An interactive method and system for real-time e-commerce transactions

Abstract

The application provides an interactive method and system for real-time e-commerce transactions, belonging to the field of computers, and solves the problem that current live interaction relies on the personal ability of anchors, is difficult to cover high-concurrency scenarios, or although reply information is generated based on a rule engine, a rule library needs to be maintained manually, and the reply content lacks context relevance and cannot handle complex problems. The method comprises: a room management intelligent agent receiving scroll information sent by a user in a live room, performing illegal risk analysis on the scroll information, shielding display of illegal scroll information, and inputting compliant scroll information to an assistant intelligent agent; the assistant intelligent agent identifies the scroll intention based on the compliant scroll information, divides the compliant scroll information into priority levels based on the scroll intention; the assistant intelligent agent determines the task resource occupation of different scroll information according to the priority of the compliant scroll information, and inputs the compliant scroll information to the anchor intelligent agent in order according to the priority of the compliant scroll information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, and in particular to an interaction method and system for real-time e-commerce transactions. BACKGROUND

[0002] Live streaming of goods is a business model that promotes and sells goods through a live streaming platform, combining the characteristics of e-commerce and live streaming, and gradually becoming an important trend in the development of e-commerce in China and even the world in recent years.

[0003] Live streaming of goods usually requires a human anchor to host live streaming interactions, which relies on the personal abilities of the anchor and is difficult to cover high-concurrency scenarios. Meanwhile, there are also live streaming scenarios that use virtual anchors to generate reply information based on a rule engine, but this requires a rule library maintained by humans, and the reply content lacks contextual relevance and cannot handle complex problems. SUMMARY

[0004] The embodiments of the present application provide an interaction method and system for real-time e-commerce transactions, which can solve the problem that current live streaming interactions rely on the personal abilities of anchors and are difficult to cover high-concurrency scenarios, or although reply information is generated based on a rule engine, a rule library maintained by humans is required, and the reply content lacks contextual relevance and cannot handle complex problems.

[0005] A first aspect of the embodiments of the present application provides an interaction method for real-time e-commerce transactions, comprising:

[0006] The housing intelligent agent receives the barrage information sent by the users in the live streaming room, analyzes the barrage information for irregular risks, screens and displays irregular barrage information, and inputs the compliant barrage information to the assistant intelligent agent;

[0007] The assistant intelligent agent identifies the barrage intent based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intent, and the barrage intent includes a purchase intent, a question intent, and a casual intent;

[0008] The assistant intelligent agent determines the processing task resource occupation of different barrage information according to the priority of the compliant barrage information, and inputs the compliant barrage information to the anchor intelligent agent in order according to the priority of the compliant barrage information, so that the anchor intelligent agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports contextual relevance.

[0009] Optionally, it further comprises:

[0010] In the case where the barrage intent is a question intent, the assistant intelligent agent classifies barrage questions;

[0011] The assistant agent narrows down the knowledge base range based on the barrage intention and question type, inputs the narrowed-down knowledge base range and the corresponding compliance barrage information to the host agent as the retrieval result.

[0012] Optionally, further comprising:

[0013] The assistant agent dynamically generates recommended commodity information and / or commodity combination information based on the user portrait of the live broadcast room user and the real-time compliance barrage information.

[0014] Optionally, further comprising:

[0015] The host agent performs live broadcast room mood analysis on the live broadcast room based on the user interaction data obtained by the room management agent, adjusts the associated live broadcast script based on the obtained live broadcast mood category, the live broadcast room mood state includes a lively live broadcast state, a cold and quiet live broadcast state, and an unattended live broadcast state, the lively live broadcast state is associated with a goods carrying script, the cold and quiet live broadcast state is associated with a preset performance script, and the unattended live broadcast state is associated with a live broadcast script for playing a preset video in a loop.

[0016] Optionally, further comprising:

[0017] After analyzing and determining the live broadcast mood category, the host agent calculates a sub-mood component based on the category of the user interaction data as a sub-channel, and performs mood contribution decomposition at a user-time-channel three-dimensional granularity to obtain a traceable mood source spectrum;

[0018] At the same time granularity, time sequence alignment is performed on the mood source spectrum and the mood source spectrum including at least one of the purchase behavior data streams of transaction, adding to cart, and receiving a coupon;

[0019] Based on the aligned mood source spectrum and purchase behavior data stream, mood-behavior causal consistency evaluation is performed, the index set of the consistency evaluation includes lag correlation coefficient, transfer entropy or Granger causality test pass rate of mood-behavior, and structured consistency score, wherein the transfer entropy or Granger causality test pass rate is used to determine the prediction power of the mood component on each behavior, and the structured consistency score is obtained by normalizing the sub-mood component indicators and summing the channel credibility weights;

[0020] When the structured consistency score is lower than a threshold value in a continuous preset time window, it is determined that there is a mood anomaly;

[0021] In the case where it is determined that there is a mood anomaly, the host agent performs abnormal source screening on the mood source spectrum to identify a hit feature source;

[0022] The sub-mood component corresponding to the hit feature source is subjected to weight reduction processing, and its contribution is inversely excluded from the mood source spectrum to obtain a purified sub-mood component.

[0023] perform live room emotion analysis on the live room based on the purified sub-emotion components to obtain a live emotion category.

[0024] Optionally, the method further comprises:

[0025] In the case of suspected emotional abnormalities, the anchor agent generates a low-friction verifiable behavior operation as a behavior probe, the low-friction verifiable behavior operation including at least one of issuing a coupon, sampling voting, or question interaction, to inversely measure the emotional abnormalities;

[0026] When the response degree of the behavior probe is lower than the historical baseline established based on membership data, the confidence of the emotional abnormalities is increased.

[0027] Optionally, the method further comprises:

[0028] The assistant agent performs sliding window monitoring on the round-trip delay, error rate, queue length, and packet loss rate of the large language model, and enters a degraded state or a disconnected state when at least one of the indicators exceeds a threshold value, and assigns a globally unique answer ID to each piece of to-be-answered bullet screen;

[0029] In the degraded state or disconnected state, the assistant agent does not wait for the large language model, and immediately generates a minimum available answer based on the commodity real-time knowledge base and associates the answer ID;

[0030] In the case of monitoring recovery to an available state or obtaining an idle computing power window, the assistant agent inputs the original bullet screen, the generated knowledge card, and the retrieved evidence to the anchor agent, so that the anchor agent calls the large language model to generate a complete answer corresponding to the answer ID.

[0031] The second aspect of the embodiments of the application provides an interactive system for real-time e-commerce transactions, comprising:

[0032] The housing management agent unit receives the bullet screen information sent by the live room users, analyzes the bullet screen information for violation risks, to shield and display the violation bullet screen information and input the compliant bullet screen information to the assistant agent;

[0033] The assistant agent unit identifies the bullet screen intent based on the compliant bullet screen information, divides the priority of the compliant bullet screen information based on the bullet screen intent, the bullet screen intent including a purchase intent, a question intent, and a casual intent, determines the different bullet screen information processing task resource occupation according to the priority of the compliant bullet screen information, and sequentially inputs the compliant bullet screen information to the anchor agent according to the priority of the compliant bullet screen information;

[0034] The anchor agent unit calls a large language model to generate feedback information of the received compliant bullet screen information, and the large language model supports context association.

