

[0128] The returned answer includes a knowledge ID because this ID is a unique identifier for the knowledge information in the knowledge base. Using the knowledge ID, the knowledge information upon which the large model's answer is based can be quickly located during subsequent checks, determining whether it accurately matches the user's question. For example, in the vector recall phase, the system selects the Top 50 candidate knowledge information for the large model to choose from. If the most relevant candidate knowledge information to the user's question is ID "5," but the large model generates an answer based on candidate knowledge information ID "7," it indicates a matching failure (see below). Figure 7 The error type is "matching error". By recording the knowledge IDs used, the cause of the error can be traced, which helps to optimize the model and improve the accuracy of question answering. Therefore, knowledge IDs not only enhance the interpretability of the answers, but also provide an important basis for system debugging and continuous optimization.

[0129] The method provided in the above embodiments, by combining user operation information and live room background information, accurately matches the target knowledge information most relevant to the user's question from the item attribute question and answer knowledge base, and generates and displays the answer based on the target knowledge information. This not only improves the semantic understanding and scene adaptation capabilities of the item attribute question and answer system, but also enhances the accuracy and context relevance of the answer, thereby effectively improving the user experience and live interaction efficiency.

[0130] See Figure 5 The diagram illustrates an optional item attribute question-and-answer method according to an embodiment of the present invention, including the following steps:

[0131] S501: Receive the item attributes labeled on the online dialogue data, and retrieve the candidate knowledge information corresponding to the labeled item attributes from the item attribute question and answer knowledge base;

[0132] S502: Using the acquired candidate knowledge information as a reference, simulate the questions that users ask about the marked item attributes, and then generate answers to the simulated questions;

[0133] S503: Generate simulated dialogue data based on simulated questions and answers.

[0134] This implementation describes how to further expand question-answer pairs based on existing knowledge information using a "user simulator agent". In the initial stage of the "item attribute question-answering agent" launch, limited online dialogue data resulted in fewer bad cases being discovered, leading to slow model and knowledge base updates. Therefore, this solution generates simulated dialogue data using a "user simulator agent" to address these issues. The specific method is as follows:

[0135] 1. Manual Annotation. First, existing online dialogue data is acquired, including target knowledge information, user questions (including attributes), responses, and background information from the live stream. Then, item attributes involved in the online dialogue data are identified through manual annotation. This process relies heavily on human classification, laying the foundation for subsequent simulated dialogues based on these attributes.

[0136] 2. Precise matching from the knowledge base. After obtaining the labeled item attributes, a precise matching strategy from the knowledge base is adopted to retrieve candidate knowledge information containing the current item attributes from the item attribute question and answer knowledge base based on these attributes.

[0137] 3. Simulate User Interaction. Among the acquired candidate knowledge information, some question-answer pairs can be directly used for simulated interaction with the "Item Attribute Question-Answer Agent." However, not all candidate knowledge information contains question-answer pairs. Furthermore, candidate knowledge information containing question-answer pairs needs to be expanded with more diverse user questions to cover more unknown situations.

[0138] Therefore, a user simulator agent is constructed to generate more high-quality simulated dialogue data by engaging in real-time dialogue with an item attribute question-and-answer agent. The user simulator agent uses "candidate knowledge information" as a reference to simulate and generate diverse user questions about item attributes, guiding the item attribute question-and-answer agent to generate the correct answers.

[0139] 4. Combine the simulated questions with the answers from the item attribute Q&A Agent to create new simulated dialogue data.

[0140] To test the effectiveness of the item attribute question-and-answer system, it is necessary to simulate real users asking questions about item attributes. Therefore, the "user simulator agent" needs a deep understanding of user behavior and questioning styles, and must be able to creatively generate natural, diverse, and targeted questions based on key information such as item titles and attributes from candidate knowledge information. Specifically, the user simulator agent needs to generate a user question that fits a real-world scenario for each candidate knowledge piece, based on a thorough understanding and analysis of the candidate knowledge information and combined with common user expression habits and questioning logic. It should be noted that the generated questions must be based on the provided candidate knowledge information and cannot introduce other irrelevant information, thus effectively verifying the accuracy and coverage of the item attribute question-and-answer agent.

[0141] Assuming the labeled item attribute is "5G network", based on this item attribute, candidate knowledge information including "5G network" is obtained from the item attribute question-and-answer knowledge base, which is "1. Question: Does this XX support 5G? || Answer: XXX is smooth || Item title: XX || Item attribute: 5G network". Based on this candidate knowledge information, the simulated dialogue data is output as [{"Item ID": "1", "User Question": "Does XX support 5G network?"}, … , {} ].

