An intelligent reply system suitable for live matrix scenarios

Abstract

The application provides a kind of intelligent reply system suitable for live matrix scene, including receiving module, algorithm processing module, reply generation module, distribution module and feedback closed loop module.Receiving module parallelly collects live room user input data and dynamic scene parameters, including audience activity, content sensitivity, real-time response level.Algorithm processing module dynamically integrates real-time sentiment analysis algorithm, semantic correlation algorithm and scene adaptation algorithm through weight distributor, to generate real-time weight.Reply generation module generates adaptive reply content according to output, distribution module configures resource allocation strategy, and feedback closed loop module monitors user interaction indicators in real time and adjusts weight distributor parameters.The application can generate reply content highly adapted to live scene in real time, effectively improving user interaction experience and live efficiency.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to an intelligent reply system suitable for live broadcast matrix scenarios. Background Art

[0002] In today's live streaming industry, with the booming development of live streaming platforms and the continuous growth in the number of live streaming rooms and user base, interaction between broadcasters and viewers has become a key factor in enhancing the live streaming experience and user retention. However, traditional live streaming interaction methods rely primarily on real-time responses from broadcasters, which has many limitations when dealing with large audiences and complex scenarios. For example, broadcasters may find it difficult to respond to every viewer's comments or gifts in a timely manner, resulting in some viewers' interaction needs not being met. Furthermore, the diversity and complexity of live streaming content makes a single response strategy difficult to adapt to different live streaming scenarios, affecting the effectiveness of interaction.

[0003] While existing livestream interactive technologies have achieved some degree of automated response automation, most rely on simple keyword matching or pre-set templates, lacking a deep understanding of livestream scenarios and the ability to dynamically adapt. These technologies are unable to flexibly adjust response strategies based on factors such as viewer activity, content sensitivity, and the need for real-time responsiveness. This can lead to responses being out of context and even triggering misunderstandings or negative feedback.

[0004] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing live broadcast interactive technology cannot effectively cope with the complex interactive needs in the scenario of multiple live broadcast rooms in parallel, lacks real-time perception and intelligent adaptation capabilities of dynamic changes in live broadcast scenes, and it is difficult to generate reply content that is highly consistent with the live broadcast scene and personalized, thereby limiting the quality of live broadcast interaction and the improvement of user experience. Summary of the Invention

[0005] The present invention provides an intelligent reply system suitable for live broadcast matrix scenarios, comprising:

[0006] A receiving module is used to collect user input data and dynamic scene parameters from multiple live broadcast rooms in parallel. The dynamic scene parameters include audience activity, content sensitivity, and real-time response level;

[0007] an algorithm processing module, connected to the receiving module, comprising a multi-algorithm fusion model, wherein the multi-algorithm fusion model dynamically integrates a real-time sentiment analysis algorithm, a semantic association algorithm, and a scene adaptation algorithm through a weight allocator, wherein the weight allocator generates a real-time weight for each algorithm based on the audience activity, content sensitivity, and real-time response level;

[0008] A reply generation module, connected to the algorithm processing module, for generating scenario-adaptive reply content based on the output of the multi-algorithm fusion model;

[0009] a distribution module, connected to the reply generation module and configured with a resource allocation strategy;

[0010] Feedback closed-loop module monitors user interaction indicators in real time and adjusts weight allocator parameters.

[0011] As a further improvement of the present application, the receiving module includes:

[0012] Data collection unit, used to obtain barrage text, gift trigger events and emoticon streams;

[0013] The scene analysis unit is used to extract the live broadcast room identifier, the anchor's behavior feature vector, and the real-time number of concurrent users;

[0014] The normalization processing unit performs the following operations:

[0015] (1) Generate word vector sequences using a bidirectional long short-term memory model;

[0016] (2) Calculate the time decay weight of the gift trigger event based on the decay factor and the time difference between the events;

[0017] (3) Output multimodal feature matrix.

[0018] As a further improvement of the present application, the weight allocator performs:

[0019] (1) Constructing a scenario parameter vector that includes audience activity, content sensitivity, and real-time response level;

[0020] (2) Calculate the intermediate variables by linearly combining the learnable parameter matrix and the scene parameter vector plus the bias term;

[0021] (3) Perform exponential normalization on the intermediate variables to generate real-time weights.

[0022] As a further improvement of this application, the real-time sentiment analysis algorithm includes:

[0023] Construct a sentiment dictionary, associate terms with extreme sentiment values ​​and confidence levels;

[0024] Calculate the weighted sentiment score based on word position weight, sentiment extreme value and confidence;

[0025] When the rate of change of sentiment scores in adjacent time exceeds the adaptive threshold, an alert is triggered.

[0026] As a further improvement of the present application, the semantic association algorithm performs:

[0027] (1) Constructing a semantic graph that represents the strength of concept associations;

[0028] (2) Extract entity sets from input text;

[0029] (3) Calculate the semantic relevance score based on the path length and path edge weight between entities.

[0030] As a further improvement of the present application, the scene adaptation algorithm performs:

[0031] Extract the live broadcast room style feature vector;

[0032] The style feature vector is concatenated with the input feature vector and then subjected to linear transformation and Sigmoid function to calculate the scene matching degree;

[0033] When the scene matching degree is lower than the matching threshold, the cross-live room reply strategy is triggered.

[0034] As a further improvement of this application, the reply generation module performs dynamic template splicing:

[0035] (1) Select a basic template;

[0036] (2) Calculate and inject dynamic variables based on the conversation context vector;

[0037] (3) Append template fragments when the user profile has tags.

