Public opinion response effect measurement method based on theme migration and emotion change recognition

Abstract

The application discloses a public opinion response effect measurement method based on theme migration and emotion change recognition. The method takes relevant microblog hot search blog posts and comment information of an event as basic data, performs theme analysis on the microblog text based on an LDA theme model, calculates the emotion value of the comment content according to a Bi-LSTM model, constructs a measurement model of the response effect under a sudden negative public opinion, and evaluates the response effect based on the recognition results of theme migration and emotion change. The application has the advantages of easy practice, scientific index, strong pertinence and the like, can be used for measuring the response effect in a sudden negative public opinion, evaluating the intervention effect on the evolution of the negative public opinion, analyzing the advantages and disadvantages of the response measures, and providing optimization strategies for optimizing the negative public opinion management work.

Description

Technical Field

[0001] This invention belongs to the field of topic recognition and sentiment recognition, and specifically relates to a technique for measuring the effectiveness of public opinion response. Background Technology

[0002] In recent years, public health emergencies have become increasingly frequent. These emergencies often attract high levels of attention. If not responded to promptly and appropriately, they can quickly escalate on social media, triggering large-scale negative public opinion. Negative public opinion refers to the negative and pessimistic social attitudes held by the public towards a specific public event. If negative public opinion is not handled in a timely and effective manner, it can easily amplify, damage public image, and exacerbate social conflicts.

[0003] Social media platforms, primarily microblogs, are the mainstream platforms for public attention and discussion of negative public opinion events. They also serve as important propaganda channels for responding to and influencing public opinion, possessing both information dissemination and the ability to represent public sentiment, thus becoming a crucial interactive channel for online public opinion. The data from microblog posts and comments, characterized by their temporal sequence, social nature, and richness, are frequently used in research on the propagation mechanisms, evolution paths, and pre-event detection of public opinion events.

[0004] In negative public opinion, both responses and public opinion interact through written symbols, meaning the interaction between these two discourse systems can be clarified based on a "discourse-emotion" mechanism. The effectiveness of responses in negative public opinion can be assessed by evaluating the shift in public attention and by measuring the positive evolution of public sentiment.

[0005] Response refers to the process of actively and continuously reacting to and replying to the demands and questions raised by the public. Existing research on response can be summarized from two perspectives: one based on theoretical mechanisms, and the other based on technical methods. This paper analyzes the impact of response on public opinion evolution using system dynamics modeling methods and establishes a mechanism encompassing public opinion evolution; based on emotion evaluation theory, it constructs the formation mechanism of secondary public opinion from the perspective of response; it reviews the mechanism of disordered response in some negative public opinion, proposing that public opinion response should focus on public sentiment; and based on survey and news data, it explores the causal relationship between public opinion and response.

[0006] With the rise of natural language processing technology, big data analytics has been applied to research on responding to public opinion. MuniZhuang et al. proposed the LDA-ARMA model for large-scale public opinion evolution analysis and analyzed the impact of responses on public opinion development based on Weibo data. Zheng Shiming et al. extracted relevant text based on the LDA topic model and analyzed the interaction mechanism of responding to public concerns during sudden public opinion events. Wen Hong used topic extraction and sentiment analysis techniques to construct a feedback graph of netizens' responses during sudden public opinion events. Deng Weihua et al. used sentiment analysis to analyze the effectiveness of responses in the evolution of public opinion.

[0007] In summary, existing technologies focus on analyzing the evolution mechanism of responses to public opinion, with little research on measurement techniques for the effectiveness of responses to negative public opinion. Summary of the Invention

[0008] The purpose of this invention is to provide a method for measuring the effectiveness of public opinion responses based on topic migration and sentiment change recognition, so as to evaluate the effectiveness of responses, analyze the advantages and disadvantages of existing responses, and provide technical method support for finding optimization strategies for response measures.

[0009] To address the aforementioned technical issues, this invention uses trending trending posts and comments related to the event as foundational data. It primarily constructs a framework for measuring the effectiveness of public opinion responses from two aspects: hot topics and sentiment characteristics. The framework employs an LDA topic model for topic extraction and a Bi-LSTM model for sentiment value calculation, exploring topic evolution and sentiment changes over time. Furthermore, it combines public opinion response measures to measure the intervention effect of the response in the evolution of public opinion. The specific technical solutions adopted are as follows.

[0010] A method for measuring the effectiveness of public opinion response based on topic shift and sentiment change recognition is characterized by the following steps:

[0011] Step 1, Information Extraction: Use trending Weibo posts and comments related to the event as the basic data.

