A method and system for real-time feedback of live broadcast bullet screen based on interactive semantic matching

Abstract

The invention relates to the field of user interaction technology, specifically a method and system for real-time feedback of live broadcast barrage based on interactive semantic matching. The method receives and caches user barrage in real time, performs density analysis and priority evaluation on the barrage stream in combination with filtering and resource optimization rules, dynamically adjusts the response frequency and filters key barrages; by introducing the intention unit of the distillation processing mechanism, extracts user intention labels from key barrages and generates vectors, and simultaneously obtains the historical behavior data of the sending user, integrates the intention label and vector with the current real-time scene data of the live broadcast room, and constructs a barrage vector; performs similarity calculation based on the live broadcast content and the user's historical behavior to obtain the first feedback content; generates the second feedback content based on the reinforcement learning sentiment analysis strategy and the user's preference memory; and embeds the second feedback content into the current live broadcast interface through the client. The present invention realizes real-time positive feedback and intelligent guidance of live broadcast barrage through interactive semantic matching.

Description

Technical Field

[0001] The present invention relates to the field of user interaction technology, and specifically to a method and system for real-time feedback of live broadcast barrage based on interactive semantic matching. Background Art

[0002] With the vigorous development of the live streaming industry, live streaming has become an important form of interaction in e-commerce, entertainment, education and other fields, among which e-commerce live streaming accounts for a significant proportion. With the surge in the number of users, the management difficulties brought about by barrage interaction have become increasingly prominent.

[0003] During e-commerce livestreams, users frequently inquire about product prices, performance, usage scenarios, inventory status, shipping times, and other key information through online comments. Livestreamers often struggle to process and respond promptly due to the overwhelming volume of comments, leading to audience loss and decreased conversion rates. Existing automated response technology has significant limitations. It primarily relies on mechanical keyword matching and generates responses based on preset rules or templates, making it incapable of addressing complex or novel inquiries. Furthermore, traditional technology cannot contextualize user history or livestream context, resulting in monotonous responses that lack emotion and personalization, leading to a stilted and illogical interactive experience. This runs counter to the need for industry upgrades, as the market urgently needs solutions that can enhance user engagement and increase purchasing decision efficiency through precise, personalized, and emotional interactions.

[0004] To this end, a live broadcast barrage real-time feedback method and system based on interactive semantic matching is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for real-time feedback of live broadcast barrage based on interactive semantic matching, so as to realize semantic understanding and intelligent response of barrage in live broadcast scene. Specifically, by receiving the user's barrage information in real time and caching it in the barrage stream data queue; constructing filtering and resource optimization rules, performing density analysis and priority evaluation on the barrage stream, dynamically adjusting the response frequency and screening key barrages; processing the key barrage by introducing the distillation processing mechanism intention unit, extracting the user intention label and generating a vector; obtaining the historical behavior data of the user who sent the barrage, and combining the user intention label, vector and the current real-time scene data of the live broadcast room to construct a barrage vector; performing interactive semantic matching processing on the barrage vector, calculating the similarity with the current live broadcast content and the user's historical behavior data, and obtaining the first feedback content; based on sentiment analysis processing, combining the user preference memory, performing emotion recognition and positive emotion tendency adjustment on the first feedback content, and generating the second feedback content with positive emotion expression; embedding the second feedback content into the current live broadcast interface through the client in a synchronous insertion manner.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] Receive user barrage information in real time and cache it in the barrage stream data queue; build filtering and resource optimization rules, perform density analysis and priority evaluation on the barrage stream, dynamically adjust response frequency and filter key barrages;

[0008] The key comments are processed using the intention unit of the distillation processing mechanism to extract the user intention label and generate a vector; the historical behavior data of the user who sent the comments is obtained, and the comment vector is constructed by combining the user intention label, vector and the current real-time scene data of the live broadcast room;

[0009] Performing interactive semantic matching on the barrage vector and calculating similarity with the current live broadcast content and the user's historical behavior data to obtain first feedback content; based on sentiment analysis and combined with user preference memory, performing emotion recognition and positive sentiment adjustment on the first feedback content to generate second feedback content with positive emotional expression;

[0010] The second feedback content is embedded into the current live broadcast interface through the client in a synchronous insertion manner.

