Background: Public engagement is a key element for mitigating pandemics, and a good understanding of public opinion could help to encourage the successful adoption of public health measures by the population. In past years, deep learning has been increasingly applied to the analysis of text from social networks. However, most of the developed approaches can only capture topics or sentiments alone but not both together.
Objective: Here, we aimed to develop a new approach, based on deep neural networks, for simultaneously capturing public topics and sentiments and applied it to tweets sent just after the announcement of the COVID-19 pandemic by the World Health Organization (WHO).
Methods: A total of 1,386,496 tweets were collected, preprocessed, and split with a ratio of 80:20 into training and validation sets, respectively. We combined lexicons and convolutional neural networks to improve sentiment prediction. The trained model achieved an overall accuracy of 81% and a precision of 82% and was able to capture simultaneously the weighted words associated with a predicted sentiment intensity score. These outputs were then visualized via an interactive and customizable web interface based on a word cloud representation. Using word cloud analysis, we captured the main topics for extreme positive and negative sentiment intensity scores.
Results: In reaction to the announcement of the pandemic by the WHO, 6 negative and 5 positive topics were discussed on Twitter. Twitter users seemed to be worried about the international situation, economic consequences, and medical situation. Conversely, they seemed to be satisfied with the commitment of medical and social workers and with the collaboration between people.
Conclusions: We propose a new method based on deep neural networks for simultaneously extracting public topics and sentiments from tweets. This method could be helpful for monitoring public opinion during crises such as pandemics.
Pandemics caused by emerging pathogens are public health emergencies. They have dramatic consequences for the population (mortality, morbidity, social life) and the economy . The number of outbreaks has increased in recent decades, and this trend is expected to intensify [ ] in the next years. In particular, when the first cases of pneumonia caused by the SARS-CoV-2 pathogen were declared in Wuhan, Hubei Province, China [ , ], the virus rapidly spread around the world, leading the World Health Organization (WHO) to declare a pandemic on March 11, 2020, and announced it on Twitter with the tweet: “BREAKING “We have therefore made the assessment that #COVID19 can be characterized as a pandemic”-@DrTedros #coronavirus.” With this declaration occurring on social media, Twitter remains an ideal medium to study public opinion on the declaration of the COVID pandemic.
Utility of Social Networks for Identifying Sentiments and Topics of the Population During Pandemics
As public engagement is a key element for mitigating pandemics [- ], several studies have already mined social media since the beginning of the COVID-19 pandemic but with distinct objectives (eg, infoveillance) [ - ] or during different periods (eg, when first important measures were taken in the United States) [ , - ]. To our knowledge, there is no study analyzing public opinion in the immediate reaction just after the WHO announcement.
Social networks have largely been used to capture public opinion, especially during outbreaks (eg, Ebola , H1N1 [ ]). The methods used to analyze texts from social networks have considerably improved over time: manual analysis first, followed by natural language processing (NLP) approaches based on syntactic-semantic or statistical techniques [ ], and more recently, deep learning approaches [ , ]. Deep learning methods provide new perspectives on text analysis since they give the possibility to (1) integrate semantic information around text (eg, with pretrained word embedding, which allows higher semantic information as the input for the neural network rather than a one-hot encoder [ ]) and (2) analyze a significantly larger corpus of text nearly in real time, making it possible to discover new evidence faster [ ]. These approaches [ , , - ] have already been used to capture topics (eg, for the Covid Infoveillance study [ ] or insulin pricing concerns in the United States [ ]) or sentiments (eg, on social network posts or on health care tweets [ , - ]).
Prior Work With Topic Extraction
Several approaches have been used for topic extraction, including qualitative analysis, descriptive analysis, and topic analysis.
Qualitative analyses [, , ] capture common themes from manual analysis, fragmentation, and labelling of text. This method has demonstrated its capacity to accurately capture new and complex topics [ ] but with some major issues: It requires human coders, time, and resource consumption and is not suitable for use with high-dimensional data.
Descriptive analyses  capture the distribution of word frequencies by studying the repetition of words among topics identified from the internet. It allows researchers to correlate the importance of a topic to the volume of searches among this peculiar topic. The main pitfall of this method is the inability to consider the context around the word.
Topic analysis is a method used to discover topics that occur in a collection of documents and has largely been used to mine social media. This method aims at identifying patterns in documents using NLP approaches. Two main categories of topic analyses are commonly used: topic classification  and topic modeling [ ].
Topic classification uses supervised learning algorithms (eg, Naïve Bayes , support vector machine [SVM] [ ]) that need to be trained beforehand with labeled documents, consequently requiring a priori knowledge of corpus topics. These algorithms can achieve variable performance, with a precision varying from 44.9% to 93.3% [ ], depending on the methods used.
On the contrary, topic modeling uses unsupervised learning algorithms that do not need to be trained beforehand. They are thus less work-intensive than supervised learning algorithms since they do not need human-labelled data but often require larger data sets and are less precise than supervised learning algorithms. Latent semantic analysis is the traditional method for topic modeling . It is based on the distributional hypothesis and assumes that words with close meaning will occur in similar pieces of text [ ]. This assumption enabled the development of algorithms such as latent Dirichlet allocation (LDA) [ , , , ], which is popular in the medical domain [ ]. This algorithm identifies latent topics from words tending to occur together and outputs n clusters of words grouped together by similarity. The topics are then manually labelled according to the interpretation of the set of words within each cluster [ , ]. However, LDA requires the investigator to predefine the number of topics and does not consider the sequence of words [ ]. Topic modeling has been poorly assessed, perhaps a result of the difficulty comparing the clusters obtained with a gold standard. To overcome this lack of evidence, Zhang et al [ ] proposed an original approach for assessing LDA: They compared the topics extracted from LDA to those collected through a national questionnaire survey and reported a kappa concordance coefficient of 0.72.
