Intelligent response method and device for commenting crisis, electronic equipment and storage medium

Abstract

The application relates to the technical field of intelligent response to comment crisis, and discloses an intelligent response method and device for comment crisis, electronic equipment and a storage medium, wherein the method comprises the following steps: through deep semantic analysis and clustering of multi-source comment data, independent crisis event entities can be accurately and efficiently identified and aggregated from noise information; through construction of a time sequence social communication network model and quantification of user node influence, scientific prediction and evaluation of crisis diffusion trend and intensity are realized, so that a crisis index can more accurately reflect the severity of an event; and in combination with cultural characteristics related to the event and the crisis index, a first-level response language suitable for the event is dynamically generated.The application has the beneficial effect that the response strategy is improved in pertinence, timeliness and cultural suitability, so that a subject can more actively and effectively manage and resolve online comment crisis, and automation and intelligentization of a main process from identification to response are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent response technology for comment crises, and more particularly to an intelligent response method, apparatus, electronic device, and storage medium for comment crises. Background Technology

[0002] With the widespread use of social media, businesses and organizations are facing increasingly frequent and complex online comment crises. Existing technological solutions have significant limitations in handling such crises: First, traditional methods rely heavily on manual monitoring or simple keyword matching, making it difficult to accurately identify potential crisis events in real time from massive, multi-source comment data. Second, the lack of quantitative analysis of comment propagation dynamics makes it impossible to effectively assess the role of individual users (especially high-influence nodes) in the spread of a crisis, leading to delayed warnings or misjudgments. Furthermore, existing response strategy generation models are often rigid, unable to dynamically and adaptively adjust according to the actual intensity of the crisis (crisis index) and the cultural background of the user groups involved, resulting in insufficient targeting and effectiveness of response strategies. Summary of the Invention

[0003] Therefore, it is necessary to propose an intelligent response method, device, electronic device, and storage medium for comment crises to address the existing problem of intelligent response to comment crises.

[0004] A smart response method for commenting on crises, the method comprising:

[0005] Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0006] Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements;

[0007] Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0008] A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0009] The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors.

[0010] Based on the commented event entity, the corresponding target cultural characteristics are obtained, and based on the crisis index and target cultural characteristics, a first-level response script is dynamically generated and released.

[0011] Furthermore, after the steps of obtaining the corresponding target cultural characteristics based on the comment event entity, and dynamically generating and publishing a first-level response script based on the crisis index and the target cultural characteristics, the method further includes:

[0012] Monitor user feedback data after the release of the first-level response script;

[0013] Analysis of changes in comment sentiment based on the aforementioned user feedback data;

[0014] Based on the changes in the sentiment of the comments, a secondary response script is generated again.

[0015] Furthermore, the step of regenerating the secondary response script based on the change in the sentiment of the comment includes:

[0016] The release operation of the first-level response script, the user feedback data obtained after release, and the changes in the sentiment of the comments are stored as a set of sample data in the experience pool;

[0017] The dialogue generation policy network is trained using the sample data in the experience pool;

[0018] The trained response generation policy network is used to generate the secondary response.

[0019] Further, the step of constructing a corresponding temporal social propagation network model based on each of the comment event entity datasets includes:

[0020] Identify all users who participated in the corresponding event discussion in the aforementioned comment event entity dataset, and form a node set;

[0021] Based on the social relationships and interaction behavior data in the comment event entity dataset, network edges are constructed between user nodes in the node set.

[0022] Extract the interaction behavior data of each network edge within a continuous time window to generate the corresponding edge temporal feature vector;

[0023] The node set, network edges, and corresponding edge temporal feature vectors are input into the temporal graph convolutional network model to construct the temporal social propagation network model.

[0024] Furthermore, the step of quantifying the influence scores of each user node in the time-series social propagation network model and calculating the crisis index in combination with the corresponding structured event vectors includes:

[0025] The influence weight of each user node is calculated using a pre-defined graph node sorting algorithm.

[0026] Sentiment analysis is performed on each published content in the aforementioned comment event entity dataset to obtain the corresponding sentiment intensity quantification value;

[0027] For each piece of published content, the corresponding emotional intensity quantification value is weighted and calculated with the influence weight of the publishing user to obtain the influence-weighted emotional value of the published content.

[0028] The influence-weighted sentiment values ​​of all published content in the comment event entity dataset are aggregated to generate the crisis index of the comment event entity.

[0029] Furthermore, the step of obtaining the corresponding target cultural characteristics based on the comment event entity, and dynamically generating and releasing a first-level response script based on the crisis index and the target cultural characteristics, includes:

[0030] Extract the region identifier information from the comment event entity dataset;

[0031] Based on the regional identification information, a set of cultural preference features associated with the regional identification information is obtained from a preset regional cultural feature database and used as the target cultural feature.

[0032] The crisis index and the target cultural characteristics are input into a speech generation engine based on a large language model to generate the first-level response speech.

[0033] Furthermore, the step of semantically parsing each of the comment data to generate a structured event vector representing event elements includes:

[0034] Identify the text content in the comment data;

[0035] A multilingual pre-trained model is used to perform unified semantic understanding on the text content and identify the core elements of the event in the text content; wherein, the core elements include the event subject, the event type, and the event object;

[0036] The core elements of the event are encoded into low-dimensional dense vectors to generate the structured event vector.

[0037] A smart response device for commenting on crises, the device comprising:

[0038] The data acquisition module is used to collect comment data from multiple users on the target social platform in real time; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0039] The parsing module is used to perform semantic parsing on each of the comment data to generate a structured event vector representing event elements;

[0040] The calculation module is used to calculate and perform real-time clustering of all the comment data based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0041] The construction module is used to construct a corresponding temporal social propagation network model based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0042] The quantization module is used to quantify the influence scores of each user node in the time-series social propagation network model and calculate the crisis index in combination with the corresponding structured event vectors.