[0142] The "User Simulator Agent" can be a large model, which mainly involves two types of task scenarios: The first type is where the user explicitly asks about the attribute of a certain item, and the corresponding knowledge information exists in the knowledge base. In this case, the large model only needs to complete information matching and extraction to give an accurate answer. This type of pattern is clear and the training effect is easy to achieve. The model can learn this type of question-and-answer pattern through repeated learning, but it is not challenging and has limited improvement on the model's capabilities.

[0143] The second category involves users whose expressions are vague or whose needs are unclear. For example, a user might ask, "What phones would you recommend for live streaming?" without mentioning specific attributes. In such cases, a large model needs to reason and make judgments based on the context. The model needs to understand the underlying needs behind the "live streaming" scenario, such as high performance and a high-quality video experience, and filter relevant answers from existing attributes. Because these types of questions cannot rely on fixed fields in a knowledge base for matching, they must rely on the model's own understanding and reasoning abilities to answer. Therefore, adding this type of training data is more helpful in improving the model's generalization ability and its ability to apply knowledge to new situations.

[0144] See the system architecture of "User Simulator Agent" for details. Figure 6 As shown, as an optimized implementation, this solution performs data filtering after generating the simulated dialogue data. To prevent the user simulator Agent from directly leaking answers during the interaction, thus causing data leakage, post-processing data filtering is required on the generated simulated dialogue data. For example, the question "User question: What is the mAh of YY's battery?" only contains the item attribute name and does not involve data leakage; however, the question "User question: YY's battery mAh is 4600" includes both the item attribute name and attribute value, which involves data leakage.

[0145] This operation can be performed by data breach detection experts, such as large-scale modeling (LLM). Their core task is to monitor and analyze the dialogue between the user simulator agent and the item attribute question-and-answer agent, ensuring that user questions generated by the user simulator agent do not result in data breaches. The output format strictly follows: {"Thinking Process": "","Detection Result":""}, where the detection result can only be "Correct" or "Incorrect". The thinking process must explain the reasons for whether a data breach occurred.

[0146] Example 1:

[0147] User simulator Agent: How many milliamps does YY's battery have?

[0148] Item Attribute Q&A Agent: YY's battery has a capacity of 4600 mAh. There's a 10% discount on YY purchases right now, don't miss it!

[0149] Reply: {"Thinking Process": "The User Simulator Agent did not directly provide the answer, but instead raised a question: There is no data breach", "Detection Result": "Correct"}.

[0150] Example 2:

[0151] User simulator Agent: YY's battery capacity is 4600 mAh.

[0152] Item Attribute Q&A Agent: YY's battery has a capacity of 4600 mAh. There's a 10% discount on YY purchases right now, don't miss it!

[0153] Reply: {"Thinking Process": "The user simulator Agent directly tells the answer that there is a data breach", "Detection Result": "Error"}.

[0154] In the training of large models, removing data leakage is of great significance, mainly in the following aspects:

[0155] 1. Improve the model's generalization ability: Data leakage refers to the inclusion of test set or future prediction target data in the training data, causing the model to "see" the target information during the training phase, thus resulting in unrealistic performance in real-world applications. Removing this data helps the model learn more generalized features, enhancing its predictive ability on unseen data.

[0156] 2. Avoid overfitting: Training sets containing leaked data can cause the model to overfit those specific samples, resulting in significantly poor performance on new samples. Removing leaked data can reduce the model's dependence on non-generalization features and better reflect the data distribution in the real-world scenario.

[0157] 3. Ensure the fairness and authenticity of the evaluation: Model evaluations (such as validation sets and test sets) should accurately reflect the model's performance in real-world applications. If data in the training set is leaked, the model may appear to perform well during evaluation, but its performance may degrade after actual deployment, thus affecting the model's reliability.

[0158] 4. Comply with data privacy and compliance requirements: Many data breaches may involve sensitive information. Removing leaked data helps to comply with laws, regulations and ethical norms, and ensures that the model training process is legal and compliant.

[0159] In summary, removing leaked data is crucial to ensuring the model's generalization ability and fairness, preventing overfitting, and improving performance and reliability in practical applications. This is a vital step in ensuring quality and reliability during model training.

[0160] The method provided in the above embodiments accurately filters candidate knowledge information from the knowledge base by utilizing labeled item attributes, and simulates user questioning behavior based on this to generate high-quality question-answer pairs. This expands the simulated dialogue data that closely resembles real-world scenarios, effectively alleviating the problem of insufficient online dialogue data and improving the diversity and relevance of the data, providing rich and reliable data support for model training and system optimization.

