A Clustering Study of Online Public Opinion Texts on Public Emergency Events Based on Sentence-Level Similarity and Sentiment Analysis

The identification of emergent word states has an important clue value for the determination of the time and content of emergencies. The main reason is that when an unexpected event occurs, the frequency of certain words highly related to the unexpected event will also increase greatly, and by recognizing the unexpected state of words in the data stream, the time and semantic content of potential unexpected events can be effectively delineated.

At present, the commonly used emergent word state recognition model in the emergent event detection method is Kleinberg state machine model. The basic idea of this model is to build a burst state recognition model for the time series of words, and identify the words that present a burst state and the duration period of the burst state of the words.

The Kleinberg state machine model assumes that the time interval x between the arrival of the same word in the data stream obeys an exponential function as follows: (1)f(x)=αe−αx$f(x) = \alpha {e^{ - \alpha x}}$

where α represents the arrival rate of the vocabulary, and the transfer behavior between bursty states of the vocabulary is simulated by modeling the state machine model of the vocabulary.

The Kleinberg state machine model is divided into the two-state Kleinberg state machine model and the infinite-state Kleinberg state machine model. The following sections describe the implementation of these two state machine models respectively. 1)

Two-state Kleinberg state machine model

The two-state Kleinberg state machine model divides the states of the vocabulary into a non-burst state q₀ and a burst state q₁. The arrival rate of the vocabulary in the non-burst state q₀ is α, and the arrival rate of the vocabulary in the burst state q₁ is s*α, s is the multiplicity, and s > 1.

The two-state Kleinberg state machine model is shown in Figure 2. The black line hollow circles in Fig. 2 represent the states of the vocabulary, the circle on the left indicates that the vocabulary is in a non-burst state, and the circle on the right indicates that the vocabulary is in a burst state. The states of the vocabulary transfer to each other with probability p.

Figure 2.

Two-state Kleinberg state machine model

The duration period of lexical bursts in the state machine model can be recognized as the time interval between transitions between non-burst and burst states over a period of time. Given the time intervals of vocabulary w, w at moment t for x_t, and w at moment t for s_t, s_t for q₀ means that w is in a non-burst state at moment t, and s_t for q₁ means that w is in a burst state at moment t.

By specifying w the sequence of time intervals {x1,⋯,xt,⋯,xT}$\left\{ {{x_1}, \cdots ,{x_t}, \cdots ,{x_T}} \right\}$, the w corresponding sequence of burst states {s1,⋯,st,⋯,sT}$\left\{ {{s_1}, \cdots ,{s_t}, \cdots ,{s_T}} \right\}$ can be obtained by the two-state Kleinberg state machine model. 2)

Infinite State Kleinberg State Machine Model

The infinite state Kleinberg state machine model extends the two-state Kleinberg state machine model by dividing the states of the vocabulary into infinite states. The arrival rate of the vocabulary in the unique non-burst state q₀ is α; among the remaining infinite burst states {q1,⋯,qi,⋯}$\left\{ {{q_1}, \cdots ,{q_i}, \cdots } \right\}$, the arrival rate of the vocabulary in burst state q_i is sⁱ*α, s is the multiplicity, and s > 1.

The infinite state Kleinberg state machine model’s sequence of burst states with the corresponding time interval sequence is shown in Fig. 3. Where the sequence of time interval observations specifying w is {x1,⋯,xt,⋯,xT}$\left\{ {{x_1}, \cdots ,{x_t}, \cdots ,{x_T}} \right\}$, the infinite state Kleinberg state machine model obtains w the corresponding sequence of burst states as {s1,⋯,st,⋯,sT}$\left\{ {{s_1}, \cdots ,{s_t}, \cdots ,{s_T}} \right\}$, where the value q_i of s_t represents the state that w is in at moment t, when the arrival rate of vocabulary w is sⁱ*α.

Figure 3.