[0043] The generation module is used to obtain the corresponding target cultural characteristics based on the comment event entity, and dynamically generate a first-level response script for release based on the crisis index and the target cultural characteristics.

[0044] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0045] Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0046] Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements;

[0047] Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0048] A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0049] The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors.

[0050] Based on the commented event entity, the corresponding target cultural characteristics are obtained, and based on the crisis index and target cultural characteristics, a first-level response script is dynamically generated and released.

[0051] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0052] Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0053] Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements;

[0054] Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0055] A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0056] The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors.

[0057] Based on the commented event entity, the corresponding target cultural characteristics are obtained, and based on the crisis index and target cultural characteristics, a first-level response script is dynamically generated and released.

[0058] The beneficial effects of this invention are as follows: By performing deep semantic analysis and clustering on multi-source comment data, it is possible to accurately and efficiently identify and aggregate independent crisis event entities from noisy information. By constructing a time-series social communication network model and quantifying the influence of user nodes, it is possible to scientifically predict and assess the trend and intensity of crisis spread, enabling the crisis index to more accurately reflect the severity of the event. By combining the cultural characteristics associated with the event with the aforementioned crisis index, it is possible to dynamically generate appropriate first-level response dialogues, which helps to improve the pertinence, timeliness, and cultural suitability of response strategies. This helps the subject to manage and resolve online comment crises more proactively and effectively, and realizes the automation and intelligence of the main process from identification to response. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] in:

[0061] Figure 1 This is a diagram illustrating the application environment of a smart response method for commenting on crises in one embodiment.

[0062] Figure 2 A flowchart of a smart response method for commenting on a crisis in one embodiment;

[0063] Figure 3 This is a structural block diagram of a smart response device for commenting on a crisis in one embodiment;

[0064] Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Figure 1 This is a diagram illustrating an intelligent response application environment for commenting on a crisis, as shown in one embodiment. (Refer to...) Figure 1 This intelligent response method for comment crises is applied to an intelligent response system for comment crises. The intelligent response system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to collect comment data, and the server 120 is used to generate primary response scripts.

[0067] like Figure 2 As shown, in one embodiment, a smart response method for comment crises is provided. This method can be applied to both terminals and servers; this embodiment uses terminal application as an example. The smart response method for comment crises specifically includes the following steps:

[0068] S1: Real-time collection of comment data from multiple users on the target social platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0069] S2: Perform semantic parsing on each of the comment data to generate a structured event vector representing the event elements;

[0070] S3: Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0071] S4: Construct a corresponding temporal social propagation network model based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0072] S5: Quantify the influence score of each user node in the time-series social propagation network model, and calculate the crisis index by combining it with the corresponding structured event vector;

[0073] S6: Obtain the corresponding target cultural characteristics based on the comment event entity, and dynamically generate and release a first-level response script based on the crisis index and target cultural characteristics.

[0074] As described in step S1 above, comment data from multiple users on the target social media platform is collected in real time. Comment data posted by users is dynamically acquired from various social media platforms. This comment data includes not only the user's text content but also various accompanying multimedia information (such as images, videos, etc.) and their corresponding social interaction behaviors. During the data collection process, an interface with the social media platform needs to be established to capture new comments in real time. In specific implementation, comment information is extracted from the target social media platform through API interfaces or web scraping technology. The acquired data should include user identity information (such as username and user ID), comment content, timestamps, and interaction relationships between users (such as likes, reposts, comments, etc.). Interaction behaviors are used to establish a relationship graph between users to further analyze the comment propagation path. The effectiveness of real-time data collection directly affects the accuracy of subsequent analysis results; therefore, data acquisition strategies should be continuously optimized to ensure the timeliness and comprehensiveness of the data.

[0075] As described in step S2 above, semantic parsing is performed on each of the comment data to generate structured event vectors representing event elements. Semantic parsing of comment data is a process of deep understanding and abstraction of collected information, capable of identifying implicit keywords, sentiment tendencies, and event elements in the comments, such as the subject (the initiator of the event), event type (e.g., complaint, praise, suggestion), and event object (the product or service involved). This can be achieved through advanced natural language processing techniques (e.g., word vector models, semantic understanding algorithms). For example, a pre-trained deep learning model (e.g., BERT or XLM-RoBERTa) can be used for semantic understanding. The model will perform word segmentation, part-of-speech tagging, and entity recognition on the comment text, thereby generating detailed structured event vectors. These event vectors convert the comment content into low-dimensional dense embedding vectors, which can be further analyzed using machine learning algorithms. Furthermore, deep learning models, such as those applying computer vision techniques and convolutional neural networks (CNNs), can enable the system to perform Optical Character Recognition (OCR) and sentiment analysis in images. This process goes beyond simply recognizing text content in images; it can also analyze the emotions conveyed by the images, such as determining whether an image expresses emotional tags like "anger," "happiness," or "sadness" through image classification models. By using convolutional neural networks and recurrent neural networks (RNNs) in deep learning, as well as models based on temporal data (such as Long Short-Term Memory networks, LSTM), the system can perform temporal analysis on video content. This includes extracting audio, dialogue, and image information from video frames to form a multi-dimensional sentiment analysis.

[0076] As described in step S3 above, the semantic similarity between the structured event vectors is calculated, and all the comment data are clustered in real time to obtain multiple comment event entity datasets. Comments with similar content themes are grouped together to form event entities. This process uses clustering analysis technology, which determines which comments can be classified into the same category by analyzing the similarity between event vectors. In practice, commonly used clustering algorithms include K-means and DBSCAN. Specifically, the system determines the similarity of comments by calculating the cosine similarity or Euclidean distance between event vectors. After the similarity calculation is completed, the system groups the comments based on a certain threshold to form multiple "comment event entity datasets." These datasets are similar in terms of theme, sentiment, etc., which can help enterprises better understand the current public opinion situation and its development dynamics.