[0038] As a further improvement of this application, the distribution module calculates the priority of the live broadcast room:

[0039] Priority is calculated based on real-time weight, audience value, content relevance, and the number of queued tasks combined with a logarithmic function;

[0040] Resources are allocated based on priority and resource tilt factors through exponential normalization.

[0041] As a further improvement of the present application, the feedback closed-loop module performs:

[0042] (1) Calculate the comprehensive effect index based on the weight coefficient, click-through rate, normalized value of dwell time and conversion rate;

[0043] (2) Adjust the trainable parameters through the policy gradient update formula based on the comprehensive effect indicators.

[0044] As a further improvement of the present application, the initialization rule of the learnable parameter matrix is:

[0045] The diagonal elements are initialized to fixed positive values;

[0046] The off-diagonal elements are initialized to fixed negative values;

[0047] The orthogonalization constraint is used to ensure that the matrix condition number is less than a fixed threshold.

[0048] The above-described embodiments of the present invention have at least the following beneficial effects: the intelligent reply system can achieve parallel collection and real-time processing of user input data and dynamic scene parameters from multiple live broadcast rooms. Through a multi-algorithm fusion model, it dynamically integrates sentiment analysis, semantic association, and scene adaptation algorithms to generate adaptive reply content based on different live broadcast scenarios, effectively improving the accuracy and relevance of replies. Furthermore, the system can dynamically adjust weight allocator parameters based on real-time monitored user interaction indicators, optimize reply strategies, and further enhance interactive effects.

[0049] Furthermore, the system can rationally allocate interactive resources within livestreaming rooms through dynamic resource allocation strategies, prioritizing high-value and highly relevant livestreaming interactions and improving resource utilization efficiency. The introduction of a closed-loop feedback module enables the system to evaluate interactive effects in real time and perform self-optimization, thereby continuously improving system performance and user experience, providing a more efficient and intelligent solution for interactive models in the livestreaming industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0051] Figure 1 A schematic diagram of the structure of an intelligent reply system suitable for a live broadcast matrix scenario provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0052] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0053] Those skilled in the art will appreciate that embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0054] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0055] Reference below Figure 1 , Figure 1 This is a schematic diagram of the structure of an intelligent reply system for a live matrix scenario provided by an embodiment of the present invention. Figure 1 As shown, an intelligent reply system 100 suitable for a live matrix scenario includes:

[0056] Receiving module 101 is used to collect user input data and dynamic scene parameters of multiple live broadcast rooms in parallel, and the dynamic scene parameters include audience activity , content sensitivity and real-time response levels ;

[0057] The algorithm processing module 102 is connected to the receiving module and includes a multi-algorithm fusion model. The multi-algorithm fusion model dynamically integrates the real-time sentiment analysis algorithm, the semantic association algorithm and the scene adaptation algorithm through a weight distributor, wherein the weight distributor is based on Generate real-time weights for each algorithm ;

[0058] A reply generation module 103, connected to the algorithm processing module, is used to generate scenario-adaptive reply content based on the output of the multi-algorithm fusion model;

[0059] a distribution module 104, connected to the response generation module and configured with a resource allocation strategy;

[0060] The feedback closed-loop module 105 monitors user interaction indicators in real time and adjusts the weight allocator parameters.

[0061] It should be noted that the intelligent reply system of the present invention collects user input data and dynamic scene parameters of multiple live broadcast rooms in parallel through a receiving module. Among them, user input data includes barrage text, gift trigger events and emoticon streams, etc., which reflect the audience's real-time feedback and willingness to interact. Dynamic scene parameters are used to describe the overall state of the live broadcast room, including audience activity (a quantitative indicator of audience participation, ranging from 0 to 1), content sensitivity (an indicator that measures whether the live broadcast content involves sensitive topics, ranging from 0 to 1) and real-time response level (a level divided according to the urgency of the live broadcast interaction, ranging from 1 to 5). The introduction of these parameters enables the system to generate more accurate and adaptive reply content according to the characteristics and needs of different live broadcast rooms.

[0062] The receiving module is responsible for concurrently collecting user input data and dynamic scene parameters from multiple livestreaming rooms. This includes user input data such as bullet comment text, gift trigger events, and emoticon streams, as well as dynamic scene parameters such as viewer activity (calculated based on viewer interaction frequency and bullet comment sending rate, ranging from 0-1), content sensitivity (using keyword matching and semantic analysis to detect whether the livestream content involves sensitive topics, ranging from 0-1), and real-time response level (based on the urgency of the interaction, ranging from 1-5). The algorithm processing module includes a multi-algorithm fusion model that dynamically integrates a real-time sentiment analysis algorithm, a semantic association algorithm, and a scene adaptation algorithm through a weight allocator. The weight allocator generates real-time weights for each algorithm based on viewer activity, content sensitivity, and real-time response level. For example, when viewer activity, content sensitivity, and real-time response level are high, the real-time sentiment analysis algorithm is assigned a higher weight, enabling timely and accurate analysis of viewer sentiment and responses. The reply generation module generates scene-adaptive replies based on the output of the multi-algorithm fusion model. For example, it generates reply text that matches the style and content of the livestream room based on sentiment analysis and semantic association analysis. The distribution module is configured with a resource allocation strategy to push generated replies to the appropriate livestream rooms and users as needed. The feedback loop module monitors user interaction metrics in real time, such as click-through rate and dwell time after replies, and adjusts the weight allocator parameters based on these metrics to optimize reply strategies and enhance engagement.

