A system and method for dynamically switching a plurality of interactive subjects in a live broadcast

Abstract

The application discloses a system and method for dynamically switching multiple interactive subjects in a cooperative live broadcast, which comprises a subject feature modeling module, a cooperative semantic perception module, a multi-subject switching module, a context migration module, a multi-source context perception module, and a candidate reply generation engine. The system learns representations based on historical text interaction samples of multiple interactive subjects to obtain expression feature models of each interactive subject. The user's text interaction information and cooperative operation information are semantically encoded to obtain cooperative interaction features. The target interactive subject identity is output by a pre-set identity matching neural network. When the interactive subject identity is switched, the inherited context features are generated, and the fused context features are obtained through attention fusion, and then the candidate replies conforming to the target interactive subject identity are generated. The system can reduce the problems of identity confusion and context break in cooperative live broadcast interaction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence interaction technology, and in particular to a collaborative live streaming multi-interactive subject dynamic switching system and method. Background Technology

[0002] With the development of the live streaming industry, collaborative live streaming has gradually become a common form of live streaming interaction. In collaborative live streaming scenarios, multiple interactive entities typically share the same live streaming account and take turns participating in the interaction according to a schedule, manual scheduling, or user requests. Different interactive entities have different expression styles and historical interaction information, and the same user may also form different interactive relationships with different interactive entities.

[0003] Existing intelligent interaction systems typically target a single interaction object, generating responses based on user input and historical context. When applied to collaborative live streaming scenarios, these systems struggle to accurately determine the appropriate interaction subject to handle the current user interaction based on the collaborative operation status, user relationships, and historical interaction preferences. Furthermore, when the interaction subject changes, the system may continue to use the expressive characteristics of the previous interaction subject or discard existing contextual information, leading to a mismatch between the generated response and the actual interaction subject.

[0004] It is evident that existing intelligent interaction systems still have certain limitations in collaborative live streaming scenarios: after the interaction subject is switched, there is a tendency for the previous context to be lost, different interaction subject expression features to be mixed, and the response content to be unnatural, which affects the continuity and accuracy of multi-interaction subject collaborative interaction. Summary of the Invention

[0005] In view of the above problems, this application provides a collaborative live streaming multi-interactive subject dynamic switching system and method, which can improve the problem of dialogue interruption and thus enhance the interactive experience during collaborative live streaming.

[0006] To achieve the above objectives, the inventors provide a collaborative live streaming multi-interactive subject dynamic switching system, comprising:

[0007] The subject feature modeling module is used to perform representation learning based on historical text interaction samples of multiple interactive subjects to obtain the expression feature model of each interactive subject;

[0008] The collaborative semantic awareness module is used to semantically encode user text interaction information and collaborative operation information to obtain collaborative interaction features.

[0009] The multi-subject switching module is used to input the collaborative interaction features and each of the expression feature models into a preset identity matching neural network and output the identity of the target interaction subject;

[0010] The context transfer module is used to generate inherited context features based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity when it is determined that a switch from the current interaction subject identity to another interaction subject identity is required based on the target interaction subject identity.

[0011] A multi-source context awareness module is used to perform attention fusion on the collaborative interaction features and the inherited context features to obtain fused context features;

[0012] A candidate response generation engine is used to input the fused context features and the expression feature model of the target interaction subject's identity into a preset response generation neural network to generate candidate responses.

[0013] To address the aforementioned issues, the inventors also provide a method for dynamic switching of multiple interactive subjects in collaborative live streaming, comprising:

[0014] Representation learning is performed based on historical text interaction samples from multiple interactive subjects to obtain the expression feature model of each interactive subject;

[0015] Semantic encoding is performed on user text interaction information and collaborative operation information to obtain collaborative interaction features;

[0016] The collaborative interaction features and each of the expression feature models are input into a preset identity matching neural network, which outputs the identity of the target interaction subject.

[0017] When it is determined that a switch from the current interaction subject identity to another interaction subject identity is needed based on the target interaction subject identity, an inherited context feature is generated based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity.

[0018] Attention fusion is performed on the collaborative interaction features and the inherited context features to obtain fused context features;

[0019] The fused context features and the expression feature model of the target interactive subject's identity are input into a preset response generation neural network to generate candidate responses.

[0020] Unlike existing technologies, the above technical solution uses representation learning based on historical text interaction samples of multiple interactive subjects to obtain the expression feature models of each interactive subject, enabling the expression styles of different interactive subjects under a collaborative interaction account to be independently represented. By semantically encoding user text interaction information and collaborative operation information, collaborative interaction features are obtained. These collaborative interaction features, along with the expression feature models of each subject, are input into a preset identity matching neural network. This network can determine the identity of the target interactive subject by combining user interaction content and the collaborative live streaming operation status, improving the accuracy of switching between multiple interactive subject identities. When switching from the current interactive subject identity to another, inherited context features are generated based on user text interaction information, historical interaction information corresponding to the target interactive subject identity, and historical interaction information corresponding to the current interactive subject identity. This ensures that the switched interactive subject identity can inherit the previous interaction content. Simultaneously, by performing attention fusion on the collaborative interaction features and inherited context features, and combining them with the expression feature model of the target interactive subject identity to generate candidate responses, the candidate responses conform to both the current interactive context and the expression style of the target interactive subject identity. This reduces the degree of confusion in expression features during collaborative live streaming interactions and improves the problem of dialogue breakage, thereby enhancing the interactive experience during collaborative live streaming.

[0021] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0022] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0023] In the accompanying drawings of the instruction manual:

[0024] Figure 1 This is a schematic diagram of the architecture of a collaborative live streaming multi-interactive subject dynamic switching system in an embodiment of the invention.

[0025] Figure 2 This is a flowchart of a collaborative live streaming multi-interactive subject dynamic switching method in an embodiment of the invention. Detailed Implementation

[0026] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0027] Example 1

[0028] This embodiment provides a collaborative live streaming multi-interactive subject dynamic switching system. The system includes a subject feature modeling module, a collaborative semantic perception module, a multi-subject switching module, a context transfer module, a multi-source context perception module, and a candidate response generation engine.

[0029] The subject feature modeling module is used to perform representation learning based on historical text interaction samples of multiple interactive subjects to obtain the expression feature model of each interactive subject.

[0030] Specifically, in collaborative live streaming scenarios, multiple interactive entities can be associated with the same main account. Each interactive entity has a corresponding entity identifier and is associated with an independent expressive feature model. Historical text interaction samples can include the corresponding interactive entity's historical real-time replies, targeted replies, comment section replies, social media posts, short video comment replies, interactive entity's selection records of candidate replies, interactive entity's editing records of candidate replies, and interactive entity's ignoring records of candidate replies. The system collects the above historical text interaction samples according to the interactive entity's identity and writes the samples of different interactive entities into the data space of the corresponding interactive entity to avoid mixing interactive samples from different interactive entities.

[0031] In some implementations, the subject feature modeling module is used to divide historical text interaction samples of the same interactive subject into active response samples, user maintenance samples, and restricted processing samples, and learn the expression stable features in the active response samples, user maintenance samples, and restricted processing samples respectively, and generate the expression feature model of the interactive subject using the expression stable features that satisfy the stability condition among the expression stable features.

