AI field control intelligent review management method and system based on live broadcasting room
By combining RPA robots with large models, interactive data from live streaming rooms is collected across platforms for intelligent semantic analysis and room control management. This solves the problems of narrow data collection scope and single strategy in existing technologies, and achieves efficient, intelligent and compliant live streaming room management.
Patent Information
- Application Number
- CN202511854754.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-02-10
AI Technical Summary
Existing live streaming room control and intelligent feedback management methods have significant shortcomings in terms of efficiency, consistency, cost, and compliance. Existing automation tools are limited by platform interfaces, making it difficult to achieve cross-platform data acquisition, intelligent semantic analysis, and unified room control strategies.
By using RPA robots to simulate human operation and collect interactive data across platforms, combined with large-scale natural language processing models, semantic recognition and violation detection are achieved. An automatic feedback module that combines template-driven and large-scale models is designed, a multi-level priority field control mechanism is established, and the system is optimized through reinforcement learning.
It enables efficient cross-platform data collection and intelligent feedback, improves the targeting and intelligence of interactive processing, reduces operating costs, and enhances user interaction experience and management efficiency.
Smart Images
Figure CN121509697A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and live streaming management technology, specifically to an AI-based intelligent feedback management method and system for live streaming rooms. Background Technology
[0002] With the rapid development of the live streaming industry, live streaming rooms have become the core carriers for e-commerce marketing, content dissemination, and user interaction. The number of interactive behaviors initiated by viewers through bullet comments, comments, gift rewards, and following is growing exponentially, which puts forward increasingly stringent requirements for the real-time control and interactive response efficiency of live streaming rooms.
[0003] In current live streaming operations, moderators and interactive feedback primarily rely on manual operation. This model has significant technical drawbacks in large-scale concurrent scenarios: First, processing efficiency is low. The delay in manually reviewing audience interactions and executing responses is generally over 5-10 seconds, which cannot meet the real-time interaction needs of millions of viewers online simultaneously, easily leading to user churn due to excessive waiting time. Second, processing consistency is poor. Human judgment is highly subjective, and different moderators have significantly different response scripts and processing standards for the same type of audience interaction (such as product inquiries and feedback suggestions), making it difficult to establish unified live streaming operation standards and affecting the consistency of user experience. Third, operating costs are high. To cover the management needs of live streaming across multiple time periods and platforms, a large number of dedicated moderators are required. Furthermore, the recruitment, training, skills maintenance, and turnover management of these personnel all incur high costs, increasing the operational burden. Fourth, compliance risks are prominent. The identification and blocking of illegal content (such as abusive language, spamming advertisements, and sensitive information) by humans is inherently delayed, easily triggering platform compliance penalties and harming the viewing experience of ordinary users.
[0004] To alleviate the pressure of manual live streaming management, some automated live streaming control tools have emerged in the existing technology. However, most of these tools rely on the open interfaces of live streaming platforms to realize data interaction and function calls, and there are still limitations that are difficult to overcome: On the one hand, platform interface permissions are usually limited, only opening basic interactive data (such as comment text), and unable to obtain in-depth data such as gift reward frequency and user historical interaction behavior, resulting in a lack of comprehensive data support for live streaming control strategies; on the other hand, the interface protocols and functional scope of different live streaming platforms vary greatly, making it difficult for automated tools to achieve cross-platform compatibility. They need to develop adaptation modules separately for a single platform, resulting in poor scalability and failing to meet the actual needs of live streaming operators to manage multiple platforms simultaneously.
[0005] In summary, existing methods for managing live streaming sessions and intelligent feedback have significant shortcomings in terms of efficiency, consistency, cost, and compliance. Existing automation tools are also limited by platform interfaces, making it difficult to achieve universal and intelligent management results. In particular, there is a lack of a technical solution based on RPA (Robotic Process Automation) technology to overcome platform data collection limitations and combine natural language processing and large-scale models to achieve intelligent semantic analysis. This makes it difficult to simultaneously address the combined problems of difficulty in cross-platform data acquisition, rigid feedback content, and simplistic session management strategies. The industry urgently needs targeted technological innovation to improve the intelligence and automation level of live streaming session management. Summary of the Invention
[0006] This invention provides an AI-powered intelligent comment management method and system for live streaming rooms. It aims to overcome the limitations of live streaming platform interface permissions by using RPA robots to simulate manual operations and automatically collect cross-platform live streaming room interaction data, solving the problems of low data collection efficiency and poor platform compatibility in existing technologies. It also aims to introduce large-scale natural language processing technology to improve the accuracy of identifying audience comment intent, sentiment, and inappropriate content, overcoming the shortcomings of weak semantic understanding and delayed violation detection in existing solutions. Furthermore, it aims to design an automatic comment module combining template-driven and large-scale models to achieve rapid response to simple interactions and personalized replies to complex inquiries, improving the rigid logic and insufficient user experience of traditional automatic comment systems. Finally, it aims to establish a multi-level priority control mechanism to prioritize responses to high-value user interactions and automatically block malicious user behavior, solving the problems of single control strategies and unreasonable resource allocation in existing systems. Simultaneously, it aims to promote continuous iteration of system comment quality and control effectiveness through reinforcement learning and user feedback optimization mechanisms, ultimately achieving efficient, intelligent, and compliant live streaming room control and comment management, reducing operating costs and improving user interaction experience.