[0012] 1.1 Data Extraction for Response Measures

[0013] Use a self-written Python web crawler to crawl relevant public opinion event announcements, and at the same time crawl the text content of relevant media reports to obtain specific response measures and timelines;

[0014] 1.2 Extraction of Weibo Public Opinion Data

[0015] From the time the event began to become a hot topic of public opinion, we crawled the Weibo trending posts and their comments, obtained the Weibo text and comments, and cleaned the data. We used Python regular expressions to remove interfering information, and also removed overly short and duplicate Weibo text. The cleaned Weibo text is the Weibo data.

[0016] Disruptive information includes HTTPS links, emoticons, tags, and @users; "too short" refers to fewer than 10 characters.

[0017] 1.3 Extraction of Key Response Measures

[0018] Based on the public opinion lifecycle of key response measures, and taking into account a series of public opinion response measures, while referring to the Baidu Index of the event, N key response measures are extracted from the reports of various response measures published by mainstream media; combined with the nodes of the event, the public opinion lifecycle is divided into N+1 stages, which are used to analyze the intervention effect of the response on the hot topics and public sentiment evolution of this public opinion event in combination with time series analysis; N is a natural number.

[0019] Step 2, Topic Transfer Recognition

[0020] Public opinion events are often a series of events consisting of a primary event and secondary events. From the perspective of the migration of hot topics, the effectiveness of the response is measured based on two indicators: "attention" and "impact". The "attention" indicator reflects the effectiveness of the response in meeting the real needs of the public, while the "impact" indicator reflects the effectiveness of the response in causing public opinion to deviate from negative topics.

[0021] LDA topic model is used to extract topic information, and LDAvis tool is used to visualize the topics. That is, the topic extraction results are analyzed in two-dimensional space so as to fine-tune the number of topics according to their overlap. LDA topic model uses some keywords to represent topic content, and LDAvis tool uses different circles to represent topics, and the diameter of the circle represents the amount of text in the topic.

[0022] The system segments key response measures within the public opinion lifecycle to obtain the impact of each response measure on public opinion, and visualizes all themes at each stage using word clouds. It determines the optimal number of themes (M) for each stage of the public opinion lifecycle, and uses several keywords from each theme as corpus to create a word cloud to represent all thematic content at that stage. If the selected M themes have no overlap in two-dimensional space and are far apart, it indicates that the pre-selected number of M themes is acceptable. Thematic information is extracted from the time series of the public opinion lifecycle to identify thematic shifts; M is a natural number.

[0023] Step 3: Recognizing Emotional Changes

[0024] From the perspective of changes in public sentiment, the effectiveness of the response is measured using two indicators: “reversal” and “mitigation”. The “reversal” indicator reflects the effect of the response in transforming negative public sentiment into positive sentiment, while the “mitigation” indicator reflects the effect of the response in smoothing out negative fluctuations in public sentiment.

[0025] Based on cleaned Weibo comment data, a line graph of public sentiment trends for the event was drawn, with one hour as the basic unit. The analysis showed whether the public sentiment was generally in a negative or positive range, whether the fluctuations were intense or gradual, the maximum and minimum values ​​of the emotional fluctuations, the time period during which the peak of public opinion discussion occurred, and the corresponding response measures released during that time period.

[0026] Step four: Measure the effectiveness of the public opinion response based on the shift in theme and changes in sentiment.

[0027] The effectiveness of a response in raising awareness or influencing public opinion is determined by the identified shift in themes: when the focus of public opinion remains consistent before and after the response, it indicates that the response has raised awareness; when there is a significant difference in the focus of public opinion before and after the response, it indicates that the response has not raised awareness; when public opinion deviates from negative topics, it indicates that the response has influenced public opinion; and when public opinion consistently revolves around negative topics, it indicates that the response has not influenced public opinion.

[0028] The effectiveness of the response is determined based on the identified changes in sentiment: when the public's negative sentiment turns into positive sentiment, it indicates a reversal effect; when the public's sentiment remains negative, it indicates no reversal effect; when the public's sentiment change value shows an upward trend, it indicates a mitigation effect; when the public's sentiment change value shows negative fluctuations, it indicates no mitigation effect.

[0029] The specific method for measuring the attention effect is as follows:

[0030] In order to better explore the effect of responding to negative public opinion in meeting the real needs of the public, i.e., the attention effect, the cleaned Weibo text was divided into two parts: before the response (i.e., the outbreak of public opinion) and after the response (i.e., the response measures). The themes of these two parts were extracted, and the "attention effect" of the response was judged based on the results.

[0031] The visualization results of the topics were repeatedly adjusted to determine the number of topics in the text before the response as M1 and the number of topics in the text after the response as M2. Word clouds were created for the keywords under the M1 topics and the M2 topics to represent the hot topics of public opinion before and after the response. When the hot topics of public opinion are consistent before and after the response, it indicates that the public's concerns have been positively addressed and the "attention" role that meets the public's real needs has been well played.