[0063] Specifically, the data acquisition unit in the receiving module is responsible for acquiring multimodal data such as barrage text, gift triggering events, and emoji streams. The scene analysis unit is responsible for extracting live broadcast room identifiers (used to distinguish different live broadcast rooms), host behavior feature vectors (reflecting the host's behavioral patterns and style), and the number of real-time concurrent users (the number of users currently watching the live broadcast). The normalization processing unit preprocesses the collected data, for example, using a bidirectional LSTM model to generate word vector sequences to better capture text semantic information. For gift triggering events, the system uses a time decay weighted formula, where the decay factor and the time difference between events are used to measure the importance and timeliness of the gift event. Ultimately, the output multimodal feature matrix integrates this data into a unified format, facilitating further analysis by the subsequent algorithm processing module.

[0064] Preferably, the weight allocator plays a key role in the algorithm processing module. It first constructs a scene parameter vector, integrating audience activity, content sensitivity, and real-time response level into a three-dimensional vector. Then, intermediate variables are calculated using a learnable parameter matrix and bias terms. These parameters can be optimized using training data to adapt to different live broadcast scenarios. Finally, normalized weights are generated to ensure that the sum of the weights of each algorithm is 1, thereby achieving dynamic integration of real-time sentiment analysis algorithms, semantic association algorithms, and scene adaptation algorithms. In practical applications, the initialization rules of the parameter matrix can be adjusted according to different live broadcast types (such as entertainment live broadcasts, educational live broadcasts, etc.). For example, the weight of content sensitivity can be increased in educational live broadcasts to better meet the needs of educational scenarios.

[0065] In some embodiments, the receiving module includes:

[0066] Data collection unit, used to obtain barrage text, gift trigger events and emoticon streams;

[0067] The scene analysis unit is used to extract the live broadcast room identifier, the anchor's behavior feature vector, and the real-time number of concurrent users;

[0068] The normalization processing unit performs the following operations:

[0069] (1) Use a bidirectional LSTM model to generate word vector sequences;

[0070] (2) The time decay weighted formula for gift triggering events is:

[0071] ;

[0072] in is the attenuation factor, The time difference between events;

[0073] (3) Output multimodal feature matrix .

[0074] It's important to note that the receiving module is a key component of the intelligent reply system. Its primary function is to obtain user input data and dynamic scene parameters from the live broadcast room. User input data includes bullet comment text, gift trigger events, and emoticon streams, which directly reflect the viewer's interaction with the live broadcast room. Dynamic scene parameters describe the overall state of the live broadcast room and include the live broadcast room identifier, the host's behavioral feature vector, and the number of concurrent users in real time. These parameters help the system better understand the real-time status of the live broadcast room, providing a foundation for subsequent intelligent processing.

[0075] In the smart reply system, the receiving module includes a data acquisition unit, a scene analysis unit, and a normalization processing unit. The data acquisition unit captures the barrage text, gift trigger events, and emoticon streams. It reads barrage messages in real time through an interface, records the time, type, and number of gifts given, and collects the type and frequency of emoticon usage. The scene analysis unit extracts the live broadcast room identifier (e.g., a unique code for each live broadcast room), the host's behavioral feature vector (a vector generated by analyzing the host's language style, interaction frequency, and movement characteristics, used to describe the host's behavioral patterns), and the real-time number of concurrent users (counting the number of users watching the live broadcast simultaneously). The normalization processing unit uses a bidirectional long short-term memory (Bi-LSTM) model to generate a sequence of word vectors to encode the barrage text and capture its semantic information. It calculates the time-decay weight of the gift trigger event based on a decay factor (e.g., 0.9, indicating that the importance of the event decreases over time) and the time difference between the events. The initial weight is set based on the gift value. Finally, it outputs a multimodal feature matrix, integrating the word vectors of the barrage text, the weight vectors of the gift event, and the emoticon vectors into a single matrix, which serves as input for subsequent processing.

[0076] Specifically, the data acquisition unit in the receiving module is responsible for obtaining the live broadcast room's barrage text, gift trigger events, and emoticon streams in real time. Barrage text refers to text messages sent by viewers in the live broadcast room. The emotional tendencies and semantic information therein can be extracted through natural language processing technology. Gift trigger events refer to the behavior of viewers giving virtual gifts. By recording the time and type of events, the system can analyze the audience's support level. The emoticon stream is the emoticons sent by the audience, reflecting the audience's emotional state. The scene analysis unit extracts the live broadcast room identifier to distinguish different live broadcast rooms; the host behavior feature vector is generated by analyzing the host's language and behavior patterns, which can reflect the host's style and interaction method; the real-time concurrent user count, that is, the number of users currently watching the live broadcast at the same time, is used to measure the popularity of the live broadcast room. The normalization processing unit preprocesses the collected data, such as using a bidirectional LSTM model to generate a word vector sequence to better capture text semantic information. The time decay weighted formula for gift trigger events is:

[0077] ;

[0078] in is the initial weight, is the attenuation factor, is the time difference between events. This formula is used to measure the importance and timeliness of gift events. Over time, the weight of an event gradually decreases. Ultimately, the output multimodal feature matrix integrates this data into a unified format, facilitating further analysis in subsequent algorithm processing modules.

[0079] Preferably, the normalization processing unit can adjust the attenuation factor according to different live broadcast scenarios when processing the gift trigger event. For example, in a live broadcast room with high activity, gift events are more time-sensitive, and the decay factor can be set to a smaller value to more quickly reflect the audience's emotional changes. For the processing of emoji streams, in addition to counting the number, sentiment analysis algorithms can also be introduced to classify the emotional tendencies of emojis, such as positive, negative, or neutral. In addition, for the output of the multimodal feature matrix, an attention mechanism can be further introduced to dynamically adjust the weights according to the importance of different features, thereby improving the adaptability and flexibility of the system.