[0032] Among them, proactive response samples can be direct response texts from the interacting entity to ordinary real-time messages, questions, interactive guidance, etc.; user maintenance samples can be relationship maintenance texts from the interacting entity to historical users, resource interaction users, team users, targeted users, etc.; restricted processing samples can be response texts from the interacting entity to avoid, transfer, or restrict content such as preset restricted questions, preset restricted topics, platform rule terms, brand restrictions, etc. Stable expression characteristics can include sentence structure, address, verbal expression, intensity of emotional expression, use of emoticons, usage of time-sensitive words, interactive guidance methods, and restricted processing expression methods. Stable conditions can include at least one of the following: repeated occurrence within a preset time range, occurrence frequency reaching a preset frequency in the same type of sample, repeated adoption by the interacting entity in candidate responses, and failure to hit any restricted expression record.

[0033] For example, if interaction subject A frequently uses short sentences, specific titles, and specific closing phrases in proactive response samples, frequently uses expressions with high emotional intensity in user maintenance samples, and frequently uses expressions that change the subject or provide gentle reminders in restricted processing samples, then the system can learn these stable expression features respectively and write the expression features that meet the stability conditions into the expression feature model of interaction subject A. If interaction subject B prefers formal sentence structures, a few emoticons, and low emotional intensity, then the system will generate an expression feature model for interaction subject B that is different from that of interaction subject A. Thus, the expression feature model of each interaction subject is derived from its own historical text interaction samples, rather than a single response style uniformly set by the main account.

[0034] In some implementations, the subject feature modeling module may include a data collector, a language style parser, expression restriction rules, a time-sensitive word update unit, an expression feature file, an A / B testing submodule, a dynamic learning submodule, and a data isolation manager.

[0035] The data collector gathers historical text interaction samples from sources such as live streaming platforms, targeted channels, comment sections, and social media, adding interaction subject identifiers, user identifiers, message timestamps, message sources, and interaction scenario identifiers to the samples. The language style parser analyzes the historical text interaction samples, extracting sentence structure features, emotional intensity, verbal expressions, symbol usage rules, and usage habits of timely words. Expression restriction rules record expressions that cannot be used by the interaction subject, platform rule terms, brand cooperation restriction terms, terms that are easily misunderstood by users, and manually marked restricted expressions. The timely word update unit periodically acquires high-frequency words or popular expressions in live streaming rooms, and after review by the interaction subject or operations personnel, adds approved timely words to the corresponding interaction subject's expression feature model. The expression feature file structurally stores the corresponding interaction subject's sentence structure feature library, emotional intensity mapping table, idiomatic expression lexicon, symbol usage rules, timely word update records, expression restriction rules, and weight parameters. The A / B testing submodule compares the actual effects of different response styles in the same or similar interaction scenarios and adjusts the expression feature model based on the interaction subject's adoption and subsequent user interactions. The dynamic learning submodule periodically reads the historical interaction records and candidate response feedback records of the interactive subjects, analyzes the changing trends of the interactive subjects' language styles, and generates style change results that can be used to update the expression feature model. The data isolation manager isolates and stores the expression feature files, user feature library, and interaction record library of each interactive subject, and records access logs.

[0036] In some implementations, the subject feature modeling module is also used to obtain the selection, editing, and ignoring results of the interactive subject on the candidate responses, and to correct the weights of the corresponding stable expression features in the expression feature model based on the selection, editing, and ignoring results.

[0037] Specifically, when an interactive subject selects a candidate response, the system can determine that the sentence structure, verbal expressions, emoticons, time-sensitive words, or restrictive expressions in that candidate response have been adopted by the interactive subject, and increase the weight of the corresponding stable expression features. When an interactive subject edits a candidate response, the system can compare the text before and after editing, using retained or added expressions as positive correction criteria and deleted or replaced expressions as negative correction criteria. When an interactive subject ignores a candidate response, the system can reduce the weight of the corresponding stable expression features in that candidate response under the same or similar interaction scenarios. Through the above methods, the expression feature model can be continuously updated with the actual usage behavior of the interactive subject, rather than relying solely on the initial samples to remain unchanged.

[0038] The collaborative semantic perception module is used to semantically encode user text interaction information and collaborative operation information to obtain collaborative interaction features.

[0039] Specifically, user text interaction information can include real-time messages, comments, targeted messages, questions, directions, support messages, resource interaction messages, and general team-related messages sent by users in the live broadcast room. Collaborative operation information can include the interaction subject's scheduling information, the current interaction subject's identity, the interaction subject's presence status, the subject's pending activation status, the subject's operating status, manual switching instructions, scheduling instructions, user direction results, and the user's historical interaction preferences, interaction density, and group selection tendencies towards the interaction subject.

[0040] In some implementations, the collaborative semantic perception module is used to extract user interaction semantics and interaction subject pointing relationships from the user text interaction information; extract manual switching instructions, user pointing results, user's historical interaction preferences for interaction subjects, and current interaction subject identity from the collaborative operation information; and encode the user interaction semantics, interaction subject pointing relationships, manual switching instructions, user pointing results, historical interaction preferences, and current interaction subject identity into the collaborative interaction features.

[0041] Among them, user interaction semantics is used to characterize the user's current text interaction purpose, such as asking questions, praising, urging, calling out names, resource interaction feedback, emotional expression, or team interaction. Interaction subject pointing relationship is used to characterize whether the user's text interaction information explicitly points to a specific interaction subject, such as whether it contains the interaction subject's name, alias, title, exclusive title, or expressions strongly associated with that interaction subject. Manual switching instructions can be structured scheduling instructions issued by dispatchers or operations personnel. User pointing result can be the interaction subject pointing result obtained after identifying the user's text interaction information. The user's historical interaction preferences for interaction subjects can be determined based on the user's interaction frequency, response rate, resource interaction behavior, dwell time, and relationship stage under different interaction subject identities. Current interaction subject identity is used to characterize the interaction subject identity currently participating in or being defaulted to by the system in the current live stream.

[0042] In its implementation, the collaborative semantic perception module can use a text semantic encoding model to encode user text interaction information, obtaining a user text semantic vector; and use a state encoding model to encode collaborative operation information, obtaining a collaborative operation state vector; then, the user text semantic vector and the collaborative operation state vector are concatenated, weighted, or fused with attention to obtain collaborative interaction features. These collaborative interaction features retain both user text content and collaborative operation information such as interaction subject rotation, manual scheduling, user input, and the current interaction subject's identity, providing an input basis for subsequent target interaction subject identity matching.

[0043] The multi-subject switching module is used to input the collaborative interaction features and each of the expression feature models into a preset identity matching neural network, and output the identity of the target interaction subject.

[0044] Specifically, the pre-defined identity matching neural network is used to determine the target interaction subject that will receive the current user's text interaction information based on the matching relationship between collaborative interaction features and the expression feature models of each interaction subject. The target interaction subject identity can be the current interaction subject identity, or it can be another interaction subject identity determined based on manual switching instructions, user pointing results, user historical interaction preferences, or team interaction patterns.

[0045] In some implementations, the multi-subject switching module is used to determine the identity of the target interactive subject according to a preset routing priority, wherein the preset routing priority includes, in order, manual switching instructions, user pointing results, user's historical interaction preferences for interactive subjects, current interactive subject identity, and team interaction mode.

[0046] The system includes several key features: manual switching commands (from dispatchers or operations personnel, indicating that a specific interaction subject will be manually designated to handle the current interaction); user-directed results (user text messages explicitly targeting a particular interaction subject); user historical interaction preferences (users showing a clear preference for a particular interaction subject in their past interactions); current interaction subject identity (handling general interactions without explicit targeting or strong historical preferences); and team interaction mode (handling general messages for the entire team, such as users expressing their affection for the group, inquiring about team activities, or initiating team-related interactions). By setting these preset routing priorities, the system can prioritize explicit scheduling and explicit referencing in complex collaborative live streaming interactions, while also considering historical relationships and the current live streaming status.