[0035] The third aspect of the embodiments of the present application provides an electronic device, comprising a memory and a processor, wherein the processor is configured to execute the computer program stored in the memory to implement the steps of the real-time e-commerce transaction interaction method.

[0036] The fourth aspect of the embodiments of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the real-time e-commerce transaction interaction method.

[0037] In summary, the real-time e-commerce transaction interaction method provided by the embodiments of the present application receives the bullet screen information sent by the user in the live room through the house management agent, analyzes the risk of violating the rules of the bullet screen information, and screens and displays the bullet screen information in violation of the rules and inputs the compliant bullet screen information into the assistant agent; the assistant agent identifies the bullet screen intent based on the compliant bullet screen information, divides the priority of the compliant bullet screen information based on the bullet screen intent, and the bullet screen intent includes purchase intent, question intent and idle chat intent; the assistant agent determines the task resource occupation of different bullet screen information according to the priority of the compliant bullet screen information, and inputs the compliant bullet screen information in order according to the priority of the compliant bullet screen information to the anchor agent, so that the anchor agent calls a large language model to generate feedback information of the received compliant bullet screen information, and the large language model supports context association. Therefore, through the cooperation of intelligent house management, assistant and anchor agents, the system can efficiently screen and process a large number of concurrent bullet screen information. The house management agent is responsible for real-time filtering of illegal content, the assistant agent schedules task resources through intent recognition and priority division, and the anchor agent generates accurate replies through a large language model. Such a process ensures that all user interaction needs in the live process can be processed in a timely manner and will not be delayed due to high concurrency. High-priority purchase and question bullet screens are responded to in a timely manner, allowing users to quickly obtain the desired purchase information or answers, effectively improving user satisfaction and purchase conversion rate. Context-based natural dialogue and intelligent replies enhance the interactive experience and improve user engagement. Through intelligent resource allocation, the system can maintain smooth operation under high concurrency and ensure that high-priority tasks are processed in a timely manner. Low-priority tasks are processed later, avoiding excessive resource consumption. Through the intelligent cooperation of the house management agent, the assistant agent and the anchor agent, the system reduces manual intervention and reduces dependence on manual customer service and rule libraries, saving human resources. In addition, the introduction of a large language model enables the system to automatically generate high-quality replies that meet the context in real time, significantly improving the level of automated processing.

[0038] Correspondingly, the system, the electronic device and the computer readable storage medium provided by the embodiments of the present application also have the above technical effects. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A possible flowchart of an interactive method of real-time e-commerce transaction provided by the embodiments of the present application;

[0040] Figure 2 A possible schematic structural block diagram of an interactive system of real-time e-commerce transaction provided by the embodiments of the present application;

[0041] Figure 3 A possible schematic structural block diagram of an interactive system of real-time e-commerce transaction provided by the embodiments of the present application;

[0042] Figure 4 A possible schematic structural block diagram of an electronic device provided by the embodiments of the present application;

[0043] Figure 5 A possible schematic structural block diagram of a computer readable storage medium provided by the embodiments of the present application. DETAILED DESCRIPTION

[0044] The embodiments of the present application provide an interactive method of real-time e-commerce transaction and related devices, which can solve the problem that the current live interaction depends on the personal ability of the host, is difficult to cover high concurrency scenarios, or although the reply information is generated based on a rule engine, the rule library needs to be maintained manually, the reply content lacks context relevance, and complex problems cannot be handled.

[0045] The terms "first", "second", "third", "fourth" and the like (if any) in the description, claims and above drawings of the present application are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices. The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0046] Please refer to Figure 1A flowchart of an interaction method of a real-time e-commerce transaction provided by an embodiment of the present application can specifically include the following steps:

[0047] S110-S130.

[0048] S110, the housing intelligent agent receives the barrage information sent by the user in the live room, analyzes the risk of the barrage information, screens and displays the illegal barrage information, and inputs the compliant barrage information into the assistant intelligent agent.

[0049] S120, the assistant intelligent agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes purchase intention, question intention and chatting intention.

[0050] S130, the assistant intelligent agent determines the processing task resource occupation of different barrage information according to the priority of the compliant barrage information, and inputs the compliant barrage information into the host intelligent agent in order according to the priority of the compliant barrage information, so that the host intelligent agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association.

[0051] It can be understood that the above method realizes intelligent processing of the barrage information in the live room by introducing the cooperation of the housing intelligent agent, the assistant intelligent agent and the host intelligent agent, and combining the large language model, so as to optimize the interactive experience in the live streaming with goods.

[0052] For example, the house management agent first receives all the barrage information from the live room. After analyzing the risk of violation, the house management agent filters out inappropriate content such as insults, advertisements, or illegal information. Using natural language processing (NLP) technology, the house management agent can quickly identify content that violates platform rules and block it. Through real-time monitoring, illegal content is removed, and only compliant barrage information continues to flow into the subsequent processing flow. This step avoids manual intervention and reduces the workload of live room management. Real-time filtering of illegal barrage improves the quality of live interaction, ensuring the safety and compliance of live content, and enhancing the credibility of the platform. For example, the barrage "This lipstick is beautiful!" is determined to be compliant and continues to be processed. The barrage "Why are you so ugly? Change your clothes!" is determined to be illegal and is blocked. Then, for the compliant barrage information transmitted by the house management agent, the assistant agent classifies it according to the barrage intent. Through keyword extraction and semantic analysis of the barrage content, the assistant agent can identify the user's specific needs, such as purchase intent, question intent, and casual intent. This step uses deep learning and natural language understanding (NLU) technology to analyze the sentiment, keywords, and context in the barrage to accurately determine the user's intent. According to the intent of the barrage, the assistant agent classifies its priority. The barrage with purchase intent has the highest priority, followed by the barrage with question intent, and finally the barrage with casual intent. Through priority classification, the system can intelligently allocate resources to ensure that users' purchase needs are responded to in a timely manner, while non-urgent casual content can be processed later. Through accurate intent recognition, the system can prioritize processing barrage with purchase and question intent according to user needs, avoiding the occupation of system resources by a large amount of low-priority casual information. This not only improves the response efficiency of the system, but also enhances the user's interactive experience and increases the conversion rate. For example, the barrage "How much is this lipstick?" is identified as purchase intent and has high priority. The barrage "How are you feeling today, host?" is identified as casual intent and has low priority. After the assistant agent completes intent recognition and priority classification, it allocates system processing resources according to the priority of each barrage. High-priority barrages such as purchase intent will occupy more resources to ensure that users' needs are responded to quickly. Low-priority barrages such as casual chat will occupy fewer resources and can be processed later according to system load. The assistant agent adjusts the task processing order flexibly according to the current resource situation to optimize resource allocation. In this way, dynamic adjustment of resource allocation ensures that the system can still run smoothly in high-concurrency scenarios, avoiding delays in responding to users' critical needs due to excessive resource concentration on low-priority tasks. Through reasonable allocation of resources, the stability and processing capacity of the system are improved. For example, the barrage "Can this bag be more affordable?" is identified as purchase intent and occupies a large amount of resources, so it is processed first. The barrage "What did everyone eat tonight?" is identified as casual intent and has low priority, so it can be processed later.Then, according to the task assigned by the assistant agent, the anchor agent inputs the compliant and high-priority fan information into the large language model. The large language model supports context-related processing, that is, it not only generates single responses, but also generates more natural and smooth feedback according to the content in the conversation history. This means that when users ask questions or communicate in the fan, the system can understand the context and avoid isolated answers. The feedback information generated by the system will be fed back to the user, such as product recommendations, price explanations, use skills, etc. The above-mentioned large language model can generate reasonable and detailed answers based on the understanding of the context for complex questions. For example, if a user asks about the details of a product multiple times, the system can provide more in-depth answers based on the previous question content. Here, the system can generate more natural and interactive feedback by using the context understanding ability of the large language model, improving the user experience of live streaming sales. Compared with the fixed answers generated by traditional rule engines, the introduction of large language models makes the interaction more flexible and personalized, and can handle more complex user questions. For example, the user sends a fan: "What is the color of this lipstick?", which is identified as a question intention, and the system generates a response according to the context. The system replies: "This lipstick is rose red, very suitable for daily and dinner use. You can try the color." The user sends a fan: "Can I use a discount on this lipstick?". It is identified as a purchase intention, and the system generates a response according to the user's intention. The system replies: "This lipstick is on sale for a limited time, originally 199 yuan, now only 159 yuan, click on the link below to purchase!"