[0161] See Figure 7 The diagram illustrates another optional item attribute question-and-answer method according to an embodiment of the present invention, including the following steps:

[0162] S701: Acquire dialogue data; wherein the dialogue data includes one or more of online dialogue data and simulated dialogue data; the online dialogue data includes the target knowledge information, the question, the answer, the attribute, and the current background information; the simulated dialogue data is generated based on the online dialogue data;

[0163] S702: Based on the candidate knowledge information and attributes corresponding to the dialogue data, evaluate the correctness of the answers in the dialogue data, and determine the error type if the evaluation result indicates that there is an error;

[0164] S703: Using the preset solution for the error type, adjust the knowledge information in the item attribute question and answer knowledge base, and then call the item attribute question and answer agent to perform a response test;

[0165] S704: In response to a failed response test, roll back the item attribute question and answer knowledge base and adjust the training set of the item attribute question and answer agent to retrain the parameters of the item attribute question and answer agent.

[0166] In the above implementation, for steps S701 and S702, regardless of whether the dialogue data is generated by the "online user in the live stream" and the "item attribute question and answer agent" or the "user simulator agent" and the "item attribute question and answer agent," the correctness of the attribute questions and answers must be evaluated, and the error type must be identified to guide subsequent model and knowledge base iteration and optimization. Therefore, this solution constructs an automatic evaluation agent based on a large model to replace manual annotation.

[0167] To ensure the accuracy and reliability of the item attribute question-and-answer agent's responses, the automatic evaluation agent needs to possess strong information analysis, logical reasoning, and detail recognition capabilities. Its core task is to determine the correctness of the agent's answers based on provided candidate knowledge information, the agent's responses, and the item attributes currently being discussed, and to identify specific error types, as shown in Figure 8(a). It should be noted that the evaluation process must be based solely on the provided candidate knowledge information, the agent's responses, and the item attributes currently being discussed; no other irrelevant information can be introduced.

[0168] Specifically, the automated evaluation agent first identifies candidate knowledge information corresponding to the dialogue data, deeply understanding key information such as item titles, item attribute names, and benefits. It then analyzes the agent's responses to item attribute questions, determining if their content aligns with the candidate knowledge information and the currently discussed item attributes. Based on the analysis results, it judges the correctness of the responses and further identifies specific error types. For simulated dialogue data, it also needs to incorporate the current background information of the live stream in the online data.

[0169] This solution pre-defines four typical error types. In practical applications, other error classifications can be expanded according to specific business needs. Here, only examples are presented:

[0170] 1. Not Recalled: The user raised a clear question, and there were corresponding answer conditions and knowledge information, but the item attribute Q&A Agent failed to provide a valid answer;

[0171] 2. False interception: The user’s question is incorrect, incomplete, or belongs to the catch-all / reverse question scenario, and should not trigger a specific answer, but the item attribute Q&A Agent still gives a positive answer, resulting in a reply that does not conform to the actual context or guiding logic;

[0172] 3. Matching error: The model should refer to a specific piece of knowledge (such as knowledge information a) to answer the question, but it incorrectly refers to other knowledge information that does not match or has a low degree of matching (such as knowledge information b).

[0173] 4. Incorrect answer: The generated answer is irrelevant to the user's question, or the guiding language used deviates from the core attributes and benefits of the item, affecting the accuracy of information delivery.

[0174] Example 1:

[0175] The question is: Does YY support 5G networks?

[0176] The item attribute Q&A agent's answer is: Of course! With 5G network, you'll have a faster surfing experience. There's a 10% discount on phone purchases right now, don't miss it!

[0177] Candidate knowledge information:

[0178] 1. Question: xx || Answer: xx || Item Title: YY || Item Attribute Name: 5G Network || Benefit: Instant discount of 1000 yuan;

[0179] ...

[0180] Assuming the output format is: {"Evaluation Result": "Correct / Incorrect", "Incorrect Error Type": "Not Recalled / Falsely Blocked / Incorrect Match / Incorrect Answer", "Evaluation Reason": "Specific Reason"}, then the output here would be: {"Evaluation Result": "Incorrect", "Incorrect Error Type": "Incorrect Answer", "Evaluation Reason": "The discount information mentioned in the answer is inconsistent with the benefits in the candidate knowledge information; it should be a discount of 1000 yuan instead of a 10% discount."}

[0181] Example 2:

[0182] User request: Which size of 65-inch screen would you recommend?

[0183] The Agent's answer to the item attribute question is: "ZZ Brand TV 65A5D 65-inch 100-level partition 4+64GB 1000nit 4K ultra-high-definition ultra-thin eye-protection flat panel LCD TV gaming TV || Price: 3799 yuan || Attribute: 65-inch".

[0184] Recommendation reason: "Item 1 is a ZZ brand TV, with a size of 65 inches, which perfectly meets the user's needs for a 65-inch TV."

