Streaming media bullet screen automatic reply system and method

Abstract

The invention discloses a streaming media bullet screen automatic reply system and method, and belongs to the technical field of streaming media interaction, and the system comprises a bullet screen collection module, a context emotion analysis module, a scene matching module, an interaction memory module, a reply generation module and a reply output module. The bullet screen acquisition module is used for acquiring user bullet screen data in real time and synchronizing the user bullet screen data to the context emotion analysis module and the interactive memory module; and the context emotion analysis module pre-constructs a semantic understanding model and an emotion classification model, and is used for outputting a bullet screen analysis result with a context label and an emotion label. According to the invention, the context emotion analysis module is combined with the semantic comprehension and emotion classification model to realize the deep analysis of the bullet screen, and the module not only can identify the core appeal and context logic of the user, but also can accurately judge the complex emotional tendency including metaphor and invert, provides an accurate intention and emotion basis for the subsequent reply generation, and improves the user experience. And the reply pertinence and the emotion matching degree are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of streaming media interaction, more particularly, to a streaming media bullet screen automatic reply system and method. BACKGROUND

[0002] With the rapid development of streaming media live platform, bullet screen has become an important real-time interaction way between users and anchors, users and content. Bullet screen interaction not only enhances the sense of participation in watching, but also provides an instant feedback channel for content creators. However, with the expansion of user scale and the surge of bullet screen quantity, it is difficult for anchors or operation personnel to respond to each bullet screen in real time and effectively, especially in high concurrency scenarios, a large number of bullet screens are ignored, leading to a decline in user interaction experience and a weakening of the sense of participation.

[0003] Currently, some existing bullet screen automatic reply schemes attempt to realize automatic response through keyword matching or simple rule engine, but these methods generally have the following defects: first, lack of deep understanding of bullet screen context and emotion, difficult to deal with irony, satire and other complex expressions; second, single reply template, unable to adapt to the differentiated needs of multiple scenarios such as games, e-commerce and education; third, lack of user historical interaction memory, reply content is isolated and repetitive, and cannot form a coherent dialogue; fourth, insufficient system scalability and real-time performance, prone to reply delay or resource overload in high concurrency scenarios; and finally, lack of continuous learning and safety control mechanism, making it difficult to guarantee the accuracy and compliance of the reply.

[0004] Based on this, the present application designs a streaming media bullet screen automatic reply system and method to solve the above problems. SUMMARY

[0005] The present application aims to provide a streaming media bullet screen automatic reply system and method to solve the problems raised in the background art.

[0006] A streaming media bullet screen automatic reply system, comprising a bullet screen collection module, a context and emotion analysis module, a scenario matching module, an interaction memory module, a reply generation module and a reply output module; The bullet screen collection module is used to collect user bullet screen data in real time and synchronize to the context and emotion analysis module and the interaction memory module; The context and emotion analysis module pre-constructs a semantic understanding model and an emotion classification model, which are used to output bullet screen analysis results with context labels and emotion labels; The scenario matching module pre-stores a multi-scenario reply strategy library, which is used to match and adapt the reply template according to the streaming media scenario type; The interaction memory module constructs an interaction conversation linked list, which is used to store historical interaction records and judge whether the new bullet screen is a follow-up question scenario; The reply generation module is configured to generate reply content in combination with a reply template, analysis result and historical interaction information. The reply output module is configured to push the reply content and update a session chain table.

