Intelligent reply system suitable for live broadcast matrix scene

Abstract

The invention provides an intelligent reply system suitable for a live broadcast matrix scene. The intelligent reply system comprises a receiving module, an algorithm processing module, a reply generation module, a distribution module and a feedback closed loop module. The receiving module collects user input data and dynamic scene parameters of the live broadcast room in parallel, wherein the parameters comprise audience activeness, content sensitivity and real-time response level. The algorithm processing module dynamically integrates a real-time sentiment analysis algorithm, a semantic association algorithm and a scene adaptation algorithm through a weight distributor to generate a real-time weight. The reply generation module generates self-adaptive reply content according to the output, the distribution module configures a resource allocation strategy, and the feedback closed loop module monitors user interaction indexes in real time and adjusts weight distributor parameters. The reply content highly matched with the live broadcast scene can be generated in real time, and the user interaction experience and the live broadcast efficiency are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. More specifically, the present invention relates to an intelligent reply system applicable to the live broadcast matrix scenario. Background Art

[0002] In today's live broadcast industry, with the booming development of live broadcast platforms, the number of live broadcast rooms and the user scale are constantly increasing. The interaction between the live broadcast host and the audience has become a key factor in enhancing the live broadcast experience and user retention. However, the traditional live broadcast interaction method mainly relies on the real-time response of the host, and there are many limitations in the face of a large number of audiences and complex scenarios. For example, it may be difficult for the host to respond to the barrage or gifts of each audience in a timely manner, resulting in the failure to meet the interaction needs of some audiences. In addition, the diversity and complexity of live broadcast content also make it difficult for a single reply strategy to adapt to different live broadcast scenarios, affecting the interaction effect.

[0003] Although the existing live broadcast interaction technologies have achieved automated replies to a certain extent, most of them are based on simple keyword matching or preset templates, lacking the in-depth understanding and dynamic adaptation ability of the live broadcast scenario. These technologies cannot flexibly adjust the reply strategy according to factors such as the audience's activity, the sensitivity of the content, and the real-time response requirements, resulting in the reply content may be disconnected from the live broadcast scenario, and even cause 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 prior art: The existing live broadcast interaction technologies cannot effectively handle the complex interaction requirements in the scenario of multiple live broadcast rooms running in parallel, lack the real-time perception and intelligent adaptation ability to the dynamic changes of the live broadcast scenario, and it is difficult to generate reply content that is highly consistent with the live broadcast scenario and personalized, thus restricting the improvement of the quality of live broadcast interaction and user experience. Summary of the Invention

[0005] The present invention provides an intelligent reply system applicable to the live broadcast matrix scenario, including:

[0006] A receiving module, configured to collect the user input data and dynamic scenario parameters of multiple live broadcast rooms in parallel, where the dynamic scenario parameters include audience activity, content sensitivity, and real-time response level;

[0007] An algorithm processing module, connected to the receiving module, including a multi-algorithm fusion model. The multi-algorithm fusion model dynamically integrates a real-time sentiment analysis algorithm, a semantic association algorithm, and a scenario adaptation algorithm through a weight allocator, where the weight allocator generates real-time weights 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 scene-adaptive reply content according to the output of the multi-algorithm fusion model;

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

[0010] A feedback closed-loop module, which monitors user interaction metrics in real time and adjusts the parameters of the weight allocator.

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

[0012] A data acquisition unit, for obtaining bullet screen text, gift trigger events and emoji streams;

[0013] A scene parsing unit, for extracting the live broadcast room identifier, the host behavior feature vector and the number of concurrent users in real time;

[0014] A normalization processing unit, performing the following operations:

[0015] (1) Using a bidirectional long short-term memory model to generate a word vector sequence;

[0016] (2) Calculating the time decay weight of the gift trigger event according to the decay factor and the time difference of the event occurrence;

[0017] (3) Outputting a multi-modal feature matrix.

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

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

[0020] (2) Calculating an intermediate variable through a linear combination of a learnable parameter matrix and the scene parameter vector plus a bias term;

[0021] (3) Performing exponential normalization processing on the intermediate variable to generate a real-time weight.

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

[0023] Constructing a sentiment dictionary, with word items associated with sentiment extreme values and confidence levels;

[0024] Calculating a weighted sentiment score according to the word position weight, the sentiment extreme value and the confidence level;

[0025] Triggering an alarm when the change rate of the sentiment score at adjacent times exceeds an adaptive threshold.

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

[0027] (1) Construct a semantic graph representing the strength of concept associations;

[0028] (2) Extract an entity set from the input text;

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

[0030] As a further improvement of this application, the scenario adaptation algorithm performs:

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

[0032] After concatenating the style feature vector and the input feature vector, calculate the scenario matching degree through linear transformation and the Sigmoid function;

[0033] When the scenario matching degree is lower than the matching threshold, trigger the cross-live broadcast room reply strategy.