[0047] In some implementations, the multi-subject switching module is used to determine the priority pointing result according to the preset routing priority when the manual switching instruction, the user pointing result, the user's historical interaction preference for the interactive subject, and the current interactive subject identity point to different interactive subject identities, and input the priority pointing result and the collaborative interaction feature into the preset identity matching neural network to output the target interactive subject identity.

[0048] Specifically, when a manual switching command points to interaction subject A, a user-directed result points to interaction subject B, historical interaction preferences point to interaction subject C, and the current interaction subject identity is interaction subject D, the system first determines the priority direction result according to a preset routing priority. This priority direction result is not a new feature generated independently of the preceding steps, but is obtained by sorting the manual switching command, user-directed result, historical interaction preferences, and current interaction subject identity extracted by the collaborative semantic perception module according to the preset routing priority. Subsequently, the system inputs the priority direction result and collaborative interaction features into a preset identity matching neural network, which combines the expressive feature models of each interaction subject to output the target interaction subject identity. This preserves the routing priority relationship in collaborative live streaming while avoiding directly determining the target interaction subject identity through fixed rules, enabling the target interaction subject identity output to be comprehensively determined by combining semantic content and expressive feature matching relationships.

[0049] In some implementations, the multi-subject switching module may further include a schedule switcher, a dispatch instruction receiver, a performance status sensor, a user pointing identifier, and an identity matching unit. The schedule switcher outputs candidate interactive subject identities based on the interactive subject rotation schedule; the dispatch instruction receiver receives structured switching instructions from dispatchers or operators; the performance status sensor identifies whether an interactive subject is on stage, waiting, performing, interacting, or leaving; the user pointing identifier identifies the interactive subject's name, alias, or unique title in user text interaction information; and the identity matching unit inputs the above results and collaborative interaction features into a preset identity matching neural network to determine the target interactive subject's identity.

[0050] The context transfer module is used to generate inherited context features based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity when it is determined that a switch from the current interaction subject identity to another interaction subject identity is required based on the target interaction subject identity.

[0051] Specifically, the current interaction subject identity refers to the identity of the interaction subject that was currently undertaking the interaction before the switch. If the target interaction subject identity is the same as the current interaction subject identity, the system can directly generate candidate responses based on the expression feature model corresponding to the target interaction subject identity; if the target interaction subject identity is different from the current interaction subject identity, it indicates that an interaction subject identity switch is required. In this case, the system does not directly discard the context formed under the current interaction subject identity, nor does it directly adopt the expression style of the current interaction subject identity, but instead generates inherited context features for semantic continuity.

[0052] In some implementations, the context transfer module is used to identify incomplete interaction content from the user text interaction information when it is necessary to switch the interaction subject identity, and encode the incomplete interaction content, the current interaction subject identity, the target interaction subject identity, and the switching reason into inherited task features; and match the inherited task features with the historical interaction information corresponding to the target interaction subject identity to generate the inherited context features.

[0053] Unfinished interactions may include unanswered user questions, commitments from the previous conversation, ongoing interactive tasks, emotions expressed by the user but not yet responded to, and requests made by the user but not yet answered. Reasons for switching may include manual switching commands, user-initiated actions, shift changes, changes in performance status, or team interaction modes. Historical interaction information corresponding to the target interaction subject may include the user's historical dialogues with the target interaction subject, user characteristic tags, relationship stage, content preferences, attention items, and interaction frequency. Historical interaction information corresponding to the current interaction subject can be used to explain the source and prior relationships of the dialogue before the switch, but it will not be directly used as a style constraint for the target interaction subject.

[0054] In some implementations, the inherited task characteristics can be encoded from the inherited task data packet. The inherited task data packet may include an inherited task identifier, the original active interaction subject identity, the target interaction subject identity, a trigger reason, pending messages, a user characteristic snapshot, dialogue history, a context summary, and a handover timestamp. The inherited task identifier is used to distinguish different inherited tasks; the original active interaction subject identity records the interaction subject identity that took over the dialogue before the switch; the target interaction subject identity records the interaction subject identity that took over the dialogue after the switch; the trigger reason records the source of the identity switch; pending messages record user messages that need further processing; the user characteristic snapshot records the necessary interactive relationships formed by the user under the original active interaction subject identity; the dialogue history records previous key dialogues; the context summary compresses and expresses the key points of previous interactions; and the handover timestamp records the time the inheritance occurred.

[0055] After receiving the inherited context features, the target interaction subject generates candidate responses using its own expressive feature model. In other words, the inherited context features are primarily used to convey incomplete interactions and preceding semantics, while the expressive feature model mainly constrains the expressive style of the generated responses. In this way, the newly activated interaction subject can understand the incomplete interactions of the previous interaction subject, but the candidate responses still conform to the expressive style of the newly activated interaction subject, avoiding inconsistent response styles.

[0056] The multi-source context awareness module is used to perform attention fusion on the collaborative interaction features and the inherited context features to obtain fused context features.

[0057] Specifically, the multi-source context awareness module can simultaneously collect the current conversation layer context, the relationship data layer context, and the group context layer context. The current conversation layer context can include recent dialogue history, current messages, dialogue rounds, current topics, and unfinished interaction content. The relationship data layer context can include user characteristic tags under the target interaction subject identity, relationship stage, resource interaction level, interaction style, content preferences, attention items, and recent emotional state. The group context layer context can include real-time message density in the live stream, interaction popularity, resource concentration periods, group selection tendencies, current performance stage, and trending topics.

[0058] In some implementations, the multi-source context awareness module includes a current conversation layer collector, a relational data layer loader, a group context layer collector, a context fusion engine, a weight dynamic adjuster, and a fusion context output interface. The current conversation layer collector is used to capture the dialogue message stream in real time; the relational data layer loader is used to read tag data from the user feature library corresponding to the target interaction subject's identity; the group context layer collector is used to calculate real-time atmosphere data through real-time message analysis in the live stream; the weight dynamic adjuster is used to adjust the fusion weights of different contexts according to the target interaction subject's identity, the live stream interaction density, and the current interaction scenario; the context fusion engine is used to perform attention fusion of the above contexts with collaborative interaction features and inherited context features; and the fusion context output interface is used to output fused context features to the candidate response generation engine.

[0059] In some implementations, the same user can maintain separate user feature information under different interaction subject identities. For example, the same user can be marked as a high-frequency interaction user in the user feature database corresponding to interaction subject A, as a regular interaction user in the user feature database corresponding to interaction subject B, and as a resource interaction user in the user feature database corresponding to interaction subject C. When the target interaction subject identity is interaction subject B, the system prioritizes calling the user's user feature information under interaction subject B, rather than directly using the user's user feature information under interaction subject A or interaction subject C. Therefore, candidate responses can adapt to the actual relationship between the target interaction subject identity and the user, reducing the probability of misjudging the interaction relationship.

[0060] In some implementations, when the message density in the live stream is high and the interaction pace is fast, the weight dynamic adjuster can increase the weight of the current session layer context and the group context layer context, and decrease the weight of long historical context, making candidate responses shorter and more timely. When the interaction density in the live stream is low and users are engaged in in-depth communication, the weight dynamic adjuster can increase the weight of the relationship data layer context and the target interaction subject's identity historical interaction context, making candidate responses more suitable for relationship maintenance.

[0061] The candidate response generation engine is used to input the fused context features and the expression feature model of the target interactive subject's identity into a preset response generation neural network to generate candidate responses.