[0011] The user's barrage information includes text barrage and voice barrage. The text barrage is text information input and sent by the user through the live broadcast platform client interface, and is obtained in real time through a browser plug-in or client program. The browser plug-in monitors the DOM structure changes of the live broadcast page console, automatically extracts and parses the JSON format barrage data embedded in the page, and caches it in the barrage stream data queue;

[0012] The voice barrage is the user's voice input content during the live broadcast, which is obtained in real time through audio transcription. The audio transcription method uses FFmpeg to pull the live audio stream, calls an external automatic speech recognition service to transcribe the voice data into text information, and the text information is pushed to the server through the Kafka message middleware and included in the barrage stream data queue.

[0013] The filtering and resource optimization rules include barrage flow detection and key barrage priority assessment. By statistically analyzing the trend of changes in the number of barrages per unit time, dynamic barrage density detection is achieved, and a barrage response threshold is set. When the barrage flow density exceeds the preset threshold, the screening threshold is raised; the indicators of each barrage content are calculated to generate a comprehensive priority score, and key barrages are selected based on the score ranking.

[0014] The indicators of the barrage content include keyword density, sentiment tendency score, semantic relevance to the current live broadcast theme, and user historical influence score; the keyword density is calculated by word frequency-inverse document frequency TF-IDF, and the sentiment tendency score is identified by a pre-trained emotion recognition model to identify the intensity and polarity of the emotions expressed in the barrage; the semantic relevance is calculated by vectorizing the barrage content with the current live video title and the host's voice summary to calculate the cosine similarity; the user historical influence score is evaluated based on the user's historical speaking frequency in the current live broadcast room.

[0015] The distillation processing mechanism intention unit introduces the knowledge distillation mechanism to train lightweight units to approach the semantic expression ability of the pre-trained units, thereby completing efficient semantic compression and intent label generation; the user intent labels include expressive intent, interactive intent, directive intent, emotional intent, content-relevance intent, social intent, and reward incentive intent; the low-dimensional but semantically rich intent vector is extracted through the lightweight intent embedding unit after distillation and used as the barrage vector for interactive semantic matching processing.

[0016] The lightweight unit is implemented through a streamlined Transformer encoding network, combined with a multi-layer semantic attention module and a sparse connection mechanism, to improve the representation ability of the semantic features of the barrage while maintaining low computing resource consumption; its training process adopts a knowledge distillation method with the BERT model as the teacher model, combined with a double distillation strategy based on soft label alignment and intermediate hidden layer representation fitting, so that the intention vector output by the lightweight unit approaches the BERT model output in the semantic space, achieving efficient semantic compression and intent label generation.

[0017] The interactive semantic matching processing includes constructing a multi-dimensional representation feature that integrates barrage features, user intent labels, user historical behavior data and current live broadcast content; matching the barrage content with the information in the video clip through keyword-level alignment; setting a joint similarity function based on user preference memory, comprehensively evaluating semantic relevance and user interest adaptability, and outputting the first feedback content.

[0018] The user preference memory is used to mine and model the long-term interest characteristics of users in live broadcast interactions, and to aggregate and analyze multi-dimensional behavioral data such as the barrage, likes, comments, and reward records sent by users in the past, and extract high-frequency keywords, common semantic themes, and interactive contexts; identify the evolution trend of user interests through time series modeling methods, and form a preference topic vector; during the interaction process, the preference topic vector is aligned with the current barrage content as auxiliary information to achieve targeted optimization.

[0019] The joint similarity function is used to comprehensively evaluate the degree of semantic matching between the barrage content and the current live broadcast segment. Its processing process includes calculating the semantic correlation between the barrage semantic vector and the current live broadcast content semantic vector to measure the degree of fit between the two at the language expression and keyword levels; combining the preference topics in the user's historical behavior data to determine whether the barrage content meets the user's interest preferences; based on the contextual information of the live broadcast site, dynamically weighted and integrated the similarity scores of the three dimensions of the current topic, the host's language style and the interactive atmosphere with the barrage content, outputting a comprehensive matching result, and generating the first feedback content.

[0020] Based on the first feedback content, the Prompt guided generation unit performs deep emotional semantic recognition of subjective emotional words, emotional expression structures, and contextual emotional orientations in the text on the basis of a small number of preset training samples. In combination with the interactive context of the current live broadcast room and the user profile, the text content identified as non-positive emotions is migrated into semantically coherent and emotionally positive expression sentences through a language style control strategy. A reinforcement learning strategy is introduced in the generation process to continuously optimize the migration degree according to the real-time emotional feedback of the user group to achieve higher-quality positive guidance barrage generation.