[0053] In summary, the real-time e-commerce transaction interaction method provided by the embodiments of the present application is as follows: the housing intelligent agent receives the barrage information sent by the live room user, analyzes the barrage information for irregular risk, screens and displays the irregular barrage information, and inputs the compliant barrage information to the assistant intelligent agent; the assistant intelligent agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes a purchase intention, a question intention, and a casual intention; the assistant intelligent agent determines the different barrage information processing task resource occupation according to the priority of the compliant barrage information, and sequentially inputs the compliant barrage information to the host intelligent agent according to the priority of the compliant barrage information, so that the host intelligent agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association. Therefore, through the irregular barrage screening of the housing intelligent agent, the intention recognition and priority division of the assistant intelligent agent, the system can quickly respond to the key needs of users in a high-concurrency situation, and avoid system overload. The system realizes context-associated answers through a large language model, making the interaction more personalized and natural, especially when dealing with complex problems, it can provide multi-round dialogue support, improving user satisfaction and participation. The assistant intelligent agent dynamically adjusts the task priority and resource allocation to ensure that high-priority tasks are processed first and low-priority tasks are reasonably delayed, optimizing the system's resource utilization and improving the system's stability. Through the multi-role intelligent barrage coordination processing mode, the participation demand of artificial customer service is reduced, the artificial cost is reduced, the frequent updating and maintenance of the rule engine are avoided, and the self-operation ability of the system is improved.

[0054] In some examples, further comprising:

[0055] In the case where the barrage intention is a question intention, the assistant intelligent agent classifies the barrage question;

[0056] The assistant intelligent agent narrows the knowledge base range based on the barrage intention and the question type, and inputs the search result or the narrowed knowledge base range together with the corresponding compliant barrage information to the host intelligent agent.

[0057] For example, when the assistant agent identifies a certain barrage as a question intent, the next task is to categorize the question. Question categorization refers to classifying the question type based on its content, such as product-related questions, after-sales service questions, and event information questions. Through natural language processing technology, the assistant agent can extract keywords and emotional colors from the question and classify it based on its semantic meaning. For example, for the question "How long does this lipstick last?", the system can categorize it as a "product attribute question", while "What is the return policy?" is categorized as an "after-sales service question". Categorized questions help the system accurately locate the problem domain and narrow down the knowledge base search range, improving search efficiency. Through question categorization, the system can more accurately locate the domain to which each question belongs and reduce irrelevant information interference. This refined processing allows the system to be more targeted and efficient in answering.

[0058] For example, after categorizing the question, the assistant agent will narrow down the knowledge base range according to the question type, thereby improving the efficiency of subsequent retrieval and answering. The knowledge base contains a large amount of product information, after-sales policies, and activity rules. The system narrows down the search range according to the type of question to avoid redundant operations of full database queries, further improving processing speed and response accuracy. For example, if the question is a "product attribute question", the system will only search in the product-related part; if the question is an "after-sales service question", the system will only search in the after-sales service part. This approach reduces unnecessary data searching, quickly locates relevant information, and then generates appropriate feedback. The narrowed search results are passed to the anchor agent along with the corresponding compliant barrage information to ensure that the system-generated response is not only accurate but also maintains coherence with the barrage content. This knowledge base range narrowing method greatly improves information retrieval efficiency and reduces resource waste. The system can quickly find the most relevant information from a vast amount of data, ensuring fast and accurate feedback generation. In addition, this also provides a better resource management solution for barrage processing in high-concurrency scenarios.

[0059] For example, after receiving the reduced knowledge base result and compliance bullet screen information, the host agent generates appropriate answers using a large language model. The large language model not only understands single questions, but also generates relevant and natural replies based on the context of the question. Because the question has been classified and the knowledge base has been reduced, the generated feedback information can more accurately respond to the user's needs. Through this process, the system can quickly and accurately return appropriate answers after the user asks a question without human intervention. The large language model provides flexible context processing capabilities, making the reply content more personalized and natural, and enabling it to handle complex user needs. For example, bullet screen: "How is the durability of this lipstick?", the system generates feedback: "This lipstick has super durability and can maintain perfect color for up to 12 hours, perfect for daily use and evening parties." Bullet screen: "How to apply for a refund?", the system generates feedback: "Dear user, you can apply for a refund through our APP or contact our customer service team to help you handle the refund matter."

[0060] It can be understood that in this way, accurate question intention recognition and problem classification can be achieved, and the assistant agent can quickly locate the problem area through problem classification technology, reduce the search range, and improve the system response speed. The knowledge base range can be reduced, and the search efficiency can be improved. The system reduces the knowledge base query range according to the problem type to ensure efficient and relevant search process and avoid redundant information. And efficiently generate feedback information, through the context processing capability of the large language model, the system can generate accurate and natural answers to improve user experience and interaction satisfaction.

[0061] In some examples, further comprising:

[0062] The assistant agent dynamically generates recommended product information and / or product combination information based on the user portrait of the live room user and the real-time compliance bullet screen information.

[0063] For example, when a user sends a compliant comment information in a live room, the assistant agent will analyze the user portrait and real-time comment information simultaneously to infer the user's interests and needs. The user portrait contains the user's historical purchase records, browsing behavior, interaction preferences, and other information, reflecting the user's consumption habits and interest points. Real-time compliant comment information provides the user's current interest points and immediate needs. The assistant agent combines these two pieces of information to dynamically generate product recommendations related to the user's interests. For example, if a user mentions a certain product or asks about related attributes in the comment, the assistant agent will recommend matching products or product combinations based on this information combined with the user's interest labels in the user portrait. This recommendation method based on user portraits and real-time interaction data can significantly improve the accuracy and personalization of recommendations, making it easier for users to discover products of interest, increasing purchase intent, and improving conversion rates. Through an intelligent recommendation system, the assistant agent can more flexibly adapt to the needs of different users, improving the effectiveness of live streaming e-commerce. Suppose a user frequently purchases skincare products and shows interest in beauty products. If the user sends a comment: "What are the features of this lipstick?" the assistant agent can recommend related lipstick or other beauty products based on the user's portrait and comment information, such as: "Based on your interests, we recommend this lip balm, which has similar color effects to the lipstick you previously purchased." If the user frequently focuses on health and sports products and asks in the live room: "Do you have any recommendations for sports shoes?" the assistant agent may recommend a pair of sports shoes based on the user's portrait and suggest related sports equipment, such as: "We recommend this running shoe for you, and you can also use this sports sock to improve comfort."