Prior Work With Sentiment Analysis
Several approaches have been used for sentiment analysis, including lexicon-based methods, supervised machine learning methods, and hybrid methods.
Lexicon-based methods are unsupervised methods that do not require training an algorithm and depend only on existing dictionaries . These methods assume that the polarity of a text (positive or negative) can be obtained by characterizing the constituent words within [ ]. A key argument for their adoption was the fact that they only compute the number of positive and negative words [ ] and thus are faster to implement. They are also easily adaptable to various languages by using language-specific dictionaries [ ]. However, they present some limitations that come with language analysis, especially regarding negation, sarcasm, or words with different meaning [ , ]. Furthermore, they are essentially limited by the size, coverage, and quality of the dictionary [ ]. Interestingly, lexicon-based methods can achieve an accuracy up to 94.6% [ ], depending on the dictionary used [ - ].
Supervised Machine Learning Methods
Supervised machine learning methods, which require time to be trained, have also been used . Naïve Bayes often better operates on well-shaped data, whereas SVM often achieves better results with low-shaped data. As social media are poor-quality data, due to very varying length of tweets, colloquial language, and numerous spelling mistakes, larger training data sets are needed to achieve good performance, and the complexity of these methods may impact training time [ ]. They can achieve variable performance, with reported accuracies ranging from 48% to 91% [ , , ], depending on the algorithm used.
Hybrid approaches combine both previous methods. In a recent literature review, Drus and Khalid  demonstrated that hybridized approaches to sentiment analysis often outperform lexicon-based or machine learning–based approaches alone. For example, Hassan et al [ ] used lexicon annotation and multinomial Naïve Bayes for depression measurement from social networks and reported an accuracy rate of 91%; Zhang et al [ ] used lexicon annotation and SVM to annotate sentiments from tweets and reported an accuracy of 85.4%.
Prior Work Aiming to Capture Both Topics and Sentiments
Few methods based on topic-sentiment models have been developed, including the joint sentiment topic (JST) model, Topic-Sentiment Mixture (TSM) model, and Time-aware Topic Sentiment (TTTS) model.
Joint Sentiment Topic Model
The JST  model is a probabilistic modelling framework that extends LDA with a new sentiment layer. JST is fully unsupervised and extracts both topics and sentiments at a document level [ ]. However, JST ignores the word ordering (bigrams or trigrams [ ]). Reverse JST [ ] is derived from JST with an inversion of the order of the topic and sentiment layers. The Aspect and Sentiment Unification Model (ASUM) [ ] is close to JST but focuses on the sentence level. These models have been poorly assessed and were essentially applied on nonmedical data sets, with an accuracy varying from 59.8% to 84.9% for JST [ , ] and 69.5% to 75.0% for reverse JST [ ].
Topic-Sentiment Mixture Model
TSM  is based on the probabilistic latent semantic indexing model and includes an extra background component and 2 sentiment subtopics. It has been assessed on various weblog data sets [ ] but suffers from problems of inferencing on new documents and overfitting data [ ] and requires postprocessing to obtain the sentiment [ ].
Time-Aware Topic Sentiment Model
More recently, the TTTS model  is a joint model for topic-sentiment evolution, based on LDA and allowing analysis of topic-sentiment evolution over time [ ].
Strengths and Weaknesses of Previous Work
Many approaches have proven useful for identifying public topics alone but without the associated sentiment. Other works, especially hybrid approaches, have proven useful for sentiment detection alone but cannot capture the topics alongside sentiment detection.
In both cases, this makes the results less informative and useful . Simultaneously capturing topics and sentiments would be more relevant for better comprehension of public opinion [ ], especially in a time of crisis. Topic-sentiment models have been proposed for the simultaneous capture of public opinion and sentiments but may require prior domain knowledge and have not been applied yet to the medical and social media domains [ , , ].
Potential for a Neural Network–Based Approach to Advance This Area of Research
Neural networks have achieved impressive performances in many NLP tasks, such as sentiment prediction [- ]. Furthermore, the probabilities generated by neural networks could be used to represent sentiment intensity through a quantitative scale leading to more precise information than basic sentiment classification into dual qualitative classes (negative or positive). Surprisingly, to our knowledge, they have not been used yet for the simultaneous capture of public topics and sentiments from social media.
Here, we propose incorporating convolutional neural networks (CNNs) in conjunction with sentiment lexica to simultaneously capture public topics and sentiments in a hybridized approach [, ]. The simultaneous capture of public topics and sentiments, without prior knowledge, would be very useful during crises, such as the COVID-19 outbreak.
Preparation of the Tweet Data Set for Use as an Input for Neural Networks
To analyze the immediate effect of the announcement of the COVID-19 pandemic by the WHO, we focused on tweets relating to coronavirus posted on Twitter the day after the announcement. We collected all tweets containing the keywords “coronavirus” or “COVID” posted in English as recognized by Twitter services on March 12, 2020 (ie, from 00:00:01 to 23:59:59). For each tweet, we extracted the tweet ID, text content, and time stamp. We also filtered them using the language parameter of Twint Python Library  to allow the extraction of English-written tweets only. We verified the absence of tweets in other language by using common stop words of these languages, resulting in only finding foreign city names or family names.