[0021] The real-time emotions of the user group include the distribution of emotional tendencies in the barrage, changes in the emotional polarity of likes and comments, fluctuations in interaction frequency, dynamic changes in user stay time, and the overall emotional heat index of the live broadcast room. Through real-time perception and comprehensive evaluation of the above emotional signals, the intensity of emotional migration and the direction of semantic control are dynamically adjusted to ensure that the generated content is highly consistent with the live broadcast atmosphere in emotional expression, thereby enhancing user resonance and positive interactive experience.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. This invention introduces a knowledge distillation mechanism, which approaches the semantic expression ability of the pre-trained model by training a lightweight intent embedding model. It can significantly reduce computing resource consumption while ensuring understanding accuracy, and achieve efficient and low-latency barrage intent recognition and processing, meeting the high real-time performance requirements of live broadcast scenarios, and is suitable for multi-terminal deployment and large-scale online user access.

[0024] 2. By constructing a multi-dimensional representation tensor that integrates barrage features, user intent labels, historical behavior data, and live content semantics, and combining it with a joint similarity function to achieve keyword-level alignment and context-level comprehensive matching, the relevance between barrage and video content is effectively improved, the accuracy of interactive feedback is enhanced, and the true intentions expressed by users are promoted to be understood and responded to in a timely and reasonable manner.

[0025] 3. This invention combines prompt-guided generation and reinforcement learning strategies to identify negative emotions and transform them into new expressions with positive emotions and semantic coherence, preserving the user's original intent while improving the quality of expression. By combining the interactive context of the current live broadcast room and the user's profile, text content identified as negative emotions is transformed into semantically coherent and emotionally positive expressions through a language style control strategy.

[0026] 4. By introducing user preference memory, recording and modeling users' past behavioral characteristics such as barrages, likes, and comments, and performing time-series modeling on the live interactive context, the present invention can realize the generation of personalized feedback content and continuously adjust the feedback strategy with the help of reinforcement learning, so that the system can dynamically optimize the output effect according to the user's real-time reaction and emotional changes, thereby improving user interaction satisfaction and the platform's intelligence level. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of a flow chart of a live broadcast barrage real-time feedback method based on interactive semantic matching of the present invention;

[0028] Figure 2 This is a schematic diagram of the unit flow of the distillation treatment mechanism of the present invention;

[0029] Figure 3 This is a structural diagram of a live broadcast barrage real-time feedback system based on interactive semantic matching according to the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Example 1:

[0032] The present invention is a method for real-time feedback of live broadcast bullet screen based on interactive semantic matching. The process is as follows: Figure 1 Shown, including:

[0033] Receive user barrage information in real time and cache it in the barrage stream data queue; build filtering and resource optimization rules, perform density analysis and priority evaluation on the barrage stream, dynamically adjust response frequency and filter key barrages;

[0034] The key comments are processed using the intention unit of the distillation processing mechanism to extract the user intention label and generate a vector; the historical behavior data of the user who sent the comments is obtained, and the comment vector is constructed by combining the user intention label, vector and the current real-time scene data of the live broadcast room;

[0035] Performing interactive semantic matching on the barrage vector and calculating similarity with the current live broadcast content and the user's historical behavior data to obtain first feedback content; based on reinforcement learning sentiment analysis, combining user preference memory, performing emotion recognition and positive sentiment adjustment on the first feedback content to generate second feedback content with positive emotional expression;

[0036] The second feedback content is embedded into the current live broadcast interface through the client in a synchronous insertion manner.

[0037] Furthermore, the user's barrage information includes text barrage and voice barrage. The text barrage is text information input and sent by the user through the live broadcast platform client interface, and is processed by a plug-in program deployed on the user's browser side; the plug-in program inserts a listening script into the live broadcast page after obtaining access rights. When adding a new barrage DOM node, the listening script can immediately capture and parse the embedded JSON format barrage data, which contains the barrage content, user information and timestamp, and cache it in the barrage stream data queue; the temporarily stored barrage data is batch packaged and sent to the back-end server through a network request at a regular time and quantity.