[0062] Specifically, contextual features are used to limit the current semantics, collaborative live streaming status, interaction relationships, and inherited contextual content that candidate responses need to inherit; the expression feature model of the target interaction subject's identity is used to limit the expression style, emotional intensity, verbal expression, use of emojis, and expression restrictions that candidate responses should conform to. The preset response generation neural network can generate at least one candidate response based on the above inputs, or it can generate multiple candidate responses from different strategy perspectives.

[0063] In some implementations, the candidate response generation engine may include a constraint receiver, a multi-strategy decoder, a primary strategy generator, an alternative strategy generator, a short strategy generator, an effect evaluator, and a rule hint generator. The constraint receiver receives expression feature constraint vectors, user feature constraint vectors, and a fusion context vector. The multi-strategy decoder generates candidate responses with different strategies using different decoding parameters. The primary strategy generator generates candidate responses that best match the target interaction subject's identity and the current context. The alternative strategy generator generates style-differentiated candidate responses, such as humorous versions, formal versions, appeasing versions, or guiding interaction versions. The short strategy generator generates short candidate responses under high-load interaction conditions. The effect evaluator calculates the expected interaction effect score for each candidate response. The rule hint generator detects whether candidate responses involve restrictive expressions, restricted topics, platform rule terms, incorrect interaction subject style, or expressions unsuitable for the current interaction relationship, and generates rule hint tags.

[0064] In some implementations, before generating candidate responses, the candidate response generation engine can constrain and filter the fused context features based on the expression restriction rules in the expression feature model of the target interaction subject's identity; after generating candidate responses, it can perform rule detection again on the candidate responses. When a candidate response hits a restricted expression, the system regenerates an alternative candidate response; when consecutively generated candidate responses all contain abnormal expressions, the system can generate safe candidate content that prompts the interaction subject for manual handling.

[0065] Example 2

[0066] This embodiment provides a collaborative live streaming method for dynamic switching of multiple interactive subjects, corresponding to Embodiment 1. This method can be executed by the system described in Embodiment 1.

[0067] The method includes the following steps:

[0068] S100 uses historical text interaction samples from multiple interactive subjects to perform representation learning, resulting in an expression feature model for each interactive subject.

[0069] In some implementations, S100 includes:

[0070] S110, the historical text interaction samples of each interactive subject are divided into active response samples, user maintenance samples and restriction processing samples according to the interaction purpose;

[0071] S120, learn the expression stability features in the active response sample, the user maintenance sample and the restriction processing sample respectively;

[0072] S130, using the expression features that satisfy the stability condition among the expression stable features to generate the expression feature model of the interactive subject.

[0073] Specifically, proactive response samples are used to learn the expression habits of interactive subjects when facing ordinary interactions; user maintenance samples are used to learn the maintenance methods of interactive subjects corresponding to different stages of interaction relationships; and restricted processing samples are used to learn the expression boundaries of interactive subjects in scenarios with preset restricted questions or where direct responses are not appropriate. When learning stable expression features, the system can identify sentence structure, oral expression, emotional intensity, emoticon usage, timely words, interaction guidance methods, and restricted processing expressions, and write expression features that meet the stability conditions and do not hit restricted expression records into the expression feature model of the corresponding interactive subject.

[0074] After generating candidate responses, the method may also include:

[0075] S140, Obtain the result of the interaction subject's selection, editing, or ignoring of the candidate response;

[0076] S150, based on the selection, editing, or ignoring results, adjust the weights of stable features in the expression feature model.

[0077] Specifically, when the interacting subject selects a candidate response, the weight of the stable features of the adopted expression in that candidate response is increased; when the interacting subject edits a candidate response, the retained, replaced, and newly added expressions are determined based on the differences between the text before and after editing, and the weights of the stable features of the expression are adjusted accordingly; when the interacting subject ignores a candidate response, the weight of the relevant stable features of the expression in the same scenario is reduced. Through S140-S150, the expression feature model can be dynamically updated based on the actual selection and editing behavior of the interacting subject.

[0078] S200 performs semantic encoding on user text interaction information and collaborative operation information to obtain collaborative interaction features.

[0079] In some implementations, S200 includes:

[0080] S210, extract user interaction semantics and interaction subject orientation relationship from the user text interaction information;

[0081] S220, extract the manual switching instruction, user pointing result, user's historical interaction preference for the interaction subject and the current identity of the interaction subject from the collaborative operation information;

[0082] S230, the user interaction semantics, interaction subject pointing relationship, manual switching instruction, user pointing result, historical interaction preference and current interaction subject identity are encoded as the collaborative interaction features.

[0083] Specifically, user text interaction information can come from live stream messages, comments, targeted messages, or other text interaction entry points. The system determines the user's current interaction intent through semantic recognition and identifies the interaction subject's relationship through the interaction subject's name, alias, title, exclusive title, or contextual keywords. Collaborative operation information can come from schedules, dispatch instructions, manual operation panels, interaction subject status detection results, and historical interaction records. The system uniformly encodes the above information into collaborative interaction features, enabling the subsequent preset identity matching neural network to simultaneously perceive user text semantics and collaborative live stream operation status.

[0084] S300, the collaborative interaction features and each of the expression feature models are input into a preset identity matching neural network, and the target interaction subject identity is output.

[0085] In some implementations, S300 includes:

[0086] S310, sort the manual switching command, the user pointing result, the user's historical interaction preferences for the interactive subject, the current interactive subject identity, and the team interaction mode according to the preset routing priority;

[0087] S320, input the sorting result and the collaborative interaction feature into the preset identity matching neural network;

[0088] S330, the target interaction subject identity is output by the preset identity matching neural network.

[0089] Specifically, the preset routing priorities can include, in order: manual switching instructions, user-directed results, user's historical interaction preferences for the interactive subject, the current interactive subject's identity, and team interaction mode. When multiple directed results conflict, the system generates a ranking result based on the preset routing priorities and uses this ranking result as one of the inputs to the preset identity matching neural network, rather than directly replacing the neural network output with the rule-based result. The preset identity matching neural network further combines collaborative interaction features and the expressive feature models of each interactive subject to output the target interactive subject's identity.

[0090] S400, when it is determined that a switch from the current interaction subject identity to another interaction subject identity is required based on the target interaction subject identity, an inherited context feature is generated based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity.

[0091] In some implementations, S400 includes:

[0092] S410, determine whether the identity of the target interaction subject is consistent with the identity of the current interaction subject;

[0093] S420: When the identity of the target interaction subject is inconsistent with the identity of the current interaction subject, identify the incomplete interaction content from the user's text interaction information;

[0094] S430, encode the incomplete interaction content, the current interaction subject identity, the target interaction subject identity, and the switching reason as inherited task features;

[0095] S440, the inherited task features are matched with the historical interaction information corresponding to the identity of the target interaction subject to generate inherited context features.

[0096] Specifically, unfinished interactions can include unanswered user questions, unfulfilled promises from previous interactions, or unresponsive emotions or requests expressed by the user. Switching reasons can include manual switching instructions, user input, scheduling triggers, changes in performance status, or team interaction patterns. Matching inherited task characteristics with historical interaction information corresponding to the target interaction subject's identity ensures that when the target interaction subject receives previous content, it can consider its existing relationship with the user and its interaction preferences, rather than simply relying on the current interaction subject's dialogue relationship.

[0097] S500, attention fusion is performed on the collaborative interaction features and the inherited context features to obtain fused context features.