[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 dialogue context vector;

[0037] (3) Append template fragments when there are labels in the user profile.

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

[0039] Calculate the priority according to the real-time weight, audience value, content relevance, and the number of queued tasks in combination with the logarithmic function;

[0040] Allocate resources through exponential normalization according to the priority and the resource skew factor.

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

[0042] (1) Calculate the comprehensive effect index according to the weight coefficient in combination with the normalized values of click-through rate, stay duration, and conversion rate;

[0043] (2) Adjust the trainable parameters through the policy gradient update formula according to the comprehensive effect index.

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

[0045] Initialize the diagonal elements to a fixed positive value;

[0046] Initialize the non-diagonal elements to a fixed negative value;

[0047] Ensure that the matrix condition number is less than a fixed threshold through orthogonality constraints.

[0048] According to the above embodiments of the present invention, it has at least the following beneficial effects: The intelligent reply system can realize parallel collection and real-time processing of input data and dynamic scene parameters of multiple live broadcast room users. Through a multi-algorithm fusion model, it dynamically integrates sentiment analysis, semantic association, and scene adaptation algorithms, and generates adaptive reply content according to different live broadcast scenarios, effectively improving the accuracy and relevance of replies. At the same time, the system can dynamically adjust the parameters of the weight allocator according to the user interaction indicators monitored in real time, optimize the reply strategy, and further improve the interaction effect.

[0049] In addition, the system can also reasonably allocate the interaction resources of the live broadcast room through a dynamic resource allocation strategy, prioritize the processing of high-value and highly relevant live broadcast room interaction requirements, and improve the resource utilization efficiency. The introduction of the feedback closed-loop module enables the system to evaluate the interaction effect in real time and perform self-optimization, thereby continuously improving the performance and user experience of the system, and providing a more efficient and intelligent solution for the interaction mode in the live broadcast industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, wherein:

[0051] Figure 1 It is a schematic structural diagram of an intelligent reply system applicable to a live broadcast matrix scenario provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[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 given only to enable those skilled in the art to better understand and then implement the present invention, and do not limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to fully convey the scope of the present invention to those skilled in the art.

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

[0054] It should be noted that any number of elements in the drawings is for illustration rather than 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 applicable to 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 broadcast 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 level ;

[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, for generating scene-adaptive reply content according to the output of the multi-algorithm fusion model;

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

[0060] The feedback closed loop module 105 monitors the user interaction index 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 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 parallelly collecting user input data and dynamic scene parameters of multiple live rooms, including user input data such as bullet screen text, gift trigger events, and emoji streams, as well as dynamic scene parameters such as audience activity (calculated based on the interaction frequency and bullet screen sending rate of the audience, with a value range of 0-1), content sensitivity (detecting whether the live content involves sensitive topics using keyword matching and semantic analysis techniques, with a value range of 0-1), and real-time response level (divided according to the urgency of the interaction, with a range of 1-5). The algorithm processing module contains a multi-algorithm fusion model that dynamically integrates real-time sentiment analysis algorithms, semantic association algorithms, and scene adaptation algorithms through a weight allocator. The weight allocator generates real-time weights for each algorithm based on audience activity, content sensitivity, and real-time response level. For example, when the audience activity is high, the content sensitivity is high, and the real-time response level is high, a larger weight is assigned to the real-time sentiment analysis algorithm to analyze the audience's emotions accurately and respond in a timely manner. The reply generation module generates scene-adaptive reply content based on the output of the multi-algorithm fusion model, such as generating reply text that conforms to the style and content of the live room according to the sentiment analysis results and semantic association analysis. The distribution module is configured with a resource allocation strategy to push the generated reply content to the corresponding live rooms and users as needed. The feedback closed-loop module monitors user interaction metrics in real time, such as the click-through rate and stay duration of users after the reply, and adjusts the parameters of the weight allocator based on these metrics to optimize the reply strategy and improve the interaction effect.

[0063] Specifically, the data acquisition unit in the receiving module is responsible for obtaining multi-modal data such as bullet screen text, gift trigger events, and emoji streams. The scene parsing unit is used to extract the live room identifier (used to distinguish different live rooms), the host behavior feature vector (reflecting the behavior pattern and style of the host), and the real-time concurrent user count (the number of users currently watching the live broadcast simultaneously). The normalization processing unit preprocesses the collected data. For example, a bidirectional LSTM model is used to generate a word vector sequence to better capture the semantic information of the text. For gift trigger events, the system uses a time decay weighted formula for processing, where the decay factor and the time difference from the event occurrence are used to measure the importance and timeliness of the gift event. Finally, the output multi-modal feature matrix integrates these data into a unified format for 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 scenario parameter vector by integrating the audience activity, content sensitivity, and real-time response level into a three-dimensional vector. Then, it calculates an intermediate variable through a learnable parameter matrix and a bias term, and these parameters can be optimized using training data to adapt to different live broadcast scenarios. Finally, it generates a normalized weight to ensure that the sum of the weights of each algorithm is 1, thereby achieving the dynamic integration of real-time sentiment analysis algorithms, semantic association algorithms, and scenario 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, in educational live broadcasts, the weight of content sensitivity can be increased to better meet the needs of the educational scenario.