[0077] As described in step S4 above, a corresponding temporal social propagation network model is constructed based on each comment event entity dataset. Based on the users and their social interaction information in each comment event entity dataset, a temporal social propagation network model is constructed for each event. This model uses a Temporal Graph Convolutional Network (TGCN) to capture the interactive relationships between users. Specifically, users in the comments are network nodes, and social interactions between users (such as comments, likes, and shares) are network edges, constructing a dynamically changing social network. TGCN is used to construct the temporal social propagation network model, where nodes represent users in the comment event entity, and edges represent the social interaction relationships between users. By inputting structured event vectors, network edges, and the temporal feature vectors of the edges, TGCN can learn user behavior and social interactions within a specific time period, predicting the propagation trend of information in the network. This enables the system to track the evolution of public opinion in real time, assess the influence of users and the ability to spread information in specific events, thereby providing accurate data support for crisis management. During the construction process, the system needs to identify the user represented by each node and construct a time-series-based network edge by analyzing their interaction behavior data, correlations, and frequency of participation. In the completed time-series social communication network, user nodes can reflect the information diffusion path and interaction intensity within a specific time period, which helps predict possible trends in public opinion evolution. Through time-series processing, the network model can be dynamically adjusted to reflect changes in user interactions in real time, providing a scientific basis for subsequent crisis index calculation and response strategies. The edge time-series feature vector is a vector used to represent the dynamic attributes of the edges (social interaction relationships between users) in the graph within a specific time window. These features include, but are not limited to, the number of interactions, the duration of interaction, and the frequency of forwarding and commenting, which collectively reflect the strength and changes in relationships between users.

[0078] As described in step S5 above, the influence score of each user node in the time-series social communication network model is quantified, and a crisis index is calculated by combining it with the corresponding structured event vector. A graph node ranking algorithm (such as PageRank) is used to calculate the influence score of each user node in the network. The influence score is based on the user's social relationships, as well as their position in the entire communication network, interaction frequency, and other factors, reflecting the user's relative importance in the comment event. Combined with the structured event vector corresponding to each user, the crisis index is calculated. By weighting each user's influence score with the emotional intensity of their published content, a comprehensive crisis index can be obtained. This index logically quantifies the degree of harm of the comment event and provides timely warnings to enterprises.

[0079] As described in step S6 above, the corresponding target cultural characteristics are obtained based on the comment event entity, and a primary response script is dynamically generated and released based on the crisis index and target cultural characteristics. After completing the crisis index calculation, the target cultural characteristics are obtained based on the comment event entity, and then the cultural characteristics are combined with the calculated crisis index. Specifically, the urgency reflected by the crisis index is combined with regional cultural characteristics, and primary response scripts adapted to different cultural scenarios are quickly generated through preset rules. The response scripts can be pre-set scripts, such as generating corresponding response scripts based on specific crisis indices and target cultural characteristics, or the crisis index and target cultural characteristics can be input into a large language model, such as GPT or Deepseek, to obtain corresponding response scripts. The primary response script is the first automatic response based on the preliminary crisis assessment and cultural characteristics.

[0080] In one embodiment, after step S6, which involves obtaining the corresponding target cultural characteristics based on the comment event entity and dynamically generating and publishing a first-level response script based on the crisis index and the target cultural characteristics, the method further includes:

[0081] S701: Monitor user feedback data after the release of the first-level response script;

[0082] S702: Analyze changes in comment sentiment based on the user feedback data;

[0083] S703: Generate a secondary response script based on the changes in the sentiment of the comments.

[0084] As described in step S701 above, user feedback data after the release of the primary response script is monitored. This involves real-time monitoring of user feedback following the release of the primary response script, specifically the continuous collection of various feedback data related to the script on social media platforms, such as user comments, likes, shares, and interactions. By applying social media monitoring tools, the system automatically captures comments related to the company's response measures, recording users' emotional reactions, attitude changes, and potential secondary comments. In practice, a specific monitoring window period can be set (e.g., within 24 hours of release) to ensure timely capture of initial user reactions. Furthermore, the system can use natural language processing technology to perform sentiment analysis on feedback comments, identifying positive, negative, and neutral sentiment tendencies in subsequent analysis. This feedback monitoring not only helps companies determine the effectiveness of the primary response script but also provides necessary data support for subsequent strategy adjustments, as well as identifying the evolution of public sentiment and potential crises. By regularly collecting and analyzing this feedback data, companies can form a closed-loop feedback mechanism, adjusting their response strategies in a timely manner to ensure continuous customer satisfaction and brand image maintenance.

[0085] As described in step S702 above, the change in sentiment of comments is analyzed based on the user feedback data. Specifically, sentiment analysis algorithms and natural language processing techniques can be used to quantify and classify user comments through a sentiment analysis model, converting the sentiment scores (such as positive, negative, or neutral) in the comments into easily understandable numerical values. In practice, each comment in the user feedback is automatically evaluated using a machine learning model (such as a support vector machine or neural network) to obtain a sentiment score, in order to identify whether the overall sentiment tone of user feedback has changed. For example, if the sentiment of most user feedback changes from negative to positive within 24 hours of publication, it can indicate that the Level 1 response dialogue has been successful to some extent. Specifically, a unified sentiment analysis model can be set up within the system. In the crisis index calculation phase, this model is used to analyze the content in the historical / current comment dataset. In the feedback monitoring phase, the same model is used to analyze newly generated user feedback data after the Level 1 response dialogue is published. Both are applications of the same technical module at different times and for different data objects.