[0007] To achieve the above objectives, the first aspect of the present invention provides an AI-based intelligent feedback management method for live streaming, comprising:
[0008] (1) Interactive data collection: Collect live room audience interaction data across platforms through process automation components. The interactive data includes at least bullet screen, comments, gift rewards and follow behavior data. Write the collected data into the cache layer for real-time access and the database for historical analysis, and generate a collection log containing collection integrity and abnormal retry records.
[0009] (2) Natural Language Processing and Classification: Using a pre-trained language model, semantic recognition, sentiment analysis and violation detection are performed on the interactive data in sequence, and the interactive data is classified into positive feedback, question consultation and violation content;
[0010] (3) Automatic response and risk control in advance: For positive feedback, a standardized thank-you reply is generated based on a template; for questions and inquiries, a large model is used in conjunction with a knowledge base to generate a personalized reply; the reply is filtered by risk control rules before being pushed to the live broadcast room;
[0011] (4) Site control management and interlocking: Based on the preset priority strategy, differentiated responses are made to the interactions of different users, and automatic protection operations are triggered for interactions that are illegal; the site control protection results are sent back to the automatic feedback stage to block the feedback push associated with the illegal interaction, and the feedback record of the automatic feedback is synchronized to the site control management.
[0012] (5) Feedback optimization: Collect audience feedback data and historical interaction data on the feedback content, and adjust natural language processing parameters, feedback strategies and event control priority rules through reinforcement learning mechanism to achieve iterative optimization.
[0013] Furthermore, in step (1), the process automation component is an RPA robot, which completes cross-platform data access by simulating human operation and supports the adaptation of interactive data formats of mainstream live streaming platforms; the cache layer is a Redis cache, used to store real-time high-frequency interactive data; the database is a MySQL database, used to store structured historical interactive data; the collection log is generated once a day, which includes the data collection volume and data format verification results of each platform in addition to the collection integrity and abnormal retry records, forming a traceable logic for the entire data collection chain.
[0014] Furthermore, in step (2), the pre-trained language model includes a BERT model or a RoBERTa model; the violation detection is achieved through the collaboration of a keyword filtering list and a deep learning detection algorithm. First, obvious violations are initially identified based on the keyword filtering list, and then the ambiguous semantic content is detected a second time through a deep learning algorithm; the classification results need to be synchronized to the automatic feedback stage and the venue control management stage, wherein the classification results synchronized to the automatic feedback stage carry semantic recognition labels and sentiment labels, and the classification results synchronized to the venue control management stage carry violation level labels.
[0015] Furthermore, in step (3), the standardized thank-you replies include thank-you replies for gift rewards and welcome replies for follower behavior. The templates are stored according to the type of interaction and can be added or deleted based on the operational needs of the live broadcast room. When the large model generates personalized replies, it first retrieves business information related to the question consultation (such as product parameters and activity rules) from the knowledge base, and then adjusts the tone of the reply based on the emotional tendency of the interaction data. A lively tone is used for positive emotional tendencies, and a soothing tone is used for negative emotional tendencies. The risk control rule filtering includes sensitive word detection and screening of illegal expressions. Replies that fail the filtering need to be returned to the automatic reply stage to be regenerated.
[0016] Furthermore, in step (4), the preset priority strategy is formulated with user value as the core. High-value users include at least VIP users, users whose gift reward amount reaches the set standard within a preset time period, and users whose historical interaction frequency is higher than the set threshold. The automatic protection operation includes content blocking, user warning, account banning, and account kicking. Warning and content blocking are performed for users who violate the rules for the first time, and account banning or kicking is performed for users who violate the rules repeatedly. The site control management also supports the configuration of scenario-based strategies. In holiday scenarios, the response priority of high-value user interactions is increased, and in promotional activity scenarios, the detection intensity of advertising-related illegal content is enhanced.
[0017] Furthermore, in step (5), the feedback data includes the like rate of the reply content, the duration of subsequent interaction after the audience receives the reply, and the number of new comments caused by the reply content; the reinforcement learning mechanism first extracts features from the feedback data and historical interaction data, and then adjusts the semantic recognition weight of natural language processing, the template call priority of automatic reply, and the violation detection threshold of the scene control management based on the extracted features; the adjusted parameters need to be verified in a small-scale test live room first, and after verification, they are updated to all live rooms according to the set cycle.
[0018] Furthermore, the interlocking logic between steps (4) and (3) also includes: when the site control management detects that the same user initiates multiple violations within a set time, in addition to triggering automatic protection operations, it will also generate a violation tag for the user and synchronize it to the interaction data collection stage. Subsequently, after the user's interaction data is collected, it will be pushed to the site control management stage for secondary verification to prevent violations from entering the natural language processing and automatic feedback stage.
[0019] Another aspect of the present invention is an AI-based intelligent feedback management system for live streaming, used to implement the above-mentioned method, comprising:
[0020] Data acquisition module: It is equipped with a process automation component for cross-platform acquisition of live stream audience interaction data. The interaction data includes at least bullet screen, comments, gift rewards and follow behavior data. The acquired data is written to the cache layer and database in the data storage module respectively to adapt to real-time call and historical analysis requirements. At the same time, it generates acquisition logs containing acquisition integrity and abnormal retry records.