[0007] A method for automatically replying to a stream media bullet screen, comprising the following steps: S1, a bullet screen collection module collects user bullet screen data of a stream media platform in real time, the bullet screen data at least including a user ID, bullet screen text and a sending timestamp, and the collection is synchronized to a context sentiment analysis module and an interactive memory module; S2, the context sentiment analysis module calls a pre-constructed semantic understanding model to analyze the context logical relationship and core appeal of the bullet screen text, and simultaneously calls a sentiment classification model to identify the sentiment tendency of the bullet screen text, and outputs a bullet screen analysis result carrying a context label, a sentiment label and a core appeal keyword; S3, a scene matching module first matches a corresponding scene sub-library in a pre-stored multi-scene reply strategy library according to a scene type preset by a host or a scene type identified by a live broadcast title, and then screens an adaptive scene-based reply template from the corresponding scene sub-library in combination with the core appeal keyword in the analysis result output in step S2; S4, the interactive memory module takes the user ID in step S1 as an index to query historical interaction records of the user in an interactive session chain table, judges whether the current bullet screen is a follow-up question scene for a historical reply through a semantic association algorithm, and if so, extracts associated question and answer information in the historical interaction records and feeds back to a reply generation module; S5, the reply generation module receives the reply template screened in step S3, the bullet screen analysis result in step S2 and the historical interaction association information in step S4, calls a natural language generation model to optimize and adjust the reply template, so that the reply content matches the context label and the sentiment label of the bullet screen, and simultaneously associates the historical information to form a coherent reply, and generates final reply content; S6, a reply output module pushes the final reply content generated in step S5 to a bullet screen area of a stream media platform in real time, and simultaneously synchronizes the reply content to the interactive memory module after associating the reply content with corresponding user bullet screen data, and updates an interactive session chain table corresponding to the user ID by the interactive memory module.

[0008] Preferably, the sentiment classification model of the context sentiment analysis module is trained by a bullet screen corpus labeled with a sentiment label, and can identify complex sentiment expressions such as metaphors and irony in the bullet screen; The multi-scene reply strategy library of the scene matching module includes game live broadcast, knowledge popularization, commodity delivery and other sub-libraries, and each sub-library is associated with a corresponding scene keyword set and a scene-based professional reply template; The conversation chain list of the interaction memory module is indexed by a user ID and stores association records of all the user's barrage in the single live broadcast and corresponding system replies.

[0009] Preferably, the model dynamic optimization module is further used to collect feedback data of the user on the system reply in the single live broadcast, the feedback data including operation records of the user's follow-up barrage emotional tendency for the system reply and manual correction of the reply by the anchor; The model dynamic optimization module calculates a reply adaptation score according to the feedback data, automatically extracts barrage reply samples with insufficient adaptability when the score is lower than a preset threshold, and performs incremental training on the semantic understanding model of the context emotional analysis module, the emotional classification model, and the natural language generation model of the reply generation module, so as to realize real-time iterative optimization of the model parameters.

[0010] Preferably, the system further comprises a user portrait construction module, which is linked with the interaction memory module and indexes the user ID, extracts interest keywords, question preferences, and emotional acceptance styles of the user based on historical interaction records, and generates a personalized user portrait; The scene matching module adjusts the tone style and content focus of the template in combination with the user portrait when screening the reply template, and the reply generation module performs individual optimization on the reply content according to the user portrait when calling the natural language generation model, so that the reply matches the interaction habits of a specific user.

[0011] Preferably, the system further comprises a security filtering module, which pre-stores a violation keyword library, a sensitive topic library, and a risk reply mode library; After the reply generation module outputs the final reply content, the security filtering module performs three checks on the reply content before the reply output module pushes the reply content: whether the reply content contains violation keywords, whether the reply content involves sensitive topics, and whether the reply content matches a risk reply mode; If the checks fail, the security filtering module triggers an interception mechanism and feeds back to the reply generation module, and the reply generation module generates a reply content that meets the security requirements based on the original analysis results, while recording violation check logs and synchronizing them to the scene matching module to update the risk avoidance templates of the corresponding scene sub-library.

[0012] Preferably, the system further comprises a module cooperative scheduling module, which is in communication connection with the context emotional analysis module, the model dynamic optimization module, and the security filtering module; The module cooperative scheduling module is internally provided with a load monitoring unit and a priority allocation unit, the load monitoring unit collects real-time running load data of each module, and the priority allocation unit allocates processing priorities according to the emergency level of the barrage; When a high-load module is detected, the processing task of low-priority bullet screen is temporarily shunted to an idle module, and the incremental training process of the model dynamic optimization module is executed in a low-peak period to avoid reply delay caused by resource preemption between modules.