[0038] The voice commentary is the user's voice input during a live broadcast, captured in real time through audio transcription. Specifically, the server uses the FFmpeg toolset to directly pull the raw audio stream from the live broadcast source address and uniformly adjust the sampling rate to the industry standard of 16kHz. The processed, clean audio data stream is directly output through a pipeline to a dedicated client program, which is responsible for communicating with the automatic speech recognition (ASR) service. The ASR divides the continuous audio data into small chunks and converts the speech into text through streaming recognition via a low-latency WebSocket connection, returning the transcription results. To ensure the stability and high availability of the entire system, the transcribed text is encapsulated into messages and pushed to a Kafka message topic. Kafka is used as the middleware to transcribe the speech and cache it in the commentary stream data queue.

[0039] By integrating text and voice interaction channels, the convenience and immersive experience of users in live broadcast environments have been significantly improved. Its text barrage acquisition path uses client plug-ins to achieve near-zero-latency real-time capture, and effectively reduces server load through a batch reporting mechanism, ensuring smooth interaction. At the same time, its voice barrage processing path leverages FFmpeg's wide compatibility and professional ASR services to ensure efficient and accurate conversion of voice information into text. The core advantage of this method is the introduction of Kafka message middleware to decouple the voice transcription link. This design not only gives the system powerful traffic peak shaving capabilities to cope with peak interactions, but also fundamentally guarantees the stability and high availability of the data link, avoiding data loss due to local failures.

[0040] The filtering and resource optimization rules include barrage traffic detection and key barrage priority evaluation. The barrage traffic detection counts the total number of barrages received within a unit time window, with each 5 seconds being a monitoring cycle. If the total number of barrages exceeds the set threshold (more than 100 per 5 seconds), the key barrage priority evaluation is triggered.

[0041] The barrage priority evaluation calculates the priority score for each barrage content in the current cycle, including the weighted summation of four dimensions: keyword density calculation, sentiment tendency score, semantic relevance calculation and user historical influence to form a comprehensive priority score for the barrage content.

[0042] By setting filtering and resource optimization rules, the efficiency of barrage processing and the quality of interaction in high-concurrency live broadcast scenarios have been effectively improved. Specifically, by detecting barrage traffic every 5 seconds and triggering the priority evaluation mechanism when the barrage density exceeds the set threshold, the system resource allocation can be dynamically adjusted to avoid information loss or processing delays caused by barrage overload. At the same time, the priority scoring mechanism combines the four dimensions of keyword density calculation, sentiment tendency score, semantic relevance calculation and user historical influence to ensure that the selected key barrages are more representative, valuable and interactive, thereby enhancing user participation and targeted feedback.

[0043] The keyword density calculation is performed by presetting a keyword document library that is highly relevant to the live broadcast theme. After the content of each bullet screen is segmented, the TF-IDF algorithm is used to calculate the importance of the keyword in the bullet screen, where the word frequency TF represents the frequency of a keyword appearing in the bullet screen, and the inverse document frequency represents the rarity of the word in all bullet screens. Finally, the keyword density score of each bullet screen is the sum of the TF-IDF of the keywords it contains, and the score is normalized to a range of 0-3 points through a normalization method; the normalization method is specifically to count the minimum and maximum values ​​of the TF-IDF sum of all bullet screens in the current monitoring period, and based on the range normalization method, the keyword density score of each bullet screen is converted to a value between 0 and 1 by subtracting the minimum value and dividing by the range (the difference between the maximum and minimum values), and then multiplying by 3 to obtain the normalized score, thereby ensuring that the keyword density score is within a uniform range of 0 to 3 points.

[0044] The emotional tendency score is obtained by using a pre-trained emotion recognition model to identify the emotional polarity of the barrage text and output the emotion category (positive, neutral, negative). If it is a positive emotion, it will get 2 points; if it is a neutral emotion, it will get 1 point; and if it is a negative emotion, it will get 0 points.

[0045] The semantic relevance calculation uses a text vectorization coding model to encode the barrage content, live broadcast room title, and host's current voice summary respectively to generate a high-dimensional semantic vector. The cosine similarity of the barrage vector is compared with the title vector and the voice summary vector, and the weighted average is taken as the semantic relevance score. A cosine similarity value greater than 0.7 is considered highly correlated (3 points), between 0.4-0.7 is moderately correlated (2 points), and below 0.4 is lowly correlated (1 point).