[0021] Natural Language Processing Module: Connected to the data acquisition module, configured with a pre-trained language model, used to sequentially perform semantic recognition, sentiment analysis and violation detection on the interactive data, and classify the interactive data into positive feedback, questions and inquiries and violation content;
[0022] Automatic response module: Connected to the natural language processing module, including a template storage unit and a large model calling unit. The template storage unit is used to generate standardized thank-you replies based on positive feedback, and the large model calling unit is used to generate personalized replies to inquiries by combining the knowledge base. The automatic response module is also equipped with a risk control filtering unit, which is used to perform risk control rules to filter the reply content before the reply is pushed to the live broadcast room.
[0023] The session control management module is connected to the natural language processing module and the automatic feedback module, and includes a priority configuration unit and an automatic protection unit. The priority configuration unit is used to respond differently to the interactions of different users based on a preset priority strategy. The automatic protection unit is used to trigger automatic protection operations for interactions that contain illegal content. The session control management module sends the protection results back to the automatic feedback module to gate and block feedback associated with illegal interactions before pushing them, and receives feedback records synchronized by the automatic feedback module.
[0024] Feedback optimization module: Connected to the automatic feedback module, the venue management module, and the data storage module, including a feedback acquisition unit and a reinforcement learning unit. The feedback acquisition unit is used to collect audience feedback data and historical interaction data on the feedback content. The reinforcement learning unit is used to adjust the parameters of the natural language processing module, the feedback strategy of the automatic feedback module, and the priority rules of the venue management module through the reinforcement learning mechanism to achieve iterative optimization of the system.
[0025] Data storage module: Connects to each of the above modules and is used to store the collected interactive data, the classification results of the natural language processing module, the response content of the automatic feedback module, and the feedback data of the feedback optimization module. The data storage module includes a cache layer and a database, which are adapted to the needs of real-time data access and historical data storage, respectively.
[0026] Furthermore, the process automation component in the data acquisition module is an RPA robot, which simulates manual operation to achieve cross-platform data access and adapts to multi-platform interactive data formats; the caching layer is Redis, used to store real-time high-frequency interactive data, and the database is MySQL, used to store structured historical interactive data; the acquisition log generated by the data acquisition module is output daily, and in addition to acquisition integrity and abnormal retry records, it also includes the data acquisition volume and data format verification results of each platform, so as to form a traceable logic for the entire data acquisition chain.
[0027] Furthermore, the natural language processing module and the automatic feedback module have the following internal units and collaborative relationships:
[0028] The pre-trained language model unit in the natural language processing module uses BERT or RoBERTa for semantic recognition and sentiment analysis; its violation detection unit is composed of a keyword filtering subunit and a deep learning detection subunit. It first identifies obvious violations based on the keyword filtering list, and then the deep learning detection subunit performs a second detection on ambiguous semantic content; its result labeling and distribution unit is used to attach semantic recognition labels and sentiment tendency labels when synchronizing with the automatic feedback module, and to attach violation level labels when synchronizing with the field control management module.
[0029] The automatic response module's template storage unit stores standardized thank-you response templates categorized by interaction type, including templates for thanking gifts and templates for welcoming followers. It also supports adding or deleting templates based on livestream operation needs. When generating personalized responses, the large model calling unit first retrieves business information related to the inquiry from the knowledge base sub-unit, then adjusts the response tone based on the sentiment of the interaction data, using a lively tone for positive sentiments and a soothing tone for negative sentiments. The risk control filtering unit's filtering rules include sensitive word detection and screening for inappropriate expressions; responses that fail filtering trigger a regeneration process.
[0030] The present invention adopts the above technical solution, and the specific beneficial effects are as follows:
[0031] In this invention, cross-platform live streaming interactive data collection is achieved through process automation components (such as RPA robots), breaking through the limitations of traditional reliance on platform open interfaces. At the same time, the layered storage of the caching layer and database, as well as the full-link collection logs, not only ensure the real-time calling efficiency of high-frequency data, but also realize the traceable management of historical data, solving the problems of narrow data collection scope, chaotic storage and difficulty in tracing the source in the existing technology.
[0032] In this invention, a pre-trained language model is used to sequentially perform semantic recognition, sentiment analysis, and violation detection on interactive data. Differentiated labels are output for the automatic response module and the venue control management module, so that the response can match the user's emotional tendency and the venue control operation can be accurately classified. This effectively improves the targeting and intelligence of interactive processing and avoids the defects of homogeneous response and low efficiency in traditional manual processing.
[0033] In this invention, the automatic response module adopts a dual mode of "standardized response generated by template and personalized response generated by large model combined with knowledge base". Before pushing, it is filtered by risk control rules, which not only ensures a rapid response to positive feedback, but also ensures the professionalism of the response to the inquiry, while avoiding the risk of non-compliant response, thus balancing the efficiency of response and the compliance of content.
