Quantification and Analysis of Group Sentiment in Electromagnetic Radiation Public Opinion Events

Abstract

This research focuses on developing a sentiment-based system to analyze public opinion on electromagnetic radiation in online networks. Issues related to EMR, such as the NIMBY effect and negative public sentiment, can lead to health crises, social conflicts, and challenges in decision-making. This study addresses limitations in existing research, including inaccurate data collection and a lack of systematic analysis. By incorporating Jieba Chinese word segmentation technology, this study introduces an innovative data collection method based on topic similarity, significantly improving data accuracy. Additionally, this research establishes a comprehensive public opinion analysis framework that integrates user follower counts, geographical distribution, and interaction data. This framework facilitates the identification of sources of negative sentiment and the development of effective response strategies. As a case study, the dissemination patterns of EMR-related public opinion on Weibo are analyzed, focusing on group sentiment and social interaction. The proposed system achieves a 65.85% improvement in data collection accuracy, demonstrating its effectiveness. Furthermore, this study provides actionable recommendations for relevant departments and governments to monitor, analyze, and respond to EMR-related public opinion. By enhancing decision-making and protecting public interests, this study highlights the role of technology in improving social governance and substantial development.

1. Introduction

In recent years, the internet industry in China has seen rapid development, with the number of internet users steadily increasing and the new media environment becoming more diverse. According to the 53rd “Statistical Report on the Development of the Internet in China”, ref. [1] as of December 2023, there were 1.092 billion Internet users in China, with an average of 26.1 h of internet use per week per user. Platforms such as Weibo, WeChat, and Douyin, which previously dominated the social media landscape, have seen new competition from platforms like Kuaishou, Zhihu, Bilibili, and Xiaohongshu, all of which have gained significant traction due to their large user bases, ease of use, real-time information sharing, and strong interactive capabilities. One particularly sensitive topic in online discussions is electromagnetic radiation (EMR), which is often linked with public concern over potential health hazards and the proliferation of EMR-emitting infrastructure such as mobile phone towers and substations.

Although scientific evidence does not conclusively confirm that long-term harm can come from everyday EMR exposure, public apprehension remains widespread. This sentiment is frequently amplified by misinformation and media sensationalism, giving rise to the “Not in My Backyard” (NIMBY) effect. In the context of EMR, this manifests as collective opposition to infrastructure projects near residential areas, which can hinder technological development and escalate into public opinion crises online. When left unaddressed, these emotional responses may intensify social tensions, delay construction, and lead to costly policy reversals.

Effective public opinion management in such scenarios requires timely identification of emerging concerns, accurate measurement of group sentiment, and appropriate response strategies. However, current approaches to analyzing online sentiment often rely on simple keyword matching, which lacks contextual understanding and leads to irrelevant or inaccurate results. Moreover, there is a lack of targeted frameworks for quantifying group emotions specific to EMR-related public discourse.

Previous studies have addressed public opinion or sentiment analysis in some respects. Social psychologist Kelman [2] proposed the Social Influence Theory (SIT), which examines how groups influence individual behavior, suggesting that a person’s thoughts, emotions, and attitudes are affected by others. Maitner A T et al. [3] explained the nature of emotions based on group identity in their group emotion theory. Nakahashi et al. [4] suggested that as group size increases, individuals capable of emotional contagion become less sensitive to their environment. These theories are also relevant to the study of network public opinion. In network public opinion events, emotions play a crucial role in driving the evolution of the events. The academic community has recognized the significance of netizens’ emotions in public opinion crises and has conducted extensive research. Researchers have constructed a network public opinion propagation model that considers individual emotion, and the main factors affecting the formation and propagation of network public opinion are discussed through simulation experiments [5]. Li Q et al. [6] used social networks to track the evolution of public emotion during the COVID-19 pandemic in China and analyze the root causes of these public emotions from an event-driven perspective.

To address the issues, this study will describe an opinion analysis and guidance system for the electromagnetic NIMBY effect, leveraging sentiment computing. As illustrated in Figure 1, our work is explained in Section 3 and can be summarized as follows: we have designed an accurate data collection module based on topic similarity and developed a public opinion analysis system for EMR events.

During the data collection process, Weibo v12.12.3 was selected as the primary data source due to its status as the largest social media platform in China, where hundreds of millions of users express their opinions daily. Using Python 3, we crawled posts related to electromagnetic radiation (EMR) published on Weibo in recent years, including associated metadata such as comments, reposts, and likes. In the next step, the collected data were stored in a structured database. To enhance the accuracy of data collection, we applied Baidu’s familia semantic similarity technology for relevance filtering, resulting in a large, high-quality dataset of authentic and valid EMR-related content.

In the public opinion analysis stage, we conducted both fine-grained public opinion analysis and domain-customized public opinion analysis. In the fine-grained analysis, the system generated insights such as group sentiment ranking, a geographic sentiment distribution map of China, cluster density ranking, a like–comment–forward ratio chart, and group sentiment trend graphs over time. In the domain-customized analysis, the system focused on hot topic mining and traceability analysis, enabling deeper understanding of discussion focal points and the origin paths of public opinion related to EMR.

This paper applies and integrates the following methods within the context of EMR-related public sentiment analysis:

• A public opinion analysis system based on sentiment computing for electromagnetic NIMBY effects is proposed. The system consists of three core functional modules: data collection, emotional intelligence analysis, and in-depth public opinion analysis. It uses a dictionary-based sentiment calculation algorithm to present sentiment and social behavior analysis results through various charts, including bar charts, trend charts, and pie charts. By extensively collecting user data and posts, the system provides a comprehensive view of public sentiment. Due to the high sensitivity and risks of EMR-related public opinion, the system includes specialized functions for hot topic mining and source tracing. These features help accurately identify negative sentiment origins and enhance public opinion monitoring and management efficiency.
• A data collection system based on topic similarity is designed: This system overcomes the limitation of traditional web crawlers that rely only on keyword searches, which do not ensure content relevance. It uses semantic matching technology instead of simple keyword matching, improving information filtering accuracy by 65.85%. This enhancement significantly boosts the precision of data collection, providing more reliable support for subsequent public opinion analysis.
• Using this system, case studies on public opinion related to EMR events have been conducted: The unique aspects have been thoroughly investigated, leading to the development of a comprehensive and effective strategy for managing network public opinion. This strategy offers valuable insights for relevant departments. Public opinion monitoring refines the monitoring system by incorporating factors such as cluster density. Public opinion analysis focuses on both the origins and dissemination processes of public opinion, which help to establish a better understanding of public needs, allow for more effective mining of opinion data, and enhance timeliness and specificity in guiding public sentiment. Public opinion response advocates for mainstream and local media to work together in addressing public opinion challenges effectively.