[0080] In some embodiments, the weight allocator performs:

[0081] (1) Constructing scene parameter vector ;

[0082] (2) Calculate intermediate variables:

[0083] ;

[0084] in is the i-th row and j-th column element of the learnable parameter matrix, is the bias term, is the parameter vector The jth component of ;

[0085] (3) Generate normalized weights:

[0086] .

[0087] It's important to note that the weight allocator is the core component of the system, used to dynamically adjust the weights of various algorithms. It constructs a scenario parameter vector and calculates intermediate variables, ultimately generating normalized weights. This allows for the dynamic integration of real-time sentiment analysis, semantic association, and scenario adaptation algorithms. This process flexibly adjusts the weights of various algorithms based on dynamic livestream parameters (such as audience activity, content sensitivity, and real-time response levels), ensuring that the system's output responses are more tailored to the needs of the current livestream scenario.

[0088] In the algorithm processing module of the smart reply system, a weight allocator constructs a scenario parameter vector containing audience activity, content sensitivity, and real-time response level. For example, if audience activity is 0.8, content sensitivity is 0.6, and real-time response level is 4, the vector is [0.8, 0.6, 4]. Intermediate variables are then calculated using a linear combination of a learnable parameter matrix (initial values ​​can be set empirically, such as 0.5 for diagonal elements and -0.1 for off-diagonal elements) and the scenario parameter vector, plus a bias term (e.g., [0.1, 0.1, 0.1]). These intermediate variables are then exponentially normalized to generate real-time weights. This involves applying an exponential function to each component of the intermediate variable and dividing it by the sum of the exponential function values ​​of all components to obtain the weights for each algorithm, ensuring that the sum of the weights is 1. For example, the real-time sentiment analysis algorithm has a weight of 0.4, the semantic association algorithm has a weight of 0.3, and the scenario adaptation algorithm has a weight of 0.3. These weights reflect the importance and contribution of each algorithm in the current scenario and are used to subsequently integrate the results of each algorithm to generate appropriate reply content.

[0089] Specifically, the weight allocator first constructs a scene parameter vector ,in Indicates audience activity. Indicates the sensitivity of the content. Indicates the real-time response level. The value ranges of these parameters are and , used to quantify the real-time status of the live broadcast room.

[0090] The weight distributor is a learnable parameter matrix and bias To calculate the intermediate variables , the specific formula is:

[0091] ;

[0092] in, is a The matrix is ​​used to adjust the impact of each scene parameter on the weight distribution; is a bias vector used to adjust the baseline of weight distribution.

[0093] Next, the final weights are generated through normalization operation , ensuring that the sum of the weights of each algorithm is 1. This process enables the system to dynamically adjust the weights of each algorithm based on the dynamic scene parameters of the current live broadcast room, thereby generating more accurate and adaptive reply content.

[0094] In some embodiments, the real-time sentiment analysis algorithm includes:

[0095] Step 41: Build a sentiment dictionary and associate terms with extreme sentiment values and confidence ;

[0096] Step 42: Calculate the weighted sentiment score:

[0097] ;

[0098] in is the word position weight, is the number of sentiment words;

[0099] Step 43: When the sentiment change rate satisfies:

[0100] ;

[0101] When the warning is triggered, is a minimum constant, is the adaptive threshold, is the emotion score at the current moment, Score the emotion for the previous moment.

[0102] It's important to note that the real-time sentiment analysis algorithm is a crucial component of the intelligent reply system. Its primary function is to analyze the emotional leaning of user input text by constructing a sentiment lexicon and calculating a weighted sentiment score. When the rate of sentiment change exceeds a set threshold, the algorithm triggers an alert, allowing the system to respond promptly to significant sentiment fluctuations. This process helps the system better understand the user's emotional state, enabling it to generate responses that are more responsive to the user's emotions.

[0103] In the intelligent reply system, the real-time sentiment analysis algorithm first constructs a sentiment lexicon. The terms in the lexicon are associated with sentiment extremes (e.g., "good" has a sentiment extreme of 0.8, "bad" has a sentiment extreme of -0.8) and confidence levels (e.g., common sentiment words have a confidence level of 0.9, and rare words have a confidence level of 0.6). A weighted sentiment score is then calculated based on word position weights (e.g., word position weights at the beginning and end of a comment are set to 1.2, and word position weights in the middle are set to 1.0), sentiment extremes, and confidence levels. An alert is triggered when the rate of change in sentiment scores between adjacent times exceeds an adaptive threshold (e.g., set to 0.5). This adaptive threshold is dynamically adjusted based on the live broadcast type and viewer activity. When the rate of change exceeds the threshold, the system issues a warning signal, prompting the host or automatically adjusting the reply strategy to address the rapid shift in viewer sentiment.

[0104] Specifically, the real-time sentiment analysis algorithm first needs to build a sentiment dictionary, in which each word is associated with a sentiment extreme value. and a confidence level The extreme value of sentiment indicates the sentiment tendency of the word, with a positive value indicating positive sentiment and a negative value indicating negative sentiment; the confidence level indicates the credibility of the sentiment tendency.

[0105] Next, the algorithm calculates the weighted sentiment score using the formula:

[0106] ;

[0107] in is the word position weight, which is used to emphasize the importance of words at different positions in the text. is the number of sentiment words.

[0108] In addition, when the emotion change rate meets:

[0109] ;

[0110] When , the warning is triggered. is a very small constant used to avoid the denominator being zero; is an adaptive threshold used to control the sensitivity of emotional changes.

[0111] Preferably, the construction of the sentiment dictionary can be done by collecting a large amount of text data with sentiment annotations and using natural language processing technology to perform part-of-speech tagging and sentiment extreme value calculation. It can be set according to the text structure. For example, for barrage text, words near the beginning and end of the text can be given higher weights because words in these positions can usually better reflect the core emotion of the text.