[0034] In this invention, the live room control management module and the automatic feedback module form an interlocking logic. The live room control protection results can be transmitted back to block the push of illegal related feedback. The automatic feedback records are synchronized to the live room control management, which avoids the generation of invalid feedback and the failure to manage illegal interactions. It realizes the coordinated linkage between live room feedback and live room control, and improves the overall management efficiency.
[0035] In this invention, the feedback optimization module uses a reinforcement learning mechanism to adjust natural language processing parameters, feedback strategies, and session control rules based on audience feedback data and historical interaction data. This enables the system to continuously iterate and optimize without frequent manual adjustments, adapting to the changing needs of different live streaming scenarios such as holidays and promotional activities, thus enhancing the system's adaptability and long-term usability.
[0036] In this invention, the data storage module adopts a combined architecture of Redis caching and MySQL database, which respectively supports real-time high-frequency data calls and structured historical data storage. With end-to-end processing latency control, it can adapt to concurrent scenarios with millions of viewers online at the same time, solving the problems of poor real-time performance and data processing lag in existing systems under high concurrency. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is the overall system architecture diagram of the present invention;
[0039] Figure 2 It is a module logic interaction process based on the AI-powered intelligent feedback management system for live streaming rooms;
[0040] Figure 3 It is based on the full-link logical interaction of the AI-powered intelligent feedback management system for live streaming rooms;
[0041] Figure 4 It is the logical interaction process between the natural language processing module and the automatic feedback module. Detailed Implementation
[0042] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.
[0043] Example 1
[0044] like Figure 1 As shown, this embodiment provides an AI-based intelligent feedback management method for live streaming rooms, including:
[0045] (1) Interactive data collection: Collect live room audience interaction data across platforms through process automation components. The interactive data includes at least bullet screen, comments, gift rewards and follow behavior data. Write the collected data into the cache layer for real-time access and the database for historical analysis, and generate a collection log containing collection integrity and abnormal retry records.
[0046] (2) Natural Language Processing and Classification: Using a pre-trained language model, semantic recognition, sentiment analysis and violation detection are performed on the interactive data in sequence, and the interactive data is classified into positive feedback, question consultation and violation content;
[0047] (3) Automatic response and risk control in advance: For positive feedback, a standardized thank-you reply is generated based on a template; for questions and inquiries, a large model is used in conjunction with a knowledge base to generate a personalized reply; the reply is filtered by risk control rules before being pushed to the live broadcast room;
[0048] (4) Site control management and interlocking: Based on the preset priority strategy, differentiated responses are made to the interactions of different users, and automatic protection operations are triggered for interactions that are illegal; the site control protection results are sent back to the automatic feedback stage to block the feedback push associated with the illegal interaction, and the feedback record of the automatic feedback is synchronized to the site control management.
[0049] (5) Feedback optimization: Collect audience feedback data and historical interaction data on the feedback content, and adjust natural language processing parameters, feedback strategies and event control priority rules through reinforcement learning mechanism to achieve iterative optimization.
[0050] As one implementation method, in step (1) of this embodiment, the process automation component is an RPA robot. The RPA robot completes cross-platform data access by simulating human operation and supports the adaptation of interactive data formats of mainstream live streaming platforms. The cache layer is a Redis cache, which is used to store real-time high-frequency interactive data. The database is a MySQL database, which is used to store structured historical interactive data. The collection log is generated once a day. In addition to the collection integrity and abnormal retry records, it also includes the data collection volume of each platform and the data format verification results, forming a traceable logic for the entire data collection chain.
[0051] As one implementation method, in step (2) of this embodiment, the pre-trained language model includes a BERT model or a RoBERTa model; the violation detection is achieved through the collaboration of a keyword filtering list and a deep learning detection algorithm. First, obvious violations are initially identified based on the keyword filtering list, and then the ambiguous semantic content is detected a second time through a deep learning algorithm; the classification results need to be synchronized to the automatic feedback stage and the venue control management stage, wherein the classification results synchronized to the automatic feedback stage carry semantic recognition labels and sentiment labels, and the classification results synchronized to the venue control management stage carry violation level labels.
[0052] As one implementation method, in step (3) of this embodiment, the standardized thank-you reply includes a thank-you reply for gift rewards and a welcome reply for follower behavior. The templates are stored according to the interaction type and can be added or deleted based on the needs of live room operation. When the big model generates personalized replies, it first retrieves business information related to the question consultation (such as product parameters and activity rules) from the knowledge base, and then adjusts the reply tone according to the emotional tendency of the interaction data. A lively tone is used for positive emotional tendency and a soothing tone is used for negative emotional tendency. The risk control rule filtering includes sensitive word detection and illegal expression screening. Replies that fail the filtering need to be returned to the automatic reply stage to be regenerated.
[0053] As one implementation method, in step (4) of this embodiment, the preset priority strategy is formulated with user value as the core. High-value users include at least VIP users, users whose gift reward amount reaches the set standard within a preset time period, and users whose historical interaction frequency is higher than the set threshold. The automatic protection operation includes content blocking, user warning, account banning, and account kicking. Warning and content blocking are performed for users who violate the rules for the first time, and account banning or kicking is performed for users who violate the rules repeatedly. The site control management also supports the configuration of scenario-based strategies. In holiday scenarios, the response priority of high-value user interactions is increased, and in promotional activity scenarios, the detection intensity of advertising-related illegal content is enhanced.