The remainder of this paper is structured as follows: Section 2 reviews previous research on online sentiment analysis and the NIMBY effect in the context of electromagnetic radiation. Section 3 describes the design and implementation of the proposed sentiment-based public opinion analysis system, including data collection and emotional modeling. Section 4 presents the research design, experimental process, and analysis of sentiment dissemination. Finally, Section 5 discusses the implications of the results from the perspectives of opinion monitoring, analysis, and response, and concludes this paper by outlining directions for future research.

2. Related Works

Since the 2016 Lianyungang “anti-nuclear waste” incident and the 2020 dismantling of the China Resources Central Park communication base station in Chongqing, NIMBY activities have become frequent in China, with each event attracting significant public attention. Public facilities are typically decided by government authorities. With rising public awareness of risks, if these NIMBY facilities threaten their interests, the public tends to conflict with the government departments overseeing such projects. They express their demands and opposition online, leading to the spread of negative emotions on the internet.

Research into the dissemination of online emotions has been conducted both domestically and internationally. Through simulation experiments, Chen T et al. [7] analyzed the influence of external information intensity, individual education level, individual stubbornness, individual initial opinion, and other factors on the formation of network public opinion. Li M et al. [8] introduced a scale-free network to develop a network public opinion propagation model based on the scale-free network to tackle the problem of the existing public opinion propagation model not being able to fully reflect the law of public opinion propagation in real life. Li C et al. [9] investigated the complex relationship between endogenous and exogenous and deterministic and stochastic stimulating factors in public opinion dynamics. An asynchronous multiagent network model is proposed to explore the interaction mechanism between individual opinions and public opinion in an online multiagent network community. Based on two datasets of more than 165,000 tweets in total, Stieglitz S et al. [10] find that emotionally charged Twitter messages tend to be retweeted more often and more quickly compared to neutral ones. Gao S et al. [11] find that after the occurrence of a public emergency, it is often easy to cause a heated public discussion on network platforms and to form public opinions about the emergency. Research on the evolution life cycle of network public opinion shows that the evolution of network public opinion in emergencies goes through three periods: formation, climax, and dissipation. Kramer A D I et al. [12] indicate that emotional states can be transferred to others via emotional contagion, leading people to experience the same emotions without their awareness. Furthermore, researchers have investigated how emotions affect the dissemination of public opinion. Illingworth V [13] studied the impact of netizens’ emotions through the dissemination of information about social hotspot events. Xu H et al. [14], based on emotion contagion and opinion dynamics, established a model for dynamic emotion diffusion. The method of computational modeling and simulation is applied to simulate real-world social interactions. Xu J et al. [15] established a novel dynamic dissemination model to systematically study the recurrence of network public opinion, and a case study from “Child abuse in Ctrip kindergarten” is used to demonstrate the validity of the proposed model.

Based on existing research on public sentiment analysis, the study of online public sentiment mainly manifests in two forms: the analysis of public sentiment using massive amounts of internet data, and software development. In the field of sentiment analysis, Arash Barfar (2019) [16] conducted a study on the cognitive and emotional responses to political misinformation on Facebook, while Oduwa Edo-Osagie et al. (2020) [17] conducted research on public health using data from Twitter. In terms of application software, Taboada et al. [18] summarized the elements of sudden online public sentiment events and conducted content induction and association analysis, discovering the structural characteristics of the evolution of online public sentiment and enriching the common understanding of the laws of online public sentiment dissemination. The large volumes and variety of expressions on social media have challenged traditional policy analysis and public sentiment assessment. Chung W et al. [19] described a framework for social media-based public policy informatics and a system called “iMood” that addresses the need for sentiment and network analyses in U.S. immigration and border security. S. Qi (2020) [20] built a big data-driven sentiment topic classification model for social network public sentiment users, providing a preliminary research framework for the visualization analysis of sentiment topic maps of social network public sentiment users across languages. In terms of application software, the “Buzz Logic Insights” service developed by Buzzlogi [21] quickly and comprehensively analyzes the current online environment, discovers sentiment trends, and provides emergency plans; Nielsen’s “Buzz Metrics” service grasps market trends and enhances the total business volume of customers. The People’s Network Public Sentiment Data Center (formerly, the People’s Network Public Sentiment Monitoring Room), mentioned in ref. [22], is the earliest institution in China to engage in the monitoring and research of internet public sentiment and is in a leading position in the country in terms of research level and capability. Baidu’s public sentiment system [22] has powerful functions, relying on Baidu’s rich data advantages and artificial intelligence capabilities, providing stable capabilities for public sentiment collection, analysis, and display.

In recent years, research on network public opinion has become more mature, and the academic community has started to explore the NIMBY effect and its role in triggering public opinion events. Scholars have focused on the evolution, dissemination, and guidance of public opinion. In terms of the evolution of NIMBY public opinion, Wei J et al. [23] found that before expressing opinions, most people usually consider the standpoint of their friends nearby to avoid being isolated, which may lead to a herding effect, and the words of celebrities in social networks usually attract public attention and affect the evolution of opinions across the entire network. For NIMBY public opinion dissemination, Zhang C et al. [24], taking the paraxylene (PX) project as an example, established an information dissemination model for NIMBY projects under stigmatization based on the SEIR model in small-world networks. They simulated and analyzed the information dissemination process and characteristics of NIMBY projects under stigmatization. Regarding NIMBY public opinion guidance, Hu S. et al. [25] found that the government affairs microblogging platform helps the government to interact effectively with the public, guiding them to align real-world demands with virtual community expressions and to achieve orderly public participation.