[0064] For example, in addition to single product recommendations, the assistant agent can also dynamically generate product combination information based on the user's interests and needs. This combination recommendation can be based on the products mentioned by the user during the live broadcast, the correlation between products, or some purchase trends analyzed through historical data. For example, a certain skincare product is often purchased together with other products, and the system can generate combination recommendations for related products based on these purchase behaviors. The assistant agent will analyze the user's historical purchase habits or combine real-time interaction information to automatically generate a set of product recommendations. For example, if the user asks about a certain lipstick in the bullet screen, the system will not only recommend the lipstick, but also recommend other makeup products that match it, such as lipsticks, lip liners, makeup mirrors, etc. Through product combination recommendations, users can easily discover other products related to their interests, increasing the opportunity for cross-category purchases and increasing sales. Dynamically generating product combination information makes the recommendation system more intelligent and flexible, and can generate the most relevant product package based on different user behaviors. For example, user bullet screen: "This lipstick is beautiful, is it suitable for summer use?", Based on the user's historical purchase data and interest tags, the assistant agent will recommend products that match it, such as summer skincare products, facial sprays, etc. System recommendation: "According to your needs, we recommend this summer sunscreen, which matches this lipstick very well. You can also use this primer to make your makeup last longer."

[0065] For example, the assistant agent not only makes recommendations based on real-time compliance bullet screens and user portraits, but also continuously optimizes the recommendation effect based on user feedback, clicks, purchase behaviors, and other data. Each time the user interacts with the recommended product, the system will re-evaluate the user's interest points based on new behavior data and update the user portrait. Through machine learning and recommendation algorithms, the assistant agent can continuously adjust the recommendation strategy to ensure that each recommendation is more accurate. As a result, continuous optimization of recommendations can adapt to changes in user interests, thereby improving user satisfaction and reducing cold start problems. Over time, the recommendation system will become more intelligent and provide more personalized product recommendations to users, improving user stickiness and purchase conversion rates. For example, if a user frequently asks about skincare products over a period of time, the system will adjust the user portrait and increase the recommendation of skincare-related products. When the user purchases a skincare cream, the system will automatically recommend products such as masks, serums, etc. that match it, ensuring that the recommended content continues to meet the user's needs.

[0066] It can be understood that by combining user portraits and real-time bullet screen information, the present scheme can intelligently recommend products that match the user's interests, improving the personalization and accuracy of recommendations. At the same time, based on real-time interaction, product combination information is generated, further increasing the user's purchase intention and conversion rate. In addition, the continuously optimized recommendation strategy ensures that the recommendation system can continuously improve as user interests change, providing more accurate services.

[0067] In some examples, further comprising:

[0068] The anchor agent performs live room mood analysis on the live room based on user interaction data obtained by the room management agent, adjusts the associated live script based on the obtained live mood category, the live room mood state includes lively live state, cold and clear live state and no one live state, the lively live state is associated with the goods carrying script, the cold and clear live state is associated with the preset performance script, and the no one live state is associated with the live script of the preset video in a loop.

[0069] For example, the room management agent is responsible for collecting and analyzing live room user interaction data, including but not limited to user sent barrage, likes, comments, sharing, participating in voting and other interactive behaviors. Through these interactive data, the room management agent can monitor the atmosphere changes of the live room in real time, collect key indicators related to audience emotions. For example, if the users in the live room frequently send positive comments and likes, and the barrage content is full of enthusiasm, the room management agent can identify that the emotional tendency at this time is positive and active; on the contrary, if the user interaction is less, and the comment content is cold or single, the system identifies the cold and clear emotional state. By collecting and analyzing interactive data in real time, the room management agent can provide accurate emotional analysis data for the anchor agent, so that the anchor agent can flexibly adjust the live broadcast strategy according to the user's participation and interaction state. This real-time feedback mechanism makes the live interaction more targeted, which can effectively improve the audience's participation and the atmosphere of the live room. After the room management agent counts the user interaction data of the live room, the anchor agent will perform emotional analysis based on these data to evaluate the current live room emotional state. The emotional state can be divided into three categories: lively live state: user interaction is frequent, barrage is active, audience emotion is high; cold and clear live state: user interaction is less, barrage content is relatively flat, lack of interactive enthusiasm; no one live state: there is almost no user interaction in the live process, and the system detects that the number of audience is extremely low. According to these emotional analysis results, the anchor agent will dynamically adjust the live script to ensure that the content display is adapted to the audience emotional state. In the lively live state, the system will associate the goods promotion script to stimulate the audience's purchase desire, and through interactive guidance, enhance the user's purchase decision. In the cold and clear live state, the system will introduce the preset performance script to activate the atmosphere through entertaining content or other interactive ways to attract the attention and participation of the audience. In the no one live state, in order to maintain the continuity of the live broadcast, the system will automatically play the preset video content to maintain the live broadcast, avoid the live room from being blank, and reduce resource occupation. After obtaining real-time emotional analysis, the anchor agent can continuously optimize the live content and interactive mode. When the live room emotional state changes, the anchor agent will immediately adjust the strategy. Lively state: in the case of high emotion, the system will continue to promote the goods promotion script to stimulate the purchase desire through multiple promotional recommendations. Cold state: when the emotion is low, the system automatically switches to the interactive entertainment link to attract audience participation and avoid long-term depression. No one state: if the audience interaction is extremely small, the system guides the audience to interact through barrage or other ways, or plays the preset content to maintain the smoothness of the live broadcast to reduce resource occupation and save cost. Through this mechanism, the system can dynamically adapt the live content to the audience emotion to ensure that the live broadcast always maintains a certain interactivity and audience participation. According to the feedback of different emotional states, the anchor agent can accurately adjust the live content to avoid the live broadcast from being in a downturn period, and improve the audience experience and conversion rate.

[0070] In some examples, further comprising:

[0071] The anchor agent calculates sub-emotion components based on the categories of the user interaction data after analyzing and determining the live broadcast emotion categories, respectively, to obtain a traceable emotion source spectrum at a user-time-channel three-dimensional granularity.

[0072] At the same time granularity, at least one of purchase behavior data streams including transaction, adding to cart, and receiving a coupon is time-aligned with the emotion source spectrum.

[0073] Based on the aligned emotion source spectrum and the purchase behavior data stream, an emotion-behavior causal consistency evaluation is performed, and the index set of the consistency evaluation includes a lag correlation coefficient of emotion-behavior, a transfer entropy or a Granger causality test pass rate, and a structured consistency score, wherein the transfer entropy or the Granger causality test pass rate is used to determine the prediction power of the emotion component on each behavior, and the structured consistency score is obtained by normalizing the sub-emotion component indicators and then summing the channel credibility weights.

[0074] When the structured consistency score is lower than a threshold value in a continuous preset time window, it is determined that there is an emotion anomaly.

[0075] In the case where it is determined that there is an emotion anomaly, the anchor agent performs an abnormal source screening on the emotion source spectrum to identify a hit feature source.

[0076] The sub-emotion component corresponding to the hit feature source is subjected to a weight reduction process, and its contribution is removed from the emotion source spectrum to obtain a purified sub-emotion component.