[0013] Preferably, it further comprises a resource dynamic adaptation module, which is linked with the bullet screen collection module and the reply generation module; The resource dynamic adaptation module preconfigures multiple resource allocation schemes, adjusts resource allocation according to the bullet screen traffic data collected by the bullet screen collection module: when the bullet screen traffic is lower than a preset low threshold, a basic resource scheme is used to reduce CPU and memory occupation; when the bullet screen traffic is in a medium threshold range, a standard resource scheme is enabled to ensure processing efficiency; when the bullet screen traffic is higher than a high threshold, an elastic expansion scheme is triggered to call spare computing resources, and a reply simplification algorithm is used to optimize the processing flow of the reply generation module, so as to ensure the real-time performance of replies in a high-concurrency scenario.

[0014] Preferably, it further comprises a special scenario adaptation module, which prestores a special scenario rule library, and the special scenario includes an anchor speech overlap scenario, a bullet screen intensive screen scenario, and a specific activity interaction scenario. When the bullet screen data collected by the bullet screen collection module meets the special scenario triggering condition, the special scenario adaptation module executes the corresponding adaptation strategy: when the anchor speech overlaps, the reply is delayed and pushed to the anchor speech gap; When the bullet screen is intensive, the bullet screen with the same appeal is aggregated to generate a unified reply and is displayed at the top; When the specific activity interacts, the activity exclusive reply template is matched and the activity entry link is associated, and is synchronized to the interactive memory module to mark the activity interaction label.

[0015] Compared with the prior art, the application has the following advantages: 1. The context emotion analysis module of the application combines semantic understanding and emotion classification model to realize deep analysis of bullet screen, which can not only identify the core appeal and context logic of the user, but also accurately judge the complex emotional tendency including metaphor and irony, and provide accurate intention and emotion basis for subsequent reply generation, thereby significantly improving the pertinence and emotional matching degree of the reply.

[0016] 2. The application realizes the continuity of cross-round dialogue through the interactive memory and conversation chain table mechanism, stores and associates the historical interaction records with the user ID as the index, can intelligently judge whether the current bullet screen is a follow-up question, and extract relevant historical information, so that each reply can effectively connect with the previous text, provide a contextually coherent dialogue experience for the user, and enhance the naturalness and depth of the interaction.

[0017] 3.The application realizes the specialization and scene adaptation of replies through the multi-scene reply strategy library and dynamic matching mechanism. The system pre-stores professional template libraries in different fields such as games, goods promotion, and science popularization, and can accurately filter and match according to the live scene and user core demands, ensuring that the reply content not only meets the scene characteristics, but also has certain professionalism and accuracy, thereby improving the interactive value.

[0018] 4.The application realizes the self-learning and continuous evolution of the system through the model dynamic optimization module. The module collects user feedback on replies in real time (such as subsequent barrage emotions and anchor correction records), calculates the adaptation score, and automatically performs incremental training on the semantic understanding, emotion classification, and generation model. This mechanism enables the system to continuously learn and optimize from actual interactions, continuously improving the accuracy and adaptability of replies.

[0019] 5.The application guarantees the stable operation and compliance of the system through the integrated security filtering and resource dynamic scheduling mechanism. The security module performs threefold verification and interception on generated replies to ensure content safety. The collaborative scheduling and resource adaptation module intelligently allocates resources and adjusts processing priorities based on system load and barrage traffic, ensuring real-time response capability of the system under high concurrency, and effectively controlling content risks. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A workflow diagram of a stream media barrage automatic reply system and method according to the application. DETAILED DESCRIPTION