[0054] As one implementation method, in step (5) of this embodiment, the feedback data includes the like rate of the reply content, the duration of subsequent interaction after the audience receives the reply, and the number of new comments caused by the reply content; the reinforcement learning mechanism first extracts features from the feedback data and historical interaction data, and then adjusts the semantic recognition weight of natural language processing, the template calling priority of automatic reply, and the violation detection threshold of the scene control management based on the extracted features; the adjusted parameters need to be verified in a small-scale test live room first, and after the verification is passed, they are updated to all live rooms according to the set period.
[0055] As one implementation method, the interlocking logic of steps (4) and (3) in this embodiment also includes: when the site control management link detects that the same user initiates multiple illegal interactions within a set time, in addition to triggering automatic protection operation, it will also generate the user's illegal tag and synchronize it to the interaction data collection link. After the user's interaction data is collected, it will be pushed to the site control management link for secondary verification to prevent illegal interactions from entering the natural language processing and automatic feedback link.
[0056] To further illustrate this point, a specific implementation example is provided:
[0057] This embodiment is a live streaming AI-driven automatic feedback management system based on RPA data collection, multimodal analysis, and large model-driven approaches. It includes the following modules:
[0058] System technical architecture modules:
[0059] The architecture adopts a distributed microservice architecture, with the backend based on Spring Boot to implement business logic and the frontend based on the Vue framework to implement real-time interactive display and configuration management.
[0060] Data storage uses a MySQL database, and the caching layer uses Redis to ensure the ability to handle millions of concurrent requests.
[0061] RPA robots are deployed at the data acquisition layer, capturing interactive data across platforms by simulating human operation, and are compatible with multiple live streaming platforms.
[0062] Data acquisition and processing module:
[0063] The RPA robot collects real-time data such as live chat messages, comments, gifts, and followers, with a latency of less than 1 second.
[0064] It provides an exception handling mechanism, automatically retrying 3 times if data collection fails, with a data missing rate of <1%.
[0065] The logging mechanism generates daily collection reports to ensure data integrity and traceability.
[0066] Natural Language Processing Module:
[0067] Semantic recognition and sentiment analysis based on large models (such as BERT and RoBERTa);
[0068] By combining keyword filtering with deep learning detection algorithms, we can identify illegal content with an accuracy rate of ≥95%.
[0069] Audience interaction is categorized into positive feedback, inquiries, and advertising / violations.
[0070] Automatic feedback module:
[0071] Template-driven: Quickly invoke standardized responses, such as "Thank you for your support";
[0072] Large model generation: Provide personalized answers to complex questions, such as "product consultation answers";
[0073] The reply content is filtered by risk control rules and then pushed to the live broadcast room.
[0074] Site control management module:
[0075] Multi-level priority strategy: respond to interactions from high-value users first, and delay processing for low-value users;
[0076] Automatic protection mechanism: Automatically triggers muting, blocking, or kicking out malicious behavior (such as spamming, insults);
[0077] Event management mechanism: Different event control strategies can be configured (such as special strategies for holiday promotional activities).
[0078] Feedback optimization module:
[0079] like Figure 2 The diagram illustrates the module logic interaction flow of the AI-powered live-streaming room-based intelligent feedback management system, as shown below:
[0080] Data collection chain: Audience interaction data sources (including bullet comments, comments, gift rewards, following behavior, etc.) are collected by the data collection module through RPA technology.
[0081] Analysis and Processing Chain: The data acquisition module transmits the collected interactive data to the natural language processing module, which performs semantic recognition, sentiment analysis, violation detection, and classification (positive feedback, inquiries, and inappropriate content). The classification results then flow to two separate modules:
[0082] Flow to the automatic response module: The automatic response module generates standardized responses by calling templates for positive feedback, and generates personalized responses by calling a large model and knowledge base for question inquiries. After being filtered by risk control, the responses are pushed to the live broadcast room.
[0083] Flow to the Live Stream Control Management Module: The Live Stream Control Management Module performs automatic protection operations against infringing content (such as content blocking, account banning / kicking, etc.) and responds differently to high-value user interactions based on priority strategies. The results of these operations are simultaneously applied to the live stream room.
[0084] Feedback and Optimization Link: Viewer feedback data (such as comment and like rate, subsequent interaction duration, etc.) and historical interaction data in the live broadcast room are collected by the feedback and optimization module. This module adjusts the parameters of the natural language processing module, the comment strategy of the automatic comment module, and the priority rules of the scene control management module through reinforcement learning mechanism, and then inputs the adjusted parameters back into the natural language processing module to achieve iterative optimization of the entire system link.
[0085] like Figure 3 As shown in the attached diagram, this diagram illustrates the full-link logical interaction of the AI-powered intelligent feedback management system for live streaming rooms. The specific text and interaction relationships involved are as follows:
[0086] Data collection phase: The "RPA robot" shown in the attached diagram serves as a process automation component, collecting interactive data such as "bullet comments, comments, gift rewards, and follow behavior data" across platforms. The collected data is written to both the "caching layer (Redis)" and the "database (MySQL)," with Redis used for real-time high-frequency data access and MySQL used for structured historical data storage. Simultaneously, a "collection log" is generated, which includes "collection integrity records, exception retry records, data collection volume for each platform, and data format verification results," and is output daily to achieve end-to-end traceability.