[0077] Based on the purified sub-emotion component, live broadcast room emotion analysis is performed on the live broadcast room to obtain a live broadcast emotion category.

[0078] It can be understood that the live room emotion is not a single quantity, but a time-varying signal formed by superposition of multiple user interaction sub-channels such as text barrage, likes, gifts, clicks and stays. The emotion channel refers to the classification dimension of user interaction data, and each channel independently calculates the sub-emotion component and contribution. The real commercial effective emotion should have a stable causal or at least predictable correlation with purchase-related behaviors such as transaction, add-to-cart and coupon redemption. If some sources such as script accounts create noise peaks on certain channels, it will lead to a higher emotion recognition but decoupling with purchase behavior. Therefore, the emotion can be first decomposed into a traceable source spectrum at the granularity of user, time and channel, and then aligned with the purchase behavior at the same granularity. The predictive power of emotion to behavior is measured by lag correlation, transfer entropy or Granger index, and aggregated into a structured consistency score. When the consistency is significantly low, the abnormal source screening is used to identify and weight the hit feature source, and after removing its contribution from the source spectrum, the emotion category is re-judged to ensure that the script switching and other decisions are based on the purified emotion consistent with the purchase behavior. Among them, the emotion source spectrum refers to the source representation of the sub-emotion component aggregated by user, time and channel within a given time window, which is used to trace the source composition of emotion peaks, locate abnormal sources and purify. The emotion source spectrum is a decomposed representation that can be held accountable, and is not equivalent to the final emotion category.

[0079] For example, the overall emotion vector can be decomposed into components of several sub-channels, and each component can be refined to a three-dimensional granularity of source account u, time slice t and channel c, so as to obtain a traceable source spectrum, which facilitates downstream noise source positioning and causal analysis. In practice, the system classifies the interaction data by channel, such as text barrage, likes, gifts, product card clicks, page stays and sharing, estimates the contribution of each interaction to the hot, cold and empty emotion dimensions within its time slice, and attaches the source account identifier and channel type to obtain the emotion source spectrum matrix. In this process, the text channel can be superimposed and cleaned, and the non-text channel can map the contribution according to the action intensity, such as high-value gifts being higher than ordinary likes. In this way, the black box emotion score is changed into a traceable and decomposed source set, providing a directly operable evidence carrier for subsequent screening and purification.

[0080] For example, to avoid false correlation or missed causality caused by the misalignment of emotion and behavior on the time axis, alignment within the same window is required, and natural lag is allowed. In practice, the purchase behavior events such as order, payment, add-to-cart and coupon redemption can be mapped to the same time slice as the emotion, and their intensity and association with SKU are recorded. At the same time, multi-scale windows such as 5s, 15s or 60s are retained to cover fast-paced and slow-heating categories. In this way, comparable and slidable data of the same scale are established for the next step of consistency evaluation, reducing the misjudgment caused by time jitter.

[0081] For example, if the emotion peak has stable prediction power on the future short-time behavior, the consistency is high. In practice, the correlation coefficient of each channel emotion component and behavior sequence at different lags is calculated respectively, and Granger or transfer entropy test is run to get the predictability score of each channel. Then, according to the channel prior credibility, such as click or stay being more credible for digital products, and gift being more credible for impulsive purchases, etc., the structured consistency score is obtained after normalization and weighting. Among them, the structured consistency score refers to a comprehensive index for measuring the predictability of emotions leading to purchase behavior. After normalization of sub-indices such as lag correlation, transfer entropy or Granger causality test, the score is obtained by weighted summation according to the channel credibility, which is used to judge the consistency of real heat and noise heat, and is not directly used as the basis for judging violations.

[0082] Therefore, high consistency means that the excitement is driven by real purchase interest, and low consistency points to emotional noise or manipulation. Then, threshold and continuous window are used for discrimination, which can avoid false positives caused by occasional jitter. The threshold can be set by the host according to the historical baseline of the category and time period, and a hysteresis interval is introduced. When the consistency score of the continuous M windows is lower than the threshold, and significantly lower than the historical same type quantile, it is determined that the emotion is abnormal. Therefore, only when the system is decoupled, the abnormality is triggered, reducing the false action in normal fluctuations. Subsequently, the above features can be calculated for each source within the window and compared with the hierarchical baseline, and the score is combined with rules, unsupervised anomaly and graph community density to obtain a comprehensive score. If it exceeds the high threshold, it is recorded as a hit feature source. Therefore, the anomaly is accurately located to the operable source set, providing specific targets for purification, and having auditability.

[0083] Exemplarily, the source contribution identified as abnormal can be physically removed or attenuated from the emotional component to obtain a purified signal closer to the true user intention. The component of the hit source can be weighted down or zeroed according to the risk level, while the evidence pointers such as the source, time slice, channel and reason are retained. After the elimination, the purified sub-emotional component is re-aggregated, and the difference before and after the purification is archived for review and offline parameter adjustment. In this way, the noise heat can be significantly reduced to pull up the overall emotion, and the accuracy of subsequent emotional rejudgment can be improved. Since the purification changes the statistical basis, the emotional probability needs to be recalculated and the hysteresis, residence and cooling strategies are used to prevent frequent jitter. The emotional probability vector and confidence can be calculated using the purified component, and the up-down switching threshold and the shortest residence time constraint can be used to switch the script. At the same time, the rejudgment result still needs to meet the consistency score after updating, and then the script can be switched from explanation / answering to goods carrying. In this way, the script switching and operation strategy are based on the purified and consistent emotion with the purchase behavior, avoiding being affected by false excitement. From a statistical point of view, the abnormal source contribution is a systematic bias term in the overall emotional signal. After the bias term is estimated and removed, the predictability of the related and causal indicators to the real purchase behavior is improved, the structured consistency score is increased, the rejudgment mode stably outputs the classification under the hysteresis and residence protection, and the strategy jitter is reduced. In this way, the live streaming room action triggered by the strong goods carrying script and other strategies occurs less in the scene where the behavior does not follow, which can bring higher order conversion, lower refund and lower risk of negative comments.

[0084] In some examples, further comprising:

[0085] In the case of suspected emotional abnormalities, the anchor intelligent agent generates a low-friction verifiable behavior operation as a behavior probe, the low-friction verifiable behavior operation including at least one of issuing a coupon, sampling voting or question interaction, to inversely measure the emotional abnormalities;

[0086] When the response degree of the behavior probe is lower than the historical baseline established based on membership data, the confidence of the emotional abnormalities is increased.

[0087] It can be understood that the operable events with very low cost, small participation threshold and weak but positive correlation with purchase decision, such as one-key coupon, single-choice voting or quick quiz, can be used to verify whether the emotional peak corresponds to the real purchase interest. If it is false excitement, the probe response will be significantly lower than the historical baseline of the crowd and time period. On the contrary, in the real interest scene, the response of the probe will have a causal or at least predictable synergistic relationship with subsequent behaviors such as clicking and adding to cart. In this way, the confidence of the emotional abnormalities is inversely measured and calibrated by the probe response, which can suppress the script switching misled by the manipulated emotion without interrupting the live streaming rhythm.