[0086] As described in step S703 above, a secondary response script is generated again based on the changes in the sentiment of the comments. After completing the user feedback sentiment analysis, the company can dynamically generate secondary response scripts based on these changes in sentiment. The goal of this process is to revise and optimize the initially released primary response scripts to better meet user expectations and needs. Specifically, by interpreting the analysis results, for example, if the sentiment analysis shows that most users expressed dissatisfaction or negative emotions, the company may need to adjust its wording to respond to user feedback more sincerely and provide appropriate solutions or remedial measures. The generated secondary response scripts can incorporate more empathetic expressions, recognition of user emotions, and even provide specific compensation solutions (such as discounts, coupons, etc.) to alleviate emotional conflict and restore user trust. Conversation generation models, preset language templates, and the company's cultural characteristics can be used to create responses tailored to specific situations. Through a continuous feedback and correction mechanism, the company can achieve adaptive customer relationship management, more effectively respond to sudden crises, improve user satisfaction, and deepen user loyalty. Secondary response dialogues are iterative responses generated through policy network optimization based on changes in user emotional feedback after primary response. The generation mechanism and input data of secondary response dialogues are different from those of primary response dialogues.

[0087] In one embodiment, step S703, which involves regenerating a secondary response script based on the change in the sentiment of the comment, includes:

[0088] S7031: The release operation of the first-level response dialogue, the user feedback data obtained after release monitoring, and the changes in the sentiment of the comments are stored as a set of sample data in the experience pool;

[0089] S7032: Train the dialogue generation policy network using the sample data in the experience pool;

[0090] S7033: Use the trained response generation policy network to generate the secondary response.

[0091] As described in step S7031 above, the publishing operation of the first-level response dialogue, the user feedback data monitored after publishing, and the changes in the sentiment of the comments are stored as a set of sample data in the experience pool. Collecting and integrating practical information related to the response dialogue allows for training and optimization when generating more accurate second-level response dialogues. The publishing operation of the first-level response dialogue includes the publishing time, platform information, and the dialogue content itself. User feedback data captures the reactions of users in the comment section to the dialogue, including the number of likes, the number of comments, and the users' sentiment (e.g., positive, negative, or neutral). Changes in the sentiment of the comments show the collective emotional response of users within a certain period after publishing, such as whether negative emotions have decreased or positive emotions have increased. Storing this information in the experience pool allows the system to quantify which responses are effective and which are not ideal, taking full account of the context, thus providing a real and accurate data foundation for subsequent model training.

[0092] As described in step S7032 above, the response dialogue generation strategy network is trained using sample data from the experience pool. This involves using deep learning to optimize generation efficiency and content relevance. The training aims to enable the network to autonomously learn how to generate more adaptive response dialogues, thereby improving overall response effectiveness. Specifically, the sample data from the experience pool serves as the training set, including published response dialogues, corresponding user feedback data, and information such as changes in sentiment. This sample data is input into a deep learning model, such as using a Long Short-Term Memory (LSTM) network or a Transformer model. These models are particularly suitable for handling text generation tasks. By continuously optimizing the model parameters during training, the model learns which factors positively influence the acceptability of dialogues and which feedback indicates the need for adjustment. In this way, when faced with similar crisis situations, the system can automatically generate targeted response dialogues that meet user expectations based on historical data and learned patterns.

[0093] As described in step S7033 above, the trained response dialogue generation strategy network is used to generate the secondary response dialogue. Specifically, the trained strategy network utilizes user feedback, comment sentiment data, and other contextual information as input to generate more detailed response dialogue. This dialogue will be more targeted and culturally adaptable. For example, the system will consider the user's emotional state and generate warm and sincere responses based on the results of sentiment analysis to alleviate user dissatisfaction. Furthermore, the generated dialogue must not only conform to the urgency level reflected by the crisis index but also consider corresponding cultural characteristics to better suit the cultural background and communication habits of the target audience. Through continuous training and updating of the strategy network, the company's response capabilities will be continuously enhanced, thereby maintaining good customer relationships and a positive corporate image in a dynamically changing environment.

[0094] In one embodiment, step S4, which involves constructing a corresponding temporal social propagation network model based on each of the comment event entity datasets, includes:

[0095] S401: Identify all users discussing the corresponding event in the comment event entity dataset and form a node set;

[0096] S402: Based on the social relationship and interaction behavior data in the comment event entity dataset, construct the network edges between user nodes in the node set;

[0097] S403: Extract the interaction behavior data of each network edge within a continuous time window, and generate the corresponding edge temporal feature vector;

[0098] S404: Input the node set, network edges and corresponding edge temporal feature vectors into the temporal graph convolutional network model to construct the temporal social propagation network model.

[0099] As described in step S401 above, all users in the comment event entity dataset corresponding to the event discussion are identified to form a node set. The content in the comment event entity is analyzed to identify users related to the event. This analysis includes not only author information in the comment text, but also attention to the interactive relationship between comments (such as comment reply chains) to ensure that all participants are captured. This is usually achieved by summarizing and organizing the comment data, for example by extracting user IDs or usernames. At the same time, the system can use the topology to determine which users have direct or indirect connections in the topic, ensuring that the node set is as comprehensive as possible and covers all key users.

[0100] As described in step S402 above, network edges are constructed between user nodes in the node set based on the social relationships and interaction behavior data in the comment event entity dataset. Network edges represent the connections between users, which can be direct interactions (such as replying, liking, or forwarding) or indirect connections (such as jointly participating in discussions). The system needs to analyze this interaction behavior data to identify which users follow each other or interact frequently, and create network edges accordingly. For example, if user A replies to or likes user B's comment, this can be considered as the construction of a social edge. Furthermore, the weight of the edge can be set according to the frequency or intensity of interaction; more frequent interactions indicate a closer connection between users. In this way, the social communication network constructed by the system will reflect social dynamics and information flow, further supporting information dissemination analysis and crisis prediction.