[0087] In the natural language processing stage: the collected interactive data flows into the "pre-trained language model (BERT / RoBERTa)", and performs "semantic recognition, sentiment analysis, and violation detection" in sequence; violation detection is completed collaboratively by the "keyword filtering list" and the "deep learning detection algorithm", first identifying obvious violations, and then detecting ambiguous semantic content; finally, the data is classified into "positive feedback, question consultation, and violation content", and the classification results are synchronized to subsequent stages, carrying "semantic recognition tags, sentiment tendency tags" (to automatic feedback) and "violation level tags" (to venue management).
[0088] Automatic response process: For "positive feedback," standardized templates such as "thank you replies (gift rewards)" and "welcome replies (attention behavior)" are called from the "template storage unit" to generate replies; for "question inquiries," the "large model calling unit" combines the "knowledge base (product parameters, activity rules)" to generate personalized replies, and adjusts the tone according to "emotional tendency" (lively tone for positive, reassuring tone for negative); all replies are processed by the "risk control filtering unit" to perform "sensitive word detection and violation expression screening." If the filtering fails, the replies are "regenerated," and if they pass, they are pushed to the live broadcast room.
[0089] In the event control management phase: Based on the "preset priority strategy" (with user value as the core, high-value users include "VIP users, users who give high gifts, and users who interact frequently"), differentiated responses are given to interactions; for "violation content", "content blocking, user warning, account banning, and account kicking" are executed through the "automatic protection unit" (first violation warning + blocking, repeated violations banning / kicking); the event control "protection results" are sent back to the automatic response phase to "block the push of violation-related responses", and the "response records" of the automatic responses are received at the same time.
[0090] Feedback optimization phase: The "feedback collection unit" collects feedback data such as "reply like rate, subsequent interaction duration, and number of new comments" as well as "historical interaction data"; after the "reinforcement learning unit" extracts features, it adjusts "semantic recognition weight, template call priority, and violation detection threshold"; the adjusted parameters are first verified in a "small-scale test live room", and after passing, they are updated to "all live rooms" according to the "set cycle", forming a closed-loop optimization.
[0091] like Figure 4 As shown in the attached diagram, the logical interaction flow between the natural language processing module and the automatic response module is as follows: The natural language processing module sequentially performs semantic recognition, sentiment analysis, and violation detection operations, and then performs result classification. After the result classification is completed, it flows to the automatic response module, which generates response content through two branches: template-driven response and personalized generation by a large model. After the generated response content is processed by the risk filtering stage, it is pushed to the live broadcast room.
[0092] Example 2
[0093] Another aspect of the present invention is an AI-based intelligent feedback management system for live streaming, used to implement the above-mentioned method, comprising:
[0094] Data acquisition module: It is equipped with a process automation component for cross-platform acquisition of live stream audience interaction data. The interaction data includes at least bullet screen, comments, gift rewards and follow behavior data. The acquired data is written to the cache layer and database in the data storage module respectively to adapt to real-time call and historical analysis requirements. At the same time, it generates acquisition logs containing acquisition integrity and abnormal retry records.
[0095] Natural Language Processing Module: Connected to the data acquisition module, configured with a pre-trained language model, used to sequentially perform semantic recognition, sentiment analysis and violation detection on the interactive data, and classify the interactive data into positive feedback, questions and inquiries and violation content;
[0096] Automatic response module: Connected to the natural language processing module, including a template storage unit and a large model calling unit. The template storage unit is used to generate standardized thank-you replies based on positive feedback, and the large model calling unit is used to generate personalized replies to inquiries by combining the knowledge base. The automatic response module is also equipped with a risk control filtering unit, which is used to perform risk control rules to filter the reply content before the reply is pushed to the live broadcast room.
[0097] The session control management module is connected to the natural language processing module and the automatic feedback module, and includes a priority configuration unit and an automatic protection unit. The priority configuration unit is used to respond differently to the interactions of different users based on a preset priority strategy. The automatic protection unit is used to trigger automatic protection operations for interactions that contain illegal content. The session control management module sends the protection results back to the automatic feedback module to gate and block feedback associated with illegal interactions before pushing them, and receives feedback records synchronized by the automatic feedback module.
[0098] Feedback optimization module: Connected to the automatic feedback module, the venue management module, and the data storage module, including a feedback acquisition unit and a reinforcement learning unit. The feedback acquisition unit is used to collect audience feedback data and historical interaction data on the feedback content. The reinforcement learning unit is used to adjust the parameters of the natural language processing module, the feedback strategy of the automatic feedback module, and the priority rules of the venue management module through the reinforcement learning mechanism to achieve iterative optimization of the system.
[0099] Data storage module: Connects to each of the above modules and is used to store the collected interactive data, the classification results of the natural language processing module, the response content of the automatic feedback module, and the feedback data of the feedback optimization module. The data storage module includes a cache layer and a database, which are adapted to the needs of real-time data access and historical data storage, respectively.