[0088] Exemplarily, in the case that the emotional-behavior consistency evaluation indicates suspected emotional abnormalities, the anchor agent generates low-friction verifiable behavior operations as behavior probes based on the current category, scene and audience crowd. The low-friction verifiable behavior operations include at least one of issuing coupons, sampling voting or question interaction. The selection of the probes follows the principles that the participation cost is below a preset threshold, has weak positive correlation with the purchase link, and has sufficient historical baseline in the current platform and category. The behavior probes are presented in floating cards, top strips or scrollable elements, and an effective response window of 5-15 seconds is set without blocking the core picture. The system records the probe exposure and effective response, and counts by user de-duplication. The response events include at least one of one-key coupon success, completed voting submission, question interaction selection or submission of valid answers. The anchor agent establishes a historical baseline based on membership data, which includes at least one dimension of anchor, category, time period, and event type such as (new, seconds or regular), crowd stratification (such as new fans, old fans, members, and first purchase in the past 30 days). To avoid bias caused by time domain and crowd structure differences, the historical baseline normalizes exposure, online number and crowd composition, and compares within the same probe type and similar price band / single interval. When the response rate of the behavior probe is lower than the historical baseline established based on the membership data, the confidence of the emotional abnormalities is increased, wherein the response rate is the ratio of the effective response of the probe to the exposure of the probe, and crowd weighting such as member weight higher than passerby can be introduced; when the response rate is lower than the quantile threshold of the baseline and continuously crosses the preset time window, the abnormal confidence is accumulated and updated according to the threshold amplitude and duration. To prevent occasional noise, the system can set a minimum sample size threshold and a double threshold strategy, such as low threshold weight reduction and high threshold shadow mute hit feature source. After the abnormal confidence is increased, the anchor agent maintains or reverts to the explanation or question script, and suppresses the triggering of strong sales scripts. In the subsequent time window, if a new round of consistency evaluation and probe response recover to above the baseline, the abnormal confidence is gradually reduced and the suppression is removed. To avoid disturbing the user with the probe, the system sets a cooling time and a frequency upper limit for the probe trigger, and switches the probe type when it does not reach the baseline for multiple times, such as changing from voting to one-key coupon. The anchor agent can launch random control probes to a small proportion of the audience to estimate the expected response when there is no abnormality at the moment. The centralized response from the same device fingerprint or high-similarity account group can be de-duplicated and weighted down to avoid interference from probe brushing. The system records the subsequent cooperation of probe-behavior, such as the increase of clicks or add-to-cart after the probe response, as the basis for offline calibration baseline and weight. Thus, when the emotional peak is caused by noise heat, the probe response as a kind of low-cost real action usually does not increase significantly, its response rate is significantly lower than the historical baseline of the membership stratification, and the abnormal confidence is increased, which is superimposed on the previous consistency indicators to form a more robust abnormal state judgment, thereby preventing the wrong strong sales script switching.When the emotional peak is driven by real interest, the probe response will approach or exceed the baseline, and in the subsequent short window, it will form a predictable synergy with clicks or adds to purchase, the system reduces the abnormal confidence, allows or promotes script switching. By stratifying the baseline and normalizing, responses under different time periods, different population compositions, and different probe types are compared fairly, reducing false positives, and by randomized control and anti-cheating de-duplication, the probability of interference by brush probes is reduced.

[0089] In some examples, further comprising:

[0090] The assistant agent monitors the round-trip delay, error rate, queue length, and packet loss rate of the large language model in a sliding window, and enters a degraded state or a disconnected state when at least one indicator exceeds the threshold value, and assigns a globally unique answer ID to each pending comment;

[0091] In the degraded state or disconnected state, the assistant agent does not wait for the large language model, and immediately generates a minimum available answer based on the real-time knowledge base of the commodity and associates it with the answer ID record;

[0092] In the case of monitoring recovery to the available state or obtaining an idle computing power window, the assistant agent inputs the original comment, the generated knowledge card, and the retrieved evidence into the host agent, so that the host agent calls the large language model to generate a complete answer corresponding to the answer ID.

[0093] For example, the pointer is for the quality of service of the large language model call service: the available state is that the delay, error rate, and queue length are all below the preset threshold; the degraded state is that at least one indicator exceeds the degraded threshold but the service is still available; the disconnected state is that the error rate or timeout or packet loss reaches the disconnection threshold, causing the service to be unavailable. The state switching adopts a hysteresis strategy.

[0094] The above describes the interaction method of real-time e-commerce transactions in the embodiments of the present application, and the following describes the interaction system of real-time e-commerce transactions in the embodiments of the present application.

[0095] In some cases, the rule-breaking comment often splits a sensitive word into fragments that are visually harmless and cannot be cut in word segmentation by using homonyms, zero-width characters, compatibility zone characters, emojis, or symbol cutting, etc. to bypass the rule library or model. However, in this case, the characters in the input layer do not match the visual framework in the rendering layer. Therefore, all comments can be first mapped to their visual framework representation, while calculating robust features such as character entropy and word segmentation stability. When the framework semantics and risk word vector angle are small, and the word segmentation is extremely unstable under different normalization strategies, it is marked as high-risk and enters the secondary review. The high-risk content is reorganized, such as the split word is spliced back, and it is reviewed whether it hits the rule-breaking word or the rule-breaking intent, thereby greatly improving the recognition rate of the split word attack without significantly increasing the average time delay.

[0096] For example, after receiving the live broadcast room barrage, the housing management agent first performs script and language rough identification on the barrage to determine the character set and language composition, which specifically includes statistics on the distribution proportion of different language systems, emojis and punctuation marks, determination of simplified and traditional Chinese, evaluation of mixed layout intensity, and rapid detection of whether there are high-frequency interlacing of emojis, extremely short sentences or abnormal character encoding. Since multi-language mixed layout and cross-script homograph are important appearance characteristics of evasive violations, only by identifying the language and script domain first can the correct homograph mapping table, compatible area regulation strategy and multi-language risk word vector library be selected in the subsequent normalization stage, thereby avoiding mapping unrelated characters to the wrong equivalence class, resulting in false positives or missed detection. The technical effect of this step is to provide accurate character domain priori for subsequent visual framework normalization, framework signature pre-screening and semantic vector verification, so that the recall rate and precision of downstream comparison can still remain stable under high concurrency conditions, and the calculation budget is concentrated on truly suspicious samples.

[0097] For example, after identifying the script and language, the system performs a hierarchical normalization pipeline on the barrage text, including NFKC compatibility regulation (such as folding full-width letters, decorative characters, combined subscripts and compatible area characters into standard forms), zero-width character and soft separator stripping (such as removing characters that are not visible to the naked eye but interfere with word segmentation), cross-script homograph mapping (such as using the confusables table to map а / α / a, etc. to a unified framework), grapheme cluster-based writing cluster segmentation to avoid being deceived at the diacritic and ligature level, and converting emojis to semantic token sequences that can participate in subsequent comparison, while retaining four outputs of original text, normalized text, skeleton sequence and emoji semantic sequence for tracing. The characters that are visually equivalent but different in encoding are regulated to the same visual framework, and the invisible word segmentation created by zero-width characters is eliminated, allowing the disassembled or disguised risk words to reassemble in the representation space. This significantly improves the stability and comparability of downstream matching, improves the initial screening hit rate of evasive text without significantly increasing the delay, while retaining the pre-post contrast for auditability to facilitate subsequent evidence chain construction.