[0046] The user's historical influence is based on the live broadcast platform's user log records, which count the user's average speaking frequency, the number of interaction citations (such as likes and replies), and the number of times the user was marked as high-quality barrage in the current live broadcast room in the past. The top 10% in speaking frequency will receive 2 points, and 1 point will be added if there are multiple historical high-interaction barrage records. Users with no historical records will be given 0 points by default.

[0047] The weighted sum of the above four scores forms a comprehensive priority score for the barrage content. During periods of high traffic, barrages are sorted based on this score, with top-scoring key barrages being prioritized, enabling dynamic filtering and optimized feedback at the content level.

[0048] This multi-dimensional, comprehensive priority scoring method, based on keyword density, sentiment, semantic relevance, and historical user influence, accurately identifies and selects key bullet comments that are highly relevant to live content and possess positive interactive value, effectively improving the quality of bullet comment information and the viewing experience. During peak bullet comment traffic periods, the system dynamically sorts and prioritizes key bullet comments, avoiding information overload and redundant content interference. This enables intelligent hierarchical filtering and optimized feedback of bullet comment content, enhancing user engagement and the real-time and effectiveness of live interaction, significantly enhancing the operational efficiency and user stickiness of the live streaming platform.

[0049] The distillation processing mechanism intention unit is built based on a lightweight Transformer encoding network, see Figure 2 The network includes multiple layers of self-attention mechanisms and feedforward neural network layers, while also introducing a sparse connection mechanism to reduce model complexity and computational resource consumption. During the training phase, the pre-trained BERT model is used as the teacher model, guiding the learning of lightweight units through knowledge distillation. The specific training process consists of two parts:

[0050] First, soft label alignment, which guides lightweight units to learn more accurate semantic expression capabilities by calculating the KL divergence between the probability distribution (soft labels) output by the teacher model and the probability distribution output by the lightweight unit;

[0051] The second is the intermediate hidden layer feature fitting. The hidden layer representations of multiple intermediate layers of the teacher model are selected as supervision signals. The output of the corresponding layer of the lightweight unit is aligned with the hidden layer representation of the teacher model by minimizing the mean square error (MSE) to ensure the effective transmission of shallow and deep semantic information.

[0052] After training, the lightweight unit can quickly extract user intent labels for the input key barrage text, including expressive intent, interactive intent, directive intent, emotional intent, content-relevance intent, social intent, and reward incentive intent, and generate intent vectors as barrage vectors for interactive semantic matching.

[0053] The intention unit of the distillation processing mechanism significantly reduces the computational complexity and resource consumption of the model by introducing knowledge distillation technology, and realizes efficient semantic compression and intent recognition; at the same time, the lightweight network can quickly and accurately extract multi-dimensional user intention labels while maintaining a high semantic expression ability, thereby improving the accuracy and real-time performance of the semantic understanding of the barrage, and greatly enhancing the interactive response efficiency and user experience of the live barrage.

[0054] The interactive semantic matching processing includes constructing a multi-dimensional representation feature that integrates barrage features, user intent labels, user historical behavior data and current live broadcast content; matching the barrage content with the information in the video clip through keyword-level alignment; setting a joint similarity function based on user preference memory, comprehensively evaluating semantic relevance and user interest adaptability, and outputting the first feedback content.

[0055] The user preference memory is used to mine and model the long-term interest characteristics of users in live broadcast interactions, and to aggregate and analyze multi-dimensional behavioral data such as the barrage, likes, comments, and reward records sent by users in the past, and extract high-frequency keywords, common semantic themes, and interactive contexts; identify the evolution trend of user interests through time series modeling methods, and form a preference topic vector; during the interaction process, the preference topic vector is aligned with the current barrage content as auxiliary information to achieve targeted optimization.

[0056] The joint similarity function is used to comprehensively evaluate the degree of semantic matching between the barrage content and the current live broadcast segment. Its processing process includes calculating the semantic correlation between the barrage semantic vector and the current live broadcast content semantic vector to measure the degree of fit between the two at the language expression and keyword levels; combining the preferred topics in the user's historical behavior data to determine whether the barrage content meets the user's interest preferences, and extracting the corresponding semantic vectors based on the contextual information of the live broadcast site; the current topic obtains the theme vector through real-time subtitles and barrage modeling; the host's language style is constructed by extracting common expressions and tone features after voice-to-text conversion; the interactive atmosphere generates an atmosphere vector through data such as barrage density and emotional distribution per unit time.