[0185] Output: {“Evaluation Result”: “Recommendation Correct”, “Thought Process”: “The user needs a 65-inch item, and the item attribute Q&A Agent recommended a suitable item that meets the user's size requirements.”}

[0186] For steps S703 and S704, after the automatic evaluation agent detects errors and specific error types in the answers provided by the item attribute question-and-answer agent, the "knowledge base iteration agent" can analyze and match corresponding solutions, thereby achieving closed-loop optimization of the entire system architecture. As shown in Figure 8(b), this solution pre-defines corresponding solutions for different types of errors, constructing a complete process from error identification to problem repair. Compared to the traditional method that relies on manual annotation and intervention, this solution, by constructing a knowledge base iteration agent with autonomous decision-making capabilities, can more efficiently and quickly locate the optimal solution, significantly improving the system's adaptability and iteration efficiency.

[0187] In terms of processing logic, the knowledge base iteration agent prioritizes the assumption that the problem originates from the knowledge base itself, and formulates targeted solutions accordingly:

[0188] 1. No recall: This is identified as a missing corpus issue, requiring the missing knowledge information to be added to the knowledge base;

[0189] 2. False blocking: This is identified as a problem of redundant corpus information, and unnecessary or interfering redundant knowledge information in the knowledge base needs to be deleted.

[0190] 3. Matching error: This is identified as a corpus confusion problem, and the ambiguous or conflicting confusing knowledge information in the knowledge base needs to be adjusted.

[0191] 4. Incorrect answer: This is identified as a corpus error and requires adjustment of the incorrect knowledge information in the knowledge base.

[0192] After adjusting the knowledge base, the item attribute question-answering agent is further invoked for response testing. If the response test result is correct, it indicates that the original problem was caused by a deficiency in the knowledge base; if the test result fails, it indicates that the problem may originate from the item attribute question-answering agent itself. In this case, the knowledge base needs to be rolled back, and the training set of the item attribute question-answering agent needs to be supplemented or optimized to retrain the model parameters and improve the understanding and reasoning ability of the item attribute question-answering agent. This mechanism effectively distinguishes the problem boundaries between the knowledge base and the model, forming a closed-loop process of "evaluation-repair-verification," promoting continuous optimization and self-improvement of the system.

[0193] The method provided in the above embodiments automatically evaluates the correctness of answers in dialogue data and identifies specific error types. It combines pre-set solutions to achieve intelligent iterative optimization of the knowledge base and provides timely feedback to the model training stage when knowledge base adjustments are ineffective. This forms a closed-loop mechanism of "evaluation-repair-verification," effectively improving the accuracy and adaptability of the question-answering system while reducing the cost of manual intervention and accelerating the collaborative optimization efficiency of the model and knowledge base.

[0194] Referring to Figure 9(a), a schematic diagram of an item attribute question-and-answer process according to an embodiment of the present invention is shown. Referring to Figure 9(b), a schematic diagram of an item attribute question-and-answer architecture according to an embodiment of the present invention is shown. This solution focuses on the construction and optimization of the item attribute question-and-answer system, proposing five core Agent modules, each undertaking key tasks such as knowledge generation, question-and-answer interaction, simulated dialogue, automatic evaluation, and knowledge base iteration, forming a closed-loop, efficient, and autonomously evolving item attribute question-and-answer system.

[0195] 1. Knowledge Mining Agent. The construction of traditional item attribute question-and-answer knowledge bases heavily relies on manual annotation and organization by livestream operators, resulting in low efficiency and limited coverage. This solution introduces a knowledge mining agent based on a large model, leveraging its powerful reasoning and reflection capabilities to automatically generate diverse and semantically coherent question-and-answer pairs, alleviating reliance on manual resources and enabling rapid construction and expansion of the knowledge base.

[0196] 2. Item Attribute Question Answering Agent. Addressing the shortcomings of traditional deep learning models in semantic understanding and dialogue generation in current e-commerce live streaming scenarios, this solution constructs an "Item Attribute Question Answering Agent" based on a large model. This agent can accurately identify the top-matching target knowledge information related to the user's question and generate natural, fluent answers that closely match the user's intent, assisting users in making more accurate shopping decisions and improving user experience and conversion rates.

[0197] 3. User Simulator Agent. In the initial stages of the item attribute question-and-answer agent's launch, the lack of sufficient online dialogue data hindered rapid iterative optimization of the agent's model. To address this, this solution constructs a "User Simulator Agent" based on a limited amount of real-world dialogue data. This user simulator agent interacts with the item attribute question-and-answer agent in real-time, simulating real user behavior and generating more high-quality simulated question-and-answer pairs for item attributes. This provides rich data support for model training and knowledge base updates, thereby accelerating the iterative optimization of the "Item Attribute Question-and-Answer Agent."