Infinite state Kleinberg state machine model

Effectively mining the sentiment information behind text data is an important task in the work of microblog text opinion analysis, in which sentiment polarity analysis is an important branch in the field of sentiment analysis. Plain Bayesian algorithm is a classical machine learning algorithm, which assumes that the conditions between features are independent of each other on the basis of Bayesian algorithm, calculates the probability of object features based on probabilistic theory and determines the object classification category. The plain Bayesian algorithm obtains the a priori probability of various features in the category through statistical analysis of corpus data, calculates the a posteriori probability based on the condition that the current event belongs to a certain category, and the category with the highest a posteriori probability can be used as the predicted category. Due to the low learning cost of the simple Bayesian algorithm, it occupies less computational resources, and has a strong ability to ignore irrelevant features in the text, so it has a better performance than the complex model that can easily lead to the phenomenon of underfitting, and it is widely used in the text sentiment analysis. The principle of the specific algorithm is as follows:

The sentiment category of text d={w1,w2,⋯,wn}$d = \left\{ {{w_1},{w_2}, \cdots ,{w_n}} \right\}$ to be analyzed belongs to one of the sentiment polarity category sets, and the sentiment polarity category set is defined as C={c1,c2,c3}$C = \left\{ {{c_1},{c_2},{c_3}} \right\}$ according to the requirements, and the elements represent positive, neutral and negative respectively. The plain Bayesian algorithm assumes that each emotion feature in the emotion feature model is independent of each other to reduce the influence of the joint probability distribution in the model, combined with the weight of the emotion feature words, according to the plain Bayesian algorithm to get the feature word emotion classification formula, see equation (2). Where P(cj)$P\left( {{c_j}} \right)$ is the prior probability of emotion polarity category c_j and P(wi,cj)$P\left( {{w_i},{c_j}} \right)$ is the posterior probability of emotion feature word w_i in emotion polarity category c_j. (2)c=argmaxcj∈C{P(cj)∏i=1nP(wi,cj)}$c = \mathop {\arg \max }\limits_{{c_j} \in C} \left\{ {P\left( {{c_j}} \right)\prod\limits_{i = 1}^n P \left( {{w_i},{c_j}} \right)} \right\}$

A priori probability P(cj)$P\left( {{c_j}} \right)$ refers to the initial probability of the sentiment polarity category c_j. The category a priori probabilities are computed from the annotated training corpus and are estimated as shown in equation (3). Where Doc(cj)$Doc\left( {{c_j}} \right)$ refers to the number of documents corresponding to category c_j in the corpus. (3)P(cj)=Doc(cj)∑cj∈CDoc(cj)$P\left( {{c_j}} \right) = \frac{{Doc\left( {{c_j}} \right)}}{{\sum\limits_{{c_j} \in C} {Doc} \left( {{c_j}} \right)}}$

The a posteriori probability P(wi,cj)$P\left( {{w_i},{c_j}} \right)$ refers to the probability of the occurrence of the sentiment feature word w_i in the category c_j, which is calculated as shown in Equation (4), where Count(wi,cj)$Count\left( {{w_i},{c_j}} \right)$ represents the number of occurrences of the word w_i in the texts belonging to the category c_j. Considering that the vector matrix of the training set does not necessarily contain all the text sentiment feature words to be analyzed, a certain word w_i does not exist in all the texts of all the categories, that is, the a posteriori probability of all the classes is 0, and the denominator of the a posteriori probability is 0 will make the classification will not be able to proceed normally. Therefore, we can consider the addition of Laplace smoothing processing, that is, for each weight value to add a smaller value of δ, to ensure that the a posteriori probability of a smaller impact and the a posteriori probability calculation is not 0, the experiment can be δ taken as 1. (4)P(wi,cj)=Count(wi,cj)+δ∑i=1n(Count(wi,cj)+δ)$P\left( {{w_i},{c_j}} \right) = \frac{{Count\left( {{w_i},{c_j}} \right) + \delta }}{{\sum\limits_{i = 1}^n {\left( {Count\left( {{w_i},{c_j}} \right) + \delta } \right)} }}$

The potential information mining of unlabeled data is usually realized by clustering, an unsupervised learning method, which is based on the principle that by analyzing the similarity of text features, similar objects are put into the same cluster, so that the data of the same class are aggregated together according to the similarity degree, so that the data within the same cluster are as similar as possible, and the differences of data in different clusters are as different as possible, so as to realize unsupervised segmentation of unlabeled data. Cluster analysis can be visualized by drawing word clouds on its analysis results.

TF-IDF (Word Frequency-Inverse Document Frequency) algorithm is a statistical method to determine the weight of a word by the number of times it appears in the bag of words. The principle is that through the statistics of the text word frequency, the higher the word frequency, the stronger the correlation between the word and the document, due to the use of common words in the process of the inevitable word frequency will be on the high side, so the use of inverse document frequency to reduce the weight of the common words, as shown in Equation (5): (5)TDIDFij=TFij*IDFij$TDID{F_{ij}} = T{F_{ij}}*ID{F_{ij}}$

Therefore, the TF-IDF value is the product of word frequency TF and inverse document frequency IDF, the larger the TF-IDF value is, the more important the word is in the text, which is used to assess the importance of a word for the text where it is located, and the importance of the word is proportional to the number of times it rises in the text, and at the same time, with the frequency of the word in the text is proportional to the inverse of the decline.