[0057] The cosine similarity between the key barrage semantic vector and the above three types of vectors is calculated separately, and the three types of similarities are fused using an average weighted method to generate a fusion score; the fusion score is jointly judged with the user intention label and preference topic vector to obtain the first feedback content. If the intention is expressive or emotional, soothing, approving or emotional resonance language templates are preferred; if the intention is directive (such as asking questions or requests), key information is extracted from the current live voice summary or knowledge base for a brief answer; if the intention is interactive or social, relaxed and anthropomorphic language can be generated to encourage more interaction; if the user's preference topic vector is highly correlated with the barrage, the user's common topics or vocabulary are preferentially cited in the reply to enhance the personalized experience.

[0058] By introducing a joint judgment mechanism that integrates fusion scores with user intent labels and preferred topic vectors, we achieve precise semantic understanding and efficient responses to live stream commentary content, significantly improving the system's interactive intelligence and personalized recommendation capabilities. Compared to traditional commentary responses based solely on keyword or template matching, this approach balances contextual adaptability with the user's long-term interest expression. While ensuring real-time responses, it effectively enhances user engagement and interactive stickiness, improving the overall live stream experience and user satisfaction.

[0059] A prompt-guided generation unit is established to receive the first feedback content and analyze the text to identify the subjective emotional words, emotional expression structure, and emotional orientation information in the context. This recognition process is driven by a small number of preset training samples. By capturing the position of subjective emotional words, their syntactic structure, and related words in the context that may influence emotional orientation, it can identify and locate non-positive emotional content.

[0060] After identifying non-positive emotions, the system migrates this type of text based on the preset language style control strategy. During the migration process, the system retains the semantic core of the original expression, and combines the interactive context of the current live broadcast room and the user portrait to select the appropriate expression to rewrite, adjusting the original negative or neutral expression to an expression with semantic coherence, positive tone, and natural style, and generating the second feedback content. For example, direct expressions with negative tendencies are reconstructed into positive expressions with guidance or constructiveness to better adapt to the overall atmosphere of the live interactive scene;

[0061] At the same time, during the generation process, a reinforcement learning strategy is introduced to continuously optimize the degree of migration based on the real-time emotional feedback of the user group to achieve higher-quality positive guidance barrage generation. Specifically, the emotional tendency distribution (positive, negative, and neutral ratios) in the barrage, the number of likes and their changes, the emotional polarity analysis of the comment content (positive or negative), the frequency of user interaction (barrage sending speed, comment frequency), and the average length of user stay are collected in real time. These are then compared with the baseline data before barrage generation, and the change in each indicator (increase or decrease) is calculated to obtain a comprehensive positive guidance effect score. This positive guidance effect score serves as a reward signal for reinforcement learning and is fed back to the Prompt guided generation unit. Based on the reward signal, the barrage text migration strategy is adjusted to strengthen the expression methods that can enhance the positive atmosphere of the live broadcast room.

[0062] By introducing a reinforcement learning strategy, we achieve dynamic optimization and emotional guidance for livestream commentary content. The system accurately identifies and effectively transforms negative emotional expressions, generating semantically coherent and positive commentary responses, significantly enhancing the positive atmosphere of livestream interactions. Furthermore, a reward mechanism based on real-time emotional feedback enables continuous self-adjustment and optimization of the generation strategy, improving the personalization and emotional relevance of commentary content, enhancing user engagement and satisfaction, and promoting livestream activity and positive interaction.

[0063] This embodiment collects and filters key comments in real time, combines user intent and historical behavior data, and uses interactive semantic matching to accurately understand the relevance of comment content to the live broadcast scene, thereby improving the pertinence and accuracy of feedback. The sentiment analysis and adjustment mechanism based on reinforcement learning effectively guides comment content towards a positive and constructive direction, enhancing the interactive atmosphere and user experience of the live broadcast room. At the same time, it dynamically optimizes response frequency and comment priority to improve system processing efficiency, ensure timely and natural feedback, and promote continuous user participation and positive interaction in the live broadcast community.

[0064] Example 2:

[0065] During e-commerce live broadcasts, viewers send barrages to ask questions about product details such as product size, material, and logistics. Due to the large number of barrages and diverse expressions, it is difficult for the anchor and customer service to capture and understand the user's specific needs in a timely and accurate manner, resulting in delayed responses or deviation from user intentions, affecting the user's shopping experience and transaction rate.