We extracted 1,386,496 tweets from Twitter’s database with the Twint Python library and stored them in the JSON format.
Ethic approval was not needed as analysis of large bodies of text written by humans on the internet and in some social media such as Twitter (eg, quantitative analysis such as infodemiology or infoveillance studies or for qualitative analysis) is not considered “human subjects research.”
We removed 241,506 (17.5%) duplicate tweets and retweets to limit the risk of overrepresentation of one person’s view. Twitter elements (URLs, links to pictures, hashtags, mentions), punctuation, isolated letters, and typographic UTF-8 characters, such as stylized commas or apostrophes, were also removed. Likewise, stop words from Porter’s list  were removed using the Python library Natural Language Toolkit (NLTK) [ ], with orthographic variations. Tweet content was then lower-cased, and “coronavirus” and “COVID” were mapped under a unique term.
provides a flow chart of tweet collection, preprocessing, and splitting into the training and testing sets.
Each tweet was automatically annotated with 3 sentiment labels from 3 different sentiment lexicons from R package tidytext  (AFINN [ ], BING [ ], and NRC [ , ]). These lexicons have largely been used in previous works [ , , ]. Each lexicon provided a numerical value for each sentiment word in the tweet, and these values were summed to annotate the general sentiment of the tweet for each lexicon considered, as described in other works [ , ]. Thus, for each annotation, the sum value could be positive, equal to 0, or negative resulting in positive, neutral, or negative annotation by the considered sentiment lexicon.
Annotation conflicts were handled using a simple rule-based algorithm to compute a single annotation for each tweet. This algorithm is based on the majority vote method and produced a unique qualitative annotation as “positive,” “neutral,” or “negative.” If a majority vote was not obtained (ie, if each algorithm returned a different statement), the tweets were excluded from the data set.
The automatic annotation of included tweets was controlled on 50 randomized tweets, using a manual revision of tweet annotation, resulting in an overall agreement of 86% between algorithm and manual annotation, resulting in a kappa coefficient score of 0.73.
Deep Neural Networks for Simultaneously Capturing Public Topics and Sentiments
Tokenization, Word Embedding, and CNN Architecture
CNN architecture was chosen as it is known to consider Ngrams, making various levels of analysis possible.
All words in each tweet were tokenized, and tweets were postpadded for use as input into the pretrained embedding layer of the neural networks, which encoded semantic properties for each token. We used a 25-dimension Global Vector for word representation (GloVe) embedding trained on 2 billion tweets to shorten training time and achieve better results. This embedding is available from the GloVe project page .
The resulting vectors were then passed to a convolutional unit composed of a convolutional layer (able to analyze unigrams, bigrams, or trigrams), global max pooling layer, dense layer, and dropout layer for regularization and prevention of overfitting. A final dense layer composed of 3 units alongside a softmax activation function computed the probabilities of the tweet belonging to each class of sentiment (positive, neutral, negative). Early stopping was used to prevent overfitting when training our models.
To perform the supervised learning step, the data set was split using stratification over sentiment annotation, allocating 80% (915,993 tweets) for training and 20% for validation (228,997 tweets;). The best model was found after 10 training iterations and used a kernel size of 2 on the convolutional layer. The accuracy was 81%, and the F1 score was 81% on the validation data set ( ).
Neural Network Outputs: Sentiment Intensity Score and Weighted Word Capture
For each tweet, we captured the dominant sentiment as a sentiment intensity score that was calculated from the 3 probabilities predicted by the CNN:
SIS = P(POSITIVE) x 1 + P(NEUTRAL) x 0 + P(NEGATIVE) x (–1)
where SIS, P(POSITIVE), P(NEUTRAL), and P(NEGATIVE) are sentiment intensity score and probabilities for a tweet to belong to the positive, neutral, and negative sentiment classes, respectively, according to the neural network.
Applying this formula allowed us to distinguish 21.82% (249,796/1,144,990) of the tweets as positive, 49.41% (565,782/1,144,990) as neutral, and 28.77% (329,412/1,144,990) as negative. The sentiment intensity score of each tweet was then represented on a scale from –100% (totally negative) to +100% (totally positive), permitted by using the softmax activation function.
As the CNN architecture alternates convolutional and pooling layers, it allows, first, aggregation of the numerical input coming from each word separately until a hidden layer and then combination of the values of this hidden layer until the output of the CNN. Hence, this hidden layer encompasses a value for each word, and this value can be seen as a contribution score (or a weight) of each word in the computation of the final output of the CNN . As the output of the CNN is used to compute the dominant sentiment intensity of the whole tweet, the intermediate values extracted from the hidden layers make it possible to associate “weighted words” to the sentiment intensity score of the tweet. summarizes the capture of the sentiment intensity score and of the weighted words.
In previous steps, the weighted words and sentiment intensity score were captured at the individual tweet level. At the tweet data set level, we computed the average weight of each word for each sentiment intensity score by gathering similar words from distinct tweets and applying a mean function. The resulting matrix contained the weighted words for each given sentiment intensity score.
Visualization of Neural Network Outputs
We developed a Shiny  application (available at [ ]) based on word cloud representation to visualize the weighted words for each sentiment intensity score. This application provides 2 panels: On the right panel, the word cloud displays the weighted words for a given sentiment intensity score. On the left panel, the word cloud can be customized through options specifying the sentiment intensity score, the number and type of words to display (coronavirus or sentiment-related terms), and the esthetics (eg, palette of colors, total percentage of vertical words, and use of a radial gradient).