K-means algorithm is a clustering algorithm based on division, by selecting the number of clusters k and n initial centers, according to the objective function and cosine similarity for the calculation of its distance to the center, each word clustered to the nearest center, and many iterations to update the number of iterations to guide to reach the maximum number of iterations or the center of clusters no longer change until the number of clusters selected for the number of clusters k can be determined by the SSE (and the variance) is determined. K-means algorithm has the advantages of fast convergence, simple algorithm, efficient processing of large data sets, has better scalability and high efficiency, suitable for the analysis of network large-scale group event data.

DBSCAN algorithm is a density-based clustering algorithm, mainly from a core point, to its center, Eps for the radius to obtain its Eps neighborhood, in the neighborhood of the data greater than the MinPts (the minimum number of) for the core object, to the expansion of the range of its density reachable, to get a maximization of the zone containing the core point and the boundary point of the city, where the density of any two points are connected. The process is repeated until all points are identified. The DBSCAN algorithm, compared to other algorithms, does not require a fixed number of clusters, is able to identify outliers as noise rather than simply categorizing them, and the shape and size of the clusters are not restricted.

1)

Crawling Scope

According to the analysis steps, the public opinion heat needs to be calculated at a later stage, i.e., the number of relevant elements in the microblogs is counted, so the number of original microblogs, comments, and retweets need to be calculated.

In terms of content, the case was selected for the purpose of theme extraction and sentiment analysis by analyzing the content of microblogs sent by users. The elements involved in microblog content are original microblogs and comments (including retweets and comments), in which the original microblog is the subjective attitude of the user towards the event, which can intuitively express the user’s own attitude and opinion towards the event, and the comments are the evaluation of others’ views, and there may be tens of thousands of comments under one original microblog, which is too large a volume of data and difficult to analyze, so only the original microblogs are selected for content analysis.

The heat of online public opinion refers to the degree of high public opinion in cyberspace, which is triggered by media reports, discussions among netizens, and specific actions taken by the government in response. By visualizing the relevant data through line graphs, the stage of public opinion can be clearly delineated. The relevant elements are the number of original tweets W_1,i, the number of retweeted tweets W_2,i, and the number of commented tweets W_3,i. The formula for calculating the daily public opinion heat O_i is: (6)Oi=b1×W1,i+b2×W2,i+b3×W3,i${O_i} = {b_1} \times {W_{1,i}} + {b_2} \times {W_{2,i}} + {b_3} \times {W_{3,i}}$

Where i denotes the ind day, b₁, b₂ and b₃ are the weights of the number of original tweets, retweets and comments respectively.

Since the number of daily microblogs, retweets and comments may reach hundreds or thousands in practice, the data need to be simplified by the extreme value method in the calculation to obtain Equation (7): (7)Oi′=Oi−Min(Oi)Max(Oi)−Min(Oi)×100%${O_i}' = \frac{{{O_i} - Min({O_i})}}{{Max({O_i}) - Min({O_i})}} \times 100\%$

For the values of weights in Equation (6), b₁ = 0.4, b₂ = 0.35, and b₃ = 0.25 are taken to get the values of O_i and O_i′. Some of the web public opinion heat values are shown in Table 1.

According to the results calculated in the previous step, take the date and draw a line graph with columns O_i and O_i′, and the trend of public opinion heat is shown in Figure 4. According to the trend of the folding line, the public opinion heat is divided into stages.

Figure 4.

Public opinion heat distribution and trend

As shown in the figure, from July 20 to July 21 was the beginning period of public opinion, with a total of 4,541 original microblogs, 89 retweets, and 6,357 comments; from July 22 to July 25 was the outbreak period, with a total of 38,770 original microblogs, 3,828 retweets, and 17,401 comments; from July 26 to July 28 was the recurrence period, with a total of 7,290 original microblogs, 126 retweets, and 1,416 comments; from July 29 to August 3 was the slowdown period, with a total of 1,498 original microblogs, 25 retweets, and 1,416 comments. From July 26 to July 28, there were 7,290 original microblogs, 126 retweets and 1,416 comments, and from July 29 to August 3, there were 1,498 original microblogs, 25 retweets and 105 comments.