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

[0066] A data collection unit for obtaining bullet screen text, gift trigger events, and emoji streams;

[0067] A scenario parsing unit for extracting the live broadcast room identifier, the host behavior feature vector, and the number of concurrent users in real time;

[0068] A normalization processing unit that performs the following operations:

[0069] (1) Generate a word vector sequence using a bidirectional LSTM model;

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

[0071] ;

[0072] where is the decay factor, is the time difference between event occurrences;

[0073] (3) Output a multi-modal feature matrix .

[0074] It should be noted that the receiving module is a key component of the intelligent reply system, and its main function is to obtain the user input data and dynamic scenario parameters in the live broadcast room. The user input data includes bullet screen text, gift trigger events, and emoji streams, which are direct manifestations of the interaction between the audience and the live broadcast room. The dynamic scenario parameters are used to describe the overall state of the live broadcast room, including the live broadcast room identifier, the host behavior feature vector, and the number of concurrent users in real time. These parameters can help the system better understand the real-time situation of the live broadcast room, thereby providing a basis for subsequent intelligent processing.

[0075] In the intelligent reply system, the receiving module includes a data acquisition unit, a scene parsing unit, and a normalization processing unit. The data acquisition unit obtains the barrage text, gift trigger events, and emoji streams. It reads the barrage messages in real time through an interface, records the time, type, and quantity of gift giving, and collects the types and usage frequencies of emojis. The scene parsing unit extracts the live room identifier (such as the unique code for each live room), the host behavior feature vector (a vector generated by analyzing the host's language style, interaction frequency, action features, etc., used to describe the host's behavior pattern), and the real-time concurrent user count (counting the number of users watching the live broadcast simultaneously). The normalization processing unit uses a bidirectional long short-term memory model (Bi-LSTM) to generate a sequence of word vectors, encodes the barrage text, and captures the semantic information of the text; calculates the time decay weight of the gift trigger event according to the decay factor (such as set to 0.9, indicating that as time goes by, the importance of the event gradually decreases) and the time difference between the event occurrences. The initial weight is set according to the gift value; finally, it outputs a multi-modal feature matrix, integrating the word vectors of the barrage text, the weight vectors of the gift events, and the vectors of the emojis into a matrix as the input for subsequent processing.

[0076] Specifically, the data acquisition unit in the receiving module is responsible for obtaining the barrage text, gift trigger events, and emoji streams in the live room in real time. The barrage text refers to the text messages sent by the audience in the live room, and through natural language processing technology, the emotional tendency and semantic information in it can be extracted. The gift trigger event refers to the behavior of the audience sending virtual gifts. By recording the occurrence time and type of the event, the system can analyze the degree of support from the audience. The emoji stream is the emojis sent by the audience, reflecting the emotional state of the audience. The scene parsing unit extracts the live room identifier to distinguish different live rooms; the host behavior feature vector, generated by analyzing the host's language and behavior patterns, 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 simultaneously, is used to measure the popularity of the live room. The normalization processing unit preprocesses the collected data, for example, using a bidirectional LSTM model to generate a sequence of word vectors to better capture the text semantic information. The time decay weighting formula for the gift trigger event is:

[0077] ;

[0078] where is the initial weight, is the decay factor, is the time difference between the event occurrences. This formula is used to measure the importance and timeliness of the gift event. As time goes by, the weight of the event will gradually decay. Finally, the output multi-modal feature matrix integrates these data into a unified format for further analysis by the subsequent algorithm processing module.

[0079] Preferably, when processing the gift trigger event, the normalization processing unit can adjust the attenuation factor according to different live broadcast scenarios. . For example, in a highly active live broadcast room, the timeliness of gift events is stronger, and the attenuation factor can be set smaller to more quickly reflect the emotional changes of the audience. For the processing of the emoji stream, in addition to counting the quantity, an emotion analysis algorithm can be introduced to classify the emotional tendency of the emojis, such as positive, negative, or neutral. In addition, for the output of the multi-modal 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) Construct a scene parameter vector ;

[0082] (2) Calculate an intermediate variable:

[0083] ;

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

[0085] (3) Generate a normalized weight:

[0086] .

