Method and device for automatically studying, judging and coping network public opinion events based on large model agent

Abstract

The invention provides an automatic research, judgment and coping method and device for a network public opinion event based on a large model agent. The method comprises the steps of obtaining public opinion data of the public opinion event at a current time node; performing situation awareness analysis on the public opinion data of the current time node to obtain situation data of the current time node; based on the situation data of the current time node, the basic information of the public opinion event and the element information of a plurality of historical key time nodes, adopting a pre-trained public opinion coping large model to generate a coping measure of the current time node; and automatically coping with the public opinion event based on the coping measure of the current time node. According to the public opinion research and judgment coping method and device, the coping measures of the public opinion event can be automatically generated and coping of the public opinion event can be automatically carried out through the public opinion large model, manual participation is not needed in the whole public opinion research and judgment coping process, automation is comprehensively achieved through the large model agent, and therefore the public opinion research and judgment coping accuracy and response speed are greatly improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a method and device for automatically analyzing and responding to network public opinion events based on a large-scale intelligent model. Background Art

[0002] With the development of the Internet, the impact of online public opinion events on society has become increasingly prominent, and they are increasingly attracting widespread attention from party and government agencies, enterprises, and institutions. Although public opinion event monitoring and analysis technology has made great progress after years of development, providing increasingly important technical support for the analysis and response of public opinion events, the current technical means of public opinion event analysis and response are mainly concentrated in the public opinion event analysis link, while the public opinion event decision-making and response links are mainly manually decided and implemented by user leaders and professionals. Faced with the ever-changing online public opinion field, the existing public opinion analysis and response technology appears to have a slow response speed. Especially when faced with sudden public opinion events, after completing manual analysis and decision-making under the existing public opinion analysis and response model, the best time to respond to public opinion events has often been missed. Therefore, how to improve the degree of automation and response speed of public opinion event response has become a technical problem that technicians in this field urgently need to solve. Summary of the Invention

[0003] In view of this, the present disclosure proposes a method and device for automatically analyzing and responding to network public opinion events based on a large model intelligent body, which can greatly improve the degree of automation and response speed of public opinion response.

[0004] According to a first aspect of the present disclosure, a method for automatically analyzing and responding to online public opinion events based on a large model agent is provided, comprising:

[0005] Obtain public opinion data of public opinion events at the current time node;

[0006] Performing situational awareness analysis on the public opinion data at the current time node to obtain situational data at the current time node;

[0007] Based on the situation data of the current time node, the basic information of the public opinion event, and the element information of multiple historical key time nodes, a pre-trained public opinion response model is used to generate response measures for the current time node;

[0008] Based on the response measures at the current time node, the public opinion event is automatically responded to.

[0009] In one possible implementation, based on the situation data at the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes, a pre-trained public opinion response model is used to generate the response measures for the current time node, including:

[0010] Constructing decision-making time series data based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes;

[0011] The decision-making basis time series data is input into the public opinion response model to generate the response measures and expected response effects of the current time node, wherein the response measures include response measure types and response measure parameters.

[0012] In one possible implementation, when constructing decision-making time series data based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes, the following steps are included:

[0013] Combine the element information of multiple historical key time nodes in chronological order to obtain the historical key node sequence data at the current time node;

[0014] The basic information of the public opinion event, the historical key node sequence data at the current time node and the situation data at the current time node are combined in sequence to construct the decision-making basis time series data.

[0015] In one possible implementation, the public opinion response model is trained based on basic information of historical public opinion events and element information of key time nodes.

[0016] In a possible implementation, after performing situational awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node, the method further includes:

[0017] Determining whether it is necessary to respond to the public opinion event at the current time node based on the situation data at the current time node;

[0018] When it is determined that a response is needed to the public opinion event at the current time node, an operation is executed based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes, using a pre-trained public opinion response model to generate response measures for the current time node.

[0019] In a possible implementation, automatically responding to the public opinion event based on the response measures at the current time node includes:

[0020] Generate an action task list based on the response measures at the current time node;

[0021] Based on the action task list, the public opinion event is automatically responded to.

[0022] In a possible implementation, when generating an action task list based on the response measures at the current time node, the following steps are included:

[0023] Analyzing the countermeasures at the current time node to obtain the countermeasure types and countermeasure parameters of the countermeasures;

[0024] An action task template matching the category is obtained, and the response measure parameters are filled into the action task template to obtain an action task list of the response measure.

[0025] According to a second aspect of the present disclosure, there is provided an automatic analysis and response device for public opinion events, comprising:

[0026] The public opinion data acquisition module is used to obtain the public opinion data of the public opinion event at the current time node;

[0027] A public opinion situation awareness module is used to perform situation awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node;

[0028] A decision-making and planning module is used to generate response measures for the current time node using a pre-trained public opinion response model based on the situation data of the current time node, the basic information of the public opinion event, and the element information of multiple historical key time nodes;

[0029] The action response module automatically responds to the public opinion event based on the response measures at the current time node.

