First, based on the API interface for the Sina microblog, data from COVID-19–related microblog comments from July 1, 2020, to June 1, 2021, which was in the middle of the COVID-19 pandemic, were obtained and preprocessed. Second, we considered the characteristic of low granularity as well as short words in the microblog comments [22]. Based on the SnowNLP Chinese sentiment vocabulary ontology, the Dalian University of Technology sentiment classification dictionary [23] is optimized to identify sentiment features, and a multidimensional analysis model of public opinion on social media during the pandemic was constructed from the perspective of spatiotemporal correlation. Third, according to the multiscale division method of public opinion, the term frequency-inverse document frequency (TF-IDF) method was used to explore the semantic features of public opinion. Fourth, according to the difference in public opinion data characteristics, a model combining LDA and Bidirectional Encoder Representations from Transformers (BERT) [24] was used to classify the topics. The number of topics was determined according to minimum perplexity, and the topic weight was established using chi-square test results. Finally, a machine learning logistic regression (ML-LR) sparse matrix fusion model was used to analyze the evolutionary trend in public opinion combined with the monthly pandemic notification information. The technical route and structure of this paper are shown in Figure 1.

This study collected data for topics related to COVID-19 in the monthly hot topic search of the People's Daily microblogs from July 2020 to June 2021, including the hot topic name, topic search volume, and date of the blog post with the most searched topic, and selected more than 100,000 comments from 12 microblogs. These 12 blogs have a high degree of heat (eg, likes, comments, and retweets), and the blog topics represent the hot events of the corresponding timeline, which can be regarded as a microcosm of the different development stages of the COVID-19 pandemic in China. The specific data from these 12 microblogs are shown in Table 1. Based on the ArcGIS interface of the geocoding package Geocoder, the geographic latitude and longitude of the check-in location coordinate system were then obtained, and the IP address geographic location information of the user was calculated. Through the Requests library of the Python language, the data interface was requested and parsed, and 2000 comments were screened for sentiment analysis. An example of the crawled microblog data is shown in Table 2.

Issues with the crawled text data included missing column values, numerical anomalies, and special symbols, which need to be supplemented and filtered. First, the Jieba word segmentation tool was used to process the text segmentation, and a custom dictionary was established for the high-frequency words and proper nouns related to the pandemic. Second, the Trie tree structure was used for efficient word graph scanning, and a stop word list was established to filter out the noisy data from the text. Third, the text information was converted into vectors (ie, feature engineering transformation). The model can only input numbers (like vectors), not text; therefore, before running the model, any signal needed to be converted into a digital signal (eg, number, vector, matrix, tensor) that can be recognized by the model. The regular word segmentation result was used as the input for the model. The data preprocessing results are shown in Table 3.

The specific text data preprocessing process included (1) removing duplicate data, (2) removing useless characters, (3) using Jieba word segmentation, (4) removing stop words, (5) removing low-frequency words, (6) removing any tags such as HTML, and (7) using the decision tree method to supplement the missing column values and integrate the data.

To reflect the characteristics of public opinion, a method based on dictionary and thesaurus matching was used. The Dalian University of Technology sentiment dictionary was used to scan the strings in the dictionary one by one and provide auxiliary annotations for sentiment classification. This method combines the characteristics of the topic and Chinese grammar and expands the original dictionary from the dimensions of “word,” “part of speech,” “word sense number,” “sentiment classification,” “intensity,” and “polarity.”

The basic dictionary selected for this paper was the emotional vocabulary library of the Dalian University of Technology, which divides emotions into 7 categories (“anger,” “disgust,” “fear,” “sadness,” “surprise,” “good,” “happy”) and 21 subcategories. The initial emotional intensity was set to 5 levels (1, 3, 5, 7, 9), which is more detailed than other dictionaries. To facilitate sentiment calculation by a computer, we divided the polarity of sentiment words into 2 categories: positive (1) and negative (0). The formula for word sentiment is given in Equation 1:

where s(w) represents the sentiment value of the term, v(w) is the sentiment intensity of the word, and p(w) denotes the sentiment polarity of the word.

First, we needed to calculate the TF-IDF values of the words in the text to obtain the feature matrix. The TF-IDF value of the feature item w_ij was calculated as shown in Equations 2-5:

Here, w_ij represents the j_th word appearing in the microblog text-based data from the i_th day. C(w_ij) is the number of occurrences of the term w_ij. D is the total number of documents in the daily microblog text-based data |D_i| for the number of words in a document D_i. denotes the bag-of-words vector of the document D_i. The function I(w_ij,D_i) takes the value 1 or 0, where 1 means that the document D_i contains the word w_ij; otherwise, it takes the value 0. The part-of-speech and sentiment classification of the words are shown in Table 4.