[0101] As described in step S403 above, the interactive behavior data of each network edge within a continuous time window is extracted to generate a corresponding edge temporal feature vector. A continuous time window is set (e.g., hourly, daily, etc.), and then the specific interactive data of each network edge (i.e., the interactive relationship between users) within this time period is analyzed. This interactive data includes, but is not limited to, the number of comments, likes, and replies within the time window. Through this behavioral data within a continuous time period, the system can construct the temporal feature vector of the edge. Each feature vector may contain indicators such as the frequency, duration, and activity level of behavior over a past period. This approach not only reflects changes in interactive behavior in social networks but also reveals the performance and influence of specific users during specific periods, thereby providing detailed temporal dimension data for subsequent propagation network modeling and sentiment analysis.

[0102] As described in step S404 above, the node set, network edges, and corresponding edge temporal feature vectors are input into the temporal graph convolutional network model to construct the temporal social propagation network model. The TGCN model combines the characteristics of graph data with the dynamic change capability of time series data, enabling it to process user nodes and their relationship information that changes over time. In this model, nodes represent users in the social network, network edges represent the interaction relationships between these users, and the input edge temporal feature vectors provide the model with dynamic information about user behavior within a specific time period, allowing the model to learn and capture the influence of user behavior on the spread of public sentiment. The completed temporal social propagation network model can reflect the patterns of information flow between users and their evolution over time, providing strong data support for subsequent crisis prediction, public opinion diffusion analysis, and the formulation of response strategies.

[0103] In one embodiment, step S5, which quantifies the influence score of each user node in the time-series social propagation network model and calculates the crisis index in combination with the corresponding structured event vector, includes:

[0104] S501: Calculate the influence weight of each user node using a preset graph node sorting algorithm;

[0105] S502: Perform sentiment analysis on each published content in the comment event entity dataset to obtain the corresponding sentiment intensity quantification value;

[0106] S503: For each piece of published content, the corresponding emotional intensity quantification value is weighted and calculated with the influence weight of the publishing user to obtain the influence-weighted emotional value of the published content;

[0107] S504: Aggregate the influence-weighted sentiment values ​​of all published content in the comment event entity dataset to generate the crisis index of the comment event entity.

[0108] As described in step S501 above, the influence weight of each user node is calculated using a preset graph node ranking algorithm. Specifically, the ranking algorithm can be PageRank, which analyzes the connections, interaction frequency, and strength of these relationships between each user node and other users. A user's influence weight depends not only on their direct interaction data (such as comments, reposts, and likes) but also on their position in the network—that is, the influence of other users around them. Through iterative calculations, the algorithm can automatically identify user nodes that play a key role in discussions of specific events and assign them corresponding influence scores. This quantitative influence assessment helps companies effectively identify key opinion leaders (KOLs) and high-influence users in public sentiment analysis, providing valuable insights for subsequent crisis management and public opinion guidance, and enabling more targeted response strategies. Specifically, user nodes and their social edges are constructed to form a social network graph. Each user node is given an initial weight, typically 1. The weight of each node is updated iteratively using the following formula until the weights of all user nodes converge (the convergence condition is that the change in weight is below a preset threshold). The weight value at this point is the user's influence weight.

[0109] As described in step S502 above, sentiment analysis is performed on each published content in the comment event entity dataset to obtain the corresponding sentiment intensity quantification value. Various sentiment analysis models and algorithms are employed, such as deep learning-based sentiment analyzers (e.g., LSTM, BERT), to analyze user comments, parse key sentiment words in the comments, and assign sentiment polarity scores to these words, thereby generating a sentiment intensity quantification value for each comment. For example, a positive comment may receive a high sentiment intensity score (e.g., 1.0), while a negative comment may receive a low score (e.g., -1.0 or 0). Furthermore, the system also incorporates contextual information, such as the degree of word modification, to further adjust the sentiment intensity value. The sentiment intensity quantification value is obtained by performing sentiment analysis on each comment, typically using a sentiment analysis model based on Natural Language Processing (NLP). Using a predefined dictionary, the sentiment scores of sentiment words in the text are accumulated or calculated. For example, one sentiment word may be labeled +1 (positive), and another may be -1 (negative). Sentiment intensity quantification values ​​are typically mapped to a specific range, usually [-1, 1] or [0, 1], depending on the method and model used. If relative sentiment scoring is used, some models may return values ​​greater than 1; therefore, normalization is generally performed in practice.

[0110] As described in step S503 above, for each piece of published content, its corresponding sentiment intensity quantification value is weighted by the influence weight of the publishing user to obtain the influence-weighted sentiment value of the published content.

[0111] In step S503, the system weights the quantified sentiment intensity value of each post derived from sentiment analysis with the influence weight of the comment author to obtain the influence-weighted sentiment value for that post. The formula for calculating the influence-weighted sentiment value is: Influence-weighted sentiment value = User influence weight × Sentiment intensity quantified value. This ensures that even users with high influence, whose comments have low sentiment intensity, can significantly impact the overall sentiment. In social media, some users may be considered more important due to their fan base or their influence on public opinion.

[0112] As described in step S504 above, the influence-weighted sentiment values ​​of all published content in the comment event entity dataset are aggregated to generate a crisis index for the comment event entity. The aggregation process sums the influence-weighted sentiment values ​​of each comment in the event to generate an overall sentiment score. This comprehensive score helps reveal whether the overall sentiment of the event is positive, negative, or neutral, and based on the sentiment score, the severity and speed of the existing crisis can be assessed. After generating the complete crisis index, enterprises can classify their responses according to the set warning level standards (such as red, orange, and yellow). This classification helps enterprises formulate corresponding crisis management strategies and take timely action. The Crisis Index (CI) is calculated as follows: CI = Σ(Wi*Si) / N, where Wi represents the influence weight of the user node corresponding to the i-th post (calculated using the PageRank algorithm, with a value range of [0,1]), Si represents the quantified sentiment intensity of the i-th post (derived through a sentiment analysis model, with a value range of [-1,1], where -1 represents extreme negativity and 1 represents extreme positivity), and N is the total number of posts in the dataset of the comment event. The larger the absolute value of CI, the higher the severity of the crisis.