To generate our word clouds, we replaced the use of word frequencies to summarize text documents by the weights calculated in our matrix. The visualization was made clearer by grouping all lexical variants of a word together, using the word lemmatizer from the R package textstem . We also implemented options allowing the user to ignore all sentiment words and emojis, to choose the word count threshold for display, and to choose the precision of the sentiment score (integer or float to 1 or 2 decimal places).
Identification of the Main Topics Discussed by the Public and Their Associated Sentiment Intensity
Using the Shiny interface, we captured the highest weighted words for the most extreme sentiment intensity scores (negative sentiment: –100; positive sentiment: +100). Author A Boukobza then manually analyzed the top 100 words for both extreme sentiments using string-matching techniques and identified main negative and positive topics within tweets. Each topic was assigned by the manual analysis of these words. Then, we calculated the number of tweets discussing each topic within the data set.
In the results section, we replaced the real names of politicians, political parties, websites, and media with anonymous epithets such as “politicianX,” “politicalPartyX,” “webX,” “mediaX.”
summarizes the general method used for extracting weighted words and their associated sentiments from Twitter data.
Visualization of Neural Network Outputs With an Interactive Interface
Neural network outputs were visualized with an interactive interface displaying a word cloud composed of the weighted words for each sentiment intensity score.
The analysis of the top 100 most important words for each class allowed us to predistinguish main themes retrieved for positive, negative, and neutral classes. In the totally positive class (ie, +100 sentiment intensity score), the top 100 words included words such as “happiness,” “democratic,” “ethical,” “quarantine,” or “expertise.” Concerning the neutral class (ie, 0 sentiment intensity score), the top 100 words included names (eg, “François,” “Eliott”), adverbs (eg, “thankfully,” “formally”), or scientific words (eg, “petri,” “aneurysm”). In the totally negative class (ie, –100 sentiment intensity score), the top 100 words included words such as “job,” “economy,” “afraid,” “panic” ().
Identification of Public Topics and Associated Sentiment Intensity
Using word cloud analysis, we captured the topics for both extreme positive and negative sentiment intensity scores that were discussed in Twitter in immediate reaction to the announcement of the pandemic by the WHO. The analysis of these topics revealed that public opinion was extremely negative about the consequences of the pandemic on the economy and health care system. Conversely, public opinion was extremely positive regarding the mutual aid and cooperation between people and the public health measures taken against the spread of COVID-19. More details are given in the following sections, and example tweets are provided in.
|ID||Topics||Weighted words identified by the neural network||Example of an original tweet||Number of tweets containing weighted words for each topic|
|1||International situation||italian, china, eu, euro, italy, politican1, politicalParty1, politicalParty2, politician2, president, government, politician3, incompetence, fascist||Italy is already today worse affected by Covid-19 than China. (...)||11,297|
|2||Economy||job, impact, industry, yougov, hire, financial, market, livelihood, diarrhea, recession, economy||(...) The markets are trash, every industry is freaking out, and people are losing their jobs because it’s stalling the economy and no one is hiring. (...)||3486|
|3||Media and social media||media1, media2, american||Signed out of my media1. (...) Media2 is a HORRIBLE thing to be on with this damn Coronavirus (...)||1428|
|4||Media and social media||media1, media2, american||Can the media be declared enemies of the people? They (...) lie to us, (...) and fail to report news/statistics that Americans need to know. (...)|
|5||Medical situation||ventilator, paramedic, triage, ration, supply||The (...) most dreadful thing we might face is rationing or triaging who gets ventilators. Emergency rooms across the U.S (...) have limited capacity and supplies (...)||1411|
|6||Public health measures||stay, senior, travel, indoor, cancel, ban||The EU travelban (...) I must admit is terrible decision extremely terrible (...)||8396|
|7||COVID-19 origin||coronavirusHoax, fake, conspiracy, propaganda||(...) the Fake News Media are fabricating the hype and panic to destroy the economy (...) #Pandumbic #coronavirusHoax||1680|
|8||International situation||italy, nhs, democracy, gov, politician4||(...) Freer and more democracy countries can do this if they take needed measures.||2178|
|9||Economy||client, colleague, customer, company||We would like to extend our heartfelt appreciation to all of our clients and partners working on the front-lines (...)||745|
|10||Medical situation||mask, research, health, healthy, resources, healthcare, doctor, applause, hero||Put all your money and resources into getting the cure for the Coronavirus you look like a hero and win the election||4803|
|11||Public health measures||stay, control, announce, interpersonal, family, canceleverything, relative, country, precaution, sanitation, icu, measures, prevention, protect||Good graphic on social distancing and how it can help healthcare capacity, especially important for a country like ours with minimal quality ICU #pandemia #coronavirus #KoronawirusWPolsce #koronavirus #koronawirus||6642|
|12||Mutual aid and cooperation||collaborative, together||(...) Communities who work together to ensure the health and well-being of their fellow neighbor will be stronger and healthier than those who don’t. #Coronavirus||470|
The 6 Main Negative Public Topics Discussed on Twitter in Immediate Reaction to the Announcement of the Pandemic by the WHO
Regarding the international situation, Twitter users were worried about the situation in Italy (eg, the number of cases exceeding those in China;, ID 1) or the risk of punishment or imprisonment for Italians not respecting lockdown. They also discussed travel bans and their consequences, such as the US decision to ban all flights to Europe at a time at which only Italy had a major COVID-19 epidemic. Crisis management and decisions taken by politicians, such as decisions relating to paramedical staff management, were also highly criticized. Regarding economy, Twitter users expressed their fears about the economic consequences of COVID-19. They were worried about the shortages induced by panic buying, such as those leading to a shortage of toilet rolls, and anxiety about the possibility of losing their jobs and being unable to pay their debts ( , ID 2). They also mentioned a potential global recession crisis, caused partly by flight limitations. Regarding media and social media, Twitter users were angry with the media and social media, which they blamed for amplifying fears and stress relating to COVID-19 ( , ID 3), and for not reporting COVID-19 statistics ( , ID 4). Regarding the medical situation, Twitter users were concerned about the medical situation, particularly the management of paramedical staff and materials. They expressed worries about the small number of ventilators available and the likely consequences in terms of equality of access to health care ( , ID 5). Regarding public health measures, Twitter users complained about the limitations of personal liberties, such as the prohibition of flights to Europe ( , ID 6) and the canceling of many events. Regarding the COVID-19 origin, Twitter users talked about “CoronavirusHoax.” They suggested that the pandemic was a hoax and that COVID-19 was a fake disease and evoked a conspiracy theory driven by economic and political motives ( , ID 7).