This study used a topical term extraction method based on the combination of LDA and BERT. First, the review text was segmented into words, and the word vector was generated based on the hybrid pretrained model. The k-means clustering method was then used to cluster the word vectors, and the words with higher word frequencies were selected as representative words in each class. At the same time, other words in the same class were replaced by representative words to reduce the amount of input data. Finally, the biterm topic model (BTM) for short text was used to extract the topics.

The BTM is aimed at the characteristics of short text and extracts more informative topics using co-occurrence patterns of word pairs in the whole corpus. The modeling process of the BTM generates a corpus of word pairs. Model training and model parameter inference are then performed based on the generated corpus. Finally, the topic distribution and word distribution on the corpus are obtained. If we suppose there are M feature words, |B| word pairs, and K topics in the corpus, then the corpus-level topic distribution is denoted by θ, the distribution of topic K is denoted by θ_k, and the word distribution is denoted by Ø, as shown in Equation 6.

where z represents a topic extracted, p(z=k) is the probability of occurrence of topic k, and p(z=k) is the probability of occurrence of words under topic k.

Perplexity is widely used to measure the fitting effect of an LDA model. Perplexity is a probabilistic model evaluation metric that evaluates the predictive ability of a trained model on new unseen data. LDA perplexity is the perplexity when the model predicts new unseen text. After understanding the calculation of the statement probability, the perplexity of the statement s=w₁, w₂, w₃, ..., w_n can be defined as seen in Equation 7:

To calculate the perplexity of LDA, we needed to utilize the generation process of the model. Specifically, we needed to split the text data into training and test sets. The training set was used to build the LDA model, and the test set was used to calculate the perplexity of the model. In the test set, we treated each document as a sequence, taking one word at a time from the sequence and inferring the topic of that word based on the current model. Once the topic of the word was predicted, the model parameters were recalculated. In this process, we needed to calculate the perplexity of the text using the topic of the predicted word and the frequency of all the words.

The microblog comment text was divided into positive and negative sentiment, and the value represented the probability that the text contained that sentiment. The values ranged from 0 to 1; closer to 1 tended to be positive, and closer to 0 tended to be negative. To improve the accuracy of sentiment prediction, the sentiment classification model needed to be retrained. The main steps are described in the following paragraphs.

First, the training method of the Bayesian model was used to train the sentiment classifier, and the classification method in the Bayesian classification was then used to predict the sentiment classification and test the accuracy of the model. Finally, the newly trained model was saved.

The daily microblog text set D={d₁, d₂, ..., d_m}. The variable m represents the number of microblog crawls per day. The sentiment score of each microblog was obtained using the newly trained sentiment analysis model. In the actual judgment, to make the visualization results more intuitive, the value of was lowered by 1. That is, when the return value was in the interval of –1 to 0, the sentiment probability value was negative; when the return value was between 0 and 1, the sentiment probability value was positive. The sentiment score S_D was calculated as shown in Equation 8:

First, the feature words were inputted, and the sentiment polarity was classified as 1 (“positive”) or 0 (“negative”). Second, the TF-IDF calculation converted the words in the text into a word frequency matrix, and the matrix element a[i][j] represented the word frequency of j words in type i text. MemoryError was used to control the parameter. The transformer was then used to calculate the TF-IDF weight of each word. The first fit_transform was used to calculate the TF-IDF, and the second fit_transform was used to convert the text into a term frequency matrix to obtain all the words in the bag-of-words model. The TF-IDF matrix was then extracted. The element w[i][j] represented the TF-IDF weight of j word in the i type text, and the data were divided by combining the sparse matrix. Finally, the logistic regression classification method was used to calculate the model benchmark indicators. The binary classification ML-LR model is given in Equations 9 and 10:

where w is the weight, b is the bias (b can be 0), and we can expand the weight and input vector a little, as shown in Equation 11:

This study did not require ethics approval because all data collected were publicly available. There is no means within this paper or its supporting materials to establish the identification of users and their corresponding tweets.

We identified 12 key dates as turning points for mood scores during this period. Figure 2 shows the general trend in the monthly sentiment changes of microblog users during the COVID-19 pandemic. In general, the overall classification results show obvious clustering according to the time series.