An analysis of both Chinese and international studies reveals that while there is significant research on online emotions and NIMBY public opinion, there is a notable gap in studies focusing on network collective emotions and NIMBY public opinion arising from the EMR produced by large-scale infrastructure projects. Thus far, a systematic framework for quantifying social group emotions and formulating public opinion early warning and response strategies has not been developed.

3. EMR Public Sentiment Analysis System Design

In this section, this study designs and implements a data collection system based on semantic matching to improve the accuracy of keyword searches by filtering out irrelevant data. Additionally, the filtering module is implemented using Baidu’s Familia open-source project, addressing the challenges of traditional topic-based web crawlers. Through evaluation experiments, the system’s performance is validated, demonstrating its ability to enhance the precision of sentiment monitoring. This study also incorporates fine-grained and domain-specific public sentiment analysis, ultimately constructing a sentiment analysis system for EMR events.

3.1. Data Collection Based on Topic Similarity

Traditional topic-based web crawlers first set a keyword and then collect all data containing this keyword. However, this includes some posts that only contain the keyword in the text but have a low relevance to the keyword. In this study, the keyword “electromagnetic radiation” was used to search on Weibo, and one post was retrieved because it mentioned “no electromagnetic radiation” while discussing the functionality of a heating blanket, even though the term “electromagnetic radiation” only appeared once in the entire text. Using a data collection algorithm based solely on keywords may reduce data accuracy. Analyzing incorrect data for online public sentiment and issuing irrelevant warnings would lower work efficiency.

To avoid the above situation, this study designed a data collection module based on semantic matching for information filtering. The specific implementation method is as follows: the system sets keywords according to the given topic and calculates the semantic similarity between “keywords and text” for the data obtained through “keyword search”, filtering out data with low similarity to ensure that the collected data are as relevant to the topic as possible.

3.1.1. Implementation of the Filtering Module

Training a high-quality topic model is costly, and researchers in academia have proposed a variety of topic models to adapt to different scenarios. However, most of this work focuses on the modeling level, that is, designing reasonable models to fit various types of data. Resources and studies are scarce, guiding the implementation of topic models in industrial scenarios. Since online public sentiment analysis is an offline task and internet buzzwords emerge rapidly, frequent updates to the training dataset are essential to maintain the quality of semantic matching. This would also slow down the processing speed of the entire public sentiment analysis system. Therefore, it is not necessary to use a complex model that requires training. This study chooses to use the Baidu’s Familia [26] open-source project to implement this module. This project directly trains and provides high-quality models to the community for direct use, which is quite convenient.

The project includes a document topic inference tool, a semantic matching calculation tool, and three industrial-grade corpus-trained topic models: Latent Dirichlet Allocation (LDA), Sentence LDA, and Topical Word Embedding (TWE) [26]. The filtering module uses the semantic matching calculation tool to implement it. Within the project’s provided APIs, semantic matching can be divided into three categories based on text length: short text–short text, short text–long text, and long text–long text [26] semantic matching. To calculate the similarity between keywords and the content of a post, short–long text matching is chosen. The model has currently been released in three versions, and this study uses the latest version—1.0.2.

The goal of the filtering module is to calculate the semantic similarity between a specified keyword (query) and the main text of a post (content). Unlike other semantic models that directly map short texts to topics, the chosen model for the filtering module calculates similarity by determining the probability that a topic distribution of long text generates the short text [27]. The calculation is as follows. Here, $\mathit{q}$ represents the query, $\mathit{c}$ represents the content, $\mathit{w}$ represents the words in $\mathit{c}$, and ${\mathit{z}}_{\mathit{k}}$ represents the $\mathit{k}$-th topic.

$$\mathit{S}\mathit{i}\mathit{m}\mathit{i}\mathit{l}\mathit{a}\mathit{r}\mathit{i}\mathit{t}\mathit{y}\left(\mathit{q},\mathit{c}\right)={\mathsf{\Pi }}_{\mathit{w}\mathit{ϵ}\mathit{q}}{\mathsf{\Sigma }}_{\mathit{k}=1}^{\mathit{K}}=\mathit{p}\left(\mathit{w}∣{\mathit{z}}_{\mathit{k}}\right)\mathit{p}\left({\mathit{z}}_{\mathit{k}}∣\mathit{c}\right)$$

(1)

After completing the semantic similarity calculation between the keyword and the text, the results are normalized. A threshold is set to filter out data with low similarity to the keyword. The size of the threshold can be adjusted according to the filtering requirements.

3.1.2. Evaluation Experiment

To verify the performance of the filtering system, this study designed an evaluation experiment, collecting data through two algorithms: one keywords-based and one semantic matching-based. The research used Recall and Precision: two important indicators used to evaluate the effectiveness of the retrieval. Recall is used to determine whether enough relevant posts have been collected during data collection, while Precision is used to evaluate the number of collected data that matches the keyword.

The experiment selected “electromagnetic radiation” as the keyword for testing, collecting a total of 700 posts on Weibo. Then, the semantic similarity between each post’s text and the keyword was calculated and normalized using MAX-MIN normalization. The criterion for judging the relevance of the collected data topics was as follows: the normalized similarity ≥ threshold r = 0.1 (chosen as a compromise based on the accuracy requirement of the capture). The experiment recorded the number of collected posts and calculated the precision and recall. Precision refers to the proportion of correctly identified relevant items among all items identified as relevant by the system. Recall measures the proportion of correctly identified relevant items among all actual relevant items in the dataset. These metrics are calculated as follows:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

where TP (True Positive) represents the number of correctly identified relevant items, FP (False Positive) refers to the number of irrelevant items incorrectly identified as relevant, and FN (False Negative) refers to the number of relevant items that were not identified by the system. High precision indicates a low false positive rate, while high recall indicates a low false negative rate. Together, these metrics provide a balanced view of the system’s data extraction performance.

The results are shown in

Table 1. Comparison of precision and recall.