[0112] Furthermore, the adaptive threshold This threshold can be adjusted based on different live broadcast scenarios. For example, in entertainment live broadcasts, where emotions fluctuate significantly, the threshold can be set relatively high, such as 0.5; whereas in educational live broadcasts, where emotions fluctuate less, the threshold can be set lower, such as 0.3. Furthermore, to further improve the accuracy of sentiment analysis, contextual information can be incorporated to adjust the sentiment score of a word by considering the sentiment of the surrounding context.

[0113] In some embodiments, the semantic association algorithm performs:

[0114] (1) Constructing semantic graph and edge weights Indicates the strength of concept association;

[0115] (2) Extract entity sets from input text ;

[0116] (3) Calculate the semantic relevance score:

[0117] ;

[0118] in is the path length between entities, is the path edge weight, i and j are entity indexes, and m is the size of the set E.

[0119] It's important to note that the semantic association algorithm is a key component of the intelligent reply system, used to analyze the semantic relevance between user input text and the live broadcast context. By constructing a semantic graph and calculating the strength of association between the input text and concepts in the graph, the algorithm quantifies the semantic relevance between the text and the live broadcast context. This process helps the system generate responses that are more relevant to the live broadcast theme and content, improving the accuracy and relevance of interactions.

[0120] The semantic association algorithm first constructs a semantic graph. Nodes in the graph represent concepts (for example, in educational live streams, "mathematics," "function," and "geometry" are nodes), and edge weights indicate the strength of concept associations (for example, the association strength between "function" and "geometry" is 0.6, determined through co-occurrence analysis or expert annotation). A set of entities is extracted from the input text. For example, if the input text is "How to solve this function equation," the extracted entity is "function." A semantic relevance score is calculated based on the path length between entities (for example, the path length from "function" to "equation" is 2) and the edge weights along the path (the sum of the edge weights along the path is 1.2). This score reflects the degree of semantic match between the input text and the live stream scenario. A high score indicates a strong relevance to the current live stream content, allowing responses to be generated that are more relevant to the topic.

[0121] Specifically, the semantic association algorithm first constructs a semantic graph, where the nodes represent concepts and the edge weights Indicates the strength of association between concepts. For the input text, the algorithm extracts the entity set through entity extraction technology. , where each entity It is a key word or phrase in the text.

[0122] Next, the algorithm calculates the semantic relevance score between the input text and the concepts in the semantic graph using the following formula:

[0123] ;

[0124] in is the path length between entities, Is with entity The number of related concepts, is the path edge weight. This score reflects the semantic match between the input text and the live broadcast scene. A higher score indicates a stronger relevance.

[0125] Preferably, semantic graphs can be constructed by analyzing large amounts of livestream-related text data, extracting keywords and concepts, and determining the strength of associations between concepts through co-occurrence analysis or knowledge graph technology. In practical applications, semantic graphs can be customized based on different livestream types (e.g., entertainment livestreams, educational livestreams, etc.). For example, in educational livestreams, the semantic graph might include concepts such as subject knowledge points and teaching terms, while in entertainment livestreams, it might include concepts such as celebrities, programs, and music.

[0126] Furthermore, in order to improve the efficiency of semantic relevance calculation, the graph neural network (GNN) technology can be introduced to further optimize the calculation of semantic relevance by learning the embedding vectors of nodes and edges in the graph. The settings can be adjusted according to actual needs. For example, when it is necessary to emphasize exact matching, a shorter path length can be set; when it is necessary to consider broader semantic associations, the path length can be appropriately increased.

[0127] In some embodiments, the scene adaptation algorithm performs:

[0128] Step 61: Extract the live broadcast room style feature vector ;

[0129] Step 62: Calculate scene matching degree:

[0130] ;

[0131] in is the Sigmoid function, is the weight matrix, is the bias term, is the input feature vector, Represents vector concatenation;

[0132] Step 63: When ( is a matching threshold), the cross-live room reply strategy is triggered.

[0133] It's important to note that the scene adaptation algorithm is a key component of the intelligent reply system, used to evaluate the match between the livestream studio's style and user input. By extracting the livestream studio's style feature vector and calculating the scene matching degree, the algorithm determines whether the current user input is suitable for the studio's style and atmosphere. If the matching degree falls below a set threshold, the system triggers a cross-studio reply strategy, ensuring that the reply aligns with the studio's overall style and enhances the user experience.

[0134] The implementation of the scene adaptation algorithm in the intelligent reply system proceeds as follows: First, a livestream room style feature vector is extracted. This feature vector is generated by performing text analysis on the livestream room's historical comments (e.g., counting word frequency and calculating sentiment) and analyzing the host's language style characteristics (e.g., speaking speed and vocabulary richness). For example, the feature vector has a dimension of 10 and includes metrics such as comment popularity and sentiment polarity. The style feature vector is then concatenated with the input feature vector (composed of the current user's input text and behavioral characteristics). The scene matching degree is calculated through a linear transformation (e.g., matrix multiplication using a weight matrix) and a sigmoid function. When the scene matching degree falls below a matching threshold (e.g., set to 0.5), a cross-livestream reply strategy is triggered. This strategy involves retrieving appropriate replies from other livestream rooms with similar styles or adjusting the current reply generation strategy to ensure that replies match the current livestream room's style. For example, the style features of the current livestream room are compared with those stored in the system to identify similar livestream rooms and then adjust the reply style and content based on their response style.