[0113] In one embodiment, step S6, which involves obtaining the corresponding target cultural characteristics based on the comment event entity and dynamically generating and publishing a first-level response script based on the crisis index and the target cultural characteristics, includes:

[0114] S601: Extract the region identification information from the comment event entity dataset;

[0115] S602: Based on the regional identification information, obtain a set of cultural preference features associated with the regional identification information from a preset regional cultural feature database, and use it as the target cultural feature;

[0116] S603: Input the crisis index and the target cultural characteristics into the speech generation engine based on the large language model to generate the first-level response speech.

[0117] As described in step S601 above, the region identifier information is extracted from the comment event entity dataset. This region identifier information can be the user's geographical location, location information provided during user registration, or explicit location names (such as city, country, etc.) extracted from the comment content. Specifically, when analyzing the content of the comment data, text parsing techniques can be used to identify explicit geographical information. Furthermore, a geographic information database can be used, combined with the user's social relationships and location fields in their personal profile, to comprehensively determine the most accurate region identifier, thus accurately identifying the associated region for each comment event entity.

[0118] As described in step S602 above, based on the regional identification information, a set of cultural preference features associated with the regional identification information is obtained from a preset regional cultural feature database, serving as the target cultural feature. The preset regional cultural feature database is a knowledge base containing various cultural features, language norms, and local customs. The system queries this database and extracts a set of features related to the regional culture based on the regional identification information. Cultural preference rules may include typical etiquette, vocabulary habits, degree of formalization, and taboo words. By integrating these cultural features, the system can adopt more targeted and culturally adaptable response strategies when dealing with user feedback, effectively reducing the risk of misunderstandings or conflicts caused by cultural differences.

[0119] As described in step S603 above, the crisis index and the target cultural features are input into a dialogue generation engine built on a large language model to generate the first-level response dialogue. The core of the generation engine is an algorithm built on a large language model (such as GPT or other pre-trained transformation models), which can generate natural and fluent text based on contextual information. During the input phase, the system combines the crisis index (reflecting the severity and urgency of the event) with the acquired cultural features to form guidance information for dialogue generation. During the generation process, the dialogue generation engine automatically generates a first-level response dialogue with semantic coherence and cultural adaptability based on the input content. The dialogue includes feedback on the user's emotions, answers to questions, and suggestions for the next steps, ensuring good interaction with the user. Furthermore, the generated dialogue will adjust the tone, word choice, and content depth in a timely manner according to the crisis index to ensure a quick and effective response to user needs.

[0120] In one embodiment, step S2, which performs semantic parsing on each of the comment data to generate a structured event vector representing event elements, includes:

[0121] S201: Identify the text content in the comment data;

[0122] S202: A multilingual pre-trained model is used to perform unified semantic understanding on the text content to identify the core elements of the event in the text content; wherein, the core elements include the event subject, the event type, and the event object;

[0123] S203: Encode the core elements of the event into a low-dimensional dense vector to generate the structured event vector.

[0124] As described in step S201 above, the text content in the comment data is identified. The text content posted by the user is extracted. The comment data can be in various forms, including plain text comments and comments with images or videos. Specifically, data cleaning and preprocessing are first required to remove irrelevant information, such as advertisements and spam comments. Next, text analysis techniques are used to locate the text portions in the comment data. It should be noted that some comments may be phrases, sentences, or long paragraphs; the system must ensure it can effectively handle texts of different lengths. Through regular expressions, natural language processing (NLP) tools, or information extraction techniques (such as syntax parsing), the system will effectively identify various comment texts. The extracted text is then standardized, for example, through word segmentation, converting the text to lowercase, or removing punctuation.

[0125] As described in step S202 above, a multilingual pre-trained model is used to perform unified semantic understanding on the text content, identifying the core elements of the event within the text content. These core elements include the event subject, event type, and event object; that is, a suitable multilingual pre-trained model (such as BERT, XLM-RoBERTa, etc.) is selected. These models are pre-trained on large datasets and possess strong capabilities for understanding natural language. The model performs comprehensive semantic analysis on each piece of text content to find the core elements of the event. These core elements typically include the event subject (such as user, brand, product, etc.), event type (such as complaint, suggestion, praise, etc.), and event object (i.e., specific targets related to the event, such as product model, service items, etc.). During this process, the model automatically learns key grammatical structures and semantic meanings in the text, effectively identifying event information implicit in the text. Extracting relationships from the text not only helps understand the content of individual comments but also provides a data foundation for overall public sentiment trend analysis.

[0126] As described in step S203 above, the core elements of the event are encoded into low-dimensional dense vectors to generate the structured event vector. The encoding process first requires numerical processing of the core elements. The system typically uses vectorization techniques, such as word embeddings, to convert textual information into numerical vectors. Each core element of the event (subject, type, object) is mapped into a low-dimensional dense space to reflect its semantic relevance. This low-dimensional dense vector representation can effectively capture the similarity and relationships between elements. Furthermore, the generated structured event vector should have sufficient expressive power to reflect the uniqueness and characteristics of the event in subsequent clustering and analysis.