[0021] The technical solutions of the application will be described below in conjunction with embodiments. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0022] Please refer to Figure 1 A stream media barrage automatic reply system, comprising a barrage collection module, a context emotion analysis module, a scene matching module, an interactive memory module, a reply generation module, and a reply output module. The barrage collection module is used to collect user barrage data in real time and synchronously to the context emotion analysis module and the interactive memory module. The barrage collection module acquires barrage data stream in real time through the WebSocket or HTTP Streaming API interface provided by the stream media platform. The context emotion analysis module pre-constructs semantic understanding models and emotion classification models, which are used to output barrage analysis results with context labels and emotion labels. The scene matching module pre-stores a multi-scene reply strategy library, and is configured to match an adaptive reply template according to a streaming media scene type; The interactive memory module constructs an interactive conversation chain table, and is configured to store historical interactive records and determine whether a new barrage is a follow-up question scene; The reply generation module is configured to generate reply content in combination with a reply template, analysis results and historical interactive information; The reply output module is configured to push the reply content and update the conversation chain table.

[0023] A streaming media barrage automatic reply method, comprising the following steps: S1, a barrage collection module collects user barrage data of a streaming media platform in real time, the barrage data at least including a user ID, a barrage text and a sending timestamp, and the collection is synchronized to a context sentiment analysis module and an interactive memory module; S2, the context sentiment analysis module calls a pre-constructed semantic understanding model to analyze a context logical relationship and a core appeal of the barrage text, and simultaneously calls a sentiment classification model to identify a sentiment tendency of the barrage text, and outputs barrage analysis results carrying a context label, a sentiment label and a core appeal keyword; S3, the scene matching module first matches a corresponding scene sub-library in a pre-stored multi-scene reply strategy library according to a scene type preset by a host or a scene type identified by a live broadcast title through the system, and then screens an adaptive scene-based reply template from the corresponding scene sub-library in combination with the core appeal keyword in the analysis results output in step S2; S4, the interactive memory module takes the user ID in step S1 as an index to query historical interactive records of the user in an interactive conversation chain table, determines whether the current barrage is a follow-up question scene for the historical reply through a semantic association algorithm, and feeds back associated question and answer information in the historical interactive records to a reply generation module if the barrage is the follow-up question scene; S5, the reply generation module receives the reply template screened in step S3, the barrage analysis results in step S2 and the historical interactive association information in step S4, calls a natural language generation model to optimize and adjust the reply template, so that the reply content matches the context label and the sentiment label of the barrage, and simultaneously associates the historical information to form a coherent reply, and generates final reply content; S6, a reply output module pushes the final reply content generated in step S5 to a barrage area of the streaming media platform in real time, and simultaneously synchronizes the reply content to the interactive memory module after associating the reply content with corresponding user barrage data, and the interactive memory module updates the interactive conversation chain table corresponding to the user ID.

[0024] The sentiment classification model of the context sentiment analysis module is trained by using barrage corpus labeled with sentiment labels (positive, negative, neutral and teasing), and can identify complex sentiment expressions such as metaphors and irony in the barrage; The multi-scene reply strategy library of the scene matching module includes game live broadcast, knowledge popularization, and commodity delivery sub-libraries, each of which is associated with a corresponding scene keyword set and a scene-based professional reply template. The conversation chain table of the interactive memory module is indexed by a user ID and stores the association records of all the bullet screens and corresponding system replies of the user in a single live broadcast.

[0025] The model dynamic optimization module is further included for collecting feedback data of the user on the system reply in a single live broadcast, the feedback data including the follow-up bullet screen emotional tendency of the user for the system reply and the operation record of the anchor manually correcting the reply. The model dynamic optimization module calculates a reply adaptation score according to the feedback data, automatically extracts the bullet screen reply sample with insufficient adaptability when the score is lower than a preset threshold, and performs incremental training on the semantic understanding model of the context emotional analysis module, the emotional classification model, and the natural language generation model of the reply generation module to realize real-time iterative optimization of the model parameters.