[0087] It should be noted that the weight allocator is the core component in the system for dynamically adjusting the weights of various algorithms. It constructs a scene parameter vector and calculates an intermediate variable, and finally generates a normalized weight, thereby realizing the dynamic integration of real-time emotion analysis algorithms, semantic association algorithms, and scene adaptation algorithms. This process can flexibly adjust the weights of various algorithms according to the dynamic scene parameters of the live broadcast room (such as audience activity, content sensitivity, and real-time response level), ensuring that the reply content output by the system is more in line with the requirements of the current live broadcast scene.

[0088] In the algorithm processing module of the intelligent reply system, the weight allocator constructs a scenario parameter vector containing audience activity, content sensitivity, and real-time response level. For example, when the audience activity is 0.8, the content sensitivity is 0.6, and the real-time response level is 4, the vector is [0.8, 0.6, 4]. Then, a linear combination of the learnable parameter matrix (the initial values can be set according to experience, such as the diagonal elements being set to 0.5 and the non-diagonal elements being set to -0.1) and the scenario parameter vector, plus a bias term (such as being set to [0.1, 0.1, 0.1]), is used to calculate an intermediate variable. Next, the intermediate variable is subjected to exponential normalization to generate real-time weights, that is, after taking the exponential function of each component of the intermediate variable, it is divided by the sum of the exponential function values of all components to obtain the weights of each algorithm, ensuring that the sum of the weights is 1. For example, the calculated real-time sentiment analysis algorithm weight is 0.4, the semantic association algorithm weight is 0.3, and the scenario adaptation algorithm weight is 0.3. These weights reflect the importance and contribution degree of each algorithm in the current scenario and are used to subsequently fuse the results of each algorithm to generate appropriate reply content.

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

[0090] The weight allocator calculates the intermediate variable through a learnable parameter matrix and a bias term , and the specific formula is:

[0091] ;

[0092] where is a matrix used to adjust the influence of each scenario parameter on weight allocation; is a bias vector used to adjust the baseline of weight allocation.

[0093] Next, the final weights are generated through a 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 according to the dynamic scenario parameters of the current live broadcast room, thereby generating more accurate and adaptable reply content.

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

[0095] Step 41: Construct an emotion dictionary, associating each term with an extreme emotion value and a confidence level ;

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

[0097] ;

[0098] where is the word position weight, is the number of emotion words;

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

[0100] ;

[0101] a warning is triggered, where is a very small constant, is the adaptive threshold, is the emotion score at the current moment, is the emotion score at the previous moment.

[0102] It should be noted that the real-time emotion analysis algorithm is an important part of the intelligent reply system. Its main function is to analyze the emotion tendency of the user input text by constructing an emotion dictionary and calculating the weighted emotion score. When the emotion change rate exceeds the set threshold, the algorithm will trigger a warning so that the system can respond in a timely manner to the situation where the emotion fluctuates greatly. This process can help the system better understand the user's emotional state, so as to generate more user-emotion-fitting reply content.

[0103] In the intelligent reply system, the real-time emotion analysis algorithm first constructs an emotion dictionary, in which each term is associated with an extreme emotion value (e.g., the extreme emotion value corresponding to "good" is 0.8, and the extreme emotion value corresponding to "bad" is -0.8) and a confidence level (e.g., the confidence level of common emotion words is 0.9, and the confidence level of rare words is 0.6). Then, according to the word position weight (e.g., the word position weights at the beginning and end of the bullet screen text are set to 1.2, and the middle part is set to 1.0), the extreme emotion value and the confidence level, the weighted emotion score is calculated. When the change rate of the emotion score between adjacent times exceeds the adaptive threshold (e.g., set to 0.5), a warning is triggered. The adaptive threshold can be dynamically adjusted according to the live broadcast type and the audience activity. When the change rate is greater than the threshold, the system sends a warning signal to prompt the anchor or automatically adjust the reply strategy to cope with the rapid change of the audience's emotions.

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

[0105] Next, the algorithm calculates the weighted emotional score, and the formula is:

[0106] ;

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

[0108] In addition, when the emotional change rate satisfies:

[0109] ;

[0110] a warning is triggered. Among them, is a very small constant, which is used to avoid the case where the denominator is zero; is the adaptive threshold, which is used to control the sensitivity of emotional changes.

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

[0112] Furthermore, the adaptive threshold can be adjusted according to different live broadcast scenarios. For example, in entertainment live broadcasts, the emotional fluctuations are relatively large, and the threshold can be set relatively high, such as 0.5; while in education live broadcasts, the emotional fluctuations are relatively small, and the threshold can be set lower, such as 0.3. In addition, in order to further improve the accuracy of emotional analysis, context information can also be introduced to adjust the emotional score of the current word by considering the emotional tendencies of the context.