[0032] The method for measuring the impact is designed to more intuitively reflect the effect of different response measures on deviating public opinion from negative topics. A time window function is used to divide the cleaned Weibo text into I parts based on key response measures. Thematic analysis and summarization are then performed on each part to measure the impact of the response. The specific procedure is as follows:

[0033] Based on the theme visualization results, the number of themes in each part is determined to be J. To more intuitively reflect the theme content, a theme word cloud is created using the high-frequency keywords of each part, thereby identifying the most important keywords at each stage of the public opinion life cycle. This allows for in-depth analysis of the theme content at each stage of response measures, clarifying the changes in public attention throughout the entire public opinion life cycle. If the attention remains unchanged, it indicates that the response has failed to effectively deflect public opinion from negative topics.

[0034] The method for measuring the reversal effect is to calculate the average public sentiment at each node from the start of the public opinion outbreak, based on the extracted N key response measures, in order to observe the reversal effect of responding to public opinion on public sentiment. The specific procedure is as follows:

[0035] The average sentiment value Q1 of all comments from the occurrence of the event to the first key response measure is used to obtain the mean sentiment value Q during the outbreak phase of public opinion. 1均 ; and so on, to obtain the mean sentiment Q of the nth key response measure node. n均 n=1~N;

[0036] A univariate linear regression equation is used to express the trend of the mean change in public sentiment value: if the mean public sentiment value shows a downward trend and the endpoint is negative sentiment, it indicates that the response has failed to transfer the public's negative sentiment to positive sentiment; if the mean public sentiment value shows an upward trend and the endpoint is positive sentiment, it indicates that the response has played a role in transferring the public's negative sentiment to positive sentiment.

[0037] The method for measuring the mitigation effect is to start with the sentiment value of the public opinion outbreak node, calculate the difference between the sentiment value of the current key response measure node and the previous node to obtain the change value of public sentiment at the current node, and then use the obtained change value of public sentiment to detect the mitigation effect of the response measures on public sentiment. The specific steps are as follows:

[0038] Plot a trend chart of public sentiment fluctuations during public opinion events. Divide the events into stages based on the upward or downward trend of public sentiment change values, identify the maximum and minimum values ​​of public sentiment change values ​​in each stage, and the corresponding key response measures. Construct a univariate linear regression equation to represent the evolution trend of public sentiment change values. If the public sentiment change value shows an upward trend, it indicates that the response measures in this event have played a role in mitigating the negative changes in public sentiment, that is, they have a mitigating effect.

[0039] The LDA topic model mentioned above is a document topic generation model proposed by Blei et al. in 2003. It includes a three-level structure of document, topic and term. It is used to identify potential topic information in the corpus, reveal the content structure of public opinion events, and explore the evolution trend of public opinion topics.

[0040] The cleaned Weibo text was analyzed using the open-source Python library Gensim. The optimal number of topics was determined using perplexity and topic consistency measures. Lower perplexity indicates better topic analysis performance. Perplexity is commonly used to evaluate the quality of a probabilistic model's predictions; lower perplexity indicates better topic analysis performance. The perplexity calculation formula is as follows:

[0041]

[0042] D represents the test set in the document, M represents the number of documents, Wd represents the words belonging to document d, Nd represents the number of words of d in each document, and P(Wd) represents the probability of Wd appearing.

[0043] Topic consistency is used to calculate the semantic relevance of words under a specific topic generated by LDA. The value ranges from -1 to 1, with a higher value indicating better topic analysis results. The formula for calculating topic consistency is as follows:

[0044] ;

[0045] D(vi,vj) represents the number of documents in which the words vi and vj appear simultaneously, D(vi) is the number of documents containing the word vi, and ε is a smoothing coefficient used to avoid the numerator being 0;

[0046] Since the effectiveness of the LDA topic model is inversely proportional to the average similarity of topic structures, the optimal number of topics is selected based on the following: low perplexity, high topic consistency, and relatively small number of topics.

[0047] The specific approach and method for identifying emotional changes are as follows:

[0048] Because Weibo comment texts are a type of text with strong emotional color in a specific context, meaning that the semantics of their words are related to the words before and after them and the context, sentiment analysis tasks for this type of comment text need to fully consider the long-term dependencies of words in order to accurately grasp the local semantics of words and the global text features, and obtain more accurate sentiment analysis.

[0049] The Bi-LSTM model, or Bi-Long Short-Term Memory model, consists of a forward LSTM and a backward LSTM recursively. It inherits the input, forget, and output gate mechanisms of LSTM. Taking all the words of the entire comment text as input, it can capture the contextual features of the comment text from both directions, which is helpful in solving the problem of short texts being difficult to classify for sentiment. The Bi-LSTM model was built using the open-source PaddleNLP library, and the model was trained and optimized using open-source Weibo comment annotation corpus, achieving an accuracy of 90.8%. The calculation result ranges from [0,1]: if the value is between 0.5 and 1, it indicates that the sentiment of the text is relatively positive, and the larger the value, the more positive it is; if the value is between 0 and 0.5, it indicates that the sentiment of the text is relatively negative, and the smaller the value, the more negative it is.