[0198] 4. Automated Agent Evaluation. Traditional evaluation methods rely on manual annotation, which is costly and inefficient. This solution uses large-scale modeling technology to build an "automatic evaluation agent" to replace manual annotation. Based on multi-dimensional inputs provided by the item attribute question-answering agent, such as "Top 1 knowledge information," "response answers," "candidate items in the shopping bag and their various attributes," and "live broadcast background information," the agent evaluates whether the answers provided by the item attribute question-answering agent are relevant to the user's question, whether the responses are correct, and identifies errors and their specific types. This guides the iterative optimization of the "item attribute question-answering agent" and the "item attribute question-answering knowledge base," thereby achieving low-cost, high-efficiency model quality monitoring and problem localization.

[0199] 5. Knowledge Base Iteration Agent. Building a knowledge base is not a one-time task; continuous optimization is needed to ensure the accuracy and unambiguity of the corpus within it. This solution constructs a "knowledge base iteration agent," which automatically evaluates the error types identified by the agent, matches pre-defined solutions, and automatically adjusts the knowledge base content (e.g., supplementing missing information, removing redundancy, correcting errors, and eliminating confusion). When necessary, it feeds back to the model training stage, driving the co-evolution of the item attribute question-answering agent and the knowledge base, ultimately achieving overall system closed-loop optimization. This optimization process plays a crucial role in the accurate response to online item attributes.

[0200] In summary, this solution offers the following advantages: it not only effectively improves the efficiency of building the item attribute question-and-answer knowledge base and significantly reduces labor costs, but also greatly improves the accuracy of item attribute question-and-answer, further enhancing users' purchasing decision efficiency. Furthermore, in the early stages of system operation when data is scarce, high-quality dialogue data can be expanded through the user simulator agent, reducing labor costs and accelerating model iteration and optimization. Moreover, by leveraging the automatic evaluation agent and the knowledge base iteration agent, accurate identification and rapid correction of error types are achieved, and solutions are quickly located based on fine-grained error analysis, realizing closed-loop optimization of the entire system architecture. Through the collaborative interaction of these five agents, this solution ultimately constructs an intelligent item attribute question-and-answer system architecture with self-driven, continuous learning, and automatic iteration capabilities.

[0201] See Figure 10 The diagram illustrates the main modules of an item attribute question-and-answer system 1000 provided in an embodiment of the present invention, including a knowledge mining agent 1001, an item attribute question-and-answer agent 1002, a user simulator agent 1003, an automatic evaluation agent 1004, and a knowledge base iteration agent 1005.

[0202] Knowledge mining agent 1001 is used to build a knowledge base for question-and-answer questions about item attributes.

[0203] Item attribute question-and-answer agent 1002, used for:

[0204] Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item;

[0205] From the item attribute question and answer knowledge base, obtain candidate knowledge information corresponding to the target item, and calculate the similarity between the question and each candidate knowledge information;

[0206] Based on the similarity, the question, and the current background information of the live stream, the target knowledge information that best matches the question is determined.

[0207] Based on the target knowledge information, an answer to the question is generated and displayed in the live broadcast room;

[0208] User simulator agent 1003 is used to generate simulated dialogue data based on online dialogue data;

[0209] An automatic evaluation agent 1004 is used to acquire dialogue data; wherein, the dialogue data includes one or more of online dialogue data and simulated dialogue data; the online dialogue data includes the target knowledge information, the question, the answer, the attributes, and the current background information;

[0210] Based on the candidate knowledge information and attributes corresponding to the dialogue data, the correctness of the answers in the dialogue data is evaluated, and if the evaluation result indicates that there is an error, the error type is determined.

[0211] Knowledge base iterative agent 1005 is used to adjust the knowledge information in the item attribute question and answer knowledge base using a preset solution for the error type, and then call item attribute question and answer agent 1002 to perform response testing.

[0212] In response to a failed response test, the item attribute question-and-answer knowledge base is rolled back, and the training set of the item attribute question-and-answer agent 1002 is adjusted to retrain the parameters of the item attribute question-and-answer agent 1002.

[0213] In the system of this invention, the knowledge mining agent 1001 is used for:

[0214] Obtain item information related to items in the live broadcast room, and extract key features from the item information;

[0215] Based on the aforementioned key features, predict the questions users will ask about the attributes of the items during the live stream, and determine the item attributes corresponding to the predicted questions.

[0216] Based on the item information, the predicted question, and the item attributes corresponding to the predicted question, an answer is generated;

[0217] Based on the predicted question, the generated answer, the item information, and the determined item attributes, knowledge information is generated and stored in the item attribute question-and-answer knowledge base. At the same time, the association between the knowledge information and the item identifier of the item is established.