Data cleansing is the processing of non-compliant content; it is the final procedure for detecting and correcting identifiable errors in data files, including the processing of mutilated data, erroneous data, and duplicate data. The ultimate goal of data cleaning is to improve the quality of data. Data cleaning in this paper is mainly twofold, duplicate data and invalid data. Duplicate data refers to the fact that due to possible errors in the operation of the data collector, the final result presents two and more completely duplicate microblog texts posted by the same user at the same time. Invalid data refers to microblog text that is completely unrelated to the topic of the study. Sina Weibo has a topic function, usually denoted by #topic#, under which netizens can post microblogs to express their opinions and views. However, the behavior of netizens posting microblogs is highly subjective, and it is inevitable that there will be the behavior of “rubbing on the topic”, i.e., posting contents completely unrelated to the topic itself, for example, publicizing the latest movie of a certain celebrity under the topic of the public emergency studied in this paper, and so on, and the emergence of these data will increase the inaccuracy and ineffectiveness of the text processing. The presence of such data will increase the inaccuracy and invalidity of text processing, so these data need to be deleted.

After manually cleaning the duplicate and invalid data, a total of 39,027 original tweets were obtained, of which 4,173 were in the starting period, 29,361 in the bursting period, 4,730 in the recurring period, and 763 in the plateauing period. The large difference between the cleaned data and the collected data is due to the fact that through the analysis of the original data in an Excel table, it is found that there is the behavior of “rubbing on the topic” in every stage of public opinion, and the number of microblogs discussing the topic itself is much larger than the number of microblogs discussing the topic itself.

The NTUSD sentiment dictionary was used as the seed, and the text sentiment analysis method in section 2.2.2 was used. In the results of the initial calculation, negative emotional words include words that are strongly related to the research topic, such as “United States”, “China”, “China and the United States”, and “trade war”, and the weight is high, so these words need to be removed. At the same time, when the negative emotion words are analyzed, it is found that the more obvious positive sentiment words, such as “come on”, are listed as negative sentiment words after calculation, which will affect the accuracy of the calculation of the emotion intensity of the subsequent text. The method adopted in this paper is to aggregate NTUSD, Hownet, BosonNLP sentiment dictionaries and other types of sentiment dictionaries on the network to calculate the positive sentiment thesaurus and negative sentiment thesaurus, respectively.

The comparison is done in the positive and negative sentiment dictionary obtained using the plain Bayesian algorithm proposed in Section 2.2.2, and if the word is a positive sentiment word in the thesaurus and a negative sentiment word after the computation, it is corrected by adding a score of 0.3 to it. If it is greater than 0 after correction, it is categorized in the positive sentiment dictionary, and if it is still less than 0 after correction, it remains in the negative sentiment dictionary. The corrected score was used as the final sentiment score for the word.

After several experiments to determine the final positive and negative sentiment words, a total of 2831 sentiment words were obtained. Among them, 1572 positive sentiment words and 1259 negative sentiment words are shown in Fig. 5 and Fig. 6 for Top 10 positive and negative sentiment words, respectively.

Figure 5.

Top 10 positive emotion words

Figure 6.

Top 10 negative emotion words

As can be seen from the above figure, in the sentiment analysis of the online public opinion text about the public emergency event “Zhengzhou 720 Heavy Rainstorm Disaster” in 2021, the top 10 positive emotional words mostly come from the praise of mutual aid, rescue, and national strength, reflecting social cohesion, among which “love”, “positive energy”, and “hero” have the highest emotional scores, which are 0.953, respectively. 0.935 and 0.915, while negative affective words focused on the destructiveness of the disaster itself and the fear and grief of the people, reflecting the severity of the disaster, among which the emotional scores of “disaster” and “ruthlessness” were the highest, -0.947 and -0.938, respectively.

Figure 7 shows the time series of microblog sentiment intensity, from which it can be seen that the overall sentiment intensity in the period of July 22-26 is higher than the sentiment intensity in other periods. And the positive sentiment intensity peaked on July 23, with 2,371 positive sentiment posts on that day. The second peak occurred on July 22, with 2,246 posts. Negative sentiment is strongest at the beginning of the public opinion period, and reaches a trough on July 21, with 1,504 negative sentiments on that day. The fluctuation of negative sentiment intensity is not big with the outbreak of the public opinion. The score of negative sentiment is about half of that of positive sentiment, so it can be seen that during the evolution of public opinion, most of the sentiments expressed on Weibo are positive and the intensity of positive sentiments is much greater than that of negative sentiments.

Figure 7.

Timing of Weibo emotion intensity