[0050] The present invention has the following beneficial effects:

[0051] This invention addresses the crucial role of social media platforms like Weibo in disseminating information about major public opinion events. Evaluating the effectiveness of responses to negative public opinion based on social media data is significant for public opinion governance. By crawling relevant Weibo trending posts and comments, this invention extracts themes from the posts using the LDA topic model and performs sentiment analysis on the comments using the Bi-LSTM model. It proposes a framework for measuring the effectiveness of responses to sudden negative public opinion events based on theme and sentiment dimensions, and conducts an empirical study using relevant Weibo data. The empirical results demonstrate that this measurement framework is easy to implement, has scientifically sound indicators, and is highly targeted. It is beneficial for analyzing the shortcomings of response measures and can provide technical support for optimizing negative public opinion governance strategies. Attached Figure Description

[0052] Figure 1 This invention provides a framework for measuring the effectiveness of public opinion response.

[0053] Figure 2 This is the Bi-LSTM model of the present invention;

[0054] Figure 3 This is the probability diagram of the LDA topic model of the present invention;

[0055] Figure 4 This invention relates to the public opinion lifecycle based on key response measures;

[0056] Figure 5 This is a visualization of the results of the invention.

[0057] Figure 6 This is a word cloud display of the subject matter of this invention;

[0058] Figure 7This is a comparison of the topic word cloud before and after the response to this invention (left side is before the response, right side is after the response);

[0059] Figure 8 A keyword cloud representing the key response measures of this invention;

[0060] Figure 9 This is a line graph showing the changes in the number of public comments and the average sentiment of this invention.

[0061] Figure 10 The Baidu Index trend for this invention event;

[0062] Figure 11 This is a line graph showing the changes in public sentiment based on key response measures in this invention.

[0063] Figure 12 This is a line graph showing the fluctuation of public sentiment based on key response measures in this invention. Detailed Implementation

[0064] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Using Weibo data related to the event as the research object, this study aims to analyze the intervention effect of responses on the thematic evolution and emotional shifts of public opinion, thereby constructing a framework for measuring the effectiveness of responses. The event sparked heated discussions on Weibo, with a relatively rich corpus that facilitates the analysis of its thematic evolution and emotional tendencies. Furthermore, the responses to this event had a predictable impact on the development of public opinion, providing a possibility for measuring the effectiveness of public opinion interaction from a holistic perspective.

[0066] Using trending Weibo posts and comments related to the event as basic data, this study constructs a framework for measuring the effectiveness of public opinion response, primarily focusing on two aspects: trending topics and sentiment characteristics. The LDA topic model is used for topic extraction, and a Bi-LSTM model is used to calculate sentiment values, exploring topic evolution and sentiment changes over time. Furthermore, combined with public opinion response measures, the intervention effect of the response in the evolution of public opinion is analyzed. The methodology is as follows: Figure 1 As shown.

[0067] The construction of the Bi-LSTM model is as follows: Figure 2 As shown. The probability diagram of LDA is as follows. Figure 3 As shown, the relevant parameters are explained in Table 1.

[0068]

[0069] Step 1: Data Acquisition and Preprocessing

[0070] 1.1 Data Extraction for Response Measures

[0071] Since the incident began, related reports have been published on mainstream media platforms such as Weibo. Therefore, this invention uses a self-developed Python web crawler to collect relevant reports about the incident and obtain specific response measures and timelines.

[0072] 1.2 Extraction of Weibo Public Opinion Data

[0073] This invention crawled 7045 Weibo posts and 34108 comments related to a specific event, using trending Weibo posts and their comments. All data was stored in XLSX format and cleaned using Python regular expressions to remove interfering information such as HTTPS links, emoticons, tags, and @mentions, while also removing excessively short and duplicate text. Examples of the final Weibo posts and comments are shown in Tables 2 and 3.

[0074] .

[0075] 1.3 Extraction of Key Response Measures

[0076] Based on the entire lifecycle of this public opinion event, and considering the series of response measures taken, along with the Baidu Index of the event, this invention extracts five key response measures—Response Measure 1, Response Measure 2, Response Measure 3, Response Measure 4, and Response Measure 5—from media reports on these measures. Combining the event's timeline, this invention divides the public opinion event lifecycle into six stages. Using time series analysis, it can more intuitively detect the intervention effect of the responses on the hot topics and the evolution of public sentiment in this public opinion event, such as... Figure 4 .

[0077] Step 2, Topic Transfer Recognition

[0078] An event is a series of public opinion events consisting of a primary event and secondary events. Therefore, from the perspective of the migration of hot topics, the measurement of response effectiveness mainly includes attention effect and impact effect.