[0135] Specifically, the scene adaptation algorithm first extracts the live broadcast room style feature vector ,in It is a feature dimension. This vector is generated by analyzing the historical data of the live broadcast room (such as the style of the barrage, the language style of the anchor, etc.), and can reflect the unique style of the live broadcast room.

[0136] Preferably, the live broadcast room style feature vector This can be generated by analyzing multi-dimensional data such as the historical comments in the live broadcast room, the host's language style, and interaction frequency. For example, text mining technology can be used to extract high-frequency words and sentiment in the comments, and then combined with the host's language style characteristics (such as speaking speed and tone) to construct a feature vector.

[0137] For the weight matrix and bias , can be optimized through training data to adapt to different live broadcast scenarios. Matching threshold The threshold can be adjusted based on the type and style of the livestream. For example, in entertainment livestreams, which have a more diverse style, the threshold can be set relatively high, such as 0.6; whereas in educational livestreams, which have a more serious style, the threshold can be set lower, such as 0.4.

[0138] Furthermore, to further improve the accuracy of scene adaptation, contextual information can be introduced to adjust the scene matching calculation by considering the context of the user input. For example, when the user input is highly relevant to the current live content, the matching weight can be appropriately increased.

[0139] In some embodiments, the reply generation module performs dynamic template splicing:

[0140] (1) Select a basic template ;

[0141] (2) Injecting dynamic variables:

[0142] ;

[0143] in is the weight matrix, is the bias term, is the conversation context vector;

[0144] (3) When user portrait Appends template fragment when tag is present.

[0145] It's important to note that the reply generation module is the core component of the intelligent reply system, responsible for generating final reply content. It uses dynamic template splicing, combining basic templates, dynamic variables, and user profile information to generate replies that are highly adapted to the current live broadcast scenario and user needs. This process ensures that replies not only match the style and atmosphere of the live broadcast room but are also customized to the user's individual needs, thereby improving user satisfaction and interactive experience.

[0146] In the reply generation module, the dynamic template assembly process is as follows: First, a base template is selected from a template library based on the type and scenario of the current live broadcast. For example, in a game live broadcast, a base template containing game operation prompts and encouragement for interaction is selected. Then, dynamic variables are calculated and injected based on the conversation context vector (a vector generated from the current conversation history and user profile, containing information such as the conversation topic and sentiment). When the user profile has labels (such as "game enthusiast" or "high-value user"), template fragments are added. For example, a fragment expressing gratitude and exclusive benefits for "high-value users" is added to personalize the reply content to meet the needs and preferences of different users.

[0147] Specifically, the reply generation module first selects a basic template , the template is a pre-designed reply framework that provides the basic structure of the reply. Then, the module injects dynamic variables through the formula Calculating dynamic variables ,in Indicates the weight matrix weight values, is the bias term, is the first An element used to reflect the context and content of the current conversation. When tags are present, the module also appends template snippets to further personalize responses. For example, if the user's profile indicates they are a devoted fan of a particular streamer, the system can include specific fan-specific content in their response, thereby enhancing the user's sense of belonging and willingness to engage.

[0148] Preferably, the base template It can be customized according to different live broadcast scenarios. For example, in educational live broadcast, the basic template can include more knowledge content and guiding questions; while in entertainment live broadcast, the basic template can be more relaxed and humorous, and include more interactive elements. and bias , which can be trained through machine learning algorithms to adapt to different conversation styles and scenario requirements.

[0149] Furthermore, to further personalize responses, we can incorporate more user profile features, such as their interests, hobbies, and viewing history. For example, if a user frequently watches live streams about technology, the system can incorporate more tech elements into their responses. Furthermore, to address complex conversation scenarios, we can design a variety of template segments and dynamically select and combine them based on the real-time conversation content to generate more natural and coherent responses.

[0150] In some embodiments, the distribution module calculates the priority of the live broadcast room:

[0151] ;

[0152] in For audience value, is the content relevance, is the number of queued tasks;

[0153] The dynamic resource allocation formula is:

[0154] ;

[0155] in is the resource tilt factor, The total number of live broadcast rooms.

[0156] It's important to note that the distribution module is a key component of the Smart Reply system for managing and allocating interactive resources. It prioritizes livestream rooms and dynamically allocates resources based on their priority, ensuring that each room receives the interactive support appropriate to its importance and needs. This process optimizes resource utilization, enhances user experience, and ensures fairness and efficiency in livestream interactions.

[0157] The distribution module's implementation in the intelligent reply system is as follows: First, the livestream room's priority is calculated using a logarithmic function based on real-time weight (generated by the algorithm processing module, reflecting the urgency of the livestream room's current interactive needs), viewer value (calculated by analyzing audience tip amounts, viewing time, and interaction frequency, e.g., a weighted sum of 0.6 for tip amounts, 0.3 for viewing time, and 0.1 for interaction frequency), content relevance (using a semantic similarity algorithm to calculate the degree of match between user input and livestream content, with values ​​ranging from 0 to 1), and the number of queued tasks (the number of interactive tasks currently awaiting processing). Resources are then allocated using exponential normalization based on the priority and a resource tilt factor (set based on the livestream room's importance and commercial value, e.g., a resource tilt factor of 1.2 for popular livestream rooms and 1.0 for standard livestream rooms) to ensure rational and efficient resource allocation, prioritizing high-priority livestream room interaction requests.

[0158] As a further improvement of this application, the distribution module first calculates the priority of each live broadcast room, and also includes the formula:

[0159] ;

[0160] in, and is the weight coefficient, The audience value reflects the audience's contribution and importance to the live broadcast room; Content relevance measures the degree of match between the current interactive content and the live broadcast room theme; is the number of queued tasks, representing the number of interactive tasks currently waiting to be processed. This formula allows the system to dynamically adjust the priority of live streaming rooms by comprehensively considering audience value, content relevance, and task backlog.

