Multi-platform AI intelligent live interaction method

Abstract

This application discloses a multi-platform AI intelligent live streaming interaction method, belonging to the field of live streaming control technology. This method adapts to the collection methods of different live streaming platform interfaces, achieving unified collection and structured preprocessing of live stream chat and user events across multiple platforms. It employs independent AI question queues, chat queues, and voice queues to hierarchically distribute different types of events. After calling a large language model to identify user intent, it generates two types of response content: text replies and voice broadcasts, which are pushed to the live stream's public screen and the broadcaster's audio device, respectively. Simultaneously, it manages the entire process's start / stop and response frequency through global control parameters. This method enables unified interaction control across multiple platform live streaming scenarios, balancing response efficiency and platform compliance, and improving the accuracy and coverage of live streaming interaction.

Description

Technical Field

[0001] This application belongs to the field of live streaming control technology, specifically involving a multi-platform AI intelligent live streaming interaction method. Background Technology

[0002] In the current live streaming industry, brands and streamers typically conduct simultaneous live streams on multiple mainstream platforms to reach a wider audience. Existing live streaming control tools are mostly platform-specific, only compatible with the messaging interface of a single platform. Simultaneous live streaming across multiple platforms requires running multiple control tools simultaneously, increasing device resource consumption and failing to achieve unified cross-platform interaction rule configuration, resulting in high maintenance costs.

[0003] Meanwhile, existing live stream control tools mostly employ a full, indiscriminate response logic to live stream comments and user events. When live stream traffic is high, a large number of messages trigger responses simultaneously, easily causing a flood of comments and triggering platform traffic limiting and muting rules. If the response frequency is reduced to avoid this flooding, user questions may go unanswered, impacting user experience. Furthermore, existing live stream control tools often use pre-set, fixed scripts for their responses, matching only fixed keywords and failing to identify the user's true intent. They also only support single-format responses to comments, making it difficult for streamers who cannot easily view comments to obtain timely user feedback, resulting in insufficient interactive coverage. Summary of the Invention

[0004] The technical problem to be solved by this invention is that existing single-platform live streaming control tools cannot adapt to multi-platform synchronous live streaming scenarios, cross-platform operation and maintenance costs are high, indiscriminate message response can easily lead to screen flooding or missed responses, response efficiency and platform compliance cannot be balanced, fixed-script replies cannot identify the user's true intent, and only supports a single reply form with poor interaction effect. This invention provides a multi-platform AI intelligent live streaming interaction method.

[0005] This invention is implemented as follows: It provides a multi-platform AI intelligent live streaming interaction method, comprising the following steps: S1. For multiple live streaming platforms, the collection methods adapted to the interfaces of each platform are used to collect the barrage messages and user events in the live streaming room in real time. The collected raw data is preprocessed in a structured manner to extract structured data containing username, message content, event type and timestamp. S2. Based on the preset event classification rules, the structured data is classified and identified, and the corresponding data is diverted to the independent AI question queue, bullet screen queue and voice queue according to the event type. S3. For the structured data that enters the AI ​​problem queue, call the pre-trained large language model to perform semantic understanding and intent recognition, and output the corresponding intent category and associated entity information; S4. Based on the identified intent category and entity information, retrieve the match in the preset matching knowledge base and generate interactive response content that includes both text reply content and voice broadcast content. S5. Push the text reply to the bullet screen queue and send it to the corresponding live room through browser automation technology; at the same time, push the voice broadcast content to the voice queue, and output it to the live audio device after text-to-speech processing. S6. Based on preset global control parameters, it performs real-time control over the start / stop and response frequency of the entire process of message collection, AI response generation, and voice broadcasting. Furthermore, the traffic splitting rule in step S2 is as follows: S21. For user messages containing questions, add them to the AI ​​question queue. The maximum capacity of the AI ​​question queue is 2 messages, and a first-in-first-out scheduling strategy is adopted. S22. When a user enters the live stream, a welcome message is generated and added to the bullet screen queue. S23. For user like events, count the number of consecutive likes from the same user, and generate a thank-you message when the preset threshold is reached, and add it to the bullet screen queue and voice queue simultaneously. Furthermore, the large language model in step S3 is either the Zhipu AI large model or the DeepSeek large model, and the semantic understanding process includes contextual association analysis and maintenance of the state of the most recent 3 rounds of dialogue. Furthermore, the specific methods for data collection on different platforms in step S1 are as follows: for Douyin and Kuaishou platforms, a WebSocket connection is used to receive live stream messages; for WeChat Video Channel and Xiaohongshu platforms, Playwright browser automation technology is used to monitor page DOM changes and extract message content. Furthermore, the browser automation technology in step S5 is implemented using ChromeDevToolsProtocol, text-to-speech processing uses LineVoiceServer service, and the generated temporary audio files are automatically deleted after playback. Furthermore, the response frequency control in step S6 is implemented through Redis caching: the number of triggers by a single user and a single IP within a preset time window is counted, and the corresponding request is blocked when the number exceeds a preset threshold. Furthermore, the method also includes a startup verification step: when the tool starts, it first performs license permission verification, and after the verification is successful, it starts an HTTP server to receive external control commands.