The 5 Main Positive Public Topics Discussed on Twitter in Immediate Reaction to the Announcement of the Pandemic by the WHO
Regarding the international situation, Twitter users expressed their satisfaction with the actions and decisions taken by some countries, such as Japan, Hong Kong, Singapore, South Korea (, ID 8), or Denmark (eg, the decision to impose a lockdown at the right timing). They also highlighted the efficient measures taken by some countries such as the United Kingdom to overcome the negative effects of lockdown (eg, National Health Service access or online courses for students). Regarding the economy, Twitter users were very grateful to all those who worked during the crisis ( , ID 9). Public workers were even described as “people working hard for ensuring population security.” Twitter users were also informed about the continuity of services ensured by some private companies despite the crisis. They were satisfied with the health measures taken by these companies (eg, social distancing, sanitizing measures, provision of masks). Regarding the medical situation, Twitter users maintained their trust and hope regarding the medical situation. They highly appreciated the work of medical and paramedical staff and their involvement in communicating reliable information about COVID-19 to the population. They highlighted the importance of developing telemedicine and evoked the possibility of a COVID-19 vaccine and its potential consequences for health policies ( , ID 10). They also discussed the production and free distribution of infographics and masks to health professionals by private companies. Regarding public health measures, Twitter users encouraged the respect of national measures, social distancing, and lockdowns to allow people to protect themselves and their families. They also appreciated the graphics providing guidance on the changes in behavior required to limit the spread of coronavirus ( , ID 11). Regarding mutual aid and cooperation, Twitter users were satisfied with the level of cooperation between people in front of the coronavirus crisis ( , ID 12). They were grateful to workers and medical and paramedical staff.
We proposed here an original new approach based on deep neural networks for the simultaneous capture of public topics and sentiments from Twitter data. We trained a CNN on a training data set of 915,993 tweets and achieved a performance of 81% for both accuracy and F1 score. The trained neural network was able to capture the weighted words and their associated sentiment intensity score. These outputs were then visualized through an interactive and customizable web interface displaying the weighted words as a word cloud representation. The trained model was then used to analyze public topics and sentiments in reaction to the announcement of the COVID-19 pandemic by the WHO.
Strengths and Limitations
Our study has several strengths. We combined lexicons and deep learning approaches to improve sentiment prediction. We used CNN to capture simultaneously weighted words associated with sentiment intensity score and to compare unigrams, bigrams, and trigrams during training. We also tried to improve the explicability of the model and to limit the black box effect [, ] by displaying the outputs of the neural networks through an interactive word cloud interface. The word cloud representation is easily understandable and made it possible to consider the outputs attributed by the neural networks to each word according to sentiment intensity score. Our study has also several limitations. First, our method was developed on a data set of tweets in English and needs to be adapted for other languages [ ] and assessed with other extensive data sets [ , ]. Another limitation is the finite set of inclusion keywords, resulting in a potential lack of information due to the total number of keywords used. Further works should concentrate on the diversification of keywords used to provide better sensibility. Furthermore, duplicate tweets and retweets were removed during preprocessing to limit the risk of overrepresenting one person’s view, but this may have also led to underestimating the weights of some words. Second, class imbalance was checked before training, and early stopping was used to prevent the neural network from overfitting the data set. This resulted in good performance, with a model accuracy of 81%. Published studies have reported accuracies ranging from 48% to 91% [ , , ] with the use of supervised learning techniques such as SVM, Naïve Bayes, logistic regression, or word2vec models. However, these performances were measured for binary sentiment classification (ie, negative vs positive sentiment). Here, we decided to consider neutral sentiments too, because it has been shown that tweets can be associated with neutral sentiments [ ]. This choice allowed us to give more explicability and granularity but remains an issue because of our inability to compare our results with those of other studies.