In the first stage (July 2020 to October 2020), strong infectious tendencies and unknown virus characteristics exacerbated public panic, and user sentiment ratings were generally low. In the second phase (November 2020 to February 2021), changes in sentiment scores were relatively stable, except for a large decrease in January 2021 and February 2021, possibly due to material shortages and population movement that accelerated virus transmission. In the third phase (March 2021 to June 2021), the proportion of positive emotions gradually increased thanks to the successful development and widespread distribution of vaccines. The data show that the changes in the number of infected people and the effectiveness of pandemic prevention measures were important factors affecting the change in the polarity of public sentiment.

In the face of the pandemic, people’s instinctive psychological reaction was panic and anxiety, and these negative emotions were particularly obvious in the first stage. In the second stage, government departments strengthened the control of public opinion on social media. As a result, the public’s confidence about fighting the pandemic was significantly enhanced, and the emotional polarity gradually improved. In the third stage, the proportion of positive emotions increased significantly, and negative emotions gradually turned into affirmation and support for antipandemic actions. Figure 3 shows the keywords in the monthly hot topic searches on the microblogs during the COVID-19 pandemic.

The SnowNLP tool was used to calculate the sentiment value, and the text was classified according to the threshold. The LDA model was then used to mine the topics. The hyperparameters of the model were set as α=50/K and γ=0.01. The number of iterations was set at 1000. The number of potential topics was determined as K=12 according to the perplexity. The same weight was given to the topics with a monthly cycle by combining the change in the intensity of the influence of the propagation of the microblog information.

We obtained 12 topics from the feature extraction of microblog topics during the COVID-19 pandemic (see Table 5). Among them, pandemic prevention and people’s livelihoods were the focus of public attention.

Multimedia Appendix 1 shows the results of topic modeling. In the first stage, the main topics were mainly focused on pandemic prevention and control. In the second stage, in addition to vaccines, food safety became the focus. The core topics of the third stage were COVID-19 vaccination and variants.

To explore the spatial distribution of public emotions, we subtracted the same types of emotions from provinces and regions in adjacent stages according to regional differences and obtained the spatiotemporal changes in negative emotions, as shown in Figure 4. Taking provincial administrative regions as the basic statistical unit and the 3 evolution stages as fixed time intervals, the natural break point method was used to divide the proportion of negative emotions.

Neighboring provinces had similar emotional characteristics; that is, the spatiotemporal similarity of emotions was high. Moreover, the proportion of negative emotions was related to the severity of the local pandemic, and a high value for negative emotions had a spatial distribution similar to that of areas with a more severe pandemic situation. In addition, provinces with many confirmed cases had a high proportion of negative emotions, while provinces with a small number of confirmed cases had a relatively low proportion of negative emotions.

Combined with China’s daily real-time notifications regarding the pandemic, the first stage of the pandemic in China occurred during the outbreak period, with the number of confirmed cases reaching a peak in August 2020 and the confirmed case rate gradually slowing down since then. As can be seen in Figure 5, during the period from July 2020 to January 2021, anger, disgust, fear, sadness, and other negative emotions prevailed. When combined with topic digging, this result shows that people’s concerns regarding the COVID-19 pandemic included fear of human-to-human transmission, a domestic outbreak, and imported cases from other countries.

During the second stage (February 2021 to June 2021), the sentiment scores basically showed positive sentiment, indicating that the government adopted effective response policies and that public sentiment tended to stabilize again. Positive emotions such as surprise, good, and happy rapidly increased, indicating that the government invested considerable human and material resources at any cost. Therefore, the public was able to observe the determination and effectiveness of the country’s antipandemic efforts. In addition, with the arrival of the inflection point during the third phase of the pandemic, the mass distribution of COVID-19 vaccines increased public confidence that the pandemic could be prevented and controlled and that production, work, and prospects could resume after the pandemic.

To visually present the words occurring at a high frequency for each emotion, we used a word cloud library to generate cloud maps of 70 words for 7 emotions corresponding to positive and negative emotions in the 3 stages, as shown in Table 6. At the same time, high-frequency words were combined with the corresponding current affairs hot spots to explore the differences in and connections of different semantic features in different sentiment classifications.

In the initial stage of the pandemic, the words occurring at a high frequency included “hopeless,” “danger,” “infect,” and other words as well as “negative,” “rumor,” “helpless,” and many other words that express emotions. In the face of the sudden COVID-19 pandemic, the public was in an extremely unstable emotional state.

In the second stage of the pandemic, the words occurring at a high frequency included “keep away,” “pandemic,” and “pathogen.” Microblog topics focused on orderly prevention and control measures and a joint fight against the pandemic. The positive sentiment of netizens gradually increased, turning from fear to solidarity and cooperation.