[0098] Illustratively, the system computes a framework signature for the output grapheme sequence (e.g., merge-compress and hash the framework sequence to get a confusable-skeleton signature), and performs hash pre-screening and nearest-neighbor signature comparison in a risk word signature library maintained in buckets by language / script. Samples that hit or nearly hit are marked as high-risk and flow into a deep determination channel, while samples that do not hit are directly and quickly released. In this way, the signature is used to filter out text that is extremely similar in form in constant time, avoiding expensive semantic operations on a large number of normal barrage in a high-concurrency scenario. Thus, most irrelevant text is eliminated within milliseconds, focusing computing power and queue resources on risk samples, thereby controlling average review time while leaving enough throughput space for subsequent statistics and semantic determination.

[0099] Illustratively, for candidate samples that pass candidate filtering, the system computes three types of robust statistical features in parallel. First, the difference between the entropy of the character distribution and the uniqueness n-gram ratio before and after normalization. High-intensity perturbation often leads to abnormal entropy difference and diversity mutation. Second, word segmentation stability. Use three sets of cutters, rule-based segmentation, statistical segmentation, and subword BPE, to segment the original text and normalized text respectively, and compare the token boundary overlap and the segmentation consistency of key fragments. The word segmentation attack will make the boundaries under different strategies highly inconsistent. Third, window-based near-duplication. Use MinHash and template library to calculate the nearest neighbor similarity ratio and the templating intensity. Since the significant statistical anomalies of evasive text exist independently of the specific word form, even if the word has not been completely restored, the structural anomaly can be captured through the combination of high entropy difference, unstable word segmentation, and high near-duplication. Therefore, without relying on accurate word matching, the recall rate of suspicious samples is improved, providing high-confidence candidates for the next semantic determination, while reducing the risk of false positives due to similarity.

[0100] Illustratively, the system encodes the normalized text and emoji semantic labels together into sentence vectors and word vectors in a multi-lingual embedding space, and performs nearest neighbor and angle threshold determination with risk word vectors maintained in domain-specific libraries, such as insults, external link diversion, and medical efficacy, etc. At the same time, the structural anomaly signal captured is used as a double-condition trigger, i.e., the semantic neighbor is established and the statistical anomaly is significant, which enters the secondary review queue. If necessary, contextual short-window features, such as the coherence of the same user's recent speech or the theme of the same cluster, can be introduced to distinguish between jokes, legitimate discussions, and real violations. In this way, the similarity in form is upgraded to semantic similarity, and the proximity in semantic space is used to verify whether the content direction falls into the restricted semantic domain, and the statistical anomaly is used as a constraint to suppress the false positives of entertainment homophones and brand words. While ensuring recall, false positives are effectively reduced, so that real semantic variants of violations, masked expressions, and cross-language replacements can also be stably identified, while normal jokes and legal discussions are released.

[0101] For example, for samples entering the secondary review, the system performs a semantic reorganization process to attempt to automatically restore fragmented sensitive fragments, such as uniformly deleting soft separators and non-semantic symbols, close-packing candidate characters within a limited window, merging adjacent homophonic or homographic clusters, and calculating editing / pronunciation distances on three levels of Chinese characters, pinyin, and phonemes to filter reliable candidates. For each candidate, a semantic vector re-determination and rule base hit are performed again to confirm whether it constitutes a violation expression. This reverses the morphological fragmentation introduced by the attacker, re-exposes the target words hidden between symbols and expressions, and converts samples that cannot be qualitatively determined by statistical anomalies into violation texts that can be confirmed by both lexical and semantic evidence. This can significantly improve the final confirmation rate of word-breaking insults, diversion, or false efficacy claims. At the same time, due to the need for distance threshold and semantic re-determination to be met simultaneously, overfitting to normal text can be effectively suppressed.

[0102] Referring to Figure 2 An embodiment of an interactive system for real-time e-commerce transactions described in the embodiments of the present application can include:

[0103] The housing management agent unit 201 receives the barrage information sent by the live broadcast room user, analyzes the barrage information for violation risk, and screens and displays the violation barrage information and inputs the compliant barrage information to the assistant agent;

[0104] The assistant agent unit 202 identifies the barrage intent based on the compliant barrage information, classifies the compliant barrage information based on the barrage intent, and determines the different barrage information processing task resource occupation according to the priority of the compliant barrage information and sequentially inputs the compliant barrage information to the host agent according to the priority of the compliant barrage information, the barrage intent including a purchase intent, a question intent, and a casual intent.

[0105] The host agent unit 203 calls a large language model to generate feedback information for the received compliant barrage information, and the large language model supports context association.

[0106] In summary, the above-mentioned embodiments provide a real-time e-commerce transaction interaction system. The house management agent receives the barrage information sent by the live room user, analyzes the barrage information for irregular risk, screens and displays the irregular barrage information, and inputs the compliant barrage information to the assistant agent. The assistant agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes purchase intention, question intention and chatting intention. The assistant agent determines the task resource occupation of different barrage information according to the priority of the compliant barrage information, and inputs the compliant barrage information to the host agent in order according to the priority of the compliant barrage information, so that the host agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association. Therefore, through the irregular barrage screening of the house management agent, the intention recognition and priority division of the assistant agent, the system can quickly respond to the key needs of users in a high-concurrency situation and avoid system overload. The system realizes context-associated answers through a large language model, making the interaction more personalized and natural, especially when dealing with complex problems, it can provide multi-round dialogue support, improving user satisfaction and participation. The assistant agent dynamically adjusts the task priority and resource allocation to ensure that high-priority tasks are processed first and low-priority tasks are reasonably delayed, optimizing resource utilization and improving system stability. Through the multi-role intelligent barrage coordination processing mode, the participation of artificial customer service is reduced, the labor cost is reduced, the frequent updating and maintenance of the rule engine are avoided, and the self-operation ability of the system is improved.

[0107] The above Figure 2 From the perspective of modular functional entities, the real-time e-commerce transaction interaction system in the embodiments of the present application is described below from the perspective of hardware processing. Please refer to Figure 3 The real-time e-commerce transaction interaction system 300 in the embodiments of the present application includes:

[0108] The input device 301, the output device 302, the processor 303 and the memory 304, wherein the number of processors 303 can be one or more, Figure 3 In some embodiments of the present application, the input device 301, the output device 302, the processor 303 and the memory 304 can be connected through a bus or other means, wherein, Figure 3 In some embodiments of the present application, the input device 301, the output device 302, the processor 303 and the memory 304 can be connected through a bus or other means, wherein,

[0109] By calling the operation instructions stored in the memory 304, the processor 303 is used to execute the following steps:

[0110] The house management agent receives the barrage information sent by the user in the live broadcast room, analyzes the risk of the barrage information, and screens and displays the illegal barrage information and inputs the compliant barrage information into the assistant agent;

[0111] The assistant agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes a purchase intention, a question intention and a casual intention;

[0112] The assistant agent determines the different barrage information processing task resource occupation according to the priority of the compliant barrage information, and sequentially inputs the compliant barrage information into the host agent according to the priority of the compliant barrage information, so that the host agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association.

[0113] By calling the operation instructions stored in the storage 304, the processor 303 is also used to execute Figure 1 Any way in the corresponding embodiment.

[0114] Please refer to Figure 4 , Figure 4 The embodiment of the electronic device provided in the embodiment of the application is shown in the embodiment of the electronic device.