	Number of Collected Posts (Number)	Recall Rate (%)	Accuracy (%)
keyword-based	700	100	41
semantic-based	440	63	68

. Before the filtering system was added, that is, when based on keyword search, all posts containing the keyword “electromagnetic radiation” within the set period were collected (with a Recall of 100%), but not all posts met the requirements (with a Precision of 41%). After the filtering system was added, many posts that only contained the keyword but were unrelated to the actual content of the keyword, were excluded, increasing the precision (with a Precision of 68%). At the same time, due to the exclusion of many posts, some posts containing relevant topics were also excluded, leading to a decrease in recall (with a Recall of 63%). Furthermore, to enhance the reliability of the dataset, we conducted manual annotation by inviting five annotators to label the data, which verified the accuracy and credibility of the collected information. Overall, the addition of the topic filter improved the usability of the data, but it also missed part of the data. Subsequently, the recall can be kept within an acceptable range by modifying the threshold, and the precision can be improved as much as possible.

This study designed an experiment to perform sentiment analysis on data before and after filtering, analyzing the accuracy of the content of the top ten posts with positive and negative sentiment values as shown in

Table 2. Positive or negative number of topic-related posts in the top ten emotional-value rankings.

	Number of Posts with Positive Emotions Related to the Topic (Number)	Number of Posts with Negative Emotions Related to the Topic (Number)
keyword-based	4	6
semantic-based	8	7

. We know that when conducting sentiment analysis without the filtering system, the results showed that most of the top ten posts with negative sentiment values were not related to the desired theme content. Furthermore, to enhance the reliability of the dataset, we conducted manual annotation by inviting two volunteers to label the data. To reduce subjectivity and ensure consistency, a double-checking process was employed whereby each data entry was reviewed by two independent annotators. This could lead to issuing public sentiment warnings for irrelevant posts. The results showed that after the addition of the filtering system, the accuracy of the content of the top ten posts with positive and negative sentiment values improved, enhancing the efficiency of monitoring emotions and providing better public sentiment warnings. Further analysis of the experimental data revealed that without information filtering, the post ranked fourth in negative sentiment was published by “Chong’er Girl”, mainly introducing a whole-house heater but only mentioning “electromagnetic radiation” once in the text, which was unrelated to the keyword theme. However, after adding the topic similarity filter, the post ranked fourth was published by “Qishan Dao Ren Lin Chengbing”, which extensively introduced the hazards of electromagnetic radiation, consistent with the keyword. The experimental results indicate that by adding keyword filtering to exclude posts with low topic similarity, the precision of emotional monitoring can be improved. This reduces the occurrence of situations where posts only contain the keyword but are unrelated in actual content, avoiding unnecessary public sentiment warnings for irrelevant posts, and thus improving work efficiency.

3.2. Public Sentiment Analysis of EMR

3.2.1. Establishing the Link Between EMR and Social Emotions

EMR is the transfer of momentum and energy through space in the form of waves, generated by the interaction of electric and magnetic fields. The interplay of these fields creates electromagnetic waves, which propagate into the air, forming EMR [28]. EMFs disturb immune function through the stimulation of various allergic and inflammatory responses, as well as effects on tissue repair processes. Such disturbances increase the risk of various diseases, including cancer. These and EMFs’ effects on other biological processes (e.g., DNA damage, neurological effects, etc.) are now widely reported to occur at exposure levels significantly below most current national and international safety limits [27]. As these research findings have emerged, public awareness of risks and rights has grown, leading to more frequent advocacy and rights protection movements. The dangers associated with EMR not only cause public health concerns and panic, but also significantly impact public opinion and emotions. This, in turn, can hinder the development of new infrastructure and technological innovation in the digital economy. These “NIMBY events” are highly sensitive, easily triggering collective emotional reactions, with negative emotions spreading quickly and potentially leading to public opinion crises online. Given these characteristics, this study aims to establish a method for quantifying social group emotions related to electromagnetic radiation. By thoroughly analyzing the characteristics of network public opinion in such events, this study seeks to propose effective strategies for relevant departments to manage network public opinion related to electromagnetic radiation. This approach aims to prevent conflicts between the public and project developers, thus avoiding losses for both parties.

Group emotion arises when individuals are influenced by verbal or non-verbal cues from others, leading to emotional contagion and the spread of a unified outward emotion. These emotions can be positive, negative, or neutral. Collective emotions driven by the NIMBY effect occur when the public fears that a project will negatively impact their health and environment, infringing on their rights, leading to strong opposition and even collective protests. In the resulting network public opinion events, netizens express their attitudes toward EMR through words or emoticons. The emotions conveyed by these words and symbols spread online, allowing readers to perceive the individual’s emotional stance, understand their view of the event, and subsequently express their own emotions. This study aims to design and construct a multi-level emotion transmission chain to accurately simulate the spread of emotions under the NIMBY effect and quantify collective emotions based on these characteristics.

3.2.2. The Group Emotions Quantification Model Based on Cluster Propagation Chains

This model for quantifying collective emotions goes beyond merely calculating the sentiment values of single-level comments. It incorporates multi-level emotion propagation chains and user behaviors to provide a more accurate measurement. When a Weibo post about EMR is published, users react by reposting, commenting, and liking the content. These actions, which directly engage with the post, are considered first-level feedback. Additionally, users can respond to first-level interactions through further reposts, comments, and likes, which are considered second-level feedback. The model integrates these multi-level interactions to form emotion propagation chains, simulating the spread of emotions in EMR discussions more realistically. Using this approach, the model identifies the original Weibo poster and first-level respondents as nodes, with the first-level feedback forming directed edges, thereby creating a macro cluster. Similarly, first-level and second-level respondents are treated as nodes, with second-level feedback forming directed edges, creating a micro cluster. By calculating the collective emotions within these two clusters and combining the results, the model provides a comprehensive group emotion value for the EMR event.

The group emotion value of a cluster is calculated by integrating cluster density ($\mathit{A}$), trustworthiness ($\mathit{T}$), and average emotional tendency ($\mathit{E}$). For each cluster node, its in-degree $\mathbf{ln}\mathit{d}$ is determined by the total number of comments and likes it receives. The formula for quantifying cluster density (A) is $\mathit{D}={\displaystyle \frac{\mathbf{ln}\mathit{d}}{\mathit{M}\mathit{a}\mathit{x}\left(\mathbf{ln}\mathit{d}\right)}}$.