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

[0114] (1) Construct a semantic graph, and the edge weight represents the concept association strength;

[0115] (2) Extract the entity set ;

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

[0117] ;

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

[0119] It should be noted that the semantic association algorithm is a key part of the intelligent reply system for analyzing the semantic relevance between the user input text and the live broadcast scenario. By constructing a semantic graph and calculating the association strength between the input text and the concepts in the graph, this algorithm can quantify the semantic relevance between the text and the live broadcast scenario. This process can help the system generate replies that are more in line with the live broadcast theme and content, improving the accuracy and relevance of interactions.

[0120] The semantic association algorithm first constructs a semantic graph. The nodes in the graph represent concepts (such as in an educational live broadcast, "mathematics", "function", "geometry", etc. are nodes), and the edge weights represent the concept association strength (such as the association strength between "function" and "geometry" is 0.6, determined by co-occurrence analysis or expert annotation). An entity set is extracted from the input text. For example, if the input text is "How to solve this function equation", the extracted entity is "function". The semantic relevance score is calculated based on the path length between entities (such as the path length from "function" to "equation" is 2) and the path edge weight (the sum of the edge weights on the path is 1.2). This score reflects the semantic matching degree between the input text and the live broadcast scenario. For example, a high score indicates a strong correlation between the text and the current live broadcast content, and more topic-appropriate reply content can be generated accordingly.

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

[0122] Then, the algorithm calculates the semantic relevance score between the input text and the concepts in the semantic graph. The formula is:

[0123] ;

[0124] Among them is the path length between entities, is the number of concepts associated with the entity , is the path edge weight. This score reflects the semantic matching degree between the input text and the live broadcast scenario. The higher the score, the stronger the correlation.

[0125] Preferably, the construction of the semantic graph can be achieved by analyzing a large amount of text data related to live broadcasts, extracting keywords and concepts therein, and determining the association strength between concepts through co-occurrence analysis or knowledge graph technology. In practical applications, the semantic graph can be customized according to different types of live broadcasts (such as entertainment live broadcasts, educational live broadcasts, etc.). For example, in an educational live broadcast, the semantic graph can include concepts such as subject knowledge points and teaching terms, while in an entertainment live broadcast, it can include concepts such as stars, programs, and music.

[0126] Furthermore, to improve the efficiency of semantic relevance calculation, 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. For the setting of the path length , it can be adjusted according to actual needs. For example, when precise matching needs to be emphasized, a shorter path length can be set; while when broader semantic associations need to be considered, the path length can be appropriately increased.

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

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

[0129] Step 62: Calculate the scenario matching degree:

[0130] ;

[0131] where 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 the matching threshold), trigger the cross-live broadcast room reply strategy.

[0133] It should be noted that the scenario adaptation algorithm is a key part of the intelligent reply system for evaluating whether the style of the live broadcast room matches the user input content. By extracting the feature vector of the live broadcast room style and calculating the scenario matching degree, this algorithm can determine whether the current user input is suitable for the style and atmosphere of the current live broadcast room. If the matching degree is lower than the set threshold, the system will trigger the cross-live broadcast room reply strategy, thereby ensuring that the reply content is consistent with the overall style of the live broadcast room and improving the user experience.

[0134] The implementation process of the scene adaptation algorithm in the intelligent reply system is as follows: First, extract the style feature vector of the live broadcast room. Generate the feature vector by performing text analysis on the historical barrage text of the live broadcast room (such as counting word frequencies and calculating sentiment tendencies) and analyzing the language style features of the anchor (such as speech rate, vocabulary richness, etc.). For example, the dimension of the feature vector is 10, including indicators such as barrage popularity and sentiment polarity. Then, splice the style feature vector with the input feature vector (composed of text features, behavioral features, etc. input by the current user), and calculate the scene matching degree through linear transformation (such as matrix multiplication using a weight matrix) and the Sigmoid function. When the scene matching degree is lower than the matching threshold (such as set to 0.5), trigger the cross-live broadcast room reply strategy, that is, retrieve appropriate reply content from other live broadcast rooms with similar styles or adjust the current reply generation strategy to ensure that the reply matches the style of the current live broadcast room. For example, compare the style features of the current live broadcast room with the style features of other live broadcast rooms pre-stored in the system, find the live broadcast rooms with high similarity, and adjust with reference to their reply styles and content.

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

[0136] Preferably, the style feature vector of the live broadcast room can be generated by analyzing multi-dimensional data such as the historical barrage text of the live broadcast room, the language style of the anchor, and the interaction frequency. For example, text mining techniques can be used to extract high-frequency words and sentiment tendencies in the barrage, and combine the language style features of the anchor (such as speech rate, tone, etc.) to construct the feature vector.