In the third stage of the pandemic, words such as “vaccine,” “recovery,” “hope,” and “sanguine” appeared frequently on the microblog. Most people were full of confidence and believed that China could win the battle against the pandemic, expressing positive feelings of positivity, unity, and hope for the future.

To verify that the proposed model and method had high accuracy, we calculated accuracy, precision, recall, and the F₁-score, as shown in Equations 12-15:

The words in the text were converted into a word frequency matrix, and the TF-IDF weight of each word was counted. The TF-IDF matrix was then extracted, and its weight was calculated. Finally, the effectiveness of the model was evaluated by extracting parameters (precision, recall, accuracy) from the confusion matrix, as shown in Figure 6.

First, the word frequency matrix was generated using the text-based data after Chinese word segmentation and data cleaning. The CountVectorizer class was called to calculate the word frequency matrix, and the generated matrix was X. Second, the TfidfTransformer class was called to calculate the TF-IDF value of the term frequency matrix X, and the Weight matrix was obtained. The Sklearn machine learning package was then called to perform the classification operation, the fit () function was called to train, and the predicted class labels were assigned to the pre array. Finally, we called the PCA () function of Sklearn to reduce these features to 2 dimensions corresponding to the X and Y axes, from which we could evaluate the algorithm.

Based on the study of the traditional double classification confusion matrix, we divided the confusion matrix into 7 emotions according to the emotional characteristics. Very few data points were misclassified from the matrix by the model. Moreover, the performance of the model in terms of accuracy, recall, and the F₁-score of the weighted average of the 7 emotions “anger,” “disgust,” “fear,” “sad,” “surprise,” “good,” and “happy” was good, as shown in Table 7.

This study constructed a multidimensional analysis model of public opinion on social media in China during the COVID-19 pandemic from the perspective of spatiotemporal correlations. We proposed an ML-LR model with a sparse matrix to analyze the evolutionary process of public opinion and interpret the dynamic relationship and multidimensional emotional characteristics of public opinion in different spatiotemporal environments. The results show that, due to the different trends in the pandemic situation and prevention and control efforts in different regions, there were differences in the emotional characteristics of public opinion. The amount of public opinion data at many different levels was similar in the temporal and spatial distributions, and the amount of public opinion data was positively correlated with the number of new cases. With the rapid spread of the COVID-19 pandemic, the monthly amount of public opinion data increased. As the COVID-19 pandemic was gradually controlled, the amount of monthly public opinion data showed a downward “zigzag” trend. The spatial distribution of the amount of public opinion data was positively correlated with the distribution of COVID-19 spread, and the provinces with more public opinion data were mostly those areas with a more serious COVID-19 situation. This study can provide theoretical support and a practical reference for government and public health safety departments to deal with public health emergencies.

In traditional research of public opinion on social media, the topic model method is often combined with a text clustering algorithm [12,23,25], but this approach is limited by the problem of insufficiently labeled data, making it difficult to reflect the changes in public mood. This paper proposes an ML-LR model with a fusion coefficient matrix to overcome these problems. In addition, social media data embedded with geographical location information provides valuable evaluation indicators for the study of public opinion characteristics, but the time span and geographical location range selected by traditional research of public opinion on social media are of limited size or are not representative [4,17]. This study captures the dynamic changes in public sentiment characteristics through multidimensional, spatiotemporal analysis of public opinion.

We will conduct further research regarding 2 aspects in the future. On one hand, when analyzing public sentiment during a pandemic, user comments may come from multiple social media platforms [25,26]. Subsequent research will consider adding data from other social media platforms, strive to include a more comprehensive user group, and describe more appropriate regional characteristics of public opinion. On the other hand, considering that the spread of the pandemic has certain geospatial heterogeneity under the joint influence of geographical proximity, transportation network, and pandemic prevention measures, taking provincial administrative regions as geospatial measurement units has certain limitations [27-29].

In future research, we will discuss how to obtain a multiscale visual analysis unit of the pandemic and public opinion according to the superposition of the transmission mode of the COVID-19 pandemic and multiscale geographic space. We will also mine the process of propagation during the pandemic from a more granular scale in terms of time and space.

The analytic method of public opinion on social media proposed in this study can effectively reflect the characteristics of public opinion in different regions and different periods of time. It can also provide theoretical support and a practical reference for the analysis of public opinion in major public health events, as well as provide correct guidance for government departments and effective control of the propagation of network public opinion.