If a post about EMR does not attract any comments or likes, indicating minimal user engagement and low cluster density, then even strong negative emotions expressed in the post will not propagate through the emotional transmission chain, thus preventing any significant adverse impact. Cluster trustworthiness ($\mathit{T}$) is measured using three indicators: cohesion ${\mathit{T}}_1$, authority ${\mathit{T}}_2$, and influence ${\mathit{T}}_3$. When mainstream media, which are authoritative and have a large and influential follower base, report on EMR events, their emotions can influence a significant number of netizens, forming large clusters and guiding the overall group’s emotional trend. The average emotional tendency ($\mathit{E}$) of a cluster is derived from the emotional value of comment texts using natural language processing. Ultimately, the group emotion value of a cluster is calculated as $\mathit{A}\ast \mathit{T}\ast \mathit{E}$.

3.2.3. Fine-Grained Public Sentiment Analysis

In public sentiment analysis, the computation of group emotions is of paramount importance. However, merely analyzing the emotions of netizens is no longer sufficient to meet current demands. Nowadays, a more detailed understanding of netizen profiles and behaviors contains a vast amount of information. Collecting these behaviors and identity information can provide more entry points for subsequent public sentiment warnings and offer more precise information for public sentiment guidance.

Therefore, this system constructs a post-publishing account database and a comment account database, including account names, the content they publish or comment on, and the emotional values triggered by the content. This method is used to analyze how different accounts guide public opinion and identify emotional disseminators in popular events. Additionally, the system collects geographical information from post-publishing and comment accounts to create a national map of group emotions. This helps to better understand the level of attention users in different regions give to electromagnetic radiation-related events. The system also records the number of likes, comments, and forwards for the posts, and separates them according to positive and negative emotional posts. By analyzing these user behaviors, one can understand users’ attitudes toward public sentiment events. For example, if the number of likes for negative emotional posts is high, it indicates that many netizens agree with the content of these negative emotional posts. If the number of discussions is also high, it means that there is a high level of attention being paid to the event, and that everyone is expressing their opinions.

3.2.4. Domain-Specific Public Sentiment Analysis

Given the high risk and sensitivity of EMR public sentiment events, this study has designed a hot-topic mining module for in-depth analysis. Hot topic mining identifies the social issues that concern the public the most from a vast amount of text. Hot topics differ from ordinary topics in that they are events and subjects widely discussed by netizens within a certain period and which have received a high degree of attention. Mining hot topics related to EMR events helps to analyze the causes of emotional transmission, thereby formulating effective and precise public sentiment intervention strategies. Additionally, the system includes source tracing analysis, which analyzes the publication, dissemination, and sources of negative emotional information to quickly and accurately identify the source of emotional transmission. Analyzing the source information helps formulate effective countermeasures to cut off the spread of negative emotions.

There are two mainstream methods for hot topic mining algorithms: one involves obtaining hot words or topics through word frequency and topic models and then clustering them according to the hot words or topics. If a topic model-based method is chosen, many studies would use LDA or BTM topic models. The LDA topic model first generates a topic distribution from the Dirichlet distribution and then generates a word distribution using different parameter Dirichlet distributions. Multiple words constitute a document. Since the Dirichlet distribution is random, if the same data are trained multiple times, the results obtained each time are different. Such errors can accumulate, causing subsequent errors to increase. Therefore, this study chooses to complete hot topic mining based on the word frequency method.

Based on the characteristics of high attention and discussion rates among netizens in online hot events, we designed a method of identifying hot words that appear frequently in published posts. The specific approach involves counting word frequency in the collected text dataset, and if a word’s frequency is high, then the post containing this hot word is likely to mention a hot event. The following is the calculation process for hot words: (1) Tokenize the text of all posts and remove stop words. (2) Calculate the frequency ratio. (3) Set a threshold to filter out words with low frequency and generate a document of hot words. After that, cluster the posts containing these hot words, categorizing repetitive discussions of the same hot event and different hot events. The specific approach involves generating feature vectors for all posts, construct a bag-of-words space, and using K-means for clustering, merging posts that discuss the same hot event. After clustering, we can obtain several clusters, each representing a topic category. By counting the number of posts included in each category, we can determine the largest topic cluster. The largest cluster represents a high number of posts published on this topic, with many people discussing it, indicating that the topic may be a hot topic. The specific implementation process is as follows: (1) Construct a feature vector for each post, with each element of the vector representing the number of times each hot word appears. (2) Use TF-IDF to transform the number of times a word appears into weights, resulting in a matrix where columns represent a set of hot words, each row represents a post, and the row vector represents the weight of the hot word in the post. (3) Use the K-means method for clustering. (4) Classify the posts according to the clustering results.

3.3. System Design and Implementation

This study ultimately constructed a public sentiment analysis system based on Streamlit, which is presented in the form of a web page for user convenience. Additionally, the data analysis results are displayed simply and clearly through graphs and tables, facilitating the user’s ability to obtain, summarize, and analyze information.

The system is mainly divided into three functional modules according to its functions: data collection and processing, sentiment calculation, and public sentiment analysis (including fine-grained public sentiment analysis and domain-specific public sentiment analysis).

We have made our dataset and related analysis code publicly available to support reproducibility and further research. It can be accessed at https://github.com/XinyiLing-cuc/Quantification-and-Analysis-of-Group-Sentiment-in-EMR-Public-Opinion-Events (accessed on 17 December 2024).

4. Research Design and Plan

4.1. Research Objective and Design

This research uses the keyword “electromagnetic radiation” to search and analyze Weibo comments, calculating the emotional values of users in various comment sections to determine the collective emotional value of the event. The experimental process is divided into data collection, data preprocessing, sentiment calculation, and result analysis. All steps are completed using the electromagnetic network public sentiment analysis system constructed in this study. Finally, this section conducts an in-depth analysis of the experimental results, summarizes the characteristics of EMR public sentiment events, and proposes response strategies for public sentiment monitoring and guidance.