[0079] LDA topic modeling is used to extract topic information, and LDAvis tools such as Sievert are used to visualize the topics. This involves analyzing the topic extraction results in a two-dimensional space to fine-tune the number of topics based on their overlap. In practice, key response measures within the public opinion lifecycle are segmented to obtain the impact of each response measure on public opinion, and the topics of each stage are visualized using word clouds. Since each stage within the public opinion lifecycle contains multiple topics, this invention uses several keywords of each topic as corpus to create a word cloud to represent all topic content of that stage. Taking the topic extraction results of response measure 2 as an example, the LDA topic model uses some keywords to represent the topic content. The keyword set for the first topic is "keyword 1-keyword 2-keyword 3-keyword 4-keyword 5-keyword 6-keyword 7-keyword 8-keyword 9-keyword 10-keyword 11-keyword 12-keyword 13-keyword 14-keyword 15-keyword 16-keyword 17-keyword 18..." etc. Figure 5 As shown. When displaying topic extraction results, the LDAvis tool uses different circles to represent topics, with the diameter of the circle representing the amount of text within that topic. It can be seen that in... Figure 5 In the selected topic cloud, the four topics have no overlap in two-dimensional space and are relatively far apart, indicating that the number of pre-selected topics is acceptable. Simultaneously, the topic cloud also highlights keywords such as "Topic 1," "Topic 2," and "Topic 3," demonstrating that the topic cloud can effectively represent the extracted topic information. Figure 6 .

[0080] 2.1 Focus on the measurement of effects

[0081] The Weibo text is divided into two parts: before the response (i.e., the outbreak of public opinion) and after the response (i.e., the response measures). The themes of these two parts are extracted, and the "attention" effect of the response is judged based on the results.

[0082] After repeated adjustments to the topic visualization results, it was determined that 3 topics were appropriate for the pre-response text and 5 topics were appropriate for the post-response text. Word clouds were then created based on the keywords under each topic before and after the response to represent the public's focus on key issues. A comparison of the post-response topic word clouds is shown below. Figure 7 As shown, it can be seen that the hot topics of public opinion before and after the response were quite consistent.

[0083] Based on the above analysis, it can be seen that the focus of public opinion remained largely consistent before and after the response. Therefore, in this instance of negative public opinion, the response positively addressed public concerns and effectively fulfilled its role of "attention" in meeting the public's genuine needs.

[0084] 2.2 Measurement of Influence Effects

[0085] Because different response measures were taken multiple times throughout the event's lifecycle, this study uses a time window function to divide the Weibo text into six parts based on key response measures. Each part is then analyzed for its own theme, and the results are summarized to measure the impact of the responses.

[0086] Based on the topic visualization results, this invention determines the number of topics for each part to be 4. To more intuitively represent the topic content, this invention uses high-frequency keywords from each part to create a word cloud, with the topic word cloud diagrams arranged from left to right and top to bottom as follows: Figure 8 As shown, it can be observed that throughout the various stages of the public opinion lifecycle, the focus of public attention consistently revolves around this event. However, this only indicates that the response is relatively aligned with the public's actual needs. To evaluate the impact of the response, a deeper analysis of the thematic content at each stage of the response measures is required.

[0087] During the initial public opinion crisis, the public's focus was on the event itself; in the first phase of response measures, the public expressed their views on the event; in the second phase of response measures, the public continued to discuss the event; in the third phase of response measures, the public paid attention to the handling of the event; in the fourth phase of response measures, the public discussed the impact of the event; and in the fifth phase of response measures, the public remained concerned about the subsequent impact of the event.

[0088] In summary, although multiple responses were taken during this negative public opinion crisis, and the focus of public attention shifted after each response, the overall focus remained on the event itself throughout its lifecycle. Therefore, the responses failed to effectively divert public attention from the negative narrative.

[0089] Step 3: Recognizing Emotional Changes

[0090] This event is a typical case of response under a crisis perspective. From the perspective of changes in public sentiment, the effectiveness of the response is mainly measured by two indicators: reversal and mitigation. Based on the obtained Weibo comment data, a line graph of the public sentiment trend of the event was plotted, using one hour as the basic unit, as shown below. Figure 9 It can be seen that public sentiment regarding this event was generally negative, with significant fluctuations, ranging from a maximum of 0.467 to a minimum of 0.007. Furthermore, the peak of public discussion occurred from January 5th to 8th, which is positively correlated with the continuous release of response measures, such as... Figure 10 .