[0098] Specifically, the system can further integrate the current conversation layer context, the relational data layer context, and the group context layer context. The current conversation layer context includes recent dialogue history and current messages; the relational data layer context includes user feature tags and historical interaction relationships under the target interaction subject's identity; and the group context layer context includes live stream message density, group sentiment, and real-time atmosphere data. The system calculates the attention weights of different contexts based on the target interaction subject's identity and the current live stream state, and generates fused contextual features.

[0099] S600, the fusion context features and the expression feature model of the target interactive subject identity are input into a preset response generation neural network to generate candidate responses.

[0100] Specifically, the preset response generation neural network uses contextual features as content constraints and the expressive feature model of the target interaction subject as style constraints to generate candidate responses. Candidate responses can include primary candidate responses, style variant candidate responses, and short candidate responses. Each candidate response can be accompanied by an expected interaction effect score and rule-based hint labels for the interaction subject to choose from.

[0101] Example 3

[0102] This embodiment, based on Embodiment 1 and Embodiment 2, further provides an implementation method for interactive subject decision control.

[0103] In some implementations, the system also includes a subject confirmation interface. This interface displays candidate responses, receives instructions from the interacting subject to select, edit, or ignore candidate responses, and only performs a message sending operation after the interacting subject actively confirms the response.

[0104] Specifically, after the candidate response generation engine generates candidate responses, it displays the candidate responses, target subject identifier, user feature summary, context inheritance hints, expected interaction effect score, and rule hint labels on the subject confirmation interface. The interacting subject can select a candidate response, edit a candidate response, or ignore all candidate responses. The system does not automatically send candidate responses; only after the interacting subject triggers a confirmation send operation will the confirmed response be sent to the corresponding user's client.

[0105] When an interactive subject selects a candidate response, the system associates that candidate response, fused contextual features, the target interactive subject's identity, and the version of the expression feature model as a positive sample. When the interactive subject edits a candidate response, the system records the text differences before and after editing and uses these differences as style correction samples for the target interactive subject's identity. When the interactive subject ignores a candidate response, the system marks that candidate response and its corresponding context as negative samples. These positive samples, style correction samples, and negative samples can be used to subsequently refine the expression feature model or train models related to response generation.

[0106] Example 4

[0107] Based on the above embodiments, this embodiment further provides a two-way synchronous implementation method for live interactive recording.

[0108] In some implementations, the system also includes a two-way synchronization module for live interaction records. This module synchronizes interaction records during the live stream to the interaction subject's dedicated backend account and the user's targeted messaging terminal, and monitors the synchronization status.

[0109] Specifically, the live stream interaction record bidirectional synchronization module can include a message capturer, a subject identifier, a data mapping router, a bidirectional synchronization engine, a synchronization status monitor, and a user-side synchronizer. The message capturer captures user messages, confirmed replies from interacting subjects, emoticons, links, images, and candidate reply confirmation records generated during the live stream. The subject identifier determines the identity of the interacting subject to which the interaction message belongs based on the target interacting subject's identity, the current interacting subject's identity, the identity matching result, and the message source. The data mapping router maps the interaction messages to the data space of the corresponding interacting subject's dedicated account. The bidirectional synchronization engine synchronizes the interaction messages to the interacting subject's backend dedicated account and the user-directed messaging client. The synchronization status monitor records the synchronization status, retries, and exception logs. The user-side synchronizer enables the user-directed messaging client to view the corresponding interaction history.

[0110] In some implementations, the synchronized content may include a complete interaction log, message timestamps, subject identifier, user identifier, message type, synchronization status, retry count, and context flags. When synchronization fails, the system can automatically perform a preset number of retries; if the retries still fail, an exception log is generated and marked as requiring manual handling. Through this implementation, the interactive data generated during the live stream can remain consistent between the interactive subject's backend dedicated account and the user-targeted messaging terminal, providing a data foundation for continued interaction after the live stream ends.

[0111] Example 5

[0112] Based on the above embodiments, this embodiment further provides an implementation method for synchronously inheriting interaction records after the live broadcast ends.

[0113] In some implementations, the system also includes a module for synchronizing and inheriting interaction records after the live stream ends. This module is used to migrate the complete interaction records, incremental update data of user features, a list of pending messages, and contextual continuity markers to the corresponding interaction subject's dedicated account after the live stream ends.

[0114] Specifically, the interaction record synchronization and inheritance module after the live stream ends can include a real-time interaction record pool, a synchronization engine, a live stream end-archiving packager, a dedicated account data writer, a context continuity validator, a dedicated account database for the interaction subject, a user feature update synchronizer, and a pending message marker. The real-time interaction record pool temporarily stores dialogue data generated during the live stream and indexes it according to the interaction subject's identity. The live stream end-archiving packager categorizes and summarizes interaction records according to the interaction subject's identity at the end of the live stream, generating an inheritance data package corresponding to the interaction subject's identity. The dedicated account data writer writes the inheritance data package to the corresponding interaction subject's dedicated account. The context continuity validator detects whether there are context breakpoints in the written dialogue and generates a context summary for dialogues with breakpoints. The user feature update synchronizer synchronizes newly added user tags, resource interaction behaviors, interaction preferences, and attention items during the live stream to the user feature database of the corresponding interaction subject's dedicated account. The pending message marker marks unprocessed user messages and can sort them according to user priority, interaction activity, or duration of no response.

[0115] Through this implementation method, user interactions generated during the collaborative live stream can be stored in the corresponding interactive entity's dedicated account. After the live stream ends, the interactive entity can continue to interact with users based on the complete interaction record, user feature update data, pending message list, and contextual continuity markers, avoiding the interruption of the dialogue relationship after the live stream ends.

[0116] Example 6

[0117] This embodiment, based on the above embodiments, further provides a cross-platform resource distribution implementation method.

[0118] In some implementations, the system also includes a cross-platform resource distribution module. This module is used to convert the live stream content, links, or resources according to the target platform's rules, perform compliance checks, and generate sharing text after the interacting entity confirms the sharing intention, and then distribute the content after confirmation by the interacting entity.

[0119] Specifically, the cross-platform resource distribution module can include a platform adaptation submodule, a highlight moment marking submodule, and a multi-interactive subject content distribution submodule. The platform adaptation submodule is used to convert the format of live stream content, links, or resources according to the target platform's community guidelines, external link rules, content review standards, character limits, and media format requirements. The highlight moment marking submodule is used to identify shareable content based on message keyword bursts, peak interaction popularity, concentrated resource periods, trending topics, and interactive subject performance nodes. The multi-interactive subject content distribution submodule is used to select the appropriate expression feature model to generate sharing copy based on the identity of the interactive subject associated with the shared content.

[0120] If the shared content is primarily associated with a specific interaction subject, the system uses the expressive feature model of that subject to generate the sharing text. If the shared content is general team content, the system can generate the sharing text based on the team's interaction patterns. After generating the sharing content, the system displays it to the interaction subject or operations personnel for confirmation. Only after receiving confirmation of the sharing instruction will the system execute the distribution operation on the corresponding platform.

[0121] Example 7

[0122] This embodiment further illustrates the data isolation and interaction relationship mapping methods for multiple interactive subjects.

[0123] In some implementations, the system configures an independent data space for each interactive subject identity. Each independent data space stores at least the corresponding interactive subject's expression feature model, user feature library, interaction record library, candidate response feedback sample library, and expression restriction library. After outputting the target interactive subject identity, the system accesses the corresponding interactive subject's data space based on the target interactive subject identity; when no inherited context features are generated, the target interactive subject identity is restricted from accessing the original interaction records and original user feature information corresponding to non-target interactive subject identities; when an interactive subject identity switch occurs, only the necessary context corresponding to the current interactive subject identity is encoded into inherited context features and provided to the target interactive subject identity for use.