The objective of this research is to evaluate how public sentiment toward electromagnetic radiation (EMR) propagates within social networks, particularly Weibo. We hypothesize that the emotional trend and diffusion of negative sentiment are influenced not only by textual content, but also by user interaction structure (e.g., cluster density and trustworthiness). To test this, we designed a multi-stage study incorporating both qualitative and quantitative components:

• Independent variables: We used post content, cluster density, trust level, and posting account type.
• Dependent variables: We used emotion score and comment count.
• Methodology: We used semantic filtering, manual annotation (with double-checking), and emotional computation models to extract metrics. Sentiment statistics were analyzed with descriptive and inferential statistics.

4.2. Analysis of Research Result

This study analyzed Weibo posts containing the keyword “electromagnetic radiation” along with user behavior data from 25 October 2021 to 10 September 2022. After cleaning the data, we quantified collective emotions and analyzed accounts for 222 Weibo posts, 222 Weibo posters, 2604 first-level comments, 1839 first-level commenters, 1809 second-level comments, and 425 second-level commenters.

Considering how user interactions and behaviors influence the spread of emotions, this study created macro- and microclusters within defined emotion propagation chains. We quantified local emotion values by calculating cluster density, cluster trustworthiness, and the average emotional tendency of clusters. Finally, we used a weighting method to evaluate the global collective emotions triggered by the posts. Our analysis showed that among the 222 Weibo posts, 158 had positive comprehensive emotion values, 9 showed no emotion value, and 55 had negative comprehensive emotion values. The mean sentiment score across 222 posts was 0.019 (SD = 0.082), with a 95% confidence interval of [0.005, 0.033], indicating a generally neutral-to-positive trend, but with large variance.

To further explain this trend, we plotted the top five posts with the highest positive and negative emotional values (Figure 2), with negative emotions displayed as absolute values. The highest absolute value for negative emotion was 0.30, while the lowest positive emotion value was 0.32, which is higher than the maximum absolute value for negative emotion. This means that the top five negative emotion values were all smaller than the positive emotion values. Thus, the overall emotion for this topic on Weibo is positive. The results suggest that discussions about EMR have not led to widespread negative emotions despite public health concerns.

The following analysis will examine the characteristics of public opinion related to EMR from the perspectives of public opinion monitoring, public opinion analysis, and public opinion response, and will propose comprehensive and effective strategies for managing network public opinion based on the experimental results.

4.2.1. Monitoring Public Opinion Related to EMR: Enhancing the Monitoring System and Considering Cluster Density

We comprehensively considered cluster density. To provide precise and effective early warnings of public opinion dissemination, intervention is unnecessary for clusters with low density or those that have not resulted in small-scale malicious dissemination. Emotional dissemination dynamics within clusters are driven by emotional resonance triggered by emotional contagion. The density and trustworthiness of clusters forming the dissemination chain during the first and second dissemination stages determine whether negative emotions can resonate. To gain deeper insights into the dissemination characteristics of negative public opinion related to electromagnetic radiation, this study analyzes cluster density.

Table 3. Negative aggregate mood value and macro cluster mood value ranking (top 5).

Ranking Table	Publisher
Overall sentiment ranking	Regarding Zhang Dodo	Fairy Zhang Xiaonao	I love Yaya Wu	Faint Lotus Fragrance	Jiangsu State Grid
Macro cluster average sentiment ranking	Teacher Jiang of Aishangke	Fragrant	Digging Chinese Cabbage	Xiuyu 1017	Yu Panxiao

lists the top five Weibo posters with the highest negative comprehensive emotion values and the top five posters with the highest average macro cluster emotion values. The average macro cluster emotion value reflects only the emotional tendency of comment texts, without considering cluster density or trustworthiness. The results indicate that the top five Weibo posts in terms of average macro cluster emotion values did not appear in the comprehensive emotion rankings. Although these posts had highly negative first-level comments, the low macro cluster density prevented even small-scale malicious dissemination. Furthermore, these posts had little second-level feedback, indicating low micro cluster density. As a result, the negative comprehensive emotion values were not high, and there was no malicious dissemination. In practical public opinion monitoring, there is no need for early warnings in such cases. The findings emphasize the importance of considering cluster density in the dissemination of public opinion related to electromagnetic radiation. By incorporating cluster density into group emotion calculations during public opinion monitoring, one can derive a comprehensive understanding of the actual emotional dissemination of posts. For clusters with low density, even if negative emotions are high, emotional guidance is unnecessary. The findings emphasize the critical role of considering cluster density in the spread of public opinion related to electromagnetic radiation. By integrating cluster density into the analysis of group emotions during public opinion monitoring, we can accurately assess the actual emotional impact of posts. For clusters with low density, even if negative emotions are prevalent, it is unnecessary to guide or intervene in these emotions.

4.2.2. Emanation Processes, Understanding Public Needs, Deep Exploration of Public Opinion Information, and Enhancing the Timeliness and Precision of Public Opinion Guidance

A comprehensive analysis should combine both the sources and the dissemination processes. When analyzing the spread of public opinion, it is crucial to consider the reasons behind the dissemination along with the process itself. Understanding the specific content of reposts, comments, and likes is essential, as this constitutes the source. This helps prevent situations where comments that support negative content are mistakenly categorized as positive emotions. For instance, the seventh-ranked post in positive emotion, published by “Rain Zihou“, had a comprehensive emotion value of 0.22. This post was a repost concerning the dangers of high-voltage lines, mentioning the harms of electromagnetic radiation: “Open-air high-voltage power grid, long-term EMR causes childhood leukemia, severely damaging health”. Despite the post’s negative content, the only comment was “Yes, being this close is harmful even to adults”, with “Yes” indicating agreement and being calculated as a positive emotion. The commenter was agreeing with a negative sentiment. This experiment’s failure to account for the actual emotional tendency expressed by the Weibo post could lead to biases in assessing the emotional dissemination it triggers. In practical public opinion analysis, both the post’s emotional tendency and the emotional tendency it generates should be evaluated comprehensively. If a post is inherently negative but elicits positive emotional responses, this discrepancy should be noted and addressed.