[0030] According to a third aspect of the present disclosure, there is provided an automatic analysis and response device for public opinion events, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute the method described in the first aspect of the present disclosure.

[0031] According to a fourth aspect of the present disclosure, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, wherein the computer program instructions, when executed by a processor, implement the method described in the first aspect of the present disclosure.

[0032] The present disclosure provides a method, device, equipment and storage medium for automatically analyzing and responding to network public opinion events based on a large-scale intelligent body. The method includes: obtaining public opinion data of the public opinion event at the current time node; performing situational awareness analysis on the public opinion data of the current time node to obtain situational data of the current time node; based on the situational data of the current time node and the basic information of the public opinion event and the element information of multiple historical key time nodes, using a pre-trained public opinion response large model to generate response measures for the current time node; based on the response measures of the current time node, automatically responding to the public opinion event. The public opinion large model in the present disclosure can automatically generate response measures for public opinion events, and automatically respond to public opinion events based on the generated response measures. The entire public opinion response process does not require human participation, and is fully automated through the large-scale intelligent body, thereby greatly improving the response speed of public opinion response.

[0033] Further features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.

[0035] Figure 1 A flowchart showing a method for automatically analyzing and responding to network public opinion events based on a large model agent according to an embodiment of the present disclosure is shown;

[0036] Figure 2 A flowchart showing an example of a method for automatically analyzing and responding to network public opinion events based on a large model agent according to an embodiment of the present disclosure;

[0037] Figure 3 A schematic block diagram of an apparatus for automatically analyzing and responding to network public opinion events based on a large model intelligent agent according to an embodiment of the present disclosure is shown;

[0038] Figure 4 A schematic block diagram of an automatic analysis and response device for network public opinion events based on a large model intelligent agent according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0039] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0040] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0041] In addition, numerous specific details are provided in the following detailed description to better illustrate the present disclosure. Those skilled in the art will appreciate that the present disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art are not described in detail in order to highlight the main points of the present disclosure.

[0042] <Method Example>

[0043] Figure 1 A flowchart of a method for automatically analyzing and responding to network public opinion events based on a large model agent according to an embodiment of the present disclosure is shown. The method is implemented by an automatic analysis and response system for public opinion events, such as Figure 1 As shown, the method includes steps S1100-S1400.

[0044] S1100, obtaining public opinion data of a public opinion event at the current time point. Public opinion data refers to a collection of information related to the current online public opinion event collected from various public opinion fields on the Internet (such as Weibo, news websites, short video platforms, etc.). The public opinion data may include: at least one of: text data, multimedia data, user interaction data, metadata, and event-related data. The text data may be, for example, news articles, Weibo posts, user comments in forums, etc.; multimedia data may be pictures, short video clips, live broadcast screen recordings, audio content, etc.; user interaction data may be the number of Weibo reposts, the number of likes, the number of followers of the user account, the region, etc.; metadata may include when the public opinion data was published, on which platform it was published, what the corresponding web link is, and the forwarding path of this public opinion data; event-related data: such as keywords related to the current public opinion event (such as "quality of a certain brand's products"), hot topic tags (such as hot search topics on Weibo), etc.

[0045] Before acquiring public opinion data, you need to set the parameters for acquiring public opinion data. These parameters can include at least one of the following: data source, data type, and update frequency. The data source specifies the platform for collecting public opinion data. The data type specifies the type of public opinion data to be collected, which can be at least one of text, images, and videos. The update frequency specifies the frequency of public opinion data collection.

[0046] After setting the parameters for obtaining public opinion data, you can obtain public opinion data according to the pre-set parameters for obtaining public opinion data. Specifically, according to the set data source, call the corresponding tool to crawl public opinion data. For example, if the data source is a platform with an open API (such as Weibo), then the public opinion data can be directly obtained through the API interface opened by the platform. For another example, if the data source is a website without an API (such as some forums), then the behavior of real people browsing the web can be simulated to "scrape" the public opinion data on the page. Among them, when crawling public opinion data, the public opinion data will be crawled according to the specified frequency and data type.