[0161] Then, the system allocates resources according to the dynamic resource allocation formula:

[0162] ;

[0163] Allocate resources, where is the total number of live broadcast rooms, Indicates the proportion of resources that should be allocated to each live broadcast room. Resource tilt factor It can be adjusted according to actual needs to further optimize the resource allocation strategy.

[0164] Preferably, audience value It can be calculated by analyzing the audience's interaction frequency, reward amount, viewing time and other data. For example, the audience's reward amount can be used as the main indicator, and the viewing time and interaction frequency can be combined for weighted calculation. The semantic similarity between the user input and the live content can be determined by analyzing the similarity, for example, by using methods such as cosine similarity and Jaccard similarity. and , can be adjusted according to different live broadcast scenarios. For example, in educational live broadcasts, the relevance of the content can be increased. To ensure that the interactive content is closely related to the teaching topic; in entertainment live broadcasts, the audience value can be appropriately increased. In order to attract more high-value audiences. In addition, the resource tilt factor It can be dynamically adjusted according to the popularity and importance of the live broadcast room. For example, for popular live broadcast rooms, the resource tilt factor can be appropriately increased to ensure that it obtains more interactive resources.

[0165] In some embodiments, the feedback closed-loop module performs:

[0166] The feedback closed loop module performs:

[0167] (1) Calculate comprehensive effect index:

[0168] ;

[0169] in and For the length of stay, is the preset threshold, CR is the conversion rate, which is the proportion of users who complete the target behavior, such as purchase, follow, etc.

[0170] (2) Policy gradient update formula:

[0171] ;

[0172] in is the policy function, is a trainable parameter.

[0173] It's important to note that the feedback loop module is a crucial component of the intelligent reply system, used to monitor and optimize interaction effectiveness in real time. By calculating comprehensive effectiveness metrics and adjusting system parameters using a policy gradient update formula, this module dynamically optimizes reply strategies based on actual feedback from user interactions, thereby continuously improving system performance and user experience.

[0174] In the feedback loop module, a comprehensive performance metric is first calculated based on weight coefficients (e.g., 0.4 for click-through rate, 0.3 for normalized dwell time, and 0.3 for conversion rate). This is combined with the click-through rate (the percentage of users clicking on a reply, calculated as clicks / impressions), the normalized dwell time (normalized duration of users staying on a page after replying, ranging from 0-1), and the conversion rate (the percentage of users completing target actions such as purchases or follow-ups, calculated as conversions / clicks). This comprehensive performance metric is then used to adjust trainable parameters using a policy gradient update formula. This approach allows the system to dynamically adjust parameters based on actual user interaction, optimize response strategies, and enhance user experience and engagement.

[0175] Furthermore, to further improve the adaptability of the system, an adaptive learning rate can be introduced to dynamically adjust the learning rate according to changes in the comprehensive effect index, thereby accelerating the system optimization speed. For example, when the comprehensive effect index decreases several times in a row, the learning rate can be appropriately increased to speed up the parameter adjustment speed.

[0176] In some embodiments, the parameter matrix initialization rule is:

[0177] The initialization rule of the learnable parameter matrix is:

[0178] The diagonal elements are initialized to fixed positive values;

[0179] The off-diagonal elements are initialized to fixed negative values;

[0180] The orthogonalization constraint is used to ensure that the matrix condition number is less than a fixed threshold.

[0181] It should be noted that the parameter matrix initialization rules and orthogonal constraints are key to ensuring the stability and efficiency of the intelligent reply system. The initialization method directly affects the system's learning efficiency and convergence speed, while orthogonalization constraints are used to ensure the numerical stability of the matrix and avoid numerical problems during the calculation process. Through reasonable initialization and constraints, the system can quickly adapt and operate stably in complex live broadcast scenarios.

[0182] During the learnable parameter matrix initialization process, the diagonal elements are first initialized to fixed positive values, such as 0.5. This enhances the independent influence of each scenario parameter (such as audience activity, content sensitivity, and real-time response level) in the initial stage, ensuring that each parameter plays a fundamental role in weight allocation. The off-diagonal elements are initialized to fixed negative values, such as -0.1, to introduce negative correlations between the parameters, reduce redundancy, and prevent mutual interference between parameters that could lead to irrational weight allocation. An orthogonalization constraint is then implemented to ensure that the matrix condition number is less than a fixed threshold (e.g., a threshold of 5). This orthogonalization constraint uses the QR decomposition method to orthogonalize the matrix, making its column vectors mutually orthogonal. This reduces the matrix condition number and improves its numerical stability, avoiding numerical issues during the calculation process. This ensures stable operation of the system in changing live broadcast scenarios and improves the accuracy and reliability of weight allocation.

[0183] Specifically,

[0184] ;

[0185] And the matrix condition number is ensured by orthogonalization constraints .

[0186] This initialization method aims to give higher initial weights to diagonal elements to enhance the independent influence of each scene parameter, while reducing the redundancy between parameters through negative correlation. Orthogonalization constraints ensure that the matrix condition number The matrix condition number is an important indicator of the matrix's numerical stability. A smaller condition number means a more stable matrix during the calculation process, which can effectively avoid the problem of error amplification in numerical calculations. Through this initialization and constraint method, the system can maintain stable performance in changing live broadcast scenarios.

[0187] Preferably, the initialization method of the parameter matrix can be adjusted according to different live broadcast scenarios. For example, in educational live broadcasts, the initial weights of content sensitivity-related parameters can be appropriately increased to better meet the rigor requirements of educational content. For orthogonalization constraints, stricter condition number restrictions can be adopted, such as , in order to further improve the numerical stability of the system.