Analyzing Dissemination Content and Understanding Public Needs. In monitoring network public opinion, special attention should be given to the spread of negative opinions. It is essential to identify early signs of malicious dissemination, fully understand the content of emotional dissemination, and address public needs precisely to resolve issues. This study specifically analyzed the top ten Weibo posts with the highest negative emotions. The content of these posts primarily involves two themes: defending rights against EMR issues caused by the construction of substations and discussing the harms of electromagnetic radiation. These themes are currently the main triggers of negative emotions regarding electromagnetic radiation. For instance, the second-ranked post in negative emotion was by “Fairy Zhang Xiaonao”, discussing concerns about a house being adjacent to a substation and fears of electromagnetic radiation. This post had a comprehensive emotion value of -0.08 and received 188 comments, leading to extensive discussion. Since the proximity of the substation to the house violated national regulations, many comments included negative terms such as “defend rights”, “speechless”, and “deceived”, spreading negative emotions. The fifth-ranked post, by “Faint Lotus Fragrance”, described the harms of EMR from radiofrequency beauty devices, with a comprehensive emotion value of −0.05. Comments mentioned “lesions” and various diseases, indicating negative sentiment. Relevant departments should collect and thoroughly analyze the content of Weibo posts with negative emotions to identify the sources of public opinion dissemination and guide emotions effectively. For example, when there is widespread discourse on the harms of electromagnetic radiation, timely dissemination of scientific information to debunk myths can help reduce public fear. If a hot topic emerges and is widely shared on Weibo, it is crucial to promptly understand the specific incident, address the public’s reasonable concerns, and resolve the issues effectively.

Thoroughly Exploring the Value of Public Opinion. Network public opinion is a crucial channel for gauging public sentiment about electromagnetic radiation. Social media should not hide the public’s genuine views on social hotspots and sensitive issues by deleting comments. Authenticity is vital for early warning systems regarding public opinion related to EMR. These systems need to extract valuable information from public discourse and effectively address public issues in conjunction with relevant government policies, rather than hiding the real emotional dissemination to control public opinion. In the top ten posts with negative emotion values, one particularly notable post was by “State Grid Jiangsu Electric Power”, which provided scientific information on the harms of EMR from substations near residences. This post was reposted 1402 times, had 1002 comments, and received 1023 likes, indicating a significant impact. However, the emotional value of this informative post was −0.04, showing a negative trend. This was because our study only identified one comment on this post, which negatively stated “Can it be the same?”, suggesting that comments might have been deleted or made invisible, thus skewing the emotional trend analysis. The experiment encountered multiple instances of comment deletion or hiding. Based on the displayed comment counts, the 2088 Weibo posts should have had a total of 17821 comments, but only 4413 were found, representing less than 30% of the actual comments. Such extensive deletion or hiding of comments decreases the accuracy of early warnings for public opinion related to EMR. Effectively addressing public opinion requires genuinely resolving public concerns, not just superficial efforts.

Secondly, the timeliness of public opinion guidance is crucial, requiring timely and targeted response strategies. Therefore, based on network public opinion analysis, improving the timeliness and specificity of public opinion guidance is necessary. To examine this, this study analyzed Weibo posts and comments related to EMR from 28 April 2022 to 10 September 2022 and calculated daily emotion values. In Figure 3 and Figure 4, the X-axis shows the date, and the Y-axis indicates the average sentiment score calculated from related Weibo posts and comments. Sentiment values range from −1 (negative) to +1 (positive), representing the overall emotional orientation of public discourse related to EMR. The daily score is the arithmetic mean of all individual sentiment values on that day. The resulting trend line helps to identify periods of emotional stability and highlights days with sudden fluctuations, which may correspond to major public events or shifts in public discourse related to EMR.

As shown in Figure 3, most of the time, these values show steady changes. This study focused on dates with sudden emotional shifts. As shown in Figure 4, sudden shifts in group emotions occurred on 19 June 2022 and 21 June 2022, with positive and negative emotions, respectively. The analysis revealed that there were 14 related Weibo posts on 19 June 2022. On this day, there were three articles about “tea effectively reducing the harm of EMR and computer radiation”, two articles about the “harm of electromagnetic radiation”, and one post from the Nanchang Ecological Environment Bureau about “conducting public science popularization activities on electromagnetic radiation”. Since the two articles on the “harm of electromagnetic radiation” had no feedback, the negative emotions they generated were not significant, leading to an overall positive emotional trend for the day. On 21 June 2022, there were eight related Weibo posts, including two about defending rights for the same issue involving both EMR pollution and other malicious events, resulting in high negative emotion values. Additionally, there were two posts about the TV adaptation of “Three-Body”, and two articles about “tea effectively reducing the harm of EMR and computer radiation”, creating an overall positive emotional trend. In the public opinion analysis process, relevant departments should not only focus on the emotional dissemination caused by EMR within a specific period but also monitor this throughout the entire duration, paying special attention to points where emotions drastically change. It is essential to understand thoroughly the current hot topics and events being discussed by the public at these critical moments, monitor emerging trends in public opinion, and provide early emotion dissemination assessments and response strategies.

4.2.3. Responding to Public Opinion Related to EMR Events: Focusing on Positive Emotions and Collaborating with Mainstream and Local Media

In public opinion, early warnings and positive emotions should also be given attention. High positive emotion values should not be ignored; rather, it is important to analyze the reasons behind these peaks to avoid overlooking widespread support for negative content that appears positive. Understanding peaks in positive emotions can provide strategies for guiding public opinion, such as increasing popular science articles to correct misconceptions about electromagnetic radiation. For example, a highly ranked positive emotion post by “Mo Wuxuan” with a group emotion value of 0.32 provided rational explanations about EMR and prevention tips. The blogger actively engaged with commenters, creating a positive trend in emotional guidance, and demonstrating the importance of popular science articles in addressing NIMBY issues like electromagnetic radiation.

When responding to EMR events, early warning departments should engage both mainstream and local media. To understand the behavior of publisher accounts and their impact on emotion dissemination, this study categorized accounts as official or unofficial for detailed analysis.