[0124] The same user can maintain separate user characteristic information under different interaction subject identities. User characteristic information can include relationship stage, resource interaction level, interaction style, content preferences, attention items, and emotional state. After the target interaction subject identity is determined, the system prioritizes calling the user characteristic information of that user under the target interaction subject identity to participate in the generation of fusion context features. In this way, the system can support multiple interaction subjects under the same collaborative interaction account to jointly undertake user interactions, while avoiding the mixing of expression feature models, user characteristics, and interaction records between different interaction subjects.

[0125] Example 8

[0126] This embodiment provides an implementation method for a storage medium. In some implementations, the aforementioned collaborative live streaming multi-interactive subject dynamic switching system can be deployed on a server, a live streaming backend system, a cloud AI service, an interactive subject client, an operation backend, or a combination thereof. Each module can be deployed on the same device or distributed across different devices, and communicate through interface calls, message queues, database read / write, or event streams.

[0127] Example 9

[0128] Please refer to Figure 2This embodiment provides a detailed process for a method of dynamically switching multiple interactive subjects in collaborative live streaming. This implementation can be applied to live streaming scenarios where multiple interactive subject identities exist under a collaborative interactive account. For example, a collaborative interactive account may include interactive subject A, interactive subject B, and interactive subject C, each with its own independent expressive feature model, historical interaction information, and user feature information. During the live stream, the system completes target interactive subject identity determination, context inheritance, candidate response generation, and continuous model optimization based on user text interaction information, collaborative operation information, the expressive feature models of each interactive subject, and the interactive subject's feedback results to candidate responses.

[0129] In practice, the system first acquires user text interaction information, collaborative operation information, manual switching instructions, user pointing results, user's historical interaction preferences for interaction subjects, and the current interaction subject's identity. User text interaction information can include live real-time messages, comments, targeted messages, questions, pointing messages, support messages, and resource interaction feedback messages. Collaborative operation information can include interaction subject schedules, scheduling instructions, current on-stage status, waiting status, performance status of interaction subjects, live room interaction density, and the current live stream segment. Manual switching instructions can be structured instructions issued by schedulers or operations personnel, such as specifying interaction subject B to take over the current interaction. User pointing results can be identified from user text interaction information and are used to characterize whether the user explicitly requests a response from a particular interaction subject. User historical interaction preferences can be determined based on the user's interaction frequency under different interaction subject identities, response acceptance status, resource interaction records, dwell behavior, and historical targeted message relationships. The current interaction subject identity is used to characterize the interaction subject identity that the system defaults to taking over the interaction before this round of processing. By acquiring the above information, the system does not generate a reply based solely on a single real-time message text, but simultaneously obtains user text, collaborative live streaming status, manual scheduling, user targeting, and historical preferences, providing a foundation for identifying the target interaction subject in the future.

[0130] After acquiring the above information, the collaborative semantic perception module performs semantic encoding on the user text interaction information and collaborative operation information to obtain collaborative interaction features. Specifically, the system extracts user interaction semantics and interaction subject orientation relationships from the user text interaction information. User interaction semantics can characterize whether the user is currently asking a question, urging, praising, calling out names, inquiring about purchases, expressing emotions, or engaging in team interaction. Interaction subject orientation relationships can be identified through the interaction subject's name, nickname, title, exclusive title, contextual pronouns, and preceding dialogue relationships.

[0131] The multi-subject switching module determines the identity of the target interaction subject based on collaborative interaction features. Specifically, the multi-subject switching module sorts the manual switching command, user pointing result, user's historical interaction preferences with the interaction subject, current interaction subject identity, and team interaction mode according to a preset routing priority. The preset routing priority includes, in order, the manual switching command, user pointing result, user's historical interaction preferences with the interaction subject, current interaction subject identity, and team interaction mode. For example, when the manual switching command specifies interaction subject C, and the user text-names interaction subject B, but the user has historically interacted more frequently with interaction subject A, the system first generates a priority pointing result according to the preset routing priority. Since the priority of the manual switching command is higher than the user pointing result and historical interaction preferences, the priority pointing result can be interaction subject C. Subsequently, the system does not directly use interaction subject C as the final target interaction subject identity, but instead inputs the priority pointing result and collaborative interaction features into a preset identity matching neural network, and outputs the target interaction subject identity by combining the expression feature models of each interaction subject. In this way, the preset routing priority is used to reflect the explicit scheduling relationship in the collaborative live streaming scenario, and the preset identity matching neural network is used to further combine semantic content and interactive subject expression feature representation for comprehensive judgment, so that the output of the target interactive subject identity considers both the scheduling results of the scheduler or operator and the matching relationship between the user's text semantics and the interactive subject expression feature.

[0132] After outputting the target interaction subject's identity, the system determines whether the target interaction subject's identity is consistent with the current interaction subject's identity. If the target interaction subject's identity is consistent with the current interaction subject's identity, it means that this round of interaction is still undertaken by the current interaction subject's identity, and the system can directly proceed to subsequent context fusion processing without generating inherited context features. If the target interaction subject's identity is inconsistent with the current interaction subject's identity, it means that it is necessary to switch from the current interaction subject's identity to another interaction subject's identity. In this case, the context transfer module performs inheritance processing.

[0133] When a switch of interaction subject identity is required, the context transfer module identifies incomplete interaction content from the user's text interaction information. Incomplete interaction content can include unanswered user questions, tasks promised by the previous interaction subject but not yet completed, requests recently made by the user but not responded to, unresolved topics from the previous round of dialogue, and emotions recently expressed by the user but not yet responded to. Subsequently, the system encodes the incomplete interaction content, the current interaction subject identity, the target interaction subject identity, and the reason for the switch into inherited task features. The reason for the switch can be a manual switch command, user-indicated result, shift rotation, change in performance status, historical interaction preference trigger, or team interaction mode trigger. The inherited task features explain which interaction subject identity the switch is transitioning from, why the switch is happening, and which incomplete content needs to be taken over.

[0134] Furthermore, the system matches the inherited task features with the historical interaction information corresponding to the target interaction subject identity to generate inherited context features. The historical interaction information corresponding to the target interaction subject identity can include the user's historical dialogues with the target interaction subject identity, user feature tags, relationship stage, interaction preferences, attention items, and historical resource interaction behaviors. By matching the inherited task features with the historical interaction information corresponding to the target interaction subject identity, the system enables the target interaction subject identity to consider its existing relationship with the user when addressing previous topics, rather than directly using the relationship between the current interaction subject identity and the user. For example, a user may have a close relationship with interaction subject A but less historical interaction with interaction subject B. When interaction subject B needs to address a question left unresolved by interaction subject A, the inherited context features generated by the system will retain the semantic content that the user is asking about size selection, but will not directly transfer the close expression style of interaction subject A to interaction subject B. Subsequent candidate responses are still style-constrained by the expression feature model of interaction subject B, thus achieving content inheritance and style isolation.