[0026] The user portrait construction module is further included for being linked with the interactive memory module, extracting the interest keywords, question preferences, and emotional acceptance styles of the user based on the historical interaction records, and generating a personalized user portrait with the user ID as the index. The scene matching module adjusts the tone style and content focus of the template when screening the reply template, and the reply generation module performs personalized optimization on the reply content according to the user portrait when calling the natural language generation model, so that the reply matches the interactive habits of a specific user.

[0027] The security filtering module is further included for pre-storing a violation keyword library, a sensitive topic library, and a risk reply mode library. After the reply generation module outputs the final reply content, the security filtering module performs three verifications on the reply content before the reply output module pushes the reply: whether the reply contains violation keywords, whether the reply involves sensitive topics, and whether the reply matches a risk reply mode. If the verification fails, the security filtering module triggers an interception mechanism and feeds back to the reply generation module, and the reply generation module generates a reply content that meets the security requirements based on the original analysis results, records the violation verification log, and synchronizes it to the scene matching module to update the risk avoidance template of the corresponding scene sub-library.

[0028] The module cooperative scheduling module is further included for being communicatively connected with the context emotional analysis module, the model dynamic optimization module, and the security filtering module. The module cooperative scheduling module is internally provided with a load monitoring unit and a priority allocation unit. The load monitoring unit collects running load data (including CPU occupancy rate and data processing delay) of each module in real time. The priority allocation unit allocates processing priority according to the emergency degree of the barrage (determined by sending frequency, user level and content importance). When a high-load module is detected, the processing task of low-priority barrage is temporarily shunted to an idle module, and the incremental training process of the model dynamic optimization module is controlled to be executed in a low-peak period, so as to avoid reply delay caused by resource preemption between modules.

[0029] The resource dynamic adaptation module is further linked with the barrage collection module and the reply generation module. The resource dynamic adaptation module is preconfigured with multiple resource allocation schemes, and adjusts resource allocation according to the barrage flow data (number of barrages sent per unit time) collected by the barrage collection module in real time: when the barrage flow is lower than a preset low threshold, the basic resource scheme is used to reduce CPU and memory occupation; when the barrage flow is in a medium threshold range, the standard resource scheme is enabled to ensure processing efficiency; when the barrage flow is higher than a high threshold, the elastic expansion scheme is triggered to call spare computing resources, and the reply generation module is optimized through the reply simplification algorithm to ensure the real-time performance of replies in a high-concurrency scenario.

[0030] The special scene adaptation module further pre-stores a special scene rule library. The special scenes include an anchor speech overlap scene, a barrage intensive screen scene and a specific activity interaction scene. When the barrage data collected by the barrage collection module meets the special scene triggering condition, the special scene adaptation module executes the corresponding adaptation strategy: when the anchor speech overlaps, the reply is delayed and pushed to the anchor speech gap; When the barrage is intensive, the barrage with the same appeal is aggregated to generate a unified reply and is displayed at the top; When the specific activity interacts, the activity exclusive reply template is matched and the activity entry link is associated, and at the same time, the activity interaction tag is marked in the interactive memory module.

[0031] The workflow of the application: after the system acquires the user's barrage data in real time through the barrage acquisition module, it is synchronously transmitted to the context sentiment analysis module and the interactive memory module. The context sentiment analysis module performs deep analysis on the barrage based on the pre-trained semantic understanding model and emotion classification model, and outputs the analysis results with context labels, emotion labels and core appeals. At the same time, the scene matching module selects the adaptive reply template from the multi-scene reply strategy library according to the live scene type, the interactive memory module queries the historical conversation records according to the user ID, judges whether the current barrage is a follow-up scene, extracts relevant historical information, the reply generation module integrates the reply template, the context sentiment analysis result and the historical interactive information, calls the natural language generation model to generate coherent and emotion-matched reply content, and the safety filtering module performs threefold verification on the reply before output to ensure the compliance of the content. Finally, the reply output module pushes the reply to the barrage area and updates the conversation records in the interactive memory module, forming a closed-loop intelligent reply process. In addition, the system also has extension modules such as model dynamic optimization, user portrait construction, module collaborative scheduling and resource dynamic adaptation, which continuously improve the accuracy, personalization and system real-time performance of the reply.