[0006] Furthermore, the method also includes a timed event control step: based on a preset event control configuration, when the timed triggering conditions are met, corresponding event control content is generated and pushed to the live broadcast room, while simultaneously recording the sending logs and interaction statistics of all content. Compared with the prior art, the present invention has the following beneficial effects: 1. By adapting to different platform collection methods, it achieves unified collection and processing of live broadcast messages across multiple platforms, eliminating the need to run multiple sets of tools when broadcasting simultaneously on multiple platforms, reducing operation and maintenance costs, and enabling unified cross-platform interaction rule configuration. 2. Three independent queues are used to process different types of events separately. Content with different priorities can be scheduled in a hierarchical manner to avoid response congestion in high-concurrency scenarios. At the same time, combined with global frequency control, it can prevent the platform's rate limiting rules from being triggered by excessive screen scrolling while ensuring that the core user questions are responded to in a timely manner. 3. By leveraging a large language model for semantic understanding and intent recognition, the system can accurately identify users' true needs and generate both text-based comments and voice broadcasts. This allows for public chat replies to users and enables broadcasters to receive real-time user feedback, improving interaction coverage and response accuracy. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the steps of the multi-platform AI intelligent live streaming interaction method described in an embodiment of the present invention; Figure 2 This is a system architecture module diagram corresponding to the multi-platform AI intelligent live streaming interaction method described in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the full-link functional modules and technology adaptation of the multi-platform AI intelligent live streaming interaction method described in this embodiment of the invention.

[0008] Attached image labels: 1-Bullet screen monitoring layer, 2-Queue layer, 3-AI reply module, 4-Bullet screen sending module, 5-Voice broadcast module. Detailed Implementation

[0009] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention. Please refer to Figure 1-3 As shown, this invention provides a multi-platform AI intelligent live streaming interaction method. This embodiment is applied to a multi-platform simultaneous live streaming scenario for a clothing brand, covering four live streaming platforms: Douyin, Kuaishou, WeChat Video Channel, and Xiaohongshu. The specific implementation process is as follows: During the tool startup phase, license and permission verification is performed first. Once the verification is successful, a local HTTP server is started to receive external control commands from operators, supporting modifications to response rules, adjustments to response frequency, and enabling or disabling of specific functions. S1, the bullet screen monitoring layer 1, uses a collection method adapted to various platforms to collect live room data in real time: For Douyin and Kuaishou platforms, it receives bullet screen messages, user entry, likes, and follow events in real time by connecting to the WebSocket message interface provided by the platform; for WeChat Video Channel and Xiaohongshu platforms, since no public message interface is provided, the Playwright browser automation tool is used to monitor changes in the DOM nodes of the live room page, and extracts the corresponding content when the DOM nodes of the corresponding bullet screen and events are updated. The collected raw data is preprocessed in a structured manner: redundant format tags and useless fields in the raw data of each platform are removed, and four core fields, username, message content, event type, and timestamp, are uniformly extracted to generate structured data with a unified format, which is then transmitted to queue layer 2. S2, Queue Layer 2 classifies and distributes structured data based on preset event classification rules: ① If it is a message sent by a user containing a question, it is diverted to the AI ​​question queue. The maximum capacity of the AI ​​question queue is set to 2 queues, and a first-in-first-out scheduling strategy is adopted to avoid queue backlog and response delay; ② If it is an event of a user entering the live broadcast room, a welcome message "Welcome XX to the live broadcast room" is directly generated and pushed to the bullet screen queue; ③ If it is a user liking event, the number of consecutive likes by the same user is counted. When the number of likes reaches the preset threshold of 100, a thank-you message "Thank you XX for liking~" is generated and pushed to both the bullet screen queue and the voice queue. S3 and AI response module 3 extract structured data sequentially from the AI ​​question queue, call Zhipu AI's GLM-4 large language model for semantic understanding, and at the same time perform contextual analysis by associating the user's previous 3 rounds of dialogue content to identify the user's intent category and entity information: for example, if a user sends "Does this dress have size M?", the intent category is identified as "product specification question", and the entity information is "dress, size M". S4. Based on the identified intent category and entity information, a search and matching process is performed in a pre-configured knowledge base. This knowledge base contains all product information, activity rules, and frequently asked questions for this live stream. Upon finding a match, a text reply is generated: "This outfit is currently available in size M. You can click on shopping cart link 1 to purchase it directly~". Simultaneously, a corresponding voice announcement is generated: "User XX inquired whether the outfit in link 1 is available in size M. The reply is yes." If no match is found in the knowledge base, a compliant general reply is generated using a large language model.