Table 4. Weibo publisher source statistics.

Account Attribute Emotional Tendency	Positive	Negative
Official	31	3
Unofficial	127	52

shows statistics on publishers discussing EMR on Weibo. Among the 222 accounts analyzed, 158 had a positive emotional tendency, including 31 official accounts (20%). Conversely, 55 accounts had a negative emotional tendency, with only 3 being official accounts (5%). This indicates that both official and unofficial media are more likely to generate positive emotions on this topic, with a higher proportion of unofficial accounts. We performed an independent samples t-test to assess differences between official and unofficial account sentiment scores. The results showed no significant difference (t = 1.21, p = 0.23), suggesting that both account types contributed similarly to emotional dissemination.

Figure 5 illustrates the top ten Weibo accounts ranked by their overall influence in the dissemination of emotional content related to electromagnetic radiation (EMR). The influence score of each account was computed using a composite index that integrates three dimensions:

• Cohesiveness—the degree to which the followers of an account form tightly connected clusters within the network, indicating the potential for localized emotional reinforcement.
• Authority—the centrality and credibility of the account in the discussion network, often linked to verification status or recognized social standing.
• Influence—the extent of the account’s reach and engagement, measured through metrics such as repost counts, comment volume, and likes.

Figure 5. Top 10 influential account names.

Figure 5. Top 10 influential account names.

These metrics were normalized and combined to create an overall influence score. The figure reveals that only one of the top ten accounts—“Yingkou Fire Services”—is an official government account. The remaining accounts are primarily unofficial or personal users, suggesting that public sentiment on this issue was driven more by grassroots voices than by institutional communication. Notably, high-impact national-level official accounts such as “CCTV News” were absent from the discussion, underscoring a gap in authoritative participation. Most of the government-related posts came from local municipal accounts, which tend to have limited follower bases and lower network centrality. As a result, unofficial accounts had a greater influence on how emotional content spread across the network, potentially shaping public perception in less controlled or more emotionally charged ways.

Considering these characteristics, this study suggests that relevant early warning departments should engage both mainstream and local media when dealing with EMR events. For mainstream media, departments should collaborate with high-quality, high-level outlets to educate the public about electromagnetic radiation. Leveraging the influence of these media outlets to spread accurate scientific information can help shift public opinion positively. When negative public opinion arises, departments should take responsibility and rely on authoritative voices to speak out. Collaborating with impartial mainstream media can help guide public opinion in a positive way and prevent crises from escalating. Construction project departments can use authoritative mainstream media to promptly express their stance, controlling the narrative and avoiding misunderstandings. For local media, collaboration should focus on addressing specific local public concerns. Using their regional advantages, local media can report hot events in real-time, ensuring timely information release and a clear understanding of the situation on the ground to prevent the spread of rumors. In terms of public education, local media should go beyond online promotion and engage directly with the public through face-to-face interactions, reducing gaps in knowledge about the harms of electromagnetic radiation. Offline promotion ensures that the information reaches a wider audience and is disseminated frequently, effectively transforming negative public opinion.

5. Conclusions

This paper presents a quantification scheme for collective emotions regarding electromagnetic radiation, calculating the group emotion value triggered by the keyword “electromagnetic radiation” on the platform Weibo. It analyzes the characteristics of public opinion events from the perspectives of monitoring, analysis, and response, and proposes comprehensive and effective strategies for managing network public opinion to serve as a reference for related efforts. In terms of public opinion monitoring, it suggests improving the system by considering cluster density. For public opinion analysis, it recommends a thorough examination of the sources and dissemination process, understanding public needs, deeply mining information, and enhancing the timeliness and specificity of guidance. For public opinion response, it emphasizes focusing on positive emotions and engaging both mainstream and local media.

In response to the high sensitivity and potentially high risk of public opinion events related to electromagnetic radiation, this study has designed a sentiment computing-based public opinion analysis system focusing on the electromagnetic NIMBY effect. The system employs an end-to-end design approach that comprehensively covers key aspects such as data collection, sentiment computation, and public opinion analysis. With this system, relevant departments can monitor the dynamic changes in public sentiment in real time, promptly identify, and effectively curb the adverse development trends in public opinion. Moreover, the system provides multi-angle and comprehensive public opinion analysis results, thoroughly exploring the sources of public opinion dissemination, thereby offering solid data support and an analytical basis for decision-makers to formulate precise and effective solutions. The advent of this system not only demonstrates the application value of technological progress in social governance but also highlights the crucial role of scientific research in serving social development and safeguarding the public interest. We anticipate that through the promotion and application of this system, we can contribute to effective management and informed decision-making regarding public opinion related to EMR.

When calculating collective emotions for public opinion related to EMR, relevant departments must comprehensively consider various emotional factors and dissemination characteristics to construct a scientific quantification model and monitor it accurately. The emotion quantification model used in this paper explores the feasibility and applicability of calculating collective emotions for EMR events in network public opinion and proposes effective response strategies. However, many issues still require further research.

In the process of monitoring public opinion, this paper only calculates group emotion values over a specific period. For future improvements, it is necessary to optimize and enhance the monitoring system to enable continuous, full-time monitoring. Based on the public opinion characteristics of EMR identified in this study, it will be essential to design analyses and visualizations for specific time points, focusing on hot events or topics. When abnormal network emotions are detected, the monitoring system should capture and analyze the hot events occurring at those times. Future work should delve deeper into the relationship between social emotions and their causes, creating an automatic analysis system to infer causes from monitoring results and provide effective guidance strategies.

In addition to monitoring public opinion, future efforts should integrate public opinions to formulate handling strategies. When relevant departments announce construction projects related to electromagnetic fields or address issues related to electromagnetic radiation, regular or irregular public opinion surveys should be conducted to gauge public sentiment and promptly adjust network public opinion strategies. Furthermore, before approving the construction of EMR projects, government departments should communicate with the public through information disclosure, popular science campaigns, public hearings, and other methods to ensure the public’s right to know and address their concerns, thereby reducing their worries. Additionally, it is recommended that government departments establish online channels or hotlines to fully understand public needs and provide effective public opinion information to monitoring departments.