[0127] Reference Figure 3The present invention also provides a smart response device for commenting on crises, the device comprising:

[0128] The data acquisition module 902 is used to collect comment data from multiple users on the target social platform in real time; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationship and interaction behavior data between users;

[0129] The parsing module 904 is used to perform semantic parsing on each of the comment data to generate a structured event vector representing event elements;

[0130] The calculation module 906 is used to calculate and perform real-time clustering of all the comment data based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0131] The construction module 908 is used to construct a corresponding temporal social propagation network model based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges.

[0132] The quantization module 910 is used to quantify the influence scores of each user node in the time-series social propagation network model and calculate the crisis index in combination with the corresponding structured event vectors.

[0133] The generation module 912 is used to obtain the corresponding target cultural characteristics based on the comment event entity, and dynamically generate a first-level response script for release based on the crisis index and the target cultural characteristics.

[0134] In one embodiment, the intelligent response device for commenting on a crisis further includes:

[0135] The feedback data monitoring module is used to monitor user feedback data after the release of the first-level response script;

[0136] The sentiment tendency change analysis module is used to analyze changes in the sentiment tendency of comments based on the user feedback data.

[0137] The secondary response dialogue generation module is used to generate secondary response dialogues again based on the changes in the sentiment of the comments.

[0138] In one embodiment, the secondary response dialogue generation module includes:

[0139] The sample data storage submodule is used to store the release operation of the first-level response dialogue, the user feedback data obtained after release, and the changes in the sentiment of the comments as a set of sample data into the experience pool.

[0140] The dialogue generation policy network training submodule is used to train the dialogue generation policy network using sample data in the experience pool.

[0141] The secondary response dialogue generation submodule is used to generate the secondary response dialogue using the trained response dialogue generation strategy network.

[0142] In one embodiment, the building module 908 includes:

[0143] The user identification submodule is used to identify all users discussing the corresponding event in the comment event entity dataset, forming a node set;

[0144] The network edge construction submodule is used to construct network edges between user nodes in the node set based on the social relationship and interaction behavior data in the comment event entity dataset;

[0145] The interaction behavior data extraction submodule is used to extract the interaction behavior data of each network edge within a continuous time window and generate the corresponding edge temporal feature vector.

[0146] The input submodule is used to input the node set, network edges and corresponding edge temporal feature vectors into the temporal graph convolutional network model to construct the temporal social propagation network model.

[0147] In one embodiment, the quantization module 910 includes:

[0148] The influence weight calculation submodule is used to calculate the influence weight of each user node using a preset graph node sorting algorithm.

[0149] The sentiment analysis submodule is used to perform sentiment analysis on each published content in the comment event entity dataset to obtain the corresponding sentiment intensity quantification value.

[0150] The weighted calculation submodule is used to calculate the influence-weighted sentiment value of each published content by weighting the corresponding sentiment intensity quantification value with the influence weight of the publishing user.

[0151] The aggregation submodule is used to aggregate the influence-weighted sentiment values ​​of all published content in the comment event entity dataset to generate the crisis index of the comment event entity.

[0152] In one embodiment, the generation module 912 includes:

[0153] The region identifier information extraction submodule is used to extract region identifier information from the comment event entity dataset;

[0154] The cultural preference feature set acquisition submodule is used to acquire a set of cultural preference features associated with the regional identification information from a preset regional cultural feature database based on the regional identification information, and use it as the target cultural feature;

[0155] The target cultural feature input submodule is used to input the crisis index and the target cultural features into the speech generation engine based on a large language model to generate the first-level response speech.

[0156] In one embodiment, the parsing module 904 includes:

[0157] The text content recognition submodule is used to recognize the text content in the comment data;

[0158] The event core element identification submodule is used to perform unified semantic understanding on the text content using a multilingual pre-trained model to identify the event core elements in the text content; wherein, the core elements include the event subject, event type, and event object;

[0159] The event core element encoding submodule is used to encode the event core elements into low-dimensional dense vectors to generate the structured event vector.

[0160] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an intelligent response method for crisis management. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement an intelligent response method for crisis management. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0161] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0162] Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0163] Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements;

[0164] Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0165] A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0166] The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors.

[0167] Based on the commented event entity, the corresponding target cultural characteristics are obtained, and based on the crisis index and target cultural characteristics, a first-level response script is dynamically generated and released.

[0168] By performing deep semantic analysis and clustering on multi-source comment data, independent crisis event entities can be accurately and efficiently identified and aggregated from noisy information. By constructing a time-series social communication network model and quantifying the influence of user nodes, scientific prediction and assessment of crisis spread trends and intensity can be achieved, enabling the crisis index to more accurately reflect the severity of the event. Combining the cultural characteristics associated with the event with the aforementioned crisis index, appropriate first-level response dialogues can be dynamically generated, which helps improve the pertinence, timeliness, and cultural suitability of response strategies. This helps entities to more proactively and effectively manage and resolve online comment crises, achieving automation and intelligence in the main process from identification to response.

[0169] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0170] Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users;

[0171] Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements;

[0172] Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets;

[0173] A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges;

[0174] The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors.

[0175] Based on the commented event entity, the corresponding target cultural characteristics are obtained, and based on the crisis index and target cultural characteristics, a first-level response script is dynamically generated and released.

[0176] By performing deep semantic analysis and clustering on multi-source comment data, independent crisis event entities can be accurately and efficiently identified and aggregated from noisy information. By constructing a time-series social communication network model and quantifying the influence of user nodes, scientific prediction and assessment of crisis spread trends and intensity can be achieved, enabling the crisis index to more accurately reflect the severity of the event. Combining the cultural characteristics associated with the event with the aforementioned crisis index, appropriate first-level response dialogues can be dynamically generated, which helps improve the pertinence, timeliness, and cultural suitability of response strategies. This helps entities to more proactively and effectively manage and resolve online comment crises, achieving automation and intelligence in the main process from identification to response.