Comparison With Prior Work
Use of Social Media to Capture Public Opinion
Approaches other than social media mining have been described. Focus groups provide a good understanding of public opinion and sentiments but are time-consuming and not necessarily representative of the whole population [, , ] as shown by Rowe et al [ ] during the avian influenza crisis. Telephone and web-based surveys are expensive and time-consuming [ ]. Systematic reviews analyze studies capturing public opinion [ ] but are inappropriate in pandemic conditions as they require multiple skill sets (eg, experts on the topic, systematic review methodologists) and are hardly usable for real-time monitoring. Unlike these approaches, social media mining captures a large range of opinions from a large sample, rapidly and for a reasonable cost [ , ]. It also has proven useful for understanding the attitudes and behavior of the public during a crisis [ ]. For example, before the COVID pandemic, Chew et al [ ] used Twitter to extract public perceptions of H1N1 during the H1N1 pandemic. However, some limitations are inherent to social media: The studied population is limited to social media users [ ], the geographic location of users cannot be assumed with absolute certainty [ ], and analyses are limited to a given language and source (eg, Twitter). Our study illustrates that, despite these issues, social media mining remains an efficient way to capture the thoughts, feelings, and fears of part of the population during a pandemic.
As the detection of topics and sentiments is directly related to neural network accuracy, more options could be explored to obtain higher scores, such as replacing word2vec embedding with Embeddings from Language Models (ELMo)  or Bidirectional Encoder Representation from Transformers (BERT) [ ], which have proven useful for aspect-based sentiment classification [ , ]. The development of a Twitter-specific version of sentiment lexicons integrating web-specific elements such as emojis, abbreviations, or hashtags might also improve results [ ]. Future research should concentrate on adding more granularity to the emotion expressed in tweets, by using emotion-specific lexicons to annotate the tweets with specific emotions such as fear, sadness, or happiness [ ]. Newly developed initiatives such as the Linguistic Inquiry and Word Count (LIWC) dictionary [ ] could also fulfill this task as they provide a dictionary able to recognize emotional words and automatically categorize them as more granular emotions in a hierarchical way (ie, each granular emotion, such as anger, is a child of a top-level emotion like a negative emotion).
Implications for Public Health
Our method could be used to guide public health decisions . Besides factual parameters such as the disease characteristics or the burden it poses to the health care system [ ], public opinion must also be considered to ensure that public health decisions are in line with the beliefs and priorities of the public [ ]. Since many people use social media to share opinions and sentiments [ ], they could provide policy makers and clinicians an opportunity to understand, in real time, the expectations, beliefs, and behaviors of the population and to adapt public health decisions accordingly [ , ]. They can also be used to communicate timely messages to the population [ ] and thus to increase the chance of successful adoption of measures by the population. The development of indicators based on the real-time tracking of health-related conversations on social media is becoming crucial [ , - ]. A major contribution of this study is to show the usefulness of deep learning methods to simultaneously capture public opinion and associated sentiments from large amounts of social media data.
We developed a new approach to conduct both sentiment and topic analyses on social media data by leveraging deep neural networks in conjunction with lexicons. We visualized the outputs of the neural network through a word cloud web interface displaying the weighted words associated with each sentiment intensity score. We demonstrated the utility of our method by applying it to a COVID-19 data set and identifying the main positive and negative topics discussed on Twitter in reaction to the announcement of the pandemic by the WHO. Future studies should concentrate on improving neural network performance and adding granularity to emotion detection. Our method may eventually prove useful for developing indicators for monitoring public opinion during pandemics.
We thank Delphine Colliot and Audrey Vignon for help with data collection. We thank the reviewers for help with manuscript improvement.
The data that support the findings of this study are not openly available due to Twitter identifying information and are available from the corresponding author upon reasonable request.
A Boukobza and RT designed the study and the methodology, implemented the study, created the visualizations, analyzed the data, and wrote the original draft of the manuscript. A Boukobza, BR, and RT collected the data. All authors performed the validation and reviewed, edited, and approval the final manuscript.
Conflicts of Interest
- Madhav N, Oppenheim B, Gallivan M, Mulembakani P, Rubin E, Wolfe N. Pandemics: Risks, Impacts, and Mitigation. In: Jamison DT, Gelband H, Horton S, Jha P, Laxminarayan R, Mock CN, et al, editors. Disease Control Priorities: Improving Health and Reducing Poverty. Washington DC: The International Bank for Reconstruction and Development and The World Bank; 2017.