[0032] The basic principles, main features and advantages of the application are shown and described above. Those skilled in the art should understand that the application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the application and are not intended to limit the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection of the application is defined by the appended claims and their equivalents.

Claims

1. A streaming media bullet screen automatic reply system, characterized in that, It includes a bullet screen collection module, a contextual sentiment analysis module, a scene matching module, an interactive memory module, a reply generation module, and a reply output module; The bullet screen acquisition module is used to collect user bullet screen data in real time and synchronize it to the context sentiment analysis module and the interactive memory module; The context sentiment analysis module pre-constructs a semantic understanding model and a sentiment classification model to output the bullet screen parsing results with context tags and sentiment tags; The scene matching module pre-stores a multi-scene response strategy library, which is used to match and adapt response templates according to the streaming media scene type; The interactive memory module constructs an interactive session linked list to store historical interaction records and determine whether a new bullet screen is a follow-up question scenario; The response generation module is used to generate response content by combining the response template, parsing results, and historical interaction information; The reply output module is used to push reply content and update the session list.

2. A method for automatically replying to bullet comments in streaming media, characterized in that, Includes the following steps: S1. The bullet screen acquisition module collects user bullet screen data from the streaming media platform in real time. The bullet screen data includes at least the user ID, bullet screen text and sending timestamp. After collection, it is synchronized to the context sentiment analysis module and the interactive memory module. S2. The context and sentiment analysis module calls a pre-built semantic understanding model to parse the contextual logic and core message of the bullet screen text. At the same time, it calls a sentiment classification model to identify the sentiment tendency of the bullet screen text and outputs the bullet screen parsing results carrying contextual tags, sentiment tags and core message keywords. S3. The scene matching module first matches the corresponding scene sub-library in the pre-stored multi-scene reply strategy library according to the scene type preset by the anchor or the scene type identified by the system through the live broadcast title. Then, combined with the core demand keywords in the parsing results output in step S2, it selects the appropriate scene-based reply template from the corresponding scene sub-library. S4. The interactive memory module uses the user ID in step S1 as an index to query the user's historical interaction records in the interactive session chain. It uses a semantic association algorithm to determine whether the current bullet screen is a follow-up question to the historical reply. If it is a follow-up question, it extracts the related question and answer information from the historical interaction records and feeds it back to the reply generation module. S5. The reply generation module receives the reply template filtered in step S3, the bullet screen parsing results in step S2, and the historical interaction association information in step S4. It calls the natural language generation model to optimize and adjust the reply template so that the reply content matches the contextual tags and sentiment tags of the bullet screen. At the same time, it associates historical information to form a coherent reply and generates the final reply content. S6. The reply output module pushes the final reply content generated in step S5 to the bullet screen area of ​​the streaming media platform in real time. At the same time, it associates the reply content with the corresponding user bullet screen data and synchronizes it to the interactive memory module, which then updates the interactive session list corresponding to the user ID.

3. The method for automatically replying to bullet comments in streaming media according to claim 2, characterized in that, The sentiment classification model of the context sentiment analysis module is trained by the bullet screen corpus labeled with sentiment tags, and can identify complex emotional expressions such as metaphor and irony in bullet screens. The multi-scenario response strategy library of the scenario matching module includes sub-libraries such as game live streaming, knowledge popularization, and product sales. Each sub-library is associated with a corresponding set of scenario keywords and scenario-based professional response templates. The interactive memory module's session list is indexed by user ID and stores all the user's bullet comments and corresponding system replies in a single live stream.