[0135] The multi-source context awareness module performs attention fusion on collaborative interaction features and inherited context features to obtain fused context features. To improve the accuracy of response generation, the system further introduces current conversation layer context, relational data layer context, and group context layer context. The current conversation layer context can include recent dialogue history, current messages, current topic, dialogue turn, incomplete interaction content, and current interaction subject status. The relational data layer context can include user characteristic information under the target interaction subject identity, such as relationship stage, interaction frequency, resource interaction level, content preferences, attention items, common titles, and historical interaction style. The same user can have different user characteristics under different interaction subject identities. After determining the target interaction subject identity, the system prioritizes calling the user characteristic information in the data space corresponding to the target interaction subject identity, rather than directly calling the user characteristic information under other interaction subject identities. The group context layer context can include live stream message density, interaction popularity, resource interaction status, group selection tendency, current performance stage, and trending topics. For example, when the message density in the live stream is high, the system can increase the weight of the current session layer context and the group context layer context, making candidate replies shorter and more suitable for rapid interaction; when the interaction density in the live stream is low and users are conducting in-depth consultations, the system can increase the weight of the relationship data layer context and the historical interaction context, making candidate replies more suitable for relationship maintenance and detailed answers. The resulting fused contextual features include not only the semantics of the current user interaction, but also the inherited information generated by the switching of the interaction subject's identity, as well as the interaction relationships under the target interaction subject's identity and the atmosphere of the live stream group.

[0136] The candidate response generation engine inputs a pre-defined response generation neural network with an expression feature model that integrates contextual features and the target interaction subject's identity. The contextual features constrain the content of the candidate responses, while the expression feature model constrains their style. The expression feature model can include the target interaction subject's sentence structure preferences, emotional intensity, restricted expression, verbal expression, emoji usage, and timely word usage. The pre-defined response generation neural network generates multiple candidate responses based on this input, such as a primary candidate response, style variant candidate responses, and short candidate responses. The primary candidate response provides the response that best matches the target interaction subject's identity, the current context, and the interaction relationship. Style variant candidate responses provide responses under different interaction strategies, such as friendly, humorous, reassuring, or guiding versions. Short candidate responses provide short text responses when the interaction pace is fast, the interaction subject is under high load, or the message density is high, allowing the interaction subject to quickly select the appropriate response.

[0137] After generating candidate responses, the system can further evaluate their effectiveness and perform rule checks. Effectiveness evaluation generates a predicted interaction effect score based on factors such as whether the response accurately addresses the user's question, aligns with the target interaction subject's identity, fits the interaction relationship, and facilitates continued interaction. Rule checks determine whether candidate responses violate preset rules based on restricted expressions in the target interaction subject's expression feature model, platform rule terms, brand rule terms, and the interaction subject's personal restrictions. If a candidate response contains an aberration, the system can regenerate an alternative candidate response; if consecutively generated candidate responses fail to meet preset rules, safe candidate content prompting the interaction subject for manual handling is generated. Through this process, candidate responses are no longer generated solely by the large language model based on text context, but rather under the combined constraints of the target interaction subject's identity, expression feature model, inherited context features, user characteristics, and expression restriction boundaries.

[0138] The system displays candidate replies on the subject confirmation interface. This interface shows the candidate reply content, target subject identifier, user characteristic summary, inherited context hints, expected interaction effect score, and rule hint tags. The interacting subject can select, edit, or ignore candidate replies. If the subject selects a candidate reply, the system sends it after confirmation. If the subject edits a candidate reply, the system uses the edited text as the reply to be sent and records the differences between the edited and unedited text. If the subject ignores all candidate replies, the system does not send a reply and marks the relevant candidate replies as negative feedback samples. The system only sends messages after receiving confirmation from the subject. The recipient can be a live stream message, a targeted messaging app, the subject's dedicated account, or other preset interaction entry points. This approach makes candidate reply generation an intelligent aid, rather than a complete automatic replacement of the subject's speech, which helps maintain a sense of real-person interaction and reduces the probability of inappropriate expression caused by automatic replies.

[0139] After a candidate reply is selected, edited, or ignored, the system continuously optimizes the expression feature model, user features, and routing strategy based on the feedback from the interacting subject. When an interacting subject selects a candidate reply, the system can increase the weight of the corresponding stable expression feature in that candidate reply and record the candidate reply, the target interacting subject's identity, the fused context features, and the version of the expression feature model as positive samples. When an interacting subject edits a candidate reply, the system analyzes the text differences before and after editing, uses the expressions retained, added, or replaced by the interacting subject as the basis for style correction, and adjusts the weight of the corresponding stable expression feature in the expression feature model. When an interacting subject ignores a candidate reply, the system can reduce the generation weight of that candidate reply in similar contexts or use it as a negative sample for subsequent model training. Simultaneously, the system can update user features based on the interaction results between the user and the target interacting subject's identity. For example, if a user continues to interact, like, or engages in resource interaction after interacting subject B sends a candidate reply, the system can increase the user's interaction preference or relationship stage in interacting subject B's user features; if the user stops interacting or shows negativity towards the reply, the system can adjust the corresponding user feature labels. The system can also optimize routing strategies based on the difference between the output result of the target interaction subject's identity and the actual result received by the interaction subject, so that subsequent identity matching is more in line with the actual operation of collaborative live streaming.

[0140] Through this embodiment, the system can complete a full processing flow from input data collection, collaborative live streaming interaction representation generation, routing and sorting, identity matching, context inheritance, context fusion, candidate response generation, and confirmation of the interaction subject to model feedback optimization. This process enables collaborative interaction accounts to accurately determine the identity of the target interaction subject in complex scenarios involving multiple interaction subject rotations, user targeting, manual scheduling, and team user interaction, maintain contextual continuity when switching interaction subjects, and generate candidate responses that conform to the expression style of the target interaction subject.

[0141] In summary, this application achieves intelligent assisted interaction in collaborative live streaming scenarios by coordinating features of interactive subjects, semantic perception of collaborative live streaming interactions, dynamic switching of multiple interactive subject identities, context inheritance and transfer, multi-source context fusion, and candidate response generation. This enables the system to move beyond simply generating generic responses based on a single user's text. Instead, it allows for comprehensive reasoning based on the differentiated representations of multiple interactive subject identities, real-time changes in the collaborative live streaming status, and the contextual relationships during subject switching. Specifically, the subject feature modeling module performs representation learning based on historical text interaction samples from multiple interactive subjects. Furthermore, it divides samples from the same interactive subject into active response samples, user maintenance samples, and restriction processing samples. This allows the system to learn stable expressive features of the interactive subject in ordinary interaction, relationship maintenance, and restriction processing scenarios, respectively. This ensures that the expressive feature model of each interactive subject reflects its authentic and relatively stable expressive habits, interaction boundaries, and restriction processing methods. Simultaneously, the system adjusts the weights of stable expressive features based on the interactive subject's selection, editing, and ignoring of candidate responses. This ensures that the expressive feature model is continuously updated based on the actual usage behavior of the interactive subject, preventing the expressive feature model from becoming static or out of touch with the current style of the interactive subject.

[0142] Furthermore, this application uses a collaborative semantic perception module to jointly semantically encode user text interaction information and collaborative operation information. This encodes user interaction semantics, interaction subject pointing relationships, manual switching instructions, user pointing results, user's historical interaction preferences for interaction subjects, and the current interaction subject identity into collaborative interaction features. This enables the subsequent identity matching process to simultaneously understand what the user said, who the user might be pointing to, and the current collaborative live streaming operation status. Based on these collaborative interaction features, the multi-subject switching module inputs them, along with the expression feature models of each interaction subject, into a preset identity matching neural network. It then combines the manual switching instructions, user pointing results, historical interaction preferences, the current interaction subject identity, and the preset routing priority between team interaction modes to output the target interaction subject identity. In principle, this process incorporates explicit scheduling relationships, user pointing relationships, historical interaction relationships, and interaction subject expression feature matching relationships in the collaborative live streaming scenario into the neural network inference, rather than simply relying on the current real-time message text or a fixed interaction subject status. Therefore, it can more reasonably determine which interaction subject identity should take over the current interaction when multiple pointing results conflict, reducing the problems of cross-platform responses from multiple interaction subjects under a collaborative interaction account and errors in target interaction subject identification.