[0137] For the weight matrix and the bias term , they can be optimized through training data to adapt to different live broadcast scenarios. The matching threshold can be adjusted according to the type and style of the live broadcast room. For example, in entertainment live broadcasts, the style is relatively diverse, and the threshold can be set relatively high, such as 0.6; while in education live broadcasts, the style is relatively serious, and the threshold can be set lower, such as 0.4.

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

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

[0140] (1) Select the basic template ;

[0141] (2) Inject dynamic variables:

[0142] ;

[0143] where is the weight matrix, is the bias term, is the dialogue context vector;

[0144] (3) Append template fragments when there are tags in the user profile Exist tags and append template fragments.

[0145] It should be noted that the reply generation module is the core part of the intelligent reply system for generating the final reply content. It generates reply content that highly adapts to the current live broadcast scenario and user needs through the method of dynamic template splicing, combining the basic template, dynamic variables, and user profile information. This process can ensure that the reply content not only conforms to the style and atmosphere of the live broadcast room but also is customized according to the personalized needs of users, thereby improving user satisfaction and interactive experience.

[0146] In the reply generation module, the process of dynamic template splicing is as follows: First, select a basic template, which is selected from the template library according to the type and scenario of the current live broadcast. For example, in a game live broadcast, select a basic template that includes game operation tips and encourages interaction. Then, calculate and inject dynamic variables based on the dialogue context vector (a vector generated from the historical records of the current dialogue, user profile, etc., containing information such as dialogue theme and sentiment tendency). When there are tags in the user profile (such as user tags being "game enthusiast", "high-value user", etc.), append template fragments. For example, append fragments expressing gratitude and exclusive welfare information for "high-value users" to make the reply content more personalized and meet the needs and preferences of different users.

[0147] Specifically, the reply generation module first selects a basic template , which is a pre-designed reply framework for providing the basic structure of the reply. Then, the module injects dynamic variables and calculates the dynamic variable through the formula , where represents the th weight value in the weight matrix, is the bias term, is the th element in the dialogue context vector, which is used to reflect the background and content of the current dialogue. When the user profile When there are tags, the module will also append template fragments to further personalize the response content. For example, if the user profile shows that the user is a loyal fan of a certain live streamer, the system can add specific fan-exclusive content to the response, thereby enhancing the user's sense of belonging and willingness to interact.

[0148] Preferably, the basic template can be customized according to different live streaming scenarios. For example, in educational live streams, the basic template can contain more knowledge-based content and guiding questions; while in entertainment live streams, the basic template can be more relaxed and humorous, containing more interactive elements. For the weight matrix and the bias term , they can be trained through machine learning algorithms to adapt to different conversation styles and scenario requirements.

[0149] Furthermore, to further improve the degree of personalization of the response, more user profile features can be introduced, such as the user's interests and hobbies, viewing history, etc. For example, if the user often watches technology live streams, the system can add more technology elements to the response. At the same time, to handle complex conversation scenarios, multiple template fragments can be designed and dynamically selected and spliced according to the real-time conversation content to generate more natural and coherent response content.

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

[0151] ;

[0152] where is the audience value degree, is the content relevance degree, is the number of queued tasks;

[0153] The dynamic resource allocation formula is:

[0154] ;

[0155] where is the resource skew factor, is the total number of live broadcast rooms.

[0156] It should be noted that the distribution module is a key part of the intelligent response system for managing and allocating interaction resources. It calculates the priority of the live broadcast room and dynamically allocates resources according to the priority, ensuring that each live broadcast room can obtain interaction support that matches its importance and needs. This process can optimize the resource utilization efficiency, improve the user experience, and at the same time ensure the fairness and efficiency of live interaction.

[0157] The implementation process of the distribution module in the intelligent reply system is as follows: First, calculate the priority of the live room. According to the real-time weight (generated by the algorithm processing module, reflecting the urgency of the current interaction needs of the live room), the audience value (calculated by analyzing the audience's reward amount, viewing duration, interaction frequency, etc. For example, the reward amount accounts for 0.6, the viewing duration accounts for 0.3, and the interaction frequency accounts for 0.1, and the weighted sum is obtained), the content relevance (calculate the matching degree between the user input and the live content using the semantic similarity algorithm, with a value range of 0-1), and the number of queued tasks (the number of interaction tasks currently waiting to be processed), and combine the logarithmic function to calculate the priority. Then, according to the priority and the resource tilt factor (set according to the importance and commercial value of the live room. For example, the resource tilt factor for a popular live room is set to 1.2, and for an ordinary live room is set to 1.0), allocate resources through exponential normalization to ensure the rationality and efficiency of resource allocation, and give priority to processing high-priority live room interaction requests.