[0218] In the system of this invention, the knowledge mining agent 1001 includes one or more of the following:

[0219] Extract live streaming information and key attributes marked on the items from the live streaming script;

[0220] Extract information related to item transactions from the item title;

[0221] Extract the specifications and characteristics of an item from its image-to-text information.

[0222] Extract the item's current attributes from its attribute description.

[0223] In the system of this invention, the knowledge mining agent 1001 is further used for:

[0224] Assess whether the predicted problem is associated with the item information. If the assessment result is negative, optimize the predicted problem using the key features to obtain the optimized problem.

[0225] The generated answer is evaluated to determine whether it is related to the item information and the predicted question. If the evaluation result is negative, the generated answer is optimized using the key features to obtain an optimized answer.

[0226] In the system of this invention, the knowledge mining agent 1001 is further used for:

[0227] Determine the attributes of the remaining items that have not generated questions, and generate knowledge information based on the item information and each remaining item attribute;

[0228] The generated knowledge information is stored in the item attribute question and answer knowledge base, and the association between the generated knowledge information and the item identifier is established.

[0229] In the system of this invention, the item attribute question-and-answer intelligent agent 1002 includes the following:

[0230] In response to the user's click to ask a question about the first item in the live stream and input of the question, the first item is identified as the target item;

[0231] The system receives a question entered by the user in the question input box in the live stream room. In response to the presence of an item name in the question, the system identifies the second item in the live stream room that corresponds to the item name as the target item.

[0232] In response to the problem that the item name does not exist, or that the item name exists but there is no item in the live stream that corresponds to the item name, all items in the live stream are identified as the target item.

[0233] In the system of this invention, the item attribute question-and-answer intelligent agent 1002 is used for:

[0234] Vector recall technology is used to calculate the text vector similarity between the question and each candidate knowledge information, and then a preset number of candidate knowledge information are selected in descending order of similarity.

[0235] In the system of this invention, the item attribute question-and-answer intelligent agent 1002 is used for:

[0236] In response to the target knowledge information including the answer, the answer in the target knowledge information is optimized without semantic invariance to obtain the answer to the question;

[0237] Since the target knowledge information does not include an answer, an answer to the question is generated based on the target knowledge information, the question, and the current scene information of the live broadcast room.

[0238] In the system of this invention, the intelligent agent 1002 for answering questions about item attributes is further used for:

[0239] Obtain the benefit information of the target item, and generate a response based on the benefit information;

[0240] Based on the given answers and the given scripted answers, a target answer is generated and displayed in the live stream.

[0241] In the system of this invention, the intelligent agent 1002 for answering questions about item attributes is further used for:

[0242] If no target knowledge information best matches the question, a preset fallback general answer is returned, or a counter-question clarification answer is generated based on the question;

[0243] The preset fallback answer or the clarification answer to the rhetorical question is displayed in the live broadcast room.

[0244] In the system of this invention, the user simulator agent 1003 is used for:

[0245] Receive item attributes labeled on online dialogue data, and retrieve candidate knowledge information corresponding to the labeled item attributes from the item attribute question and answer knowledge base;

[0246] Using the acquired candidate knowledge information as a reference, the system simulates questions posed by users to the labeled item attributes, and then generates answers to the simulated questions.

[0247] Simulated dialogue data is generated based on simulated questions and responses.

[0248] In the system of this invention, the user simulator agent 1003 is further used for:

[0249] Determine if there are attribute values ​​corresponding to item attributes in the simulated problem. If attribute values ​​exist, perform data filtering processing on the simulated dialogue data.

[0250] In the system of this invention, the automatic evaluation agent 1004 is used for:

[0251] Based on the target knowledge information and the attributes, the correctness of answers in online dialogue data is evaluated.

[0252] Based on the current background information, the labeled item attributes, and the candidate knowledge information corresponding to the labeled item attributes, the correctness of the answers in the simulated dialogue data is evaluated.

[0253] In the system of this invention, the automatic evaluation agent 1004 is used for:

[0254] Based on the target knowledge information, the attributes, and the benefit information corresponding to the target knowledge information, the correctness of the answers in the online dialogue data is evaluated.

[0255] Based on the current background information, the labeled item attributes, the candidate knowledge information corresponding to the labeled item attributes, and the benefit information corresponding to the candidate knowledge information, the correctness of the answers in the simulated dialogue data is evaluated.

[0256] Furthermore, the specific implementation details of the device described in the embodiments of the present invention have been described in detail in the above-described method, so the details will not be repeated here.

[0257] Figure 11 An exemplary system architecture 1100 to which embodiments of the present invention can be applied is shown, including terminal devices 1101, 1102, 1103, network 1104, and server 1105 (this is merely an example).