[0143] Furthermore, when the identity of the target interaction subject is inconsistent with the identity of the current interaction subject, this application identifies incomplete interaction content from the user's text interaction information through a context transfer module. The incomplete interaction content, the current interaction subject identity, the target interaction subject identity, and the reason for switching are encoded as inherited task features. These inherited task features are then matched with the historical interaction information corresponding to the target interaction subject identity to generate inherited context features. The principle of this mechanism is to transfer the semantics of the interaction that has been formed but not yet completed by the previous interaction subject to the target interaction subject identity in the form of inherited context. At the same time, it uses the target interaction subject identity's own historical interaction information and expression feature model to constrain subsequent responses, enabling the target interaction subject to understand the preceding dialogue without directly copying the expression style of the previous interaction subject. Therefore, in scenarios such as interaction subject rotation, manual scheduling, user-directed input, or manual switching, the system can maintain both the continuity of dialogue content and the consistency of the target interaction subject identity, reducing dialogue breaks, relationship misjudgments, and style mismatches caused by interaction subject switching.

[0144] Meanwhile, this application uses a multi-source context awareness module to perform attention fusion on collaborative interaction features and inherited context features to obtain fused context features. This allows the system to dynamically balance the importance of the user's current session, user history relationships, collaborative live streaming status, and inherited context based on the current interaction scenario. When the interaction density in the live stream is high or the current session changes rapidly, the system can focus more on the current session and the real-time status of the collaborative live stream. In scenarios involving maintaining interaction relationships or inheriting from the end of the live stream, the system can focus more on the historical interaction information and user feature information corresponding to the target interaction subject's identity. Subsequently, the candidate reply generation engine inputs the expression feature model of the fused context features and the target interaction subject's identity into a preset reply generation neural network to generate candidate replies. This ensures that the candidate replies can both inherit the current user interaction and switch inherited content, and conform to the expression style and expression constraints of the target interaction subject's identity. In summary, this application can improve the accuracy of determining the target interaction subject's identity in collaborative live streaming scenarios, the continuity of context when switching interaction subjects, the matching of candidate replies with the target interaction subject's expression features, and the controllability of multi-interaction subject collaborative interaction, thereby enhancing the naturalness, stability, and practical value of intelligent assisted interaction under collaborative interaction accounts.

[0145] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A collaborative live streaming multi-interactive subject dynamic switching system, characterized in that, include: The subject feature modeling module is used to perform representation learning based on historical text interaction samples of multiple interactive subjects to obtain the expression feature model of each interactive subject; The collaborative semantic awareness module is used to semantically encode user text interaction information and collaborative operation information to obtain collaborative interaction features. The multi-subject switching module is used to input the collaborative interaction features and each of the expression feature models into a preset identity matching neural network and output the identity of the target interaction subject; The context transfer module is used to generate inherited context features based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity when it is determined that a switch from the current interaction subject identity to another interaction subject identity is required based on the target interaction subject identity. A multi-source context awareness module is used to perform attention fusion on the collaborative interaction features and the inherited context features to obtain fused context features; A candidate response generation engine is used to input the fused context features and the expression feature model of the target interaction subject's identity into a preset response generation neural network to generate candidate responses.

2. The collaborative live streaming multi-interactive subject dynamic switching system according to claim 1, characterized in that, The subject feature modeling module is used to divide the historical text interaction samples of the same interactive subject into active response samples, user maintenance samples and restricted processing samples, and learn the expression stable features in the active response samples, user maintenance samples and restricted processing samples respectively, and generate the expression feature model of the interactive subject using the expression stable features that meet the stability conditions among the expression stable features.

3. The collaborative live streaming multi-interactive subject dynamic switching system according to claim 2, characterized in that, The subject feature modeling module is also used to obtain the selection, editing, and ignoring results of the interactive subject on the candidate responses, and to correct the weights of the corresponding stable expression features in the expression feature model based on the selection, editing, and ignoring results.

4. The collaborative live streaming multi-interactive subject dynamic switching system according to claim 1, characterized in that, The multi-subject switching module is used to determine the identity of the target interactive subject according to a preset routing priority. The preset routing priority includes, in order, manual switching instructions, user pointing results, user's historical interaction preferences for interactive subjects, current interactive subject identity, and team interaction mode.

5. The collaborative live streaming multi-interactive subject dynamic switching system according to claim 4, characterized in that, The multi-subject switching module is used to determine the priority pointing result according to the preset routing priority when the manual switching instruction, the user pointing result, the user's historical interaction preference for the interactive subject, and the current interactive subject identity point to different interactive subject identities, and inputs the priority pointing result and the collaborative interaction feature into the preset identity matching neural network to output the target interactive subject identity.

6. A method for dynamic switching of multiple interactive subjects in collaborative live streaming, characterized in that, include: Representation learning is performed based on historical text interaction samples from multiple interactive subjects to obtain the expression feature model of each interactive subject; Semantic encoding is performed on user text interaction information and collaborative operation information to obtain collaborative interaction features; The collaborative interaction features and each of the expression feature models are input into a preset identity matching neural network, which outputs the identity of the target interaction subject. When it is determined that a switch from the current interaction subject identity to another interaction subject identity is needed based on the target interaction subject identity, an inherited context feature is generated based on the user text interaction information, the historical interaction information corresponding to the target interaction subject identity, and the historical interaction information corresponding to the current interaction subject identity. Attention fusion is performed on the collaborative interaction features and the inherited context features to obtain fused context features; The fused context features and the expression feature model of the target interactive subject's identity are input into a preset response generation neural network to generate candidate responses.

7. The collaborative live streaming multi-interactive subject dynamic switching method according to claim 6, characterized in that, The representation learning based on historical text interaction samples from multiple interacting entities yields an expressive feature model for each interacting entity, including: The historical text interaction samples of each interactive subject are divided into active response samples, user maintenance samples, and restricted processing samples according to the interaction purpose; The expression-stable features in the active response samples, the user-maintained samples, and the restricted processing samples are learned respectively; An expression feature model of the interactive subject is generated using expression features that satisfy the stability condition among the expression stable features.

8. The collaborative live streaming multi-interactive subject dynamic switching method according to claim 7, characterized in that, After generating candidate responses, the following is also included: Obtain the interaction subject's selection, editing, or ignoring results for the candidate responses; Based on the selection, editing, or ignoring results, the weights of stable features in the expression feature model are adjusted.

9. A method for dynamic switching of multiple interactive subjects in collaborative live streaming according to claim 6, characterized in that, The semantic encoding of user text interaction information and collaborative operation information to obtain collaborative interaction features includes: Extract user interaction semantics and interaction subject orientation relationships from the user text interaction information; Extract manual switching instructions, user pointing results, user's historical interaction preferences for the interaction subject, and the current identity of the interaction subject from the collaborative operation information; The user interaction semantics, interaction subject pointing relationship, manual switching command, user pointing result, historical interaction preference and current interaction subject identity are encoded as the collaborative interaction features.

10. The collaborative live streaming multi-interactive subject dynamic switching method according to claim 9, characterized in that, The step of inputting the collaborative interaction features and each of the expression feature models into a preset identity matching neural network and outputting the identity of the target interaction subject includes: The manual switching command, the user's pointing result, the user's historical interaction preferences for the interactive subject, the current interactive subject's identity, and the team interaction mode are sorted according to a preset routing priority. The sorting results and the collaborative interaction features are input into the preset identity matching neural network; The identity of the target interactive subject is output by the preset identity matching neural network.

Images