[0047] In one possible implementation, the parameters for acquiring public opinion data can be adaptively adjusted based on the analysis results of the public opinion response model. Specifically, when the analysis results of the public opinion response model are that the public opinion event needs to be focused on or responded to, it will generate: expanding the data source, increasing the collection of related data sources such as accounts associated with the public opinion event, linked web pages, etc.; expanding the data collection type, and comprehensively collecting multiple types of data such as text, pictures and videos associated with the public opinion event; increasing the update frequency of data such as comments and likes on the public opinion event. At least one of the first public opinion data acquisition parameter adjustment strategies will be generated. When the analysis results of the public opinion response model are that the public opinion event no longer needs to be focused on or the response action has ended, it will generate: restoring the data source, data type and update frequency to the system default values. In this way, the adaptive adjustment of the public opinion data acquisition parameters can be achieved by executing the first public opinion data acquisition parameter adjustment strategy or the second public opinion data acquisition parameter adjustment strategy.

[0048] After obtaining the public opinion data at the current time node, step S1200 can be executed to perform situational awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node. The situation data at the current time node includes basic situation information of the public opinion event at the current time node, the index value of each situation indicator in the pre-established situation indicator system, the index value of the comprehensive situation indicator that can comprehensively reflect the urgency of the public opinion event at the current time node, and the development trend prediction results of each situation indicator and the comprehensive situation indicator.

[0049] When performing situational awareness analysis on public opinion data at the current time node to obtain situational data at the current time node, the following steps may be included:

[0050] First, extract and quantitatively analyze public opinion data at the current time point to obtain basic situational information and the values ​​of various situational indicators for the public opinion event at the current time point. This basic situational information may include at least one of the causes of public opinion changes, content evolution, changes in popularity, changes in sentiment, dissemination paths, and user group characteristics. The pre-established situational indicator system may include at least one of the following situational indicators: spread, popularity, and sensitivity.

[0051] Second, based on the values ​​of each situation indicator at the current time node, the value of the comprehensive situation indicator at the current time node is calculated. Specifically, the system pre-sets a corresponding weight coefficient for each situation indicator. After calculating the value of each situation indicator at the current time node, the value of each situation indicator is weighted and summed according to the weight coefficient of each situation indicator. The summed result is the value of the comprehensive situation indicator at the current time node.

[0052] In one possible implementation, the weight coefficient corresponding to each situation indicator in the system can be dynamically adjusted and updated through linear regression or reinforcement learning based on continuously accumulated data such as manual feedback, response plans, and response effect evaluations.

[0053] Third, the development trends of each situation indicator and the comprehensive situation indicator at the current time node are predicted to obtain the development trend prediction results of each situation indicator and the comprehensive situation indicator. Specifically, for each situation indicator that changes over time, the indicator values ​​of the situation indicator at the current time node and the previous time nodes are sorted and combined in chronological order to obtain the time series of each situation indicator; then, a fitting algorithm is used to perform data fitting on the time series of each situation indicator to obtain the prediction results of the development trend of each situation indicator. Similarly, the indicator values ​​of the comprehensive situation indicator at the current time node and the previous time nodes are sorted and combined in chronological order to obtain the time series of the comprehensive situation indicator; and a fitting algorithm is used to perform data fitting on the time series of the comprehensive situation indicator to obtain the prediction results of the development trend of the comprehensive situation indicator. The fitting algorithm can be implemented using algorithms such as the least squares method, a time series model, and a neural network model, which are not specifically limited here.

[0054] The basic situation information of the public opinion event at the current time node, the index values ​​of each situation indicator, the index value of the comprehensive situation indicator, and the development trend prediction results of each situation indicator and the comprehensive situation indicator together constitute the situation data of the public opinion event at the current time node.

[0055] In one possible implementation, after obtaining the situation data at the current time node, the following step is further included: determining whether it is necessary to respond to the public opinion event at the current time node based on the situation data at the current time node. If it is determined that it is necessary to respond to the public opinion event at the current time node, step S1300 is then executed.

[0056] In this embodiment, when judging whether it is necessary to respond to the public opinion event at the current time node based on the situation data at the current time node, it can be achieved based on a pre-set response threshold. Among them, the response threshold may include at least one of the response threshold corresponding to each situation indicator and the response threshold corresponding to the comprehensive situation indicator. Specifically, when the index value of any situation indicator exceeds the corresponding response threshold, or the index value of the comprehensive situation indicator exceeds the corresponding response threshold, it is judged that it is necessary to respond to the public opinion event at the current time node; otherwise, it is not necessary to respond to the public opinion event at the current time node.

[0057] Furthermore, when it is determined that a response is needed to the public opinion event at the current time node, the current time node is identified as a key time node. For the identified key time node, not only the situation data corresponding to the key time node will be recorded, but also the response measures and response effects taken at the key time node will be recorded after the response to the key time node is completed, and the situation data, response measures and response effects of the key time node will be associated with the basic information of the public opinion event as the element information of the key time node and stored in the short-term memory module of the system. That is, during the entire process of responding to the public opinion event, the system will continuously identify key time points, and associate the basic information of the public opinion event and the element information of each key time node into the short-term memory module, so that the response measures for the current time node can be generated based on the data stored in the short-term memory module and combined with the situation data of the current time node. In a specific example, for a public opinion event that is currently occurring, the short-term memory module records the basic information of the public opinion event, the situation data, response measures and response effects of key time node 1, the situation data, response measures and response effects of key time node 2,..., the situation data, response measures and response effects of key time node i, where key time node i is the previous key time node at the current moment.