[0115] As Figure 4 shown, the embodiment of the application provides an electronic device, which includes a storage 410, a processor 420, and a computer program 411 stored in the storage 420 and executable on the processor 420, and the processor 420 implements the following steps when executing the computer program 411:

[0116] The house management agent receives the barrage information sent by the user in the live broadcast room, analyzes the risk of the barrage information, and screens and displays the illegal barrage information and inputs the compliant barrage information into the assistant agent;

[0117] The assistant agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes a purchase intention, a question intention and a casual intention;

[0118] The assistant agent determines the different barrage information processing task resource occupation according to the priority of the compliant barrage information, and sequentially inputs the compliant barrage information into the host agent according to the priority of the compliant barrage information, so that the host agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association.

[0119] In the specific implementation process, when the processor 420 executes the computer program 411, the following steps can be implemented Figure 1Any of the embodiments of the corresponding embodiments.

[0120] Since the electronic device introduced in the embodiment is the device used in the implementation of the interactive system for real-time electronic commerce transaction in the embodiment, based on the method introduced in the embodiment, those skilled in the art can understand the specific implementation of the electronic device in the embodiment and its various forms, so the implementation of the method in the embodiment by the electronic device will not be introduced in detail, as long as the device used in the implementation of the method in the embodiment by those skilled in the art belongs to the scope of protection of the present application.

[0121] Please refer to Figure 5 , Figure 5 An embodiment of a computer readable storage medium provided in the embodiment is shown in the figure.

[0122] As Figure 5 shown, the embodiment provides a computer readable storage medium 500, which stores a computer program 511, and the computer program 511 is executed by a processor to implement the following steps:

[0123] The housing intelligent agent receives the barrage information sent by the user in the live broadcast room, analyzes the risk of violation of the barrage information, screens and displays the illegal barrage information, and inputs the compliant barrage information into the assistant intelligent agent;

[0124] The assistant intelligent agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes purchase intention, question intention and chatting intention;

[0125] The assistant intelligent agent determines the processing task resource occupation of different barrage information according to the priority of the compliant barrage information, and inputs the compliant barrage information to the host intelligent agent in order according to the priority of the compliant barrage information, so that the host intelligent agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association.

[0126] By calling the operation instructions stored in the memory 304, the processor 303 is also used to execute Figure 1 Any of the embodiments of the corresponding embodiments.

[0127] The embodiments of the present application provide a computer program product, which comprises one or more computer instructions. When the computer instructions are loaded and executed on a computer, the computer instructions produce, wholly or partially, the flow or function described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be stored by the computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)), etc.

[0128] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0129] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0130] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiments of the present application.

[0131] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0132] When the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions, can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods according to the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various other media that can store program codes.

[0133] The above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the technical solutions thereof; even though the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still make modifications to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features thereof; and these modifications or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of interaction for real-time e-commerce transactions, characterized in that, The method comprises the following steps: The house management agent receives the barrage information sent by the user in the live broadcast room, analyzes the risk of the barrage information, and shields the display of the illegal barrage information and inputs the compliant barrage information into the assistant agent; The assistant agent identifies the barrage intention based on the compliant barrage information, divides the priority of the compliant barrage information based on the barrage intention, and the barrage intention includes purchase intention, question intention and casual intention; The assistant agent determines the processing task resource occupation of different barrage information according to the priority of the compliant barrage information, and inputs the compliant barrage information into the host agent in order according to the priority of the compliant barrage information, so that the host agent calls a large language model to generate feedback information of the received compliant barrage information, and the large language model supports context association; Further comprising: The host agent performs live room emotion analysis on the live room based on the user interaction data obtained by the house management agent, adjusts the associated live script based on the obtained live emotion category, the live room emotion state includes lively live state, cold live state and no one live state, the lively live state is associated with the goods carrying script, the cold live state is associated with the preset performance script, and the no one live state is associated with the live script of the loop playing preset video; After analyzing and determining the live emotion category, the host agent calculates the sub-emotion component based on the category of the user interaction data as a sub-channel, and performs emotion contribution decomposition at a user-time-channel three-dimensional granularity to obtain a traceable emotion source spectrum; Align the emotion source spectrum with the purchase behavior data stream including at least one of transaction, add-to-cart and receive a coupon in the same time granularity; Perform emotion-behavior causal consistency evaluation based on the aligned emotion source spectrum and purchase behavior data stream, the index set of the consistency evaluation includes lag correlation coefficient, transfer entropy or Granger causality test pass rate of emotion-behavior, and structured consistency score, wherein the transfer entropy or Granger causality test pass rate is used to determine the prediction power of the emotion component to each behavior, and the structured consistency score is obtained by normalizing the sub-emotion component index and weighting and summing according to the channel credibility; When the structured consistency score is lower than the threshold value in a continuous preset time window, it is determined that there is an emotion anomaly; In the case where it is determined that there is an emotion anomaly, the host agent performs abnormal source screening on the emotion source spectrum to identify the hit feature source; Perform weight reduction processing on the sub-emotion component corresponding to the hit feature source, and remove its contribution from the emotion source spectrum in reverse to obtain the purified sub-emotion component; Perform live room emotion analysis on the live room based on the purified sub-emotion component to obtain the live emotion category.

2. The method of claim 1, wherein, Further comprising: In the case where the barrage intention is a question intention, the assistant agent classifies the barrage question; The assistant agent narrows down the knowledge base range based on the barrage intention and question type, and inputs the search result or the narrowed down knowledge base range together with the corresponding compliant barrage information into the host agent.

3. The method of claim 1, wherein, Further comprising: The assistant agent dynamically generates recommended commodity information and / or commodity combination information based on the user portrait of the live room user and real-time compliance bullet screen information.

4. The method of claim 1, wherein, Also included are: In the case of suspected emotional abnormalities, the host agent generates a low-friction verifiable behavior operation as a behavior probe, including at least one of issuing a coupon, a sample vote, or a question interaction, to counter-pressure the emotional abnormalities; When the response of the behavior probe is lower than the historical baseline established based on membership data, the confidence of the emotional abnormalities is increased.

5. An interactive system for real-time e-commerce transactions, characterized by, The system is used to perform the method of any one of claims 1 to 4, and the system comprises: A housing management agent unit, the housing management agent receives the bullet screen information sent by the live room user, analyzes the bullet screen information for violation risk, and screens out and displays the violation bullet screen information and inputs the compliance bullet screen information to the assistant agent; An assistant agent unit, the assistant agent identifies the bullet screen intent based on the compliance bullet screen information, divides the priority of the compliance bullet screen information based on the bullet screen intent, and determines the different bullet screen information processing task resource occupation according to the priority of the compliance bullet screen information and sequentially inputs the compliance bullet screen information to the host agent according to the priority of the compliance bullet screen information, the bullet screen intent including purchase intent, question intent and casual intent; A host agent unit, the host agent calls a large language model to generate feedback information of the received compliance bullet screen information, and the large language model supports context association.

6. An electronic device, comprising: The electronic device includes at least one processor and at least one memory connected to the processor, wherein the processor is configured to call program instructions in the memory and execute the interactive method of real-time e-commerce transactions according to any one of claims 1 to 4.

7. A storage medium, characterized by The storage medium includes a stored program, wherein the program controls the device where the storage medium is located to execute the interactive method of real-time e-commerce transactions according to any one of claims 1 to 4 when the program is executed.

Images