[0091] 3.1 Measurement of Reversal Effect

[0092] according to Figure 9 It can be seen that the public opinion response measures during the event had a significant impact on public sentiment, causing it to fluctuate dramatically. However, it is difficult to analyze the "reversal" and "alleviation" effects of these response measures on public sentiment. Therefore, this invention, based on the five key response measures selected above, calculates the average public sentiment at each stage from the outbreak of public opinion, in order to observe the "reversal" effect of the response on public sentiment. Specifically, the sentiment values ​​of all comments from the occurrence of the event to response measure 1 are averaged to obtain the average sentiment value during the outbreak stage of public opinion. This process is repeated to obtain the average sentiment value at all key response measure stages.

[0093] Figure 11 This reflects the changing trend of public sentiment during the event, which can be divided into three stages. The first stage shows a downward trend in public sentiment, from the outbreak of the public opinion crisis to response measure 2, a total of two response measures. The outbreak of the public opinion crisis marks the highest sentiment value at 0.137, while the lowest sentiment value occurs at response measure 2 at 0.108, which is also the lowest public sentiment value for the entire public opinion event. The second stage, from response measure 2 to response measure 4, shows an upward trend. The highest sentiment value occurs at response measure 4 at 0.147, which is also the highest point of public sentiment value for the entire public opinion event. The final stage, response measure 4, uses a univariate linear regression equation to express the changing trend of public sentiment value (y = -0.0003x + 0.1263), with a slope of -0.0003.

[0094] Overall, despite the intervention of the response measures, public sentiment in this public opinion event showed a downward trend, with the final sentiment value slightly lower than the initial value. Therefore, the response measures failed to successfully "reverse" negative sentiment into positive sentiment in this event, meaning the "reversal" effect was not achieved.

[0095] 3.2 Measurement of Relief Effect

[0096] Starting with the sentiment value at the point where public opinion erupts, the change value of the current key response measure node is obtained by subtracting the sentiment value of the previous node from the current key response measure node. Based on the obtained change value of public sentiment, the "alleviation" effect of the response measures on public sentiment is detected.

[0097] The trend of public sentiment fluctuation in this public opinion event is as follows: Figure 12 As shown, the event can be divided into two main phases. The first phase, from response measure 1 to response measure 4, shows an upward trend in public sentiment change, with a maximum value of 0.022, located at the node of response measure 4. The second phase, namely response measure 5, shows a downward trend in public sentiment change, reaching the lowest point of the entire public opinion event at -0.032.

[0098] Similarly, a univariate linear regression equation was constructed to represent the evolution trend of public sentiment change (y = 0.0004x - 0.0057), with a slope of 0.0004. It can be seen that the intervention of the response measures caused the public sentiment change value of the entire public opinion event to show an upward trend. Therefore, in this event, the response measures played a role in "mitigating" the negative change in public sentiment to a certain extent.

[0099] Using the method of this invention, this study measures the intervention effect of responses on public opinion based on event-related Weibo data, and draws the following main conclusions: ① In terms of theme shift, public opinion remained focused on the event under the influence of responses, with occasional related themes emerging. Although responses failed to effectively shift public opinion away from negative topics, they effectively captured public attention in a way that met the public's real needs. ② In terms of emotional changes, the public's overall sentiment was relatively negative and pessimistic, but positive emotions such as concern, hope, and gratitude also emerged. Overall, the frequency and amplitude of emotional fluctuations were high. Although responses failed to successfully reverse the public's negative emotions into positive ones, they did, to some extent, alleviate the negative changes in public sentiment. ③ The framework for measuring the effect of responses under sudden negative public opinion based on theme and emotional dimensions can effectively explore the intervention effect of responses on the evolution of negative public opinion.

[0100] Based on the above conclusions, the shortcomings of the response measures are analyzed, and the following optimization strategies are provided for the governance of negative public opinion.

[0101] ① When reporting negative news events, the principles of information disclosure and truthfulness should be upheld. Event updates should be provided promptly, and targeted emergency measures should be taken as early as possible. Effective responses to the public's genuine needs should be made in the early stages of an event's escalation to gain public support and understanding. Simultaneously, public attention should be shifted to the approval, recognition, and promotion of positive cases. ② When responding to negative public opinion, the "golden four hours" must be seized to swiftly address the issue, curbing its continued deterioration. Simultaneously, timely investigations and accountability measures should be implemented, related work improved, and the progress of public opinion management continuously updated. These actions should effectively "reverse" negative sentiment and "influence" the shift in public opinion focus. ③ In major public opinion events, opinion leaders play a crucial role in guiding netizens' attention. Their influence in the dissemination and discussion of public opinion should be valued, and "two-way communication" channels should be established with them to prevent the trajectory of negative public opinion events from becoming uncontrollable.