[0100] Furthermore, the process automation component in the data acquisition module is an RPA robot, which simulates manual operation to achieve cross-platform data access and adapts to multi-platform interactive data formats; the caching layer is Redis, used to store real-time high-frequency interactive data, and the database is MySQL, used to store structured historical interactive data; the acquisition log generated by the data acquisition module is output daily, and in addition to acquisition integrity and abnormal retry records, it also includes the data acquisition volume and data format verification results of each platform, so as to form a traceable logic for the entire data acquisition chain.
[0101] Furthermore, the natural language processing module and the automatic feedback module have the following internal units and collaborative relationships:
[0102] The pre-trained language model unit in the natural language processing module uses BERT or RoBERTa for semantic recognition and sentiment analysis; its violation detection unit is composed of a keyword filtering subunit and a deep learning detection subunit. It first identifies obvious violations based on the keyword filtering list, and then the deep learning detection subunit performs a second detection on ambiguous semantic content; its result labeling and distribution unit is used to attach semantic recognition labels and sentiment tendency labels when synchronizing with the automatic feedback module, and to attach violation level labels when synchronizing with the field control management module.
[0103] The automatic response module's template storage unit stores standardized thank-you response templates categorized by interaction type, including templates for thanking gifts and templates for welcoming followers. It also supports adding or deleting templates based on livestream operation needs. When generating personalized responses, the large model calling unit first retrieves business information related to the inquiry from the knowledge base sub-unit, then adjusts the response tone based on the sentiment of the interaction data, using a lively tone for positive sentiments and a soothing tone for negative sentiments. The risk control filtering unit's filtering rules include sensitive word detection and screening for inappropriate expressions; responses that fail filtering trigger a regeneration process.
[0104] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for intelligent feedback management based on AI-powered live streaming room control, characterized in that: include: (1) Interactive data collection: Collect live room audience interaction data across platforms through process automation components. The interactive data includes at least bullet screen, comments, gift rewards and follow behavior data. Write the collected data into the cache layer for real-time access and the database for historical analysis, and generate a collection log containing collection integrity and abnormal retry records. (2) Natural Language Processing and Classification: Using a pre-trained language model, semantic recognition, sentiment analysis and violation detection are performed on the interactive data in sequence, and the interactive data is classified into positive feedback, question consultation and violation content; (3) Automatic feedback and proactive risk control: For positive feedback, standardized thank-you replies are generated based on templates; for inquiries, a large model is used in conjunction with a knowledge base to generate personalized replies; The responses were filtered by risk control rules before being pushed to the live stream. (4) Site control management and interlocking: Based on the preset priority strategy, differentiated responses are made to the interactions of different users, and automatic protection operations are triggered for interactions that are illegal; the site control protection results are sent back to the automatic feedback stage to block the feedback push associated with the illegal interaction, and the feedback record of the automatic feedback is synchronized to the site control management. (5) Feedback optimization: Collect audience feedback data and historical interaction data on the feedback content, and adjust natural language processing parameters, feedback strategies and event control priority rules through reinforcement learning mechanism to achieve iterative optimization.
2. The method according to claim 1, characterized in that: In step (1), the process automation component is an RPA robot, which completes cross-platform data access by simulating human operation and supports the adaptation of interactive data formats of mainstream live streaming platforms; the caching layer is a Redis cache, used to store real-time high-frequency interactive data; the database is a MySQL database, used to store structured historical interactive data; the collection log is generated once a day, which includes data collection volume and data format verification results of each platform in addition to collection integrity and abnormal retry records, forming a traceable logic for the entire data collection chain.
3. The method according to claim 1, characterized in that: In step (2), the pre-trained language model includes the BERT model or the RoBERTa model; the violation detection is achieved through the collaboration of a keyword filtering list and a deep learning detection algorithm. First, obvious violations are initially identified based on the keyword filtering list, and then the ambiguous semantic content is detected a second time through the deep learning algorithm; the classification results need to be synchronized to the automatic feedback stage and the venue control management stage. The classification results synchronized to the automatic feedback stage carry semantic recognition labels and sentiment labels, and the classification results synchronized to the venue control management stage carry violation level labels.
4. The method according to claim 1, characterized in that: In step (3), the standardized thank-you replies include thank-you replies for gift rewards and welcome replies for follower behavior. The templates are stored according to the type of interaction and can be added or deleted based on the operational needs of the live broadcast room. When the big model generates personalized replies, it first retrieves business information related to the question consultation (such as product parameters and activity rules) from the knowledge base, and then adjusts the tone of the reply based on the emotional tendency of the interaction data. A lively tone is used for positive emotional tendencies, and a soothing tone is used for negative emotional tendencies. The risk control rule filtering includes sensitive word detection and screening of illegal expressions. Replies that fail the filtering need to be returned to the automatic reply stage to be regenerated.
5. The method according to claim 1, characterized in that: In step (4), the preset priority strategy is formulated with user value as the core. High-value users include at least VIP users, users whose gift reward amount reaches the set standard within a preset time period, and users whose historical interaction frequency is higher than the set threshold. The automatic protection operation includes content blocking, user warning, account banning, and account kicking. Warning and content blocking are performed for users who violate the rules for the first time, and account banning or kicking is performed for users who violate the rules repeatedly. The site control management also supports the configuration of scenario-based strategies. In holiday scenarios, the response priority of high-value user interactions is increased, and in promotional activity scenarios, the detection intensity of advertising-related illegal content is enhanced.