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

[0159] ;

[0160] Among them, and are weight coefficients, is the audience value, reflecting the contribution and importance of the audience to the live room; is the content relevance, measuring the matching degree between the current interaction content and the live room theme; is the number of queued tasks, indicating the number of interaction tasks currently waiting to be processed. Through this formula, the system can comprehensively consider the audience value, content relevance, and task backlog situation, and dynamically adjust the priority of the live room.

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

[0162] ;

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

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

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

[0166] The feedback closed-loop module performs:

[0167] (1) Calculate the comprehensive effect index:

[0168] ;

[0169] where and is the retention duration, is the preset threshold, is the click-through rate, which is the proportion of users clicking on the reply content; CR is the conversion rate, that is, the proportion of users completing the target behavior, such as purchasing, following, etc.

[0170] (2) Policy gradient update formula:

[0171] ;

[0172] where is the policy function, is the trainable parameter.

[0173] It should be noted that the feedback closed-loop module is an important part of the intelligent reply system for real-time monitoring and optimizing the interactive effect. By calculating the comprehensive effect index and using the policy gradient update formula to adjust the system parameters, this module can dynamically optimize the reply strategy according to the actual feedback of user interaction, thereby continuously improving the performance of the system and the user experience.

[0174] In the feedback closed-loop module, first, according to the weight coefficients (such as the click-through rate weight set to 0.4, the normalized dwell time weight set to 0.3, and the conversion rate weight set to 0.3), combined with the click-through rate (the proportion of users clicking on the response content, and the calculation formula is the number of clicks / the number of impressions), the normalized dwell time (normalizing the dwell time of users after the response, with a range of 0 - 1), and the conversion rate (the proportion of users completing target behaviors such as purchases or follows, and the calculation formula is the number of conversions / the number of clicks), the comprehensive effect index is calculated. Then, according to the comprehensive effect index, the trainable parameters are adjusted through the policy gradient update formula. In this way, the system can dynamically adjust the parameters according to the actual user interaction effect, optimize the response strategy, and improve the user experience and interaction effect.

[0175] Furthermore, in order to further improve the adaptability of the system, an adaptive learning rate can be introduced to dynamically adjust the learning rate according to the change of the comprehensive effect index, thereby accelerating the optimization speed of the system. For example, when the comprehensive effect index drops continuously for multiple times, the learning rate can be appropriately increased to accelerate the parameter adjustment speed.

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

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

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

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

[0180] The matrix condition number is ensured to be less than a fixed threshold through orthogonality constraints.

[0181] It should be noted that the initialization rule of the parameter matrix and the orthogonality constraints are the key links to ensure the stability and efficiency of the intelligent response system. The initialization method of the parameter matrix directly affects the learning efficiency and convergence speed of the system, while the orthogonality 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 initialization process of the learnable parameter matrix, first, the diagonal elements are initialized to a fixed positive value, such as 0.5, to enhance the independent influence of each scenario parameter (such as audience activity, content sensitivity, real-time response level) in the initial stage, enabling each parameter to play a fundamental role in weight allocation. The non-diagonal elements are initialized to a fixed negative value, such as -0.1, to introduce negative correlations between parameters, reduce redundancy, and avoid unreasonable weight allocation caused by interference between parameters. Then, through orthogonality constraints, it is ensured that the condition number of the matrix is less than a fixed threshold (such as setting the threshold to 5). The orthogonality constraint uses the QR decomposition method to orthogonalize the matrix, making the column vectors of the matrix orthogonal to each other, thereby reducing the condition number of the matrix, improving the numerical stability of the matrix, avoiding numerical problems during calculations, ensuring the stable operation of the system in diverse live broadcast scenarios, and improving the accuracy and reliability of weight allocation.

[0183] Specifically,

[0184] ;

[0185] And through orthogonality constraints, it is ensured that the condition number of the matrix .

[0186] This initialization method aims to assign higher initial weights to the diagonal elements to enhance the independent influence of each scenario parameter, while reducing redundancy between parameters through negative correlations. The orthogonality constraint ensures that the condition number of the matrix . The condition number of the matrix is an important indicator to measure the numerical stability of the matrix. A smaller condition number means the matrix is more stable during calculations and can effectively avoid the problem of error magnification in numerical calculations. Through this initialization and constraint method, the system can maintain stable performance in diverse 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 parameters related to content sensitivity can be appropriately increased to better meet the requirements of the rigor of educational content. For the orthogonality constraint, more stringent condition number limits can be adopted, such as , to further improve the numerical stability of the system.