4. The method for automatically replying to bullet comments in streaming media according to claim 3, characterized in that, It also includes a model dynamic optimization module, which is used to collect user feedback data on system replies during a single live stream. The feedback data includes the emotional tendency of users' follow-up comments on system replies and the operation records of the anchor manually correcting replies. The model dynamic optimization module calculates the response adaptation score based on the feedback data. When the score is lower than the preset threshold, it automatically extracts the bullet screen response samples with insufficient adaptation and performs incremental training on the semantic understanding model, sentiment classification model of the context sentiment analysis module, and natural language generation model of the response generation module to achieve real-time iterative optimization of model parameters.

5. The method for automatically replying to bullet comments in streaming media according to claim 4, characterized in that, It also includes a user profile building module, which is linked with the interaction memory module. Using the user ID as an index, it extracts the user's interest keywords, question preferences, and emotional acceptance style based on historical interaction records to generate a personalized user profile. When filtering response templates, the scene matching module adjusts the tone and content focus of the templates based on user profiles. When the response generation module calls the natural language generation model, it personalizes the response content according to user profiles to match the interaction habits of specific users.

6. The method for automatically replying to bullet comments in streaming media according to claim 5, characterized in that, It also includes a security filtering module, which pre-stores a database of prohibited keywords, a database of sensitive topics, and a database of risk response patterns; After the response generation module outputs the final response content and before the response output module pushes it, the security filtering module performs triple verification on the response content: verifying whether it contains prohibited keywords, whether it involves sensitive topics, and whether it matches the risk response mode; If the verification fails, the security filtering module triggers the interception mechanism and sends a response to the response generation module. The response generation module then regenerates a response that meets the security requirements based on the original parsing result. At the same time, it records the violation verification log and synchronizes it to the scenario matching module, updating the risk avoidance template in the corresponding scenario sub-library.

7. The method for automatically replying to bullet comments in streaming media according to claim 6, characterized in that, It also includes a module collaborative scheduling module, which is communicatively connected to the context sentiment analysis module, the model dynamic optimization module, and the security filtering module, respectively. The module collaborative scheduling module has a built-in load monitoring unit and a priority allocation unit. The load monitoring unit collects the operating load data of each module in real time, and the priority allocation unit allocates processing priorities according to the urgency of the bullet comments. When a high-load module is detected, the processing tasks of low-priority bullet comments are automatically temporarily diverted to an idle module. At the same time, the incremental training process of the model dynamic optimization module is controlled to be executed during off-peak hours to avoid response delays caused by resource contention between modules.

8. The method for automatically replying to bullet comments in streaming media according to claim 7, characterized in that, It also includes a dynamic resource adaptation module, which is linked with the bullet screen acquisition module and the reply generation module; The resource dynamic adaptation module is pre-configured with multiple resource allocation schemes and adjusts resource allocation based on the bullet screen traffic data collected in real time by the bullet screen acquisition module: when the bullet screen traffic is lower than the preset low threshold, the basic resource scheme is used to reduce CPU and memory usage; when the bullet screen traffic is in the middle threshold range, the standard resource scheme is enabled to ensure processing efficiency. When the volume of bullet comments exceeds a high threshold, an elastic scaling solution is triggered to call up backup computing resources. At the same time, the processing flow of the reply generation module is optimized through a reply simplification algorithm to ensure the real-time performance of replies in high-concurrency scenarios.

9. The method for automatically replying to bullet comments in streaming media according to claim 8, characterized in that, It also includes a special scene adaptation module, which pre-stores a special scene rule library. The special scenes include overlapping anchor speeches, dense barrage of bullet comments, and specific activity interaction scenarios. When the bullet screen data collected by the bullet screen collection module meets the special scenario triggering conditions, the special scenario adaptation module executes the corresponding adaptation strategy: when the broadcaster's speech overlaps, the reply is delayed and pushed to the gap between the broadcaster's speech. When the barrage of comments is dense, a unified reply is generated by aggregating comments with the same message and displayed at the top; When interacting with specific activities, a unique response template is matched with the activity entry link, and the activity interaction tag is simultaneously synced to the interaction memory module.

Images