[0058] Among them, the basic information of the public opinion event may include the type of public opinion event, content label, time and location and other basic information.

[0059] The response measures may include the response measures type and specific response measures parameters. The response measures type may include at least one of the following types: publishing promotional materials (product promotional soft articles, statements about public opinion events, product marketing copy, etc.), publishing online comments and replies, generating public opinion response command and coordination instructions, generating public opinion reports, etc. The response measures parameters are parameters that indicate the specific response content. For example, when the response measure type is publishing online comments and replies, the corresponding response measures parameters may include the specific website platform, link address, comment content, number of comments, comment account, etc.

[0060] The expected response effect can be characterized by at least one response effect indicator among the change rate of the comprehensive situation indicator, the change rate of the heat, and the degree of emotional improvement. The calculation formula for each response effect indicator can be as follows:

[0061] Change rate of comprehensive situation index = (index value of current comprehensive situation index - index value of comprehensive situation index before execution) / index value of comprehensive situation index before execution;

[0062] Heat change rate = (current heat - heat before execution) / heat before execution;

[0063] Emotional improvement = the increase in the proportion of positive emotions.

[0064] S1300, based on the situation data of the current time node, the basic information of the public opinion event and the element information of multiple historical key time nodes, a pre-trained public opinion response model is used to generate response measures for the current time node.

[0065] Before executing the method disclosed in this disclosure, it is necessary to first train a public opinion response model based on the basic information of each historical public opinion event and the element information of each key time node. The specific training steps are as follows:

[0066] First, obtain the basic information of each historical public opinion event and the element information of each key time node. Specifically, in the process of responding to public opinion events, the system will store the basic information of each historical public opinion event that has been responded to and the element information of each key time node through the long-term memory module. In this way, the basic information of each historical public opinion event and the element information of each key time node can be obtained from the long-term memory module. Among them, the element information of each key time node is the same as above and will not be repeated here.

[0067] Second, based on the basic information of each historical public opinion event and the element information of each key time node, the decision-making basis time series data of each historical public opinion event is constructed. Specifically, for each historical public opinion event, the element information of each key time node of the historical public opinion event is combined in chronological order to obtain the historical key node sequence data; the basic information of the historical public opinion event and the historical key node sequence data are combined in sequence to construct the decision-making basis time series data of each historical public opinion event.

[0068] For example, for public opinion event A, its basic information is B, and the element information of the five key time nodes are C, D, E, F, and G respectively. The historical key node sequence data composed of the element information of the five key time nodes is {C, D, E, F, G}. By combining the basic information B and the historical key node sequence data {C, D, E, F, G} in sequence, the decision-making basis time series data of historical public opinion event A is {B, C, D, E, F, G}.

[0069] Third, we construct training data for the large language model based on the time series data of the decision-making basis of each historical public opinion event. Specifically, we vectorize the time series data of the decision-making basis of each historical public opinion event and perform high-order feature extraction to obtain the high-order feature sequence of the decision-making basis of each historical public opinion event. The high-order feature sequence of the decision-making basis of all historical public opinion events constitutes the training data of the large language model.

[0070] When vectorizing the decision-making basis time series data of historical public opinion events, each data in the decision-making basis time series data of historical public opinion events is sequentially mapped to the corresponding vector representation. Continuing with the above example, the decision-making basis time series data S = {B, C, D, E, F, G} of historical public opinion event A, then vectorizing S means sequentially mapping the data B, C, D, E, F, G in S to the corresponding vector representation X. B ,X c ,X D ,X E ,X F ,X G The process of X B ,X c ,X D ,X E ,X F ,X G By combining them in order, we can get the vector representation sequence X of S. The calculation formula of the vector representation sequence X is as follows:

[0071] X=f(s) s∈S ={X B ,X c ,X D ,X E,X F ,X G}

[0072] Where f(s) is a vectorized function used to map each data s in the decision-making time series data S into a corresponding vector representation.

[0073] It should be noted here that when the data s in the decision-making basis time series data S is the indicator data, the indicator value of the indicator data can be directly used as the vector representation of the indicator data.

[0074] After vectorizing the decision-making basis time series data S of historical public opinion events to obtain the corresponding vector representation sequence X, high-order features will be extracted from the vector representation sequence X to obtain the high-order feature sequence Z of the decision-making basis of historical public opinion events.