[0066] See Figure 3 This embodiment proposes a live broadcast barrage real-time feedback system based on interactive semantic matching, including a barrage collection module, an intention recognition module, a semantic matching module, an emotion regulation module, and a feedback insertion module:

[0067] The system uses the barrage collection module to receive all user barrages in the live broadcast room in real time and caches the barrage information in the barrage stream data queue. It combines the barrage density and priority evaluation mechanism to screen out key barrages and process them first.

[0068] For the selected key comments, the system uses a distillation mechanism to analyze the intent unit text, extract the user intent label, and generate a corresponding intent vector. Simultaneously, the system retrieves the user's historical purchasing behavior data from the Redis cache or database, such as information on other sizes of the same item the user has purchased, to form a user behavior vector.

[0069] Combine the user intention vector with the historical behavior vector and the current product information status of the live broadcast room to construct a comprehensive bullet comment vector representation;

[0070] The bullet comment vector is input into the semantic matching module, where it performs multi-dimensional similarity calculations against the product description database of the current live broadcast content and historical user behavior data. This accurately matches the product inventory and size information corresponding to the user's question, generating preliminary feedback content. Emotion recognition is then performed on the preliminary feedback content to determine whether its emotional tendency aligns with the positive atmosphere of the live broadcast room. If the feedback content lacks emotional expression or has a bland tone, a language style control strategy is invoked. Combining the current live broadcast interaction environment and user profile, the feedback text is transformed into a semantically coherent and emotionally positive expression. For example, "in stock" can be changed to "sufficient stock, welcome to purchase," thereby improving user acceptance.

[0071] The adjusted feedback content is formed into the second feedback content, which is pushed to the live broadcast room in real time in the form of barrage through the feedback insertion module of the client, and automatically presented in the user's field of view to ensure that the information is conveyed instantly; combined with the reinforcement learning mechanism, the intensity of the language style migration strategy is adjusted according to the positive feedback signal to achieve dynamic optimization, ensure that the generated content continues to adapt to the live broadcast atmosphere, and improve user satisfaction and conversion rate.

[0072] The present invention realizes the efficient collection, accurate intention recognition and deep semantic matching of live broadcast barrage through the collaborative work of multiple modules. It can combine the user's historical behavior and the real-time scene of the live broadcast room to dynamically generate feedback content that meets user needs and is semantically coherent. The emotional regulation module further performs emotional recognition and positive adjustment on the feedback content, effectively improving the enthusiasm and interactive experience of barrage replies. This system not only enhances the user's sense of participation and emotional resonance in the live broadcast room, but also improves the efficiency of customer service response, promotes the healthy development of user purchase decisions and live broadcast atmosphere, and significantly improves the problems of slow response, single content and emotional imbalance in traditional barrage interaction.

[0073] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time feedback of live broadcast bullet screen based on interactive semantic matching, characterized in that: include: Receive user's bullet message information in real time and cache it in the bullet message stream data queue; Construct filtering and resource optimization rules, perform density analysis and priority evaluation on the barrage stream data queue, and screen key barrages; The key comments are processed using the intention unit of the distillation processing mechanism to extract the user intention label and generate a vector; the historical behavior data of the user who sent the comments is obtained, and the comment vector is constructed by combining the user intention label, the vector and the current real-time scene data of the live broadcast room; Performing interactive semantic matching on the barrage vector and calculating similarity with the current live broadcast content and historical user behavior data to obtain first feedback content; the interactive semantic matching process constructs a multi-dimensional representation feature that integrates barrage features, user intent labels, historical user behavior data, and the current live broadcast content; matching the barrage content with information in the video clip through keyword-level alignment; setting a joint similarity function based on user preference memory, comprehensively evaluating semantic relevance and user interest adaptability, and outputting the first feedback content; Based on reinforcement learning sentiment analysis processing, combined with user preference memory, emotion recognition and positive sentiment tendency adjustment are performed on the first feedback content to generate second feedback content with positive emotional expression; The reinforcement learning sentiment analysis process is to perform deep emotional semantic recognition on the first feedback content by using a prompt-guided generation unit built based on a small number of training samples. In combination with the current live broadcast room interaction context and user profile, the feedback content with negative emotions is transformed into new sentences with positive emotions. Introducing reinforcement learning strategies to dynamically adjust the emotions of barrage comments; The second feedback is synchronously inserted into the live broadcast interface.