[0188] Furthermore, other matrix initialization methods, such as random initialization or small value initialization, can be introduced in combination with orthogonalization to explore initialization strategies that are more suitable for specific scenarios. For example, the matrix can be initialized using the Xavier initialization method and then orthogonalized using QR decomposition or other orthogonalization techniques, thereby improving the adaptability and flexibility of the system while ensuring numerical stability.

[0189] The above-mentioned embodiments of the present invention have the following beneficial effects: The intelligent reply system collects user input data and dynamic scene parameters from multiple live broadcast rooms in parallel through a receiving module, achieving comprehensive perception of complex live broadcast scenes and providing a rich and accurate data foundation for subsequent intelligent processing. The algorithm processing module dynamically integrates multiple algorithms based on a multi-algorithm fusion model and can dynamically adjust the weights of each algorithm according to different live broadcast scene parameters, thereby generating more accurate and adaptive reply content. The reply generation module uses dynamic template splicing technology to flexibly generate personalized reply content based on conversation context and user profiles, further enhancing the user experience. The distribution module optimizes resource utilization efficiency by calculating live broadcast room priorities and performing dynamic resource allocation, ensuring that high-value and high-relevance live broadcast rooms receive more interactive resources. The feedback closed-loop module monitors user interaction indicators in real time and adjusts weight allocator parameters, enabling system self-optimization and continuously improving interactive effects. In addition, the system uses specific rules when initializing the parameter matrix and ensures the matrix condition number through orthogonalization constraints, which can improve system stability and computational efficiency.

[0190] In summary, this system effectively enhances the intelligence of livestreaming interactions, increasing user engagement and satisfaction. It can adapt to various livestreaming scenarios, generating precise and personalized responses while optimizing resource allocation to ensure efficient and fair interactions. Through a closed-loop feedback mechanism, the system continuously learns and optimizes, further improving performance and user experience, providing innovative solutions for interactive models in the livestreaming industry.

[0191] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0192] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An intelligent reply system suitable for live broadcast matrix scenarios, characterized in that: include: A receiving module is used to collect user input data and dynamic scene parameters from multiple live broadcast rooms in parallel. The dynamic scene parameters include audience activity, content sensitivity, and real-time response level; an algorithm processing module, connected to the receiving module, comprising a multi-algorithm fusion model, wherein the multi-algorithm fusion model dynamically integrates a real-time sentiment analysis algorithm, a semantic association algorithm, and a scene adaptation algorithm through a weight allocator, wherein the weight allocator generates a real-time weight for each algorithm based on the audience activity, content sensitivity, and real-time response level; A reply generation module, connected to the algorithm processing module, for generating scenario-adaptive reply content based on the output of the multi-algorithm fusion model; a distribution module, connected to the reply generation module and configured with a resource allocation strategy; Feedback closed-loop module, which monitors user interaction indicators in real time and adjusts weight allocator parameters; The weight allocator performs: (1) Constructing a scenario parameter vector that includes audience activity, content sensitivity, and real-time response level; (2) Calculate the intermediate variables by linearly combining the learnable parameter matrix and the scene parameter vector plus the bias term; (3) Perform exponential normalization on intermediate variables to generate real-time weights; The initialization rule of the learnable parameter matrix is: The diagonal elements are initialized to fixed positive values; The off-diagonal elements are initialized to fixed negative values; The orthogonalization constraint is used to ensure that the matrix condition number is less than a fixed threshold.

2. The system according to claim 1, wherein: The receiving module includes: Data collection unit, used to obtain barrage text, gift trigger events and emoticon streams; The scene analysis unit is used to extract the live broadcast room identifier, the anchor's behavior feature vector, and the real-time number of concurrent users; The normalization processing unit performs the following operations: (1) Generate word vector sequences using a bidirectional long short-term memory model; (2) Calculate the time decay weight of the gift trigger event based on the decay factor and the time difference between the events; (3) Output multimodal feature matrix.

3. The system according to claim 1, wherein: The real-time sentiment analysis algorithm includes: Construct a sentiment dictionary, associate terms with extreme sentiment values ​​and confidence levels; Calculate the weighted sentiment score based on word position weight, sentiment extreme value and confidence; When the rate of change of sentiment scores in adjacent time exceeds the adaptive threshold, an alert is triggered.

4. The system according to claim 1, wherein: The semantic association algorithm performs: (1) Constructing a semantic graph that represents the strength of concept associations; (2) Extract entity sets from input text; (3) Calculate the semantic relevance score based on the path length and path edge weight between entities.

5. The system according to claim 1, wherein: The scene adaptation algorithm performs: Extract the live broadcast room style feature vector; The style feature vector is concatenated with the input feature vector and then subjected to linear transformation and Sigmoid function to calculate the scene matching degree; When the scene matching degree is lower than the matching threshold, the cross-live room reply strategy is triggered.

6. The system according to claim 1, wherein: The reply generation module performs dynamic template splicing: (1) Select a basic template; (2) Calculate and inject dynamic variables based on the conversation context vector; (3) Append template fragments when the user profile has tags.

7. The system according to claim 1, wherein: The distribution module calculates the priority of the live broadcast room: Priority is calculated based on real-time weight, audience value, content relevance, and the number of queued tasks combined with a logarithmic function; Resources are allocated based on priority and resource tilt factors through exponential normalization.

8. The system according to claim 1, wherein: The feedback closed loop module performs: (1) Calculate the comprehensive effect index based on the weight coefficient, click-through rate, normalized value of dwell time and conversion rate; (2) Adjust the trainable parameters through the policy gradient update formula based on the comprehensive effect indicators.

Images