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

[0189] The above embodiments of the present invention have the following beneficial effects: The intelligent reply system can comprehensively perceive complex live broadcast scenarios by using the receiving module to collect user input data and dynamic scenario parameters of multiple live broadcast rooms in parallel, providing a rich and accurate data basis 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 various algorithms according to different live broadcast scenario parameters, thereby generating more accurate and adaptable reply content. The reply generation module uses dynamic template splicing technology to flexibly generate personalized reply content according to the conversation context and user profile, further enhancing the user experience. The distribution module can optimize the resource utilization efficiency by calculating the live broadcast room priority and performing dynamic resource allocation, ensuring that high-value and highly relevant live broadcast rooms obtain more interaction resources. The feedback closed-loop module can realize the self-optimization of the system and continuously improve the interaction effect by monitoring user interaction indicators in real time and adjusting the parameters of the weight allocator. In addition, when initializing the parameter matrix, the system adopts specific rules and ensures the matrix condition number through orthogonalization constraints, which can improve the stability and calculation efficiency of the system.

[0190] In summary, the system can effectively improve the intelligent level of live broadcast interaction, enhance user participation and satisfaction. It can adapt to various live broadcast scenarios, generate accurate and personalized reply content, while optimizing resource allocation to ensure the efficiency and fairness of interaction. Through the feedback closed-loop mechanism, the system can also continuously learn and optimize, further improving performance and user experience, providing an innovative solution for the interaction mode in the live broadcast industry.

[0191] Furthermore, the storage medium of the embodiment of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0192] The above description is only some preferred embodiments of the present invention and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present invention.

Claims

1. An intelligent reply system applicable to the live broadcast matrix scenario, characterized in that, Including: A receiving module for parallelly collecting user input data and dynamic scene parameters of multiple live rooms, where the dynamic scene parameters include audience activity, content sensitivity, and real-time response level; An algorithm processing module connected to the receiving module, including a multi-algorithm fusion model. 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, where the weight allocator generates real-time weights 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 scene-adaptive reply content according to the output of the multi-algorithm fusion model; A distribution module connected to the reply generation module, configured with a resource allocation strategy; A feedback closed-loop module for real-time monitoring of user interaction metrics and adjusting the parameters of the weight allocator.

2. The system according to claim 1, characterized in that, The receiving module includes: A data collection unit for obtaining bullet screen text, gift trigger events, and emoji streams; A scene parsing unit for extracting live room identifiers, host behavior feature vectors, and real-time concurrent user numbers; A normalization processing unit that performs the following operations: (1) Generating a word vector sequence using a bidirectional long short-term memory model; (2) Calculating the time decay weight of the gift trigger event according to the decay factor and the time difference between events; (3) Outputting a multi-modal feature matrix.

3. The system according to claim 1, characterized in that The weight allocator performs: (1) Constructing a scene parameter vector including audience activity, content sensitivity, and real-time response level; (2) Calculating an intermediate variable through a linear combination of a learnable parameter matrix and the scene parameter vector plus a bias term; (3) Performing exponential normalization on the intermediate variable to generate real-time weights.

4. The system according to claim 1, characterized in that, The real-time sentiment analysis algorithm includes: Constructing a sentiment dictionary, where the terms are associated with sentiment extremes and confidence levels; Calculating a weighted sentiment score according to the word position weight, sentiment extreme, and confidence level; Triggering an alarm when the change rate of the sentiment score between adjacent times exceeds an adaptive threshold.

5. The system according to claim 1, characterized in that The semantic association algorithm performs: (1) Constructing a semantic graph representing the strength of concept associations; (2) Extracting an entity set from the input text; (3) Calculating a semantic relevance score according to the path length between entities and the path edge weights.

6. The system according to claim 1, wherein The scene adaptation algorithm performs: Extracting the live room style feature vector; Calculating the scene matching degree by concatenating the style feature vector and the input feature vector and then performing a linear transformation and a Sigmoid function; Triggering a cross-live-room reply strategy when the scene matching degree is lower than the matching threshold.

7. The system according to claim 1, wherein The reply generation module performs dynamic template splicing: (1) Selecting a basic template; (2) Calculating and injecting dynamic variables based on the dialogue context vector; (3) Appending template fragments when there are labels in the user profile.

8. The system according to claim 1, wherein The distribution module calculates the priority of the live room: Calculating the priority by combining the real-time weight, audience value, content relevance, and the number of queued tasks with a logarithmic function; Allocating resources through exponential normalization according to the priority and the resource tilt factor.

9. The system according to claim 1, wherein The feedback closed-loop module performs: (1) Calculating a comprehensive effect index according to the weight coefficient combined with the normalized values of click-through rate, dwell time, and conversion rate; (2) Adjust the trainable parameters according to the comprehensive effect index through the policy gradient update formula.

10. The system according to claim 3, wherein The initialization rule of the learnable parameter matrix is as follows: The diagonal elements are initialized to a fixed positive value; The non-diagonal elements are initialized to a fixed negative value; Ensure that the matrix condition number is less than a fixed threshold through orthogonality constraints.

Images