[0075] In one possible implementation, high-order feature extraction is performed on the vector representation sequence X based on the Transformer encoder. Specifically, the vector representation of each sequence position in the vector representation sequence X is concatenated with the corresponding position information through the Transformer encoder and then fused and output to obtain a high-order vector representation of each sequence position. In this way, the high-order vector representation of each sequence position retains both the semantic information of the vector representation and the temporal information of the corresponding position information. The high-order vector representations of all sequence positions are sequentially combined to obtain a high-order feature sequence Z as the basis for decision-making, thereby realizing the mapping from the input space to the model representation space. The calculation formula of the high-order feature sequence Z is as follows:

[0076] Z=TransformerEncoder(X+PE)={z1,z2,...,z T}

[0077] Where PE is the position encoding matrix, which includes the position information of each sequence position. The position information of each sequence position is obtained by position encoding the temporal information of its position using sine-cosine. The absolute temporal relationship is retained in this position information. Z is the high-order feature sequence output by the encoder as the decision basis, which is a comprehensive representation of event semantic information and spatiotemporal features.

[0078] It should be noted here that the Transformer encoder captures the long-term dependencies between historical nodes through the self-attention mechanism. The formula is as follows:

[0079]

[0080] Where Q, K, V are the query, key, and value matrices of the input vector representation sequence X, d k is the scaling factor.

[0081] Fourth, the large language model can be trained using training data to obtain a large public opinion response model. Specifically, the large language model is trained using training data from each historical public opinion event. This allows the large language model to plan response measures for the current time node and predict the effectiveness of the response based on the basic information of the public opinion event, the element information of each key time node that has occurred, and the situation data at the current time node. A large language model with such capabilities is the large public opinion response model in this disclosure.

[0082] In one possible implementation, to improve the accuracy of large language model training, a multi-task loss function is designed. This multi-task loss function is derived by weighted summing the losses of the type of public opinion response measure, response measure parameters, and expected response effect. The type of response measure is a typical classification problem, and the loss function is designed based on the cross-entropy loss function, measuring the difference between the probability distribution of the model output and the probability distribution of the true classification label. The expected response effect is a typical regression problem, and the loss function is designed based on the mean squared error loss function, measuring the average of the squared differences between the model's predicted value and the true value.

[0083] In one possible implementation, the multi-task loss function The calculation formula is as follows:

[0084]

[0085] Where, To deal with the cross entropy loss of the action type, To cope with the loss of measure parameters, is the mean square error loss of the expected response effect, λ1, λ2, λ3 are task weight coefficients, and their sizes are set according to specific needs. Among them, the cross entropy loss of the response type The calculation formula is as follows:

[0086]

[0087] Where y i is the true label of the response type of sample i, p i is the predicted probability.

[0088] Countermeasures parameter loss The calculation formula is as follows:

[0089]

[0090] Where t is the time step, y t is the target word at time step t, y <tare all target words before time step t, H is the model hidden state (including context information), P(y t │y <t ,H) indicates that the model is based on the historical sequence y <t Predict the probability of the current word.

[0091] Mean squared error loss for expected response effect prediction The calculation formula is as follows:

[0092]

[0093] Where S pred,j is the situation index predicted by the model, S true,j is the true value.

[0094] In one possible implementation, during the training and optimization of large language models, the optimizer uses the AdamW optimizer to update model parameters, decoupling weight decay from gradient updates to improve generalization capabilities. Layer-by-layer learning rates are set, with smaller learning rates (such as 1e-5) used for the encoder's bottom-layer parameters and larger learning rates (such as 1e-3) for the top and output layers. The cosine annealing scheduler is used to dynamically adjust the learning rate to accelerate convergence.

[0095] After training the public opinion response model, you can use it to generate response measures for the current time point. This can include the following steps:

[0096] First, based on the situation data at the current time node, the basic information of the public opinion event, and the element information of historical key time nodes, the decision-making basis time series data is constructed. Specifically, the basic information of the public opinion event and the element information of historical key time nodes are read from the short-term memory module; the element information of multiple historical key time nodes is combined in chronological order to obtain the historical key node sequence data at the current time node; the basic information of the public opinion event, the historical key node sequence data at the current time node, and the situation data at the current time node are sequentially combined to construct the decision-making basis time series data.

[0097] For example, for public opinion event A, its basic information is B, the element information of the five key time nodes that have occurred are C, D, E, F, and G, the situation data at the current time node is H, and the historical key node sequence data at the current time node composed of the element information of the five key time nodes is {C, D, E, F, G}. The basic information B and the historical key node sequence data {C, D, E, F, G} at the current time node and the situation data H at the current time node are combined in sequence to obtain the decision-making basis time series data {B, C, D, E, F, G, H} at the current time node.