2. A method for real-time feedback of live barrage based on interactive semantic matching according to claim 1, characterized in that: The user's barrage information includes text barrage and voice barrage. The text barrage is the text information entered and sent by the user through the live broadcast platform client interface, and is obtained in real time through the browser plug-in. The browser plug-in monitors the DOM structure changes of the live broadcast page console based on the Chrome Extension API, automatically extracts and parses the JSON format barrage data embedded in the live broadcast page, and caches it in the barrage stream data queue; The voice barrage is the user's voice input content during the live broadcast, which is obtained in real time through audio transcription. The audio transcription method uses FFmpeg to pull the live audio stream, calls an external automatic speech recognition service to transcribe the voice data into text information, and the text information is pushed to the server through the Kafka message middleware and included in the barrage stream data queue.

3. The method for real-time feedback of live barrage based on interactive semantic matching according to claim 1, characterized in that: The filtering and resource optimization rules include barrage flow detection and key barrage priority evaluation. The barrage flow detection realizes dynamic detection of barrage density by counting the changing trend of the number of barrages in unit time, sets the barrage response threshold, and triggers the key barrage priority evaluation when the barrage flow density exceeds the preset threshold; the key barrage priority evaluation calculates the indicators of each barrage content, generates a comprehensive priority score, and selects key barrages based on the comprehensive priority score.

4. The method for real-time feedback of live barrage based on interactive semantic matching according to claim 3 is characterized in that: The specific calculation method for the indicators of each barrage content is as follows: The keyword density, sentiment tendency score, semantic relevance of the current live broadcast theme, and user historical influence score of the barrage content are extracted; the keyword density is calculated by term frequency-inverse document frequency TF-IDF, and the sentiment tendency score is identified by a pre-trained emotion recognition model to identify the intensity and polarity of the emotions expressed in the barrage; the semantic relevance is calculated by vectorizing the barrage content, the current live broadcast video title, and the host's voice summary to calculate the cosine similarity.

5. The method for real-time feedback of live barrage based on interactive semantic matching according to claim 1 is characterized in that: The specific steps of using the distillation processing mechanism intent unit to extract user intent labels and generate vectors are as follows: The intention unit of the distillation processing mechanism includes an encoding network based on a lightweight Transformer structure, which is trained using a knowledge distillation method using a teacher model as a pre-trained BERT. The output expression is optimized through soft label alignment and intermediate layer feature fitting. After training, the user intention labels of key comments can be extracted and semantic vectors can be generated. The user intention labels include expressive intention, interactive intention, directive intention, emotional intention, content relevance intention, social intention and reward incentive intention. The intention vector is extracted through the distilled intention embedding unit and used as the barrage vector for interactive semantic matching.

6. A live broadcast barrage real-time feedback system based on interactive semantic matching, characterized by: Danmu collection module: Receives user Danmu information in real time, caches it in the Danmu stream data queue, and builds filtering and resource optimization rules; Intent Recognition Module: This module processes key comments using the intent unit introduced in the distillation processing mechanism, extracts user intent labels and generates vectors. It also obtains historical behavior data of users who send comments, and combines the user intent labels and vectors with the current real-time scene data of the live broadcast room to construct the comment vector. Semantic matching module: performs interactive semantic matching processing on the barrage vector, calculates similarity with the current live broadcast content and the historical behavior data, and obtains first feedback content; The interactive semantic matching process is to construct a multi-dimensional representation feature that integrates the bullet comment features, user intent labels, user historical behavior data, and the current live broadcast content; match the bullet comment content with the information in the video clip through keyword-level alignment; set a joint similarity function based on user preference memory, conduct a comprehensive evaluation of semantic relevance and user interest adaptability, and output the first feedback content; Emotion Regulation Module: Based on reinforcement learning emotion analysis and processing, combined with user preference memory, it performs emotion recognition and positive emotion tendency adjustment on the first feedback content to generate second feedback content with positive emotion expression; The reinforcement learning sentiment analysis process is to perform deep emotional semantic recognition on the first feedback content by using a prompt-guided generation unit built based on a small number of training samples. In combination with the current live broadcast room interaction context and user profile, the feedback content with negative emotions is transformed into new sentences with positive emotions. Introducing reinforcement learning strategies to dynamically adjust the emotions of barrage comments; Embed the second feedback content into the current live broadcast interface.

Images