- Novel coronavirus (2019-nCoV) situation report. World Health Organization. 2020 Jan 21. URL: https://apps.who.int/iris/handle/10665/330760 [accessed 2022-05-15]
- Zhu N, Zhang D, Wang W, Li X, Yang B, Song J, China Novel Coronavirus Investigating and Research Team. A novel coronavirus from patients with pneumonia in China, 2019. N Engl J Med 2020 Feb 20;382(8):727-733 [FREE Full text] [CrossRef] [Medline]
- Teasdale E, Yardley L. Understanding responses to government health recommendations: public perceptions of government advice for managing the H1N1 (swine flu) influenza pandemic. Patient Educ Couns 2011 Dec;85(3):413-418. [CrossRef] [Medline]
- Henrich N, Holmes B. The public's acceptance of novel vaccines during a pandemic: a focus group study and its application to influenza H1N1. Emerg Health Threats J 2009;2:e8 [FREE Full text] [CrossRef] [Medline]
- Morrison LG, Yardley L. What infection control measures will people carry out to reduce transmission of pandemic influenza? A focus group study. BMC Public Health 2009 Jul 23;9:258 [FREE Full text] [CrossRef] [Medline]
- Abd-Alrazaq A, Alhuwail D, Househ M, Hamdi M, Shah Z. Top concerns of tweeters during the COVID-19 pandemic: infoveillance study. J Med Internet Res 2020 Apr 21;22(4):e19016 [FREE Full text] [CrossRef] [Medline]
- Zhao Y, Cheng S, Yu X, Xu H. Chinese public's attention to the COVID-19 epidemic on social media: observational descriptive study. J Med Internet Res 2020 May 04;22(5):e18825 [FREE Full text] [CrossRef] [Medline]
- Sarker A, Lakamana S, Hogg-Bremer W, Xie A, Al-Garadi MA, Yang Y. Self-reported COVID-19 symptoms on Twitter: an analysis and a research resource. J Am Med Inform Assoc 2020 Aug 01;27(8):1310-1315 [FREE Full text] [CrossRef] [Medline]
- Wang X, Zou C, Xie Z, Li D. Public opinions towards COVID-19 in California and New York on Twitter. medRxiv 2020 Jul 14:1 [FREE Full text] [CrossRef] [Medline]
- Xue J, Chen J, Hu R, Chen C, Zheng C, Su Y, et al. Twitter discussions and emotions about the COVID-19 pandemic: machine learning approach. J Med Internet Res 2020 Nov 25;22(11):e20550 [FREE Full text] [CrossRef] [Medline]
- Zou C, Wang X, Xie Z, Li D. Public reactions towards the COVID-19 pandemic on Twitter in the United Kingdom and the United States. medRxiv 2020 Jul 28:1 [FREE Full text] [CrossRef] [Medline]
- Hung M, Lauren E, Hon ES, Birmingham WC, Xu J, Su S, et al. Social network analysis of COVID-19 sentiments: application of artificial intelligence. J Med Internet Res 2020 Aug 18;22(8):e22590 [FREE Full text] [CrossRef] [Medline]
- Chandrasekaran R, Mehta V, Valkunde T, Moustakas E. Topics, trends, and sentiments of Tweets about the COVID-19 pandemic: temporal infoveillance study. J Med Internet Res 2020 Oct 23;22(10):e22624 [FREE Full text] [CrossRef] [Medline]
- Fung IC, Duke CH, Finch KC, Snook KR, Tseng P, Hernandez AC, et al. Ebola virus disease and social media: A systematic review. Am J Infect Control 2016 Dec 01;44(12):1660-1671. [CrossRef] [Medline]
- Chew C, Eysenbach G. Pandemics in the age of Twitter: content analysis of Tweets during the 2009 H1N1 outbreak. PLoS One 2010 Nov 29;5(11):e14118 [FREE Full text] [CrossRef] [Medline]
- Akter S, Aziz MT. Sentiment analysis on facebook group using lexicon based approach. 2017 Presented at: 3rd International Conference on Electrical Engineering and Information Communication Technology (ICEEICT); September 22-24, 2016; Dhaka, Bangladesh. [CrossRef]
- Ghiassi M, Skinner J, Zimbra D. Twitter brand sentiment analysis: A hybrid system using n-gram analysis and dynamic artificial neural network. Expert Systems with Applications 2013 Nov;40(16):6266-6282. [CrossRef]
- Ji X, Chun SA, Geller J. Monitoring Public Health Concerns Using Twitter Sentiment Classifications. 2013 Presented at: IEEE International Conference on Healthcare Informatics; September 9-11, 2013; Philadelphia, PA. [CrossRef]
- Devlin J, Chang MW, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for language understanding. arXiv. Preprint posted online on May 24, 2019 [FREE Full text] [CrossRef]
- Geirhos R, Janssen D, Schütt H, Rauber J, Bethge M, Wichmann F. Comparing deep neural networks against humans: object recognition when the signal gets weaker. arXiv. Preprint posted online on December 11, 2018 [FREE Full text] [CrossRef]
- Wahbeh A, Nasralah T, Al-Ramahi M, El-Gayar O. Mining physicians' opinions on social media to obtain insights into COVID-19: mixed methods analysis. JMIR Public Health Surveill 2020 Jun 18;6(2):e19276 [FREE Full text] [CrossRef] [Medline]
- Park HW, Park S, Chong M. Conversations and medical news frames on Twitter: infodemiological study on COVID-19 in South Korea. J Med Internet Res 2020 May 05;22(5):e18897 [FREE Full text] [CrossRef] [Medline]
- Wang T, Huang Z, Gan C. On mining latent topics from healthcare chat logs. J Biomed Inform 2016 Jun;61:247-259 [FREE Full text] [CrossRef] [Medline]
- Liu Q, Zheng Z, Zheng J, Chen Q, Liu G, Chen S, et al. Health communication through news media during the early stage of the COVID-19 outbreak in China: digital topic modeling approach. J Med Internet Res 2020 Apr 28;22(4):e19118 [FREE Full text] [CrossRef] [Medline]
- Jo W, Lee J, Park J, Kim Y. Online information exchange and anxiety spread in the early stage of the novel coronavirus (COVID-19) outbreak in South Korea: structural topic model and network analysis. J Med Internet Res 2020 Jun 02;22(6):e19455 [FREE Full text] [CrossRef] [Medline]