[0098] Second, the decision-making time series data is vectorized and high-level features are extracted before being fed into a pre-trained public opinion response model. This model, based on the context vector, generates a structured response measure and expected response effect at the current time point through a multi-task decoding mechanism. The response measure can be one or more.

[0099] In one possible implementation, the public opinion response model outputs the response measures of the high-order feature sequence through a pre-built first Softmax function. The calculation formula of the first Softmax function is as follows:

[0100] y measures =Softmax(W a ·Z)

[0101] Where y measures is the corresponding countermeasure for the high-order feature sequence, W a is the classification weight matrix of various response measures, |A| is the number of action categories (such as "comment guidance", "offline processing", etc.), Z is the high-order feature sequence output by TransformerEncoder, n is the length of the feature sequence Z, d model is the characteristic dimension of the large model.

[0102] The first Softmax function above maps the high-order feature sequence output by the Transformer Encoder into a probability distribution for each response type, indicating the recommendation probability for each response type. For example, the model might output a probability of 0.7 for "comment guidance" and a probability of 0.3 for "offline action," ultimately recommending the top K responses with the highest probabilities.

[0103] In one possible implementation, the method for generating the countermeasure parameters is as follows:

[0104] The response parameters (platform, link, content) in this example are a piece of text generated by autoregressive decoding. The Softmax function is used to project the decoder output into a probability distribution, which represents the probability of choosing the word at the current step. After being processed by n layers of Transformer blocks, the model uses the final hidden state h n To predict the next word u i The probability distribution of .

[0105] The probability distribution function designed in this example is to combine the final hidden state with another weight matrix W n Multiply them together, then use the Softmax function to get the probability distribution. Finally, use the argMax function to get the word with the highest probability as the prediction result of the current word.

[0106] u i =argMax(Softmax(h n W n ))

[0107] Here u i Represents the parameter of the response measure after a given context, the i-th word, h n is the hidden state of the model, W n is the weight matrix.

[0108] In one possible implementation, the method for generating the expected response effect prediction is as follows:

[0109] Through joint probability modeling, predict the action y measures For example, the model might predict that "comment guidance" will cause a decrease in popularity by ΔH = -10% and an increase in sentiment by ΔS = +0.2.

[0110] y efffect =W e Concat(Z,y measures )+b e

[0111] Input: high-level features Z output by the encoder and recommended countermeasures y measures .

[0112] Output: Z and y measures The features of the node are spliced ​​or weighted fused; the mean of the effect is predicted by the multi-layer perceptron, where Z is the high-order feature of the current node, y measures is the embedding vector of the classification result, y efffect is the predicted mean response effect.

[0113] In one possible implementation, each response measure may include various response measures of varying degrees, such as radical, moderate, and conservative, and the user may select one based on specific needs. In this implementation, the response measure selected by the user is automatically executed as the final response measure.

[0114] In order to improve the accuracy of the output response measures of the public opinion response model, during the public opinion response process, new training data will be regularly constructed based on the data of new public opinion events that have been recently responded to, and the public opinion response model will be continuously optimized and trained based on the new training data.

[0115] Based on supervised pre-training, reinforcement learning fine-tuning is performed to maximize the expected cumulative effect of the generated actions (i.e., responses) for a given public opinion event. This example uses an algorithm-based optimization model based on PPO (Proximal Policy Optimization) to maximize the long-term rewards of the generated responses. The model first uses random trial and error (exploration), recording which responses have high scores, gradually reducing low-scoring actions and increasing high-scoring actions. Rewards are given based on the degree of match with the true answer, learning to select the optimal response in a complex public opinion environment.

[0116] Through the model fine-tuned by reinforcement learning, the reinforcement learning fine-tuning of the large model transforms abstract human preferences into the driving force for updating model parameters through the closed loop of "human feedback → reward model → strategy optimization", and can automatically select the optimal measures (with the best effect and lowest cost) based on the real-time public opinion status.

[0117] S1400: Automatically respond to the public opinion event based on the response measures at the current time node. This may include the following steps:

[0118] First, based on the response measures at the current time node, an action task list is generated. Specifically, the response measures at the current time node are parsed to obtain the response type and response parameters. An action task template that matches the type is obtained, and the response parameter is filled into the action task template to obtain the action task list.

[0119] It should be noted here that the system has pre-built a matching action task template for each type of response measure, and established an association relationship between the type and the corresponding action task template. In this way, after parsing the type from the response measures of the current time node, the action task template that matches the type can be obtained based on the pre-built association relationship, and then the parsed response measure parameters can be filled into the corresponding position of the action task template to obtain an action task list for the response measure.