[0258] Terminal devices 1101, 1102, and 1103 can be various electronic devices with displays and support web browsing, and have various communication client applications installed. Users can use terminal devices 1101, 1102, and 1103 to interact with server 1105 through network 1104 to receive or send messages, etc.

[0259] Network 1104 is a medium used to provide a communication link between terminal devices 1101, 1102, 1103 and server 1105. Network 1104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0260] Server 1105 can be a server providing various services, such as a backend management server supporting shopping websites browsed by users using terminal devices 1101, 1102, and 1103 (this is just an example). The backend management server can analyze and process received data such as product information query requests, and feed back the processing results (such as target push information and product information—this is just an example) to the terminal devices. It should be noted that the method provided in this embodiment of the invention is generally executed by server 1105, and correspondingly, the apparatus is generally set in server 1105.

[0261] It should be understood that Figure 11 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0262] The following is for reference. Figure 12 It shows a schematic diagram of the structure of a computer system 1200 suitable for implementing a terminal device of the present invention. Figure 12 The terminal device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0263] like Figure 12 As shown, the computer system 1200 includes a central processing unit (CPU) 1201, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1202 or programs loaded from storage section 1208 into random access memory (RAM) 1203. The RAM 1203 also stores various programs and data required for the operation of the system 1200. The CPU 1201, ROM 1202, and RAM 1203 are interconnected via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0264] The following components are connected to I / O interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card such as a LAN card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to I / O interface 1205 as needed. Removable media 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1210 as needed so that computer programs read from them can be installed into storage section 1208 as needed.

[0265] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by central processing unit (CPU) 1201, it performs the functions defined above in the system of this invention.

[0266] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0267] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0268] The modules described in the embodiments of this invention can be implemented in software or hardware. The described modules can also be housed in a processor; for example, a processor can be described as including a knowledge mining agent, an item attribute question-answering agent, a user simulator agent, an automatic evaluation agent, and a knowledge base iteration agent. The names of these modules do not necessarily limit the module itself; for example, the user simulator agent can also be described as a "user simulator module."

[0269] In another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to perform any of the above-described item attribute question-and-answer methods.

[0270] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the item attribute question-and-answer method in the embodiments of the present invention.

[0271] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An item attribute question answering method, characterized by, include: Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item; From the item attribute question-and-answer knowledge base, candidate knowledge information corresponding to the target item is obtained, and the similarity between the question and each candidate knowledge information is calculated; wherein, the item attribute question-and-answer knowledge base stores knowledge information of questions, answers and item attributes generated based on item information prediction; Based on the similarity, the question, and the current background information of the live stream, the target knowledge information that best matches the question is determined. Based on the target knowledge information, an answer to the question is generated and displayed in the live broadcast room.

2. The method according to claim 1, characterized in that, The method also includes a process for constructing an item attribute question-and-answer knowledge base, the construction process including: Obtain item information related to items in the live broadcast room, and extract key features from the item information; Based on the aforementioned key features, predict the questions users will ask about the attributes of the items during the live stream, and determine the item attributes corresponding to the predicted questions. Based on the item information, the predicted question, and the item attributes corresponding to the predicted question, an answer is generated; Based on the predicted question, the generated answer, the item information, and the determined item attributes, knowledge information is generated and stored in the item attribute question-and-answer knowledge base. At the same time, the association between the knowledge information and the item identifier of the item is established.

3. The method according to claim 2, characterized in that, The extraction of key features from the item information includes one or more of the following: Extract live streaming information and key attributes marked on the items from the live streaming script; Extract information related to item transactions from the item title; Extract the specifications and characteristics of an item from its image-to-text information. Extract the item's current attributes from its attribute description.

4. The method according to claim 2 or 3, characterized in that, The method further includes: Assess whether the predicted problem is associated with the item information. If the assessment result is negative, optimize the predicted problem using the key features to obtain the optimized problem. The generated answer is evaluated to determine whether it is related to the item information and the predicted question. If the evaluation result is negative, the generated answer is optimized using the key features to obtain an optimized answer.

5. The method according to claim 2 or 3, characterized in that, The method further includes: Determine the attributes of the remaining items that have not generated questions, and generate knowledge information based on the item information and each remaining item attribute; The generated knowledge information is stored in the item attribute question and answer knowledge base, and the association between the generated knowledge information and the item identifier is established.

6. The method according to claim 1, characterized in that, Receiving user operation information in the live stream and determining the target item corresponding to the operation information includes one of the following situations: In response to the user's click to ask a question about the first item in the live stream and input of the question, the first item is identified as the target item; The system receives a question entered by the user in the question input box in the live stream room. In response to the presence of an item name in the question, the system identifies the second item in the live stream room that corresponds to the item name as the target item. In response to the problem that the item name does not exist, or that the item name exists but there is no item in the live stream that corresponds to the item name, all items in the live stream are identified as the target item.