- Ahne A, Orchard F, Tannier X, Perchoux C, Balkau B, Pagoto S, et al. Insulin pricing and other major diabetes-related concerns in the USA: a study of 46 407 tweets between 2017 and 2019. BMJ Open Diabetes Res Care 2020 Jun;8(1) [FREE Full text] [CrossRef] [Medline]
- Khan MT, Durrani M, Ali A, Inayat I, Khalid S, Khan KH. Sentiment analysis and the complex natural language. Complex Adapt Syst Model 2016 Feb 03;4(1). [CrossRef]
- Drus Z, Khalid H. Sentiment analysis in social media and its application: systematic literature review. Procedia Computer Science 2019;161:707-714. [CrossRef]
- Gohil S, Vuik S, Darzi A. Sentiment analysis of health care tweets: review of the methods used. JMIR Public Health Surveill 2018 Apr 23;4(2):e43 [FREE Full text] [CrossRef] [Medline]
- Hays R, Daker-White G. The care.data consensus? A qualitative analysis of opinions expressed on Twitter. BMC Public Health 2015 Sep 02;15:838 [FREE Full text] [CrossRef] [Medline]
- Huston P, Rowan M. Qualitative studies. Their role in medical research. Can Fam Physician 1998 Nov;44:2453-2458 [FREE Full text] [Medline]
- Boyack KW, Newman D, Duhon RJ, Klavans R, Patek M, Biberstine JR, et al. Clustering more than two million biomedical publications: comparing the accuracies of nine text-based similarity approaches. PLoS One 2011 Mar 17;6(3):e18029 [FREE Full text] [CrossRef] [Medline]
- Karami A, Gangopadhyay A, Zhou B, Kharrazi H. Fuzzy approach topic discovery in health and medical corpora. Int. J. Fuzzy Syst 2017 May 17;20(4):1334-1345. [CrossRef]
- Armouty B, Tedmori S. Automated Keyword Extraction using Support Vector Machine from Arabic News Documents. 2019 IEEE Jordan International Joint Conference on Electrical Engineering and Information Technology (JEEIT) Internet Amman, Jordan: IEEE; 2019 Presented at: IEEE Jordan International Joint Conference on Electrical Engineering and Information Technology (JEEIT); April 9-11, 2019; Amman, Jordan. [CrossRef]
- Dumais ST. Latent semantic analysis. Ann. Rev. Info. Sci. Tech 2005 Sep 22;38(1):188-230. [CrossRef]
- Dumais ST, Furnas GW, Landauer TK, Deerwester S, Harshman R. Using latent semantic analysis to improve access to textual information. 1988 Presented at: CHI88: Human Factors in Computing Systems; May 15-19, 1988; Washington, DC. [CrossRef]
- Zhang H, Wheldon C, Dunn AG, Tao C, Huo J, Zhang R, et al. Mining Twitter to assess the determinants of health behavior toward human papillomavirus vaccination in the United States. J Am Med Inform Assoc 2020 Feb 01;27(2):225-235 [FREE Full text] [CrossRef] [Medline]
- Blei DM, Ng AY, Jordan MI. Latent Dirichlet Allocation. Journal of Machine Learning Research 2003;3:993-1022 [FREE Full text] [CrossRef]
- Wang S, Ding Y, Zhao W, Huang Y, Perkins R, Zou W, et al. Text mining for identifying topics in the literatures about adolescent substance use and depression. BMC Public Health 2016 Mar 19;16(1):279 [FREE Full text] [CrossRef] [Medline]
- Schmidt T, Burghardt M. An Evaluation of Lexicon-based Sentiment Analysis Techniques for the Plays of Gotthold Ephraim Lessing. Proceedings of the Second Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature 2018:139-149 [FREE Full text] [CrossRef]
- Mohammad S, Salameh M, Kiritchenko S. Sentiment Lexicons for Arabic Social Media. Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16) 2016:33-37 [FREE Full text]
- Hu M, Liu B. Mining and Summarizing Customer Reviews. 2004 Presented at: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; August 22-25, 2004; Seattle, WA. [CrossRef]
- Nielsen FÅ. A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. arXiv. Preprint posted online March 15, 2011 [FREE Full text] [CrossRef]
- Mohammad S, Turney P. Crowdsourcing a Word-Emotion Association Lexicon. arXiv. Preprint posted online on August 28, 2013 [FREE Full text] [CrossRef]
- Mohammad SM, Turney PD. Emotions Evoked by Common Words and Phrases: Using Mechanical Turk to Create an Emotion Lexicon. 2010 Presented at: Workshop on Computational Approaches to Analysis and Generation of Emotion in Text; June 5, 2010; Los Angeles, CA. [CrossRef]
- Hassan AUI, Hussain J, Hussain M, Sadiq M, Lee S. Sentiment analysis of social networking sites (SNS) data using machine learning approach for the measurement of depression. 2017 Presented at: International Conference on Information and Communication Technology Convergence (ICTC); December 14, 2017; Jeju, South Korea. [CrossRef]
- Chekima K, Alfred R. Sentiment Analysis of Malay Social Media Text. In: Alfred R, Iida H, Ag Ibrahim A, Lim Y, editors. Computational Science and Technology. Singapore: Springer; 2018:205-219.
- Mamidi R, Miller M, Banerjee T, Romine W, Sheth A. Identifying key topics bearing negative sentiment on Twitter: insights concerning the 2015-2016 Zika epidemic. JMIR Public Health Surveill 2019 Jun 04;5(2):e11036 [FREE Full text] [CrossRef] [Medline]
- Daniulaityte R, Chen L, Lamy FR, Carlson RG, Thirunarayan K, Sheth A. "When 'Bad' is 'Good'": identifying personal communication and sentiment in drug-related tweets. JMIR Public Health Surveill 2016 Oct 24;2(2):e162 [FREE Full text] [CrossRef] [Medline]
- Zhang L, Ghosh R, Dekhil M, Hsu M, Liu B. Combining Lexicon-based and Learning-based Methods for Twitter Sentiment Analysis. HP Laboratories. 2011. URL: https://www.hpl.hp.com/techreports/2011/HPL-2011-89.pdf [accessed 2022-05-15]
- