Analysis Method and Device for Network Public Opinion Control Measures Based on Intelligent Robots

Abstract

This application provides a method and apparatus for analyzing online public opinion control measures based on intelligent robots. The method includes: acquiring target public opinion text related to a target public opinion event published by users on a target online platform; analyzing the target public opinion text using a sentiment analysis model and a topic extraction model to obtain sentiment tags and topic tags for the target public opinion text; determining the user category of each user at different times within a preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user at different times within a preset sampling period, and then calculating at least one propagation dynamics parameter based on the user category at different times; and conducting simulation analysis on public opinion control measures based on the propagation dynamics parameter and the user category to which each user currently belongs to, in order to evaluate the effectiveness of the public opinion control measures. This scheme introduces the mutual influence between users with different emotions into the simulation analysis process through a propagation dynamics model, making the analysis results more accurate.

Description

Technical Field

[0001] This invention relates to the field of online public opinion control technology, and in particular to a method and apparatus for analyzing online public opinion control measures based on intelligent robots. Background Technology

[0002] With the development of the internet, more and more users are expressing their opinions or feelings about recent public opinion events by posting comments on online platforms, while also browsing the comments posted by other users. To prevent the spread of inappropriate remarks, online platform operators sometimes need to take necessary public opinion control measures. Consequently, it is particularly important to assess the effectiveness of these measures in order to decide which measures to take based on their effectiveness.

[0003] Current research in this area mainly employs analytical methods such as sentiment dictionaries, machine learning, and deep learning to analyze public sentiment during public opinion crises and predict the effectiveness of public opinion control measures based on the analysis results.

[0004] The problem with existing solutions is that they ignore the impact of different users' emotions on each other during the duration of a public opinion event. Therefore, the analysis results of public sentiment and the effectiveness of the public opinion control measures predicted based on the analysis results are not accurate enough. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for analyzing online public opinion control measures based on intelligent robots, thereby offering a more accurate method for evaluating the effectiveness of online public opinion control measures.

[0006] The first aspect of this application provides a method for analyzing online public opinion control measures based on intelligent robots, applied to intelligent robots, the method comprising:

[0007] Acquire target public opinion texts published by users of the target network platform within a preset sampling period; wherein, the target public opinion texts refer to texts related to the preset target public opinion events;

[0008] The target public opinion texts are analyzed using sentiment analysis models and topic extraction models respectively to obtain the sentiment tags and topic tags corresponding to each target public opinion text;

[0009] Based on the sentiment tags and topic tags of the target public opinion texts published by each user in each time period within the preset sampling period on the target network platform, the user category of each user in each time period within the preset sampling period is determined; wherein, the user category is divided according to the propagation dynamics model, and the user category includes emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals.

[0010] At least one propagation dynamics parameter is calculated based on the user category of each user in the target network platform at different times.

[0011] Based on the propagation dynamics parameters and the user category of each user in the current target network platform, simulation analysis is performed on the public opinion control measures to evaluate their effectiveness.

[0012] A second aspect of this application provides a network public opinion control and analysis device based on an intelligent robot, applied to an intelligent robot, the device comprising:

[0013] The sampling unit is used to acquire target public opinion texts published by users of the target network platform within a preset sampling period; wherein, the target public opinion texts refer to texts related to the preset target public opinion events;

[0014] The analysis unit is used to analyze the target public opinion text using a sentiment analysis model and a topic extraction model respectively, and obtain the sentiment tag and topic tag corresponding to each target public opinion text;

[0015] The determining unit is used to determine the user category of each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user in each time period within the preset sampling period on the target network platform; wherein, the user category is divided according to the propagation dynamics model, and the user category includes emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals.

[0016] The calculation unit is used to calculate at least one propagation dynamics parameter based on the user category of each user in the target network platform at different times.

[0017] The simulation unit is used to perform simulation analysis on the public opinion control measures based on the propagation dynamics parameters and the user category to which each user belongs in the current target network platform, so as to evaluate the effectiveness of the public opinion control measures.

[0018] This application provides a method and apparatus for analyzing online public opinion control measures based on intelligent robots. The method includes: acquiring target public opinion text related to a target public opinion event published by users on a target online platform within a preset sampling period; analyzing the target public opinion text using a sentiment analysis model and a topic extraction model to obtain sentiment tags and topic tags; determining the user category of each user at different times within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user at different times, and then calculating at least one propagation dynamics parameter based on the user category at different times; and performing simulation analysis on public opinion control measures based on the propagation dynamics parameter and the current user category of each user to evaluate the effectiveness of the public opinion control measures. This solution incorporates the mutual influence between users with different emotions into the simulation analysis process through a propagation dynamics model, making the analysis results more accurate. Attached Figure Description

[0019] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 A schematic diagram illustrating a network public opinion control measure analysis method based on intelligent robots, provided as an embodiment of this application;

[0021] Figure 2 A flowchart illustrating a network public opinion control measure analysis method based on intelligent robots, provided as an embodiment of this application;

[0022] Figure 3 A schematic diagram of a bidirectional encoder representation model based on a converter is provided in an embodiment of this application;

[0023] Figure 4 A schematic diagram of the structure of a bidirectional long short-term memory model provided in an embodiment of this application;

[0024] Figure 5 This is a schematic diagram of the structure of a sentiment analysis model provided in an embodiment of this application;

[0025] Figure 6 A schematic diagram of a propagation dynamics model provided in an embodiment of this application;

[0026] Figure 7 This is a schematic diagram of a network public opinion control measure analysis device based on an intelligent robot, provided in an embodiment of this application. Detailed Implementation

[0027] 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.

[0028] Please see Figure 1 The diagram illustrates the network public opinion control measures analysis method based on intelligent robots provided in this application embodiment. It can be seen that the method can be divided into three stages: data capture and preprocessing, public opinion analysis process, and public opinion prevention and control suggestions.

[0029] In the data crawling and preprocessing stage, operations such as crawling Weibo data (or data crawling from other online platforms, taking Weibo as an example here) can be performed, data cleaning, stop word removal and word segmentation, and building a text corpus can be carried out.

[0030] In the process of public opinion analysis, a deep neural network model can be constructed by fusing a Bidirectional Encoder Representations from Transformer (BERT) model with a Bidirectional Long Short-Term Memory (BiLSTM) model. Figure 1 The BERT-BiLSTM sentiment analysis model shown is referred to as the sentiment analysis model below, and the Latent Dirichlet Allocation (LDA) topic extraction model is constructed, referred to as the topic extraction model below. The sentiment analysis model and the topic extraction model are used to analyze the public opinion events of the target public opinion texts obtained in the data crawling and preprocessing stage, and to determine the sentiment tag and topic tag of each target public opinion text.

[0031] The sentiment analysis model can be trained on a corpus first. The trained model generates word vectors for the target public opinion text, and then performs sentiment analysis based on these word vectors to obtain the sentiment distribution for different time periods within the sampling period, such as daily sentiment distribution. The sentiment distribution for a specific time period can be understood as the number of different sentiment tags in the target public opinion text published within that time period. Therefore, determining the sentiment distribution for different time periods is essentially determining the sentiment tags of the target public opinion text published within those time periods.

[0032] When determining topic tags, the topic extraction model can first select the number of topics to determine the appropriate number of topic tags. Then, it can extract topic features from the word vector features of the target public opinion texts published at each stage within the sampling period, thereby determining the corresponding topic tags for the target public opinion texts published at each stage.

[0033] In the public opinion control recommendation stage, a Susceptible, Exposed, Infected, and Recovered (SEIR) model based on propagation dynamics can be implemented. Specifically, a validation dataset can be selected first to verify the effectiveness of the SEIR model. After successful validation, users on the target online platform (e.g., Weibo) are categorized according to the sentiment and topic tags of the target public opinion text, that is, determining which of the four user categories each user belongs to: Susceptible, Exposed, Infected, or Recovered. Then, propagation dynamics parameters are determined, which may include conversion rate and related parameters. Finally, based on the conversion rate and related parameters, a public opinion recommendation validation analysis is conducted, that is, an evaluation of the effectiveness of the public opinion control measures to be taken.

[0034] The analytical method provided in this application will be explained in detail below with reference to the accompanying drawings. Please refer to... Figure 2 The flowchart below illustrates a method for analyzing online public opinion control measures based on intelligent robots, which includes the following steps.

[0035] First, it should be noted that the method provided in this embodiment can be executed by an existing intelligent robot platform (such as a Raspberry Pi intelligent robot). The advantage of having this embodiment executed by an intelligent robot platform is that it can improve the applicability of the method.

[0036] S201, Obtain target public opinion texts published by users on the target network platform within a preset sampling period.

[0037] Among them, target public opinion text refers to text related to the pre-set target public opinion event.

[0038] The target online platform can be any pre-defined open online platform where registered users can express their opinions through articles, discussions with other users, and other means. For example, a target online platform could be microblogs.

[0039] Public opinion events can be defined as events that generate significant buzz on a target online platform. The buzz surrounding an event can be measured by the number of texts about that event on the target online platform within a certain period (e.g., within one day). The more texts about an event within a certain period, the higher the buzz surrounding that event. When the buzz surrounding an event exceeds a preset buzz threshold, the event is considered a public opinion event.

[0040] For example, if a city issues a notice and the popularity of the text about the notice on a target online platform exceeds a popularity threshold within one day, then the issuance of the notice is considered a public opinion event.

[0041] A target public opinion event is any pre-specified public opinion event among the existing public opinion events on the target online platform.

[0042] Optionally, the execution process of step S201 may include:

[0043] A1, acquire massive amounts of text published on the target network platform within a preset sampling period;

[0044] A2 filters a massive amount of text to remove texts that are irrelevant to the preset target public opinion event, thus obtaining the target public opinion text.

[0045] In step A1, a web crawler can be used to crawl a large amount of text published on the target network platform within a preset sampling period. The preset sampling period can be determined based on the occurrence time of the target public opinion event. For example, if the target public opinion event occurred 5 days ago, the preset sampling period can be set to the most recent 5 days. In step A1, the web crawler will crawl all the text published within the most recent 5 days from the target network platform to obtain a large amount of text.

[0046] The text crawled in step A1 may include text directly posted by users on their personal pages on the target network platform, or text in which users comment on content posted by other users. In other words, it may include original text and comment text on original text.

[0047] In step A2, the filtering method can be to identify whether each crawled text contains keywords corresponding to the target public opinion event. If a text contains at least one keyword, then the text belongs to the target public opinion text. If a text does not contain any keyword, then the text does not belong to the target public opinion text.

[0048] Understandably, after filtering through step A2, multiple target public opinion texts can be obtained, meaning there are multiple target public opinion texts.

[0049] Keywords corresponding to a target public opinion event can be pre-defined manually or determined using machine learning algorithms. Continuing the previous example, assuming the target public opinion event is a notice B issued by City A, then the keywords corresponding to this target public opinion event can be set as "City A", "Notice B", and "Spokesperson C", where Spokesperson C is the employee responsible for issuing Notice B. Additionally, keywords that frequently appear on the target online platform can also be included. For example, if Notice B is about road maintenance on Road D in City A, then the keywords corresponding to this target public opinion event could also include "Road D".

[0050] In some optional embodiments, in order to improve the accuracy of subsequent analysis results, multiple target public opinion texts obtained after step A2 are cleaned.

[0051] During the text cleaning process, some content that is not relevant to the subsequent sentiment and theme analysis can be removed from the target public opinion text, including but not limited to stop words, function words (which can be understood as interjections, such as "ya" and "ba"), Uniform Resource Identifier (i.e., URL) links, punctuation marks, and special characters (such as emoticons customized by the target network platform).

[0052] The above cleaning process can be implemented using existing text processing frameworks, such as the Nltk framework. This embodiment does not limit the specific implementation method.

[0053] Step S201 is equivalent to Figure 1 The data capture and preprocessing process is shown.

[0054] S202, using sentiment analysis model and topic extraction model respectively to analyze the target public opinion text, and obtain the sentiment tag and topic tag corresponding to each target public opinion text.

[0055] Step S202 is equivalent to Figure 1 The process of public opinion analysis is shown.

[0056] Optionally, the sentiment analysis model is a deep neural network model obtained by fusing a transducer-based bidirectional encoder representation model with a bidirectional long short-term memory model;

[0057] The specific implementation of step S202 may include:

[0058] B1 extracts word vectors from the target public opinion text using a converter-based bidirectional encoder representation model;

[0059] B2, using a bidirectional long short-term memory model to analyze the word vectors of the target public opinion text, to obtain the sentiment labels of the target public opinion text;

[0060] B3, Calculate the perplexity curve of the implicit Dirichlet distribution topic extraction model; where the horizontal axis of the perplexity curve represents the number of topic tags, and the vertical axis represents the perplexity.

[0061] B4. The word vectors of the target public opinion text are analyzed using the Latent Dirichlet Distribution Topic Extraction Model to obtain N topic tags of the target public opinion text and determine the topic tags corresponding to each target public opinion text; where N is the number of topic tags corresponding to the lowest point in the perplexity curve.

[0062] Please see Figure 3 , Figure 4 and Figure 5 ,in Figure 5 The schematic diagram of the sentiment analysis model formed by fusing the BERT model and the BiLSTM model provided in the embodiments of this application shows that the sentiment analysis model includes an input layer, a BERT layer (i.e., the BERT model), word vectors output by the BERT layer, a BiLSTM layer, a Softmax layer (also called a fully connected layer), and an output layer.

[0063] The structure of the BERT model is as follows: Figure 3 As shown, after the target public opinion text is input into the BERT model, it is processed sequentially through the token embeddings, segment embeddings and position embeddings of the BERT model to obtain the word vectors of the target public opinion text.

[0064] The structure of the BiLSTM model can be found in [link to BiLSTM model]. Figure 4 As can be seen, the BiLETM model consists of multiple repeating substructures. Each substructure includes an input layer, embeddings, a Long Short-Term Memory (LSTM) layer, and an output layer. The LSTM layer of each substructure comprises two LSTM units, which are the word vectors of the target sentiment text input into the BiLETM model. Figure 4 The values ​​of x1, x2, etc. shown will be processed by each substructure in the BiLSTM model to obtain the sentiment label of each target public opinion text.

[0065] In step B2, the determined sentiment tag for each target sentiment text can be any one of a number of preset sentiment tags.

[0066] The division of emotion tags can be based on any existing emotion classification method, without any restrictions.

[0067] As an example, in this embodiment, the emotions that users may experience on the target network platform can be finely divided and corresponding emotion tags can be set according to the emotion lexicon ontology released by the Dalian University of Technology Information Retrieval (DUTIR) Laboratory. That is, in step B2, seven emotion tags such as happiness, good, anger, sadness, fear, disgust, and surprise can be preset. For each target public opinion text, when executing B2, an emotion tag corresponding to the target public opinion text can be determined from the aforementioned seven emotion tags. For example, the emotion tag corresponding to target public opinion text 1 is determined to be "anger", and the emotion tag corresponding to target public opinion text 2 is determined to be "surprise".

[0068] The sentiment tag of a target public opinion text can be understood as the emotion of the user who published the target public opinion text at the time of publication. For example, if the sentiment tag of the target public opinion text 1 published by user A is "anger", then it can be assumed that user A was angry when publishing the target public opinion text.

[0069] It should be noted that each target public opinion text may correspond to any of the seven emotion tags mentioned above, or it may not correspond to any emotion tag. If a target public opinion text corresponds to any of the above emotion tags, it is called a target public opinion text with an emotion tag; if a target public opinion text does not correspond to any of the above emotion tags, it is called a target public opinion text without an emotion tag.

[0070] The advantage of classifying emotion tags in the above manner is that:

[0071] Most existing public opinion analysis solutions categorize sentiment tags into three types: neutral, positive, and negative. However, this generalized classification clearly fails to accurately reflect the true emotions of users when publishing target public opinion texts. In contrast, this solution sets seven sentiment tags—happiness, goodwill, anger, sorrow, fear, disgust, and surprise—to better reflect the true emotions of users when publishing target public opinion texts, resulting in more accurate subsequent analysis results.

[0072] In step B3, several candidate quantities can be set first, such as 5, 10, 15, 20...100. For each candidate quantity, the topic tags of the candidate quantity are extracted from the public opinion text set composed of the aforementioned multiple target public opinion texts using the Latent Dirichlet Allocation (LDA) topic extraction model, thus obtaining several topic tag sets.

[0073] For example, based on the above candidate quantities of 5, 10, 15, 20...100, the LDA model is used to extract 5 topic tags from the public opinion text set to obtain topic tag set 1. The LDA model is used to extract 10 topic tags from the public opinion text set to obtain topic tag set 2, and so on, until the LDA model is used to extract 100 topic tags from the public opinion text set to obtain topic tag set 20.

[0074] Based on this, for each number of alternatives, the perplexity corresponding to that number of alternatives is calculated according to the following definition of perplexity:

[0075] .

[0076] In the above formula, Perlexity represents the degree of perplexity, M represents the number of alternatives, and w i p(w) represents the i-th topic in the set of topic tags corresponding to this number of alternatives. i The tag w indicates that in the collection of public opinion texts, the topic tag w i The probability of the corresponding target sentiment text appearing, N i This represents the number of target public opinion texts corresponding to the topic tags in the public opinion text set. It can be seen that the denominator of the above formula is equal to the total number of target public opinion texts in the public opinion text set.

[0077] Through the above process, we can obtain the perplexity corresponding to each of the previously set candidate quantities. For example, the perplexity of candidate quantity 5 is Perlexity-5, the perplexity of candidate quantity 10 is Perlexity-10, and so on. Then, we can perform curve fitting based on these candidate quantities and their corresponding perplexities to obtain a perplexity curve with the number of topic tags as the horizontal axis and the perplexity corresponding to that number of topic tags as the vertical axis.

[0078] After obtaining the perplexity curve, the lowest point of the perplexity curve can be determined, and the number N corresponding to the lowest point is determined as the number of topic tags to be extracted. For example, assuming the number corresponding to the lowest point is 33, then in step B4, the LDA topic extraction model can be used to extract topics from the target public opinion text, obtain 33 topic tags, and determine the topic tags corresponding to each target public opinion text.

[0079] In some alternative embodiments, the number N corresponding to the inflection point of the confusion curve can also be determined as the number of topic tags to be extracted in step B4.

[0080] In some optional embodiments, to improve accuracy, when executing B4, topic tags can be extracted from multiple target public opinion texts in stages. Specifically, the entire sampling period can be pre-divided into multiple time periods (also known as multiple stages), and then for the target public opinion texts published in each stage, N topic tags for the target public opinion texts published in that stage can be extracted respectively. Finally, N topic tags can be obtained for each stage.

[0081] For example, if the sampling period is the last 5 days, then the first two days of the last 5 days can be divided into the first time period (i.e., the first stage), the third and fourth days into the second time period (i.e., the second stage), and the last day into the third time period (i.e., the third stage).

[0082] Then, for multiple target public opinion texts published within the first phase, step B4 is executed to extract N topic tags of the target public opinion texts published in the first phase, and to determine the corresponding topic tags for each target public opinion text published in the first phase. For multiple target public opinion texts published within the second phase, step B4 is executed to extract N topic tags of the target public opinion texts published in the second phase, and to determine the corresponding topic tags for each target public opinion text published in the second phase. The third phase is the same and will not be described in detail here.

[0083] The advantage of extracting topic tags in stages is that, for the same target public opinion event, as the duration of the event extends, the topics discussed by users on the target online platform may change. For example, topics that frequently appeared in the first stage may no longer appear in the second stage. Classifying the target public opinion texts throughout the entire sampling period using the same set of topic tags obviously cannot reflect the changes in topics at different stages. However, extracting topic tags in stages can reflect the changes in topics discussed by users over time. Therefore, extracting and determining the topic tags of target public opinion texts in this way is obviously more accurate.

[0084] It should be noted that a single target public opinion text can have one or more topic tags.

[0085] The LDA topic extraction model is a mature existing model. The specific process of using this model to extract topic tags from a text set and determine the topic tags corresponding to each text can be found in relevant existing technologies, and will not be elaborated here.

[0086] S203, based on the sentiment tags and topic tags of the target public opinion texts published by each user in each time period within the preset sampling period on the target network platform, determine the user category of each user in each time period within the preset sampling period.

[0087] The user categories are divided according to the propagation dynamics model, including emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals.

[0088] Optionally, the specific execution process of step S203 includes:

[0089] C1, for each user on the target network platform, statistically analyze the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period;

[0090] C2, for each user on the target network platform, determines the user category for each time period within the preset sampling period based on the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period.

[0091] In this embodiment, emotionally susceptible individuals are represented by S, which is defined as users whose proportion of target public opinion texts with emotional tags is less than a preset first proportion threshold and who have a clear thematic tendency within a detection period.

[0092] For example, the first percentage threshold can be set to 50%, but it can also be set to other values ​​as needed, without limitation.

[0093] The detection period is shorter than the sampling period. For example, if the sampling period is the most recent 5 days, then each day within the sampling period can be defined as a detection period.

[0094] Having a clear thematic tendency means that, within a detection period, the proportion of target public opinion texts corresponding to the same topic tag is greater than the preset second proportion threshold.

[0095] The second percentage threshold can also be set to 50%, or other values ​​as needed, without limitation.

[0096] Based on the above example, suppose User A publishes 10 targeted public opinion texts in one day. Only three of these targeted public opinion texts have emotion tags, and seven of them correspond to the same topic tag. In other words, the proportion of targeted public opinion texts published by User A with emotion tags in this day is 30%, which is less than the first proportion threshold. At the same time, the proportion of targeted public opinion texts corresponding to the same topic tag reaches 70%. This indicates that the targeted public opinion texts published by User A in this day have a clear thematic tendency. Therefore, it is determined that User A is an emotionally susceptible person in this day.

[0097] Emotional latent users are denoted by E and are defined as users whose proportion of target public opinion texts with emotion tags is greater than or equal to a first proportion threshold, but whose proportion of target public opinion texts with the same emotion tag is less than the first proportion threshold, within a detection period.

[0098] Taking a one-day detection period as an example, if User A publishes 10 target sentiment texts in one day, 8 of which have emotion tags, but most of the target sentiment texts have different emotion tags. For example, 2 target sentiment texts have the emotion tag "good", 3 target sentiment texts have the emotion tag "angry", and 3 target sentiment texts have the emotion tag "fear". It can be found that the proportion of target sentiment texts with the same emotion tag is less than the first proportion threshold. Therefore, it is determined that User A is an emotion lurker in this day.

[0099] An "emotional infected person" is denoted by I and is defined as a user whose target public opinion texts with emotion tags account for a proportion greater than or equal to a first proportion threshold within a detection period, and whose target public opinion texts with the same emotion tags account for a proportion greater than or equal to the first proportion threshold. In this case, the emotion tags carried by the target public opinion texts with a proportion greater than or equal to the first proportion threshold can be identified as the emotions infected by the emotional infected person.

[0100] Taking a one-day detection period as an example, if User A publishes 10 target public opinion texts in one day, 9 of which have emotion tags, and 7 of which have emotion tags of "good", then it is determined that User A is an emotion infected person on that day, and the infected emotion is "good".

[0101] Specifically, based on the different emotions that infected individuals experience, they are divided into seven categories, numbered I1, I2, I3...I7, respectively. The emotions experienced by individuals in categories I to I7 are, in order, happiness, goodness, anger, sorrow, fear, disgust, and shock.

[0102] Emotional recovery users are represented by R, defined as users whose proportion of target public opinion texts with emotion tags is less than a preset first proportion threshold and who do not have a clear thematic orientation within a detection period.

[0103] Based on the definition of having a clear thematic tendency mentioned above, it can be understood that if the proportion of target public opinion texts with the same theme tag published by a user within a detection period is less than the second proportion threshold, then it can be considered that the user does not have a clear thematic tendency within that detection period.

[0104] Taking a one-day detection period as an example, if User A publishes 10 target public opinion texts in one day, 7 of which do not have emotion tags, and most of the target public opinion texts correspond to different topic tags, and the proportion of target public opinion texts corresponding to the same topic tags is less than 50%, then it can be determined that User A is an emotional recovery person on that day.

[0105] Based on the above definition, when performing step C1, the sampling period can be divided into multiple detection periods. For example, if the sampling period is the last 5 days and each day is designated as a detection period, then the last 5 days can be divided into 5 detection periods.

[0106] After dividing the detection period, we can then analyze the distribution of sentiment tags and topic tags for each user's target public opinion text published on the target network platform within that detection period.

[0107] Next, in step C2, it is possible to identify which of the above definitions the distribution of sentiment tags and topic tags of the target public opinion text published by the user conforms to in each detection period, and thus determine the user category to which the user belongs in that detection period.

[0108] The time periods in steps C1 and C2 are the detection periods mentioned in the definition of each user category above.

[0109] S204, at least one propagation dynamics parameter is calculated based on the user category of each user at different times in the target network platform.

[0110] In step S204, it is first necessary to determine the conversion paths between different user categories. In this embodiment, the following conversion paths can be determined:

[0111] In public opinion events, those who are emotionally latent have a certain probability of transforming into a certain state of emotionally infected person I. Therefore, we determine the path 2 to increase the transformation of emotional latent persons into emotionally infected persons.

[0112] Since the spread of information in public opinion is different from the spread of viruses, emotionally susceptible individuals may not necessarily experience a long incubation period after being exposed to relevant information. Susceptible individuals have a certain probability of directly expressing their opinions on public opinion events online. Therefore, we have determined the path 5 for emotionally susceptible individuals to directly become emotionally infected individuals.

[0113] The difference between a person who is a latent in the spread of public opinion and a person who has been exposed to the virus is that a person who has been exposed to the virus cannot decide whether he or she wants to be infected or directly become a recovered person. However, a person who is a latent in the spread of public opinion can change into an infected person or a recovered person of various emotions according to his or her own will. Therefore, the path from an infected person of emotions to a recovered person of emotions is determined 3.

[0114] In addition to the above-mentioned pathways, based on existing principles of transmission dynamics, pathway 1 can be determined to transform emotionally susceptible individuals into emotionally latent individuals, and pathway 4 can be determined to transform emotionally infected individuals into emotionally recovered individuals.

[0115] The finalized conversion path and the relationship between the different user categories can be found in [reference needed]. Figure 6 The diagram shows a schematic of the SEIR model.

[0116] After determining the above conversion paths, the conversion rate corresponding to each conversion path can be calculated based on the number of users of different user categories in each detection period. The multiple conversion rates corresponding to the multiple conversion paths are the propagation dynamics parameters in step S204.

[0117] To facilitate the explanation of how conversion rates are calculated, the user category before conversion in a conversion path is defined as the front-end user category of that conversion path, and the user category after conversion is defined as the back-end user category of that conversion path.

[0118] For example, for Figure 6 The path shown is Path 2, which transforms users from emotional latenters to emotional infected users. Before the transformation occurs, the user category is emotional latenters. After the transformation, the user category is emotional infected users. Therefore, for Path 2, the front-end user category is emotional latenters and the back-end user category is emotional infected users.

[0119] For any conversion path, the conversion rate corresponding to that path can be calculated as follows:

[0120] For every two adjacent detection periods within the sampling period, calculate the number of users belonging to the front-end user category of the conversion path in the previous detection period. Then, in the next detection period, count how many users who belonged to the front-end user category in the previous detection period converted to the back-end user category of the conversion path. The proportion of the counted number to the number of users belonging to the front-end user category in the previous detection period is determined as the reference conversion rate for these two adjacent detection periods. Finally, the maximum value among the calculated reference conversion rates is determined as the conversion rate of the conversion path.

[0121] Continuing with the previous example, when calculating the conversion rate of path 2, suppose that for the adjacent first and second detection periods, statistics show that there were 100 users who were emotionally latent in the first detection period, and in the second detection period, 40 of these users became emotionally influenced. Therefore, for these two adjacent detection periods, the reference conversion rate for path 2 is 40 divided by 100, which is 40%. Similarly, continue calculating the reference conversion rate for path 2 in the second and third detection periods, the third and fourth detection periods, and so on, until all combinations of adjacent detection periods are traversed. Finally, the maximum value of the reference conversion rate for all paths 2 is found to be 40%, thus determining the conversion rate of path 2 to be 40%.

[0122] S205. Based on propagation dynamics parameters and the user category of each user in the current target network platform, simulation analysis is conducted on public opinion control measures to evaluate their effectiveness.

[0123] Steps S203 to S205 are equivalent to Figure 1 The process of public opinion control recommendations is shown below.

[0124] Optionally, the specific execution process of step S205 may include:

[0125] D1, determine the predicted update amount of the propagation dynamics parameters after the implementation of public opinion control measures;

[0126] D2, update the propagation dynamics parameters based on the predicted update amount;

[0127] D3. Based on the updated propagation dynamics parameters and the user category of each user in the current target network platform, a propagation dynamics model simulation is performed to obtain simulation results; the simulation results include at least the distribution of each user category in the target network platform after a preset control period.

[0128] D4 determines the effectiveness of public opinion control measures by the degree of similarity between the simulation results and the expected control results of the measures.

[0129] The predicted update of the propagation dynamics parameters after the implementation of public opinion control measures refers to the possible change in the conversion rate of at least one path in the propagation dynamics parameters after the implementation of a certain public opinion control measure.

[0130] For example, anticipating the effects of implementing a certain public opinion control measure X Figure 6 The conversion rate corresponding to path 3 shown may decrease by 30% from the current value. Therefore, the predicted update of the propagation dynamics parameters after implementing this public opinion control measure X is that the conversion rate of path 3 decreases by 30%.

[0131] In step D1, the predicted update of the propagation dynamics parameters after implementing public opinion control measures can be determined based on experience from previous implementations of the same public opinion control measures.

[0132] The public opinion control measures in step D1 are those whose effectiveness needs to be evaluated. For example, if the effectiveness of public opinion control measure X needs to be evaluated, then in step D1, it is necessary to determine the predicted update amount of the propagation dynamics parameters after the implementation of public opinion control measure X.

[0133] The specific content of public opinion control measures can be determined based on the actual situation and is not limited.

[0134] For example, one possible public opinion control measure is to limit the exposure of a target public opinion text to no more than a specific maximum exposure. For instance, if the maximum exposure is 1000, this measure means that for each piece of text related to the target public opinion event published on the target online platform, if the cumulative number of users who have viewed the text is less than 1000, the text is allowed to be viewed normally. However, if the cumulative number of users who have viewed the text reaches 1000, the text is prohibited from being viewed by other users, and the text will be set to be visible only to the publisher.

[0135] In step D2, the conversion rate of the corresponding path can be updated according to the predicted update amount determined in D1.

[0136] Continuing with the previous example, suppose that after implementing public opinion control measure X, the predicted update of the propagation dynamics parameters is that the conversion rate of path 3 decreases by 30%. And the conversion rate of path 3 calculated in S204 is 40%. Then, in step D2, firstly, the original conversion rate of path 3, 40%, is multiplied by the reduction of 30% to obtain the absolute update amount, which is 12%. Then, the original conversion rate of path 3, 40%, is subtracted from the absolute update amount (representing the reduction in conversion rate) to obtain the updated conversion rate of path 3, which is 40% minus 12%, equal to 28%.

[0137] The example above indicates that if public opinion control measure X is implemented, the conversion rate of path 3 is expected to decrease from 40% before implementation to 28% after implementation.

[0138] In step D3, the updated propagation dynamics parameters, namely the conversion rates corresponding to each conversion path after the update, and the number of users of different user categories in the current target network platform, can be substituted into the computer simulation program designed based on the propagation dynamics model to obtain simulation results.

[0139] The number of users of different user categories in the current target network platform can be understood as the number of users of different user categories in the most recent detection period.

[0140] In the field of propagation dynamics research, there are already a number of computer simulation programs designed based on propagation dynamics models. When executing step D3, these existing simulation programs can be directly applied. The specific execution process will not be described in detail.

[0141] The length of the control cycle is an adjustable parameter in the computer simulation program; for example, the control cycle can be set to 5 days.

[0142] Continuing with the example above, taking the public opinion control measure X that needs to be evaluated, and the control period set to 5 days, the meaning of the simulation result obtained after executing step D3 is: if public opinion control measure X is continuously implemented for the next 5 days, then how many users of different user categories will be on the target network platform after 5 days.

[0143] For example, the simulation results can be expressed as follows: after implementing public opinion control measures for 5 days, for the target public opinion event, the number of emotionally susceptible individuals is 100, the number of emotionally latent individuals is 50, the number of emotionally infected individuals is 20, and the number of emotionally recovered individuals is 600.

[0144] The expected control results of public opinion control measures are artificially set by the operation and maintenance personnel of the target network platform based on the current number of users in each user category. The form can be the same as the simulation results, but the specific values ​​may differ.

[0145] Continuing with the previous example, the expected control results of public opinion control measure X can be expressed as follows: 5 days after the implementation of public opinion control measure X, for the target public opinion event, the number of emotionally susceptible individuals is 50, the number of emotionally latent individuals is 10, the number of emotionally infected individuals is 10, and the number of emotionally recovered individuals is 700.

[0146] In step D4, the simulation results and the expected control results can be represented by vectors respectively. Then, the cosine similarity between the two vectors is calculated as the similarity level in step D4. The higher the similarity level, the more likely the expected control result will be achieved after the implementation of the public opinion control measures, and the higher the effectiveness of the current evaluation of the public opinion control measures. Conversely, the lower the similarity level, the higher the effectiveness of the current evaluation of the public opinion control measures.

[0147] Continuing with the previous example, the simulation results can be expressed as follows: After implementing public opinion control measures for 5 days, for the target public opinion event, the number of emotionally susceptible individuals is 100, the number of emotionally latent individuals is 50, the number of emotionally infected individuals is 20, and the number of emotionally recovered individuals is 600, with the corresponding vector representation being (100, 50, 20, 600).

[0148] The expected control result of public opinion control measure X can be expressed as follows: Five days after the implementation of public opinion control measure X, for the target public opinion event, the number of emotionally susceptible individuals is 50, the number of emotionally latent individuals is 10, the number of emotionally infected individuals is 10, and the number of emotionally recovered individuals is 700. The corresponding vector is represented as (50, 10, 10, 700). The cosine similarity between these two vectors can be regarded as the degree of similarity between the simulation result and the expected control result of the public opinion control measure, which is equivalent to the effectiveness of public opinion control measure X.

[0149] For example, if the cosine similarity between the vector corresponding to the expected control result and the vector corresponding to the simulation result is 80%, then the effectiveness of the current public opinion control measures can be considered to be 80%.

[0150] This application provides a method for analyzing online public opinion control measures based on intelligent robots. The method includes: acquiring target public opinion texts related to target public opinion events published by users on a target online platform within a preset sampling period; analyzing the target public opinion texts using a sentiment analysis model and a topic extraction model to obtain sentiment tags and topic tags; determining the user category of each user at different times within the preset sampling period based on the sentiment tags and topic tags of the target public opinion texts published by each user at different times, and then calculating at least one propagation dynamics parameter based on the user categories at different times; and conducting simulation analysis on public opinion control measures based on the propagation dynamics parameter and the current user category of each user to evaluate the effectiveness of the public opinion control measures. This scheme incorporates the mutual influence between users with different emotions into the simulation analysis process through a propagation dynamics model, making the analysis results more accurate.

[0151] In addition, this embodiment also has the following beneficial effects:

[0152] Existing solutions perform sentiment analysis on public opinion texts, but when analyzing netizens' sentiment, they do not classify sentiment tags in a fine-grained manner, only roughly dividing them into positive, negative, and neutral. The solution of this invention performs user sentiment analysis based on the aforementioned seven more refined sentiment tags, which has higher accuracy.

[0153] Existing solutions lack simulation verification of the evolution of user emotions after the implementation of public opinion control measures. This solution, however, combines the SEIR model from communication dynamics to simulate the evolution of user emotions after the implementation of public opinion control measures, which can more accurately evaluate the effectiveness of public opinion control measures.

[0154] Please see Figure 7 The diagram below is a structural schematic of a network public opinion control measure analysis device based on an intelligent robot, which may include the following units.

[0155] This device can be applied to intelligent robots; specifically, it can be considered a virtual device (program) running on the intelligent robot.

[0156] The sampling unit 701 is used to acquire target public opinion texts published by users of the target network platform within a preset sampling period; wherein, the target public opinion texts refer to texts related to the preset target public opinion events;

[0157] Analysis unit 702 is used to analyze target public opinion texts using sentiment analysis model and topic extraction model respectively, and obtain sentiment tags and topic tags corresponding to each target public opinion text;

[0158] The determination unit 703 is used to determine the user category of each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion texts published by each user in each time period within the preset sampling period on the target network platform; wherein, the user category is divided according to the propagation dynamics model, and the user category includes emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals.

[0159] The calculation unit 704 is used to calculate at least one propagation dynamics parameter based on the user category of each user in the target network platform at different times.

[0160] Simulation unit 705 is used to perform simulation analysis on public opinion control measures based on propagation dynamics parameters and the user category to which each user belongs in the current target network platform, in order to evaluate the effectiveness of public opinion control measures.

[0161] Optionally, when sampling unit 701 acquires target public opinion text published by users of the target network platform within a preset sampling period, it is specifically used for:

[0162] Acquire massive amounts of text published on the target network platform within a preset sampling period;

[0163] The massive amount of text is filtered to remove texts that are irrelevant to the preset target public opinion event, thus obtaining the target public opinion text.

[0164] Optionally, the sentiment analysis model is a deep neural network model obtained by fusing a transducer-based bidirectional encoder representation model with a bidirectional long short-term memory model;

[0165] Analysis unit 702 analyzes the target public opinion text using both a sentiment analysis model and a topic extraction model to obtain the sentiment tags and topic tags corresponding to each target public opinion text. Specifically, it is used for:

[0166] Word vectors of the target public opinion text are extracted using a converter-based bidirectional encoder representation model;

[0167] By analyzing the word vectors of the target public opinion text using a bidirectional long short-term memory model, the sentiment labels of the target public opinion text can be obtained.

[0168] Calculate the perplexity curve of the implicit Dirichlet distribution topic extraction model; where the horizontal axis of the perplexity curve represents the number of topic tags and the vertical axis represents the perplexity.

[0169] The word vectors of target public opinion texts are analyzed using the Latent Dirichlet Distribution Topic Extraction Model to obtain N topic tags for the target public opinion texts and determine the topic tags corresponding to each target public opinion text; where N is the number of topic tags corresponding to the lowest point in the perplexity curve.

[0170] Optionally, when determining the user category of each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user in each time period within the preset sampling period on the target network platform, the determining unit 703 is specifically used for:

[0171] For each user on the target network platform, the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period is statistically analyzed.

[0172] For each user on the target network platform, the user category for each time period within the preset sampling period is determined based on the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period.

[0173] Optionally, the simulation unit 705, based on propagation dynamics parameters and the user category of each user in the current target network platform, performs simulation analysis on public opinion control measures to evaluate their effectiveness. Specifically, it is used for:

[0174] Determine the predicted update amount of the propagation dynamics parameters after the implementation of public opinion control measures;

[0175] Update the propagation dynamics parameters based on the predicted update amount;

[0176] Based on the updated propagation dynamics parameters and the user category of each user in the current target network platform, a propagation dynamics model simulation is performed to obtain simulation results; the simulation results include at least the distribution of each user category in the target network platform after a preset control period.

[0177] The degree of similarity between the simulation results and the expected control results of the public opinion control measures is used to determine the effectiveness of the public opinion control measures.

[0178] The specific principles and beneficial effects of the network public opinion control measures analysis device based on intelligent robots provided in this embodiment can be found in the relevant steps and beneficial effects of the network public opinion control measures analysis method based on intelligent robots provided in this application embodiment, and will not be repeated here.

[0179] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0180] It should be noted that the concepts of "first" and "second" mentioned in this invention are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0181] Those skilled in the art will be able to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for analyzing online public opinion control measures based on intelligent robots, characterized in that, Applied to intelligent robots, the method includes: Acquire target public opinion texts published by users of the target network platform within a preset sampling period; wherein, the target public opinion texts refer to texts related to the preset target public opinion events; The target public opinion texts are analyzed using sentiment analysis models and topic extraction models respectively to obtain the sentiment tags and topic tags corresponding to each target public opinion text; Based on the sentiment tags and topic tags of the target public opinion texts published by each user in each time period within the preset sampling period on the target network platform, the user category of each user in each time period within the preset sampling period is determined; wherein, the user category is divided according to the propagation dynamics model, and the user category includes emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals. At least one propagation dynamics parameter is calculated based on the user category of each user in the target network platform at different times. Based on the propagation dynamics parameters and the user category of each user in the current target network platform, simulation analysis is performed on the public opinion control measures to evaluate their effectiveness. The sentiment analysis model is a deep neural network model obtained by fusing a converter-based bidirectional encoder representation model with a bidirectional long short-term memory model. The analysis of the target public opinion texts using sentiment analysis models and topic extraction models, respectively, to obtain sentiment tags and topic tags corresponding to each target public opinion text, includes: The word vectors of the target public opinion text are extracted using the converter-based bidirectional encoder representation model; The word vectors of the target public opinion text are analyzed using the bidirectional long short-term memory model to obtain the sentiment tags of the target public opinion text; Calculate the perplexity curve of the implicit Dirichlet distribution topic extraction model; wherein, the horizontal axis of the perplexity curve is the number of topic tags, and the vertical axis is the perplexity. The word vectors of the target public opinion text are analyzed using the Latent Dirichlet Distribution Topic Extraction Model to obtain N topic tags for the target public opinion text, and the topic tags corresponding to each target public opinion text are determined; wherein, N is the number of topic tags corresponding to the lowest point in the perplexity curve.

2. The method according to claim 1, characterized in that, The acquisition of target public opinion texts published by users of the target network platform within a preset sampling period includes: Acquire massive amounts of text published on the target network platform within a preset sampling period; The massive amount of text is filtered to remove texts that are not related to the preset target public opinion event, thus obtaining the target public opinion text.

3. The method according to claim 1, characterized in that, The step of determining the user category for each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user in each time period within the preset sampling period on the target network platform includes: For each user on the target network platform, the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period is statistically analyzed. For each user on the target network platform, the user category for each time period within the preset sampling period is determined based on the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period.

4. The method according to claim 1, characterized in that, The simulation analysis of public opinion control measures based on the propagation dynamics parameters and the user category of each user in the current target network platform, in order to evaluate the effectiveness of the public opinion control measures, includes: Determine the predicted update amount of the propagation dynamics parameters after the implementation of public opinion control measures; The propagation dynamics parameters are updated based on the predicted update amount; Based on the updated propagation dynamics parameters and the user category to which each user belongs in the current target network platform, a propagation dynamics model simulation is performed to obtain simulation results; wherein, the simulation results include at least the distribution of each user category in the target network platform after a preset control period; The degree of similarity between the simulation results and the expected control results of the public opinion control measures is used to determine the effectiveness of the public opinion control measures.

5. A network public opinion control and analysis device based on intelligent robots, characterized in that, The device, used in intelligent robots, includes: The sampling unit is used to acquire target public opinion texts published by users of the target network platform within a preset sampling period; wherein, the target public opinion texts refer to texts related to the preset target public opinion events; The analysis unit is used to analyze the target public opinion text using a sentiment analysis model and a topic extraction model respectively, and obtain the sentiment tag and topic tag corresponding to each target public opinion text; The determining unit is used to determine the user category of each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user in each time period within the preset sampling period on the target network platform; wherein, the user category is divided according to the propagation dynamics model, and the user category includes emotionally susceptible individuals, emotionally latent individuals, emotionally infected individuals, and emotionally recovered individuals. The calculation unit is used to calculate at least one propagation dynamics parameter based on the user category of each user in the target network platform at different times. The simulation unit is used to perform simulation analysis on the public opinion control measures based on the propagation dynamics parameters and the user category to which each user belongs in the current target network platform, so as to evaluate the effectiveness of the public opinion control measures. The sentiment analysis model is a deep neural network model obtained by fusing a converter-based bidirectional encoder representation model with a bidirectional long short-term memory model. The analysis unit analyzes the target public opinion text using a sentiment analysis model and a topic extraction model, respectively, to obtain the sentiment tag and topic tag corresponding to each target public opinion text. Specifically, this is used for: The word vectors of the target public opinion text are extracted using the converter-based bidirectional encoder representation model; The word vectors of the target public opinion text are analyzed using the bidirectional long short-term memory model to obtain the sentiment tags of the target public opinion text; Calculate the perplexity curve of the implicit Dirichlet distribution topic extraction model; wherein, the horizontal axis of the perplexity curve is the number of topic tags, and the vertical axis is the perplexity. The word vectors of the target public opinion text are analyzed using the Latent Dirichlet Distribution Topic Extraction Model to obtain N topic tags for the target public opinion text, and the topic tags corresponding to each target public opinion text are determined; wherein, N is the number of topic tags corresponding to the lowest point in the perplexity curve.

6. The apparatus according to claim 5, characterized in that, When the sampling unit acquires target public opinion texts published by users of the target network platform within a preset sampling period, it is specifically used for: Acquire massive amounts of text published on the target network platform within a preset sampling period; The massive amount of text is filtered to remove texts that are not related to the preset target public opinion event, thus obtaining the target public opinion text.

7. The apparatus according to claim 5, characterized in that, When determining the user category for each user in each time period within the preset sampling period based on the sentiment tags and topic tags of the target public opinion text published by each user in the target network platform within each time period, the determining unit is specifically used for: For each user on the target network platform, the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period is statistically analyzed. For each user on the target network platform, the user category for each time period within the preset sampling period is determined based on the distribution of sentiment tags and topic tags of the target public opinion texts published by the user in each time period within the preset sampling period.

8. The apparatus according to claim 5, characterized in that, The simulation unit, based on the propagation dynamics parameters and the user category of each user in the current target network platform, performs simulation analysis on public opinion control measures to evaluate their effectiveness. Specifically, it is used for: Determine the predicted update amount of the propagation dynamics parameters after the implementation of public opinion control measures; The propagation dynamics parameters are updated based on the predicted update amount; Based on the updated propagation dynamics parameters and the user category to which each user belongs in the current target network platform, a propagation dynamics model simulation is performed to obtain simulation results; wherein, the simulation results include at least the distribution of each user category in the target network platform after a preset control period; The degree of similarity between the simulation results and the expected control results of the public opinion control measures is used to determine the effectiveness of the public opinion control measures.

Images