[0010] S5. The generated text reply content is pushed to the bullet screen queue. The bullet screen sending module 4 extracts the content from the bullet screen queue in turn, and uses ChromeDevToolsProtocol to control the browser to input the content in the public screen input box of the corresponding live room and send it. At the same time, the voice broadcast content is pushed to the voice queue. The voice broadcast module 5 extracts the content from the voice queue, calls the LineVoiceServer service to convert the text into a temporary audio file, plays it to the headphone device worn by the anchor, and deletes the temporary audio file immediately after playback to avoid disk space occupation.

[0011] S6. The entire process is managed based on the global control parameters preset by the operations personnel: The global control parameters include the monitoring start / stop switch, AI reply switch, voice broadcast switch, the maximum number of responses per user per minute is 3, and the maximum number of responses per IP per minute is 10. The number of responses for each user and IP is counted through Redis cache. When the threshold is exceeded, the corresponding trigger request is blocked to avoid repeated replies and screen refresh.

[0012] In addition, based on the preset live stream settings, the system automatically generates content such as "Today, you can enjoy a discount of 30 RMB off orders over 200 RMB in the live stream. Don't miss it!" every 15 minutes. This content is then pushed to the chat queue and recorded, along with logs of all sent content and statistics on the number of responses, for later review by operations staff.

[0013] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-platform AI intelligent live streaming interaction method, characterized in that, Includes the following steps: S1. For multiple live streaming platforms, the collection methods adapted to the interfaces of each platform are used to collect the barrage messages and user events in the live streaming room in real time. The collected raw data is preprocessed in a structured manner to extract structured data containing username, message content, event type and timestamp. S2. Based on the preset event classification rules, the structured data is classified and identified, and the corresponding data is diverted to the independent AI question queue, bullet screen queue and voice queue according to the event type. S3. For the structured data that enters the AI ​​problem queue, call the pre-trained large language model to perform semantic understanding and intent recognition, and output the corresponding intent category and associated entity information; S4. Based on the identified intent category and entity information, retrieve the match in the preset matching knowledge base and generate interactive response content that includes both text reply content and voice broadcast content. S5. Push the text reply to the bullet screen queue and send it to the corresponding live room through browser automation technology; at the same time, push the voice broadcast content to the voice queue, and output it to the live audio device after text-to-speech processing. S6. Based on preset global control parameters, it performs real-time control over the start / stop and response frequency of the entire process of message collection, AI response generation, and voice broadcasting.

2. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The specific traffic splitting rules for step S2 are as follows: S21. For user messages containing questions, add them to the AI ​​question queue. The maximum capacity of the AI ​​question queue is 2 messages, and a first-in-first-out scheduling strategy is adopted. S22. When a user enters the live stream, a welcome message is generated and added to the bullet screen queue. S23. For user like events, count the number of consecutive likes from the same user, and generate a thank-you message when the preset threshold is reached, and add it to the bullet screen queue and voice queue simultaneously.

3. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The large language model in step S3 is either the Zhipu AI large model or the DeepSeek large model. The semantic understanding process includes contextual association analysis and maintenance of the state of the most recent three rounds of dialogue.

4. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The specific data collection methods for different platforms in step S1 are as follows: for Douyin and Kuaishou platforms, WebSocket connection is used to receive live room messages; for WeChat Video Channel and Xiaohongshu platforms, Playwright browser automation technology is used to monitor page DOM changes and extract message content.

5. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The browser automation technology in step S5 is implemented using ChromeDevToolsProtocol, the text-to-speech processing uses LineVoiceServer service, and the generated temporary audio files are automatically deleted after playback is complete.

6. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The response frequency control in step S6 is implemented through Redis caching: the number of triggers for a single user and a single IP within a preset time window is counted, and the corresponding request is blocked when the number exceeds a preset threshold.

7. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The method also includes a startup verification step: when the tool starts, it first performs license permission verification, and after the verification is successful, it starts an HTTP server to receive external control commands.

8. The multi-platform AI intelligent live streaming interaction method according to claim 1, characterized in that, The method also includes a timed event control step: based on a preset event control configuration, when the timed triggering conditions are met, corresponding event control content is generated and pushed to the live broadcast room, while recording the sending logs and interaction statistics of all content.

Images