7. The method according to claim 1, characterized in that, The calculation of the similarity between the question and each candidate knowledge information includes: Vector recall technology is used to calculate the text vector similarity between the question and each candidate knowledge information, and then a preset number of candidate knowledge information are selected in descending order of similarity.

8. The method according to claim 1, characterized in that, The process of generating an answer to the question based on the target knowledge information includes: In response to the target knowledge information including the answer, the answer in the target knowledge information is optimized without semantic invariance to obtain the answer to the question; Since the target knowledge information does not include an answer, an answer to the question is generated based on the target knowledge information, the question, and the current scene information of the live broadcast room.

9. The method according to any one of claims 1, 6-8, characterized in that, The method further includes: Obtain the benefit information of the target item, and generate a response based on the benefit information; Based on the given answers and the given scripted answers, a target answer is generated and displayed in the live stream.

10. The method according to any one of claims 1, 6-8, characterized in that, The method further includes: If no target knowledge information best matches the question, a preset fallback general answer is returned, or a counter-question clarification answer is generated based on the question; The preset fallback answer or the clarification answer to the rhetorical question is displayed in the live broadcast room.

11. The method according to claim 1, characterized in that, The method further includes: Acquire dialogue data; wherein the dialogue data includes one or more of online dialogue data and simulated dialogue data; the online dialogue data includes the target knowledge information, the question, the answer, the attributes, and the current background information; the simulated dialogue data is generated based on the online dialogue data; Based on the candidate knowledge information and attributes corresponding to the dialogue data, the correctness of the answers in the dialogue data is evaluated, and if the evaluation result indicates that there is an error, the error type is determined. Using the pre-defined solutions for the error types, the knowledge information in the item attribute question and answer knowledge base is adjusted, and then the item attribute question and answer agent is invoked to perform a response test. In response to a failed response test, the item attribute question-and-answer knowledge base is rolled back, and the training set of the item attribute question-and-answer agent is adjusted to retrain the parameters of the item attribute question-and-answer agent.

12. The method according to claim 11, characterized in that, The process of generating simulated dialogue data based on online dialogue data includes: Receive item attributes labeled on online dialogue data, and retrieve candidate knowledge information corresponding to the labeled item attributes from the item attribute question and answer knowledge base; Using the acquired candidate knowledge information as a reference, the system simulates questions posed by users to the labeled item attributes, and then generates answers to the simulated questions. Simulated dialogue data is generated based on simulated questions and responses.

13. The method according to claim 12, characterized in that, The method further includes: Determine if there are attribute values ​​corresponding to item attributes in the simulated problem. If attribute values ​​exist, perform data filtering processing on the simulated dialogue data.

14. The method according to claim 12 or 13, characterized in that, The step of evaluating the correctness of answers in the dialogue data based on candidate knowledge information and attributes corresponding to the dialogue data includes: Based on the target knowledge information and the attributes, the correctness of answers in online dialogue data is evaluated. Based on the current background information, the labeled item attributes, and the candidate knowledge information corresponding to the labeled item attributes, the correctness of the answers in the simulated dialogue data is evaluated.

15. The method according to claim 12 or 13, characterized in that, The step of evaluating the correctness of the answers in the dialogue data based on candidate knowledge information and attributes corresponding to the dialogue data includes: Based on the target knowledge information, the attributes, and the benefit information corresponding to the target knowledge information, the correctness of the answers in the online dialogue data is evaluated. Based on the current background information, the labeled item attributes, the candidate knowledge information corresponding to the labeled item attributes, and the benefit information corresponding to the candidate knowledge information, the correctness of the answers in the simulated dialogue data is evaluated.

16. An item attribute question-and-answer system, characterized in that, This includes knowledge mining agents and item attribute question-answering agents: A knowledge mining intelligent agent is used to construct an item attribute question-and-answer knowledge base; wherein, the item attribute question-and-answer knowledge base stores questions, answers, and knowledge information of item attributes that are predicted and generated based on item information; An intelligent agent for answering questions about item attributes, used for: Receive user operation information for the live stream and determine the target item corresponding to the operation information; wherein, the operation information includes the operation method and the input questions related to the attributes of the target item; From the item attribute question and answer knowledge base, obtain candidate knowledge information corresponding to the target item, and calculate the similarity between the question and each candidate knowledge information; Based on the similarity, the question, and the current background information of the live stream, the target knowledge information that best matches the question is determined. Based on the target knowledge information, an answer to the question is generated and displayed in the live broadcast room.

17. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-15.

18. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-15.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-15.

Images