[0120] For example, if the response type at the current moment is posting a comment on an online post, the action task list generated based on the response parameters is as follows:

[0121] {

[0122] "Publishing platform": "Sina News", "Sina Weibo";

[0123] "Original post title": "Clarification statement regarding incident X";

[0124] "Review content": "XXXXXXXXX";

[0125] "Execution priority": "P0";

[0126] "Expected effect": "Positive sentiment increased by 20%, and popularity decreased by 35% within 3 hours";

[0127] }

[0128] Second, based on the action list, public opinion events are automatically responded to. Specifically, specific execution tools are configured for each type of response measure. This way, after the action list is generated, the corresponding execution tool can be automatically called up to execute each action in the action list, thereby automatically responding to public opinion events.

[0129] When response measures include publishing promotional materials, posting online comments, generating public opinion response command and coordination instructions, and generating public opinion reports: For publishing promotional materials, configure an automatic tool for publishing promotional materials; for posting online comments, configure an automatic tool for posting online comments; for generating public opinion reports, configure an automatic tool for generating public opinion reports; and for generating public opinion response command and coordination instructions, configure the internet management department's command and coordination tool. This allows you to automatically execute various action lists by calling the corresponding execution tools.

[0130] Furthermore, the public opinion response model can not only predict the response measures at the current time node, but also predict the expected effects of the response measures. In this way, after the above-mentioned response measures are automatically executed, the execution effect of the response measures can be evaluated to obtain the actual execution effect of the response measures. By comparing the actual execution effect with the expected response effect, the effectiveness of the response measures can be evaluated, and then the model parameters of the public opinion response model can be continuously optimized based on the effectiveness of the response measures, thereby continuously improving the effectiveness of the response measures generated by the public opinion response model.

[0131] In order to clearly illustrate the method of the present disclosure, the following Figure 2The method disclosed herein is further illustrated by an example. Specifically, the method for automatically analyzing and responding to public opinion events includes the following steps:

[0132] Step 1: Public opinion data acquisition: The intelligent agent uses the public opinion data acquisition module to continuously obtain public opinion information from the Internet and input it into the public opinion situation awareness module of the system.

[0133] Step 2: Public opinion situation perception: The intelligent agent public opinion situation perception module analyzes and processes the network public opinion data input into the system and conducts situation assessment.

[0134] Step 3: Planning and decision-making: The intelligent agent's planning and decision-making module uses the reasoning and planning capabilities of the public opinion response model based on the public opinion situation perception assessment results to determine whether public opinion response measures need to be taken. If public opinion response is required, the planning and decision-making module calls the public opinion response model to generate a network public opinion response decision plan, and automatically decomposes the decision plan to form a specific action task list.

[0135] Step 4: Response action: The response action module of the intelligent agent actively discovers and calls appropriate external tools and corresponding resources to complete the response action task based on the public opinion response decision plan and action task list.

[0136] Return to step 1: continuously obtain public opinion data from the Internet, perceive changes in public opinion trends and evaluate the effectiveness of response actions, continuously optimize planning decisions and response actions, and ultimately achieve the goal of public opinion response.

[0137] From this example, we can see that the entire public opinion response process is entirely Figure 2 The implementation of the large-scale intelligent agent in the system means that all the above steps can be realized through a large-scale intelligent agent, and the analysis and response of the entire public opinion event can be completed automatically without human participation, thereby greatly improving the accuracy and response speed of public opinion analysis and response.

[0138] The present disclosure provides a method for automatically analyzing and responding to online public opinion events based on a large-scale intelligent agent, including: obtaining public opinion data of the public opinion event at the current time node; performing situational awareness analysis on the public opinion data of the current time node to obtain situational data of the current time node; using a pre-trained public opinion response large-scale model to generate response measures for the current time node based on the situational data of the current time node and the basic information of the public opinion event and the element information of multiple historical key time nodes; and automatically responding to the public opinion event based on the response measures of the current time node. The public opinion large-scale model in the present disclosure can automatically generate response measures for public opinion events, and automatically respond to public opinion events based on the generated response measures. The entire public opinion response process does not require human participation, and is fully automated through the large-scale intelligent agent, thereby greatly improving the accuracy and response speed of public opinion response.

[0139] <Device Example>

[0140] Figure 3 The following is a schematic block diagram of an automatic analysis and response device for network public opinion events based on a large model intelligent agent according to an embodiment of the present disclosure. Figure 3 As shown, the device 100 includes:

[0141] The public opinion data acquisition module 110 is used to obtain the public opinion data of the public opinion event at the current time node;

[0142] The public opinion situation awareness module 120 is used to perform situation awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node;

[0143] The decision-making planning module 130 is used to generate response measures for the current time node using a pre-trained public opinion response model based on the situation data of the current time node, basic information of public opinion events, and factor information of multiple historical key time nodes;

[0144] The action response module 140 automatically responds to public opinion events based on the response measures at the current time node.