Claims

1. A method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition, characterized by the following steps: Step 1, Information Extraction: Use trending Weibo posts and comments related to the event as the basic data. 1.1 Data Extraction for Response Measures Use Python web crawlers to crawl relevant public opinion event announcements, and also crawl the text content of relevant reports from mainstream media to obtain specific response measures and timelines; 1.2 Extraction of Weibo Public Opinion Data From the time the event began to become a hot topic of public opinion, we crawled the Weibo trending posts and their comments, obtained the Weibo text and comments, and cleaned the data. We used Python regular expressions to remove interfering information, and also removed overly short and duplicate Weibo text. The cleaned Weibo text is the Weibo data. Disruptive information includes HTTPS links, emoticons, tags, and @users; "too short" refers to fewer than 10 characters. 1.3 Extraction of Key Response Measures Based on the public opinion lifecycle of key response measures, and taking into account a series of public opinion response measures, while referring to the Baidu Index of the event, N key response measures are extracted from the reports of various response measures published by mainstream media; combined with the original post time that triggered the public opinion, the public opinion lifecycle is divided into N+1 stages, which are used to analyze the intervention effect of the response on the hot topics and public sentiment evolution of this public opinion event in combination with time series analysis; N is a natural number. Step 2, Topic Transfer Recognition Public opinion events are often a series of events consisting of a primary event and secondary events. From the perspective of the migration of hot topics, the effectiveness of the response is measured based on two indicators: "attention" and "impact". The "attention" indicator reflects the effectiveness of the response in meeting the real needs of the public, while the "impact" indicator reflects the effectiveness of the response in causing public opinion to deviate from negative topics. LDA topic model is used to extract topic information, and LDAvis tool is used to visualize the topics. That is, the topic extraction results are analyzed in two-dimensional space so as to fine-tune the number of topics according to their overlap. LDA topic model uses some keywords to represent topic content, and LDAvis tool uses different circles to represent topics, and the diameter of the circle represents the amount of text in the topic. The system segments key response measures within the public opinion lifecycle to obtain the impact of each response measure on public opinion, and visualizes all themes at each stage using word clouds. It determines the optimal number of themes (M) for each stage of the public opinion lifecycle, and uses several keywords from each theme as corpus to create a word cloud to represent all thematic content at that stage. If the selected M themes have no overlap in two-dimensional space and are far apart, it indicates that the pre-selected number of M themes is acceptable. Thematic information is extracted from the time series of the public opinion lifecycle to identify thematic shifts; M is a natural number. Step 3: Recognizing Emotional Changes From the perspective of changes in public sentiment, the effectiveness of the response is measured using two indicators: "reversal" and "mitigation." The "reversal" indicator reflects the effect of the response in transforming negative public sentiment into positive sentiment, while the "mitigation" indicator reflects the effect of the response in smoothing out negative fluctuations in public sentiment. Based on the cleaned Weibo comment data, a line graph of public sentiment trend of the public opinion event is drawn with one hour as the basic unit. The analysis is to see whether the public sentiment is generally in a negative or positive range, whether the fluctuation is intense or gentle, the maximum and minimum values ​​of sentiment fluctuation, the time period in which the public opinion hot discussion peaks, and the release of response measures during the corresponding time period. Step four: Measure the effectiveness of the public opinion response based on the shift in theme and changes in sentiment. The effectiveness of a response in raising awareness or influencing public opinion is determined by the identified shift in themes: when the focus of public opinion remains consistent before and after the response, it indicates that the response has raised awareness; when there is a significant difference in the focus of public opinion before and after the response, it indicates that the response has not raised awareness; when public opinion deviates from negative topics, it indicates that the response has influenced public opinion; and when public opinion focuses on negative topics, it indicates that the response has not influenced public opinion. The effectiveness of the response is determined based on the identified changes in sentiment: when the public's negative sentiment turns into positive sentiment, it indicates a reversal effect; when the public's sentiment remains negative, it indicates no reversal effect; when the public's sentiment change value shows an upward trend, it indicates a mitigation effect; when the public's sentiment change value shows negative fluctuations, it indicates no mitigation effect.

2. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The specific method for measuring the attention effect is as follows: In order to better explore the effect of responding to negative public opinion in meeting the real needs of the public, i.e., the attention effect, the cleaned Weibo text was divided into two parts: before the response (i.e., the outbreak of public opinion) and after the response (i.e., the response measures). The themes of these two parts were extracted, and the "attention effect" of the response was judged based on the results. The visualization results of the topics were repeatedly adjusted to determine the number of topics in the text before the response as M1 and the number of topics in the text after the response as M2. Word clouds were created for the keywords under the M1 topics and the M2 topics to represent the hot topics of public opinion before and after the response. When the hot topics of public opinion are consistent before and after the response, it indicates that the public's concerns have been positively addressed and the "attention" role that meets the public's real needs has been well played.

3. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The method for measuring the impact is designed to more intuitively reflect the effect of different response measures on deviating public opinion from negative topics. A time window function is used to divide the cleaned Weibo text into I parts based on key response measures. Thematic analysis and summarization are then performed on each part to measure the impact of the response. The specific procedure is as follows: Based on the theme visualization results, the number of themes in each part is determined to be J. To more intuitively reflect the theme content, a theme word cloud is created using the high-frequency keywords of each part, thereby identifying the most important keywords at each stage of the public opinion life cycle. This allows for in-depth analysis of the theme content at each stage of response measures, clarifying the changes in public attention throughout the entire public opinion life cycle. If the attention remains unchanged, it indicates that the response has failed to effectively deflect public opinion from negative topics.

4. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The method for measuring the reversal effect is to calculate the average public sentiment at each node from the start of the public opinion outbreak, based on the extracted N key response measures, in order to observe the reversal effect of responding to public opinion on public sentiment. The specific procedure is as follows: The average sentiment value Q1 of all comments from the occurrence of the event to the first key response measure is used to obtain the mean sentiment value Q during the outbreak phase of public opinion. 1均 ; and so on, to obtain the mean sentiment Q of the nth key response measure node. n均 n=1~N; A univariate linear regression equation is used to express the trend of the mean change in public sentiment value: if the mean public sentiment value shows a downward trend and the endpoint is negative sentiment, it indicates that the response has failed to transfer the public's negative sentiment to positive sentiment; if the mean public sentiment value shows an upward trend and the endpoint is positive sentiment, it indicates that the response has played a role in transferring the public's negative sentiment to positive sentiment.

5. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The method for measuring the mitigation effect is as follows: starting with the sentiment value of the public opinion outbreak node, the difference between the sentiment value of the current key response measure node and the previous node is calculated to obtain the change value of public sentiment at the current node. Based on the obtained change value of public sentiment, the mitigation effect of the key response measures on public sentiment is detected. The specific steps are as follows: Draw a trend chart of public sentiment fluctuations during public opinion events, divide the stages according to the rising or falling trend of public sentiment change values, and find the maximum and minimum values ​​of public sentiment change values ​​in each stage and the corresponding key response measures nodes. A univariate linear regression equation is constructed to represent the evolution trend of public sentiment change value. If the public sentiment change value shows an upward trend, it indicates that the key response measures in this event have played a role in mitigating the negative change of public sentiment, that is, they have a mitigating effect.

6. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The LDA topic model includes a three-level structure of documents, topics, and terms, which is used to identify potential topic information in corpora, reveal the content structure of public opinion events, and explore the evolution trend of public opinion topics. The cleaned Weibo text was analyzed using the open-source Python library Gensim. The optimal number of topics was determined using perplexity and topic consistency measures. Lower perplexity indicates better topic analysis performance. Perplexity is commonly used to evaluate the quality of a probabilistic model's predictions; lower perplexity indicates better topic analysis performance. The perplexity calculation formula is as follows: ； D represents the test set in the document, M represents the number of documents, Wd represents the words belonging to document d, Nd represents the number of words of d in each document, and P(Wd) represents the probability of Wd appearing. Topic consistency is used to calculate the semantic association of words under a certain topic generated by LDA. The value ranges from -1 to 1. The larger the value, the better the topic analysis effect. The formula for calculating topic consistency is as follows: ； The formula for the score function is as follows: ； D(vi,vj) represents the number of documents in which the words vi and vj appear simultaneously, D(vi) is the number of documents containing the word vi, and ε is a smoothing coefficient used to avoid the numerator being 0; Since the effectiveness of the LDA topic model is inversely proportional to the average similarity of topic structures, the optimal number of topics is selected based on the following: low perplexity, high topic consistency, and relatively small number of topics.

7. The method for measuring the effectiveness of public opinion response based on topic migration and sentiment change recognition according to claim 1, characterized in that... The specific approach and method for identifying emotional changes are as follows: Because Weibo comment texts are a type of text with strong emotional color in a specific context, meaning that the semantics of their words are related to the words before and after them and the context, sentiment analysis tasks for this type of comment text need to fully consider the long-term dependencies of words in order to accurately grasp the local semantics of words and the global text features, and obtain more accurate sentiment analysis. The Bi-LSTM model, or Bi-Long Short-Term Memory model, consists of a forward LSTM and a backward LSTM recursively. It inherits the input, forget, and output gate mechanisms of LSTM. Taking all the words of the entire comment text as input, it can capture the contextual features of the comment text from both directions, which is helpful in solving the problem of short texts being difficult to classify for sentiment. The Bi-LSTM model was built using the open-source PaddleNLP library, and the model was trained and optimized using open-source Weibo comment annotation corpus, achieving an accuracy of 90.8%. The calculation result ranges from [0,1]: if the value is between 0.5 and 1, it indicates that the sentiment of the text is relatively positive, and the larger the value, the more positive it is; if the value is between 0 and 0.5, it indicates that the sentiment of the text is relatively negative, and the smaller the value, the more negative it is.

Images