6. The method according to claim 1, characterized in that: In step (5), the feedback data includes the like rate of the reply content, the duration of subsequent interaction after the audience receives the reply, and the number of new comments caused by the reply content; the reinforcement learning mechanism first extracts features from the feedback data and historical interaction data, and then adjusts the semantic recognition weight of natural language processing, the template calling priority of automatic reply, and the violation detection threshold of the scene control management based on the extracted features; the adjusted parameters need to be verified in a small-scale test live room first, and after verification, they are updated to all live rooms according to the set period.
7. The method according to any one of claims 1 to 6, characterized in that: The interlocking logic between steps (4) and (3) also includes: when the venue control management detects that the same user initiates multiple violations within a set time, in addition to triggering automatic protection operations, it will also generate a violation tag for the user and synchronize it to the interaction data collection stage. After the user's interaction data is collected, it will be pushed to the venue control management stage for secondary verification to prevent violations from entering the natural language processing and automatic feedback stage.
8. An AI-powered intelligent feedback management system for live streaming, characterized in that: A method for implementing any one of rights 1 to 7 includes: Data acquisition module: It is equipped with a process automation component for cross-platform acquisition of live stream audience interaction data. The interaction data includes at least bullet screen, comments, gift rewards and follow behavior data. The acquired data is written to the cache layer and database in the data storage module respectively to adapt to real-time call and historical analysis requirements. At the same time, it generates acquisition logs containing acquisition integrity and abnormal retry records. Natural Language Processing Module: Connected to the data acquisition module, configured with a pre-trained language model, used to sequentially perform semantic recognition, sentiment analysis and violation detection on the interactive data, and classify the interactive data into positive feedback, questions and inquiries and violation content; Automatic response module: Connected to the natural language processing module, including a template storage unit and a large model calling unit. The template storage unit is used to generate standardized thank-you replies based on positive feedback, and the large model calling unit is used to generate personalized replies to inquiries by combining the knowledge base. The automatic response module is also equipped with a risk control filtering unit, which is used to perform risk control rules to filter the reply content before the reply is pushed to the live broadcast room. The session control management module is connected to the natural language processing module and the automatic feedback module, and includes a priority configuration unit and an automatic protection unit. The priority configuration unit is used to respond differently to the interactions of different users based on a preset priority strategy. The automatic protection unit is used to trigger automatic protection operations for interactions that contain illegal content. The session control management module sends the protection results back to the automatic feedback module to gate and block feedback associated with illegal interactions before pushing them, and receives feedback records synchronized by the automatic feedback module. Feedback optimization module: Connected to the automatic feedback module, the venue management module, and the data storage module, including a feedback acquisition unit and a reinforcement learning unit. The feedback acquisition unit is used to collect audience feedback data and historical interaction data on the feedback content. The reinforcement learning unit is used to adjust the parameters of the natural language processing module, the feedback strategy of the automatic feedback module, and the priority rules of the venue management module through the reinforcement learning mechanism to achieve iterative optimization of the system. Data storage module: Connects to each of the above modules and is used to store the collected interactive data, the classification results of the natural language processing module, the response content of the automatic feedback module, and the feedback data of the feedback optimization module. The data storage module includes a cache layer and a database, which are adapted to the needs of real-time data access and historical data storage, respectively.
9. The system according to claim 8, characterized in that: The process automation component in the data acquisition module is an RPA robot, which simulates human operation to achieve cross-platform data access and adapts to multi-platform interactive data formats; the caching layer is Redis, used to store real-time high-frequency interactive data, and the database is MySQL, used to store structured historical interactive data; the acquisition log generated by the data acquisition module is output daily, and in addition to acquisition integrity and abnormal retry records, it also includes the data acquisition volume and data format verification results of each platform, so as to form a traceable logic for the entire data acquisition chain.
10. The system according to claim 8, characterized in that: The natural language processing module and the automatic feedback module have the following internal units and collaborative relationships: The pre-trained language model unit in the natural language processing module uses BERT or RoBERTa for semantic recognition and sentiment analysis; its violation detection unit is composed of a keyword filtering subunit and a deep learning detection subunit. First, obvious violations are initially identified based on the keyword filtering list, and then the deep learning detection subunit performs a second detection on ambiguous semantic content. Its result labeling and distribution unit is used to attach semantic recognition tags and sentiment tags when synchronizing with the automatic feedback module, and to attach violation level tags when synchronizing with the field control management module; The automatic response module's template storage unit stores standardized thank-you response templates categorized by interaction type, including templates for thanking gifts and templates for welcoming followers. It also supports adding or deleting templates based on livestream operation needs. When generating personalized responses, the large model calling unit first retrieves business information related to the inquiry from the knowledge base sub-unit, then adjusts the response tone based on the sentiment of the interaction data, using a lively tone for positive sentiments and a soothing tone for negative sentiments. The risk control filtering unit's filtering rules include sensitive word detection and screening for inappropriate expressions; responses that fail filtering trigger a regeneration process.