[0177] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0178] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0179] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A smart response method for commenting on crises, characterized in that, The method includes: Real-time collection of comment data from multiple users on the target social media platform; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users; Semantic parsing is performed on each of the aforementioned comment data to generate structured event vectors representing event elements; Calculate and cluster all the comment data in real time based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets; A corresponding temporal social propagation network model is constructed based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed with each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges; the temporal social propagation network model is used to learn the behavior and social interaction of users in a specific time period, predict the propagation trend of comment events, and the propagation trend provides a basis for the calculation of crisis index; The influence score of each user node in the time-series social propagation network model is quantified, and the crisis index is calculated by combining the corresponding structured event vectors. Based on the comment event entity, obtain the corresponding target cultural characteristics, and based on the crisis index and target cultural characteristics, dynamically generate and publish a first-level response script; The steps of quantifying the influence scores of each user node in the time-series social propagation network model and calculating the crisis index in combination with the corresponding structured event vectors include: The influence weight of each user node is calculated using a pre-defined graph node sorting algorithm. Sentiment analysis is performed on each published content in the aforementioned comment event entity dataset to obtain the corresponding sentiment intensity quantification value; For each piece of published content, the corresponding emotional intensity quantification value is weighted and calculated with the influence weight of the publishing user to obtain the influence-weighted emotional value of the published content. The influence-weighted sentiment values ​​of all published content in the comment event entity dataset are aggregated to generate the crisis index of the comment event entity.

2. The intelligent response method for comment crisis according to claim 1, characterized in that, After the steps of obtaining the corresponding target cultural characteristics based on the comment event entity, and dynamically generating and releasing a first-level response script based on the crisis index and the target cultural characteristics, the method further includes: Monitor user feedback data after the release of the first-level response script; Analysis of changes in comment sentiment based on the aforementioned user feedback data; Based on the changes in the sentiment of the comments, a secondary response script is generated again.

3. The intelligent response method for comment crisis according to claim 2, characterized in that, The step of generating a secondary response script based on the change in the sentiment of the comment includes: The release operation of the first-level response script, the user feedback data obtained after release, and the changes in the sentiment of the comments are stored as a set of sample data in the experience pool; The dialogue generation policy network is trained using the sample data in the experience pool; The trained response generation policy network is used to generate the secondary response.

4. The intelligent response method for comment crises according to claim 1, characterized in that, The step of constructing a corresponding temporal social propagation network model based on each of the comment event entity datasets includes: Identify all users who participated in the corresponding event discussion in the aforementioned comment event entity dataset, and form a node set; Based on the social relationships and interaction behavior data in the comment event entity dataset, network edges are constructed between user nodes in the node set. Extract the interaction behavior data of each network edge within a continuous time window to generate the corresponding edge temporal feature vector; The node set, network edges, and corresponding edge temporal feature vectors are input into the temporal graph convolutional network model to construct the temporal social propagation network model.

5. The intelligent response method for comment crises according to claim 1, characterized in that, The steps of obtaining the corresponding target cultural characteristics based on the comment event entity, and dynamically generating and releasing a first-level response script based on the crisis index and the target cultural characteristics, include: Extract the region identifier information from the comment event entity dataset; Based on the regional identification information, a set of cultural preference features associated with the regional identification information is obtained from a preset regional cultural feature database and used as the target cultural feature. The crisis index and the target cultural characteristics are input into a speech generation engine based on a large language model to generate the first-level response speech.

6. The intelligent response method for comment crisis according to claim 1, characterized in that, The step of semantically parsing each of the comment data to generate a structured event vector representing event elements includes: Identify the text content in the comment data; A multilingual pre-trained model is used to perform unified semantic understanding on the text content and identify the core elements of the event in the text content; wherein, the core elements include the event subject, the event type, and the event object; The core elements of the event are encoded into low-dimensional dense vectors to generate the structured event vector.

7. A smart response device for commenting on crises, characterized in that, The device includes: The data acquisition module is used to collect comment data from multiple users on the target social platform in real time; wherein, the comment data includes the content data posted by the corresponding user, as well as the social relationships and interaction behavior data between users; The parsing module is used to perform semantic parsing on each of the comment data to generate a structured event vector representing event elements; The calculation module is used to calculate and perform real-time clustering of all the comment data based on the semantic similarity between the structured event vectors to obtain multiple comment event entity datasets; The construction module is used to construct a corresponding temporal social propagation network model based on each of the comment event entity datasets; wherein, the social propagation network model is a temporal graph convolutional network model, which is constructed by taking each user in the comment event entity dataset as user nodes and the social interaction relationship between users as network edges; the temporal social propagation network model is used to learn the behavior and social interaction of users in a specific time period, predict the propagation trend of comment events, and the propagation trend provides a basis for the calculation of the crisis index; The quantization module is used to quantify the influence scores of each user node in the time-series social propagation network model and calculate the crisis index in combination with the corresponding structured event vectors. The generation module is used to obtain the corresponding target cultural characteristics based on the comment event entity, and dynamically generate a first-level response script for release based on the crisis index and the target cultural characteristics. The quantization module includes: The influence weight calculation submodule is used to calculate the influence weight of each user node using a preset graph node sorting algorithm. The sentiment analysis submodule is used to perform sentiment analysis on each published content in the comment event entity dataset to obtain the corresponding sentiment intensity quantification value. The weighted calculation submodule is used to calculate the influence-weighted sentiment value of each published content by weighting the corresponding sentiment intensity quantification value with the influence weight of the publishing user. The aggregation submodule is used to aggregate the influence-weighted sentiment values ​​of all published content in the comment event entity dataset to generate the crisis index of the comment event entity.

8. A computer-readable storage medium, characterized in that, The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the intelligent response method for a comment crisis as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the intelligent response method for commenting on a crisis as described in any one of claims 1 to 6.

Images