[0145] <Equipment Example>

[0146] Figure 4 The following is a schematic block diagram of an automatic analysis and response device for network public opinion events based on a large model intelligent agent according to an embodiment of the present disclosure. Figure 4 As shown, the automatic analysis and response device 200 for public opinion events includes: a processor 210 and a memory 220 for storing executable instructions of the processor 210. The processor 210 is configured to implement any of the above-mentioned automatic analysis and response methods for public opinion events when executing the executable instructions.

[0147] It should be noted that there can be one or more processors 210. Furthermore, the automatic analysis and response device 200 for public opinion events according to the embodiment of the present disclosure may also include an input device 230 and an output device 240. The processor 210, memory 220, input device 230, and output device 240 may be connected via a bus or other means, which are not specifically limited herein.

[0148] Memory 220, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and various modules, such as the programs or modules corresponding to the method for automatically analyzing and responding to public opinion events in the embodiments of the present disclosure. Processor 210 executes the software programs or modules stored in memory 220 to perform various functional applications and data processing of the device 200 for automatically analyzing and responding to public opinion events.

[0149] The input device 230 may be used to receive input numbers or signals. The signals may be key signals related to user settings and function control of the device / terminal / server. The output device 240 may include a display device such as a display screen.

[0150] <Storage Medium Embodiment>

[0151] According to the fourth aspect of the present disclosure, a non-volatile computer-readable storage medium is also provided, on which computer program instructions are stored. When the computer program instructions are executed by the processor 210, any of the above-mentioned methods for automatically analyzing and responding to public opinion events is implemented.

[0152] While various embodiments of the present disclosure have been described above, the foregoing description is intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for automatically analyzing and responding to online public opinion events based on a large model agent, characterized in that: include: Obtain public opinion data of public opinion events at the current time node; Performing situational awareness analysis on the public opinion data at the current time node to obtain situational data at the current time node; Based on the situation data of the current time node, the basic information of the public opinion event, and the element information of multiple historical key time nodes, a pre-trained public opinion response model is used to generate response measures for the current time node; Based on the response measures at the current time node, the public opinion event is automatically responded to.

2. The method according to claim 1, characterized in that When generating response measures for the current time node using a pre-trained public opinion response model based on the situation data of the current time node, the basic information of the public opinion event, and the element information of multiple historical key time nodes, the following steps are included: Constructing decision-making time series data based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes; The decision-making basis time series data is input into the public opinion response model to generate the response measures and expected response effects of the current time node, wherein the response measures include response measure types and response measure parameters.

3. The method according to claim 2, characterized in that When constructing decision-making based on time series data based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes, it includes: Combine the element information of multiple historical key time nodes in chronological order to obtain the historical key node sequence data at the current time node; The basic information of the public opinion event, the historical key node sequence data at the current time node and the situation data at the current time node are combined in sequence to construct the decision-making basis time series data.

4. The method according to claim 1, wherein The public opinion response model is trained based on the basic information of historical public opinion events and the element information of each key time node.

5. The method according to claim 1, wherein After performing situation awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node, the method further includes: Determining whether it is necessary to respond to the public opinion event at the current time node based on the situation data at the current time node; When it is determined that a response is needed to the public opinion event at the current time node, an operation is executed based on the situation data of the current time node, the basic information of the public opinion event, and the element information of the historical key time nodes, using a pre-trained public opinion response model to generate response measures for the current time node.

6. The method according to claim 1, wherein When automatically responding to the public opinion event based on the response measures at the current time node, it includes: Generate an action task list based on the response measures at the current time node; Based on the action task list, the public opinion event is automatically responded to.

7. The method according to claim 6, characterized in that When generating an action task list based on the response measures at the current time node, it includes: Analyzing the countermeasures at the current time node to obtain the countermeasure types and countermeasure parameters of the countermeasures; An action task template matching the category is obtained, and the response measure parameters are filled into the action task template to obtain an action task list of the response measure.

8. A device for automatically analyzing and responding to online public opinion events based on a large model intelligent agent, characterized in that: include: The public opinion data acquisition module is used to obtain the public opinion data of the public opinion event at the current time node; A public opinion situation awareness module is used to perform situation awareness analysis on the public opinion data at the current time node to obtain the situation data at the current time node; A decision-making and planning module is used to generate response measures for the current time node using a pre-trained public opinion response model based on the situation data of the current time node, the basic information of the public opinion event, and the element information of multiple historical key time nodes; The action response module automatically responds to the public opinion event based on the response measures at the current time node.

9. A device for automatically analyzing and responding to online public opinion events based on a large model intelligent agent, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to implement the method according to any one of claims 1 to 7 when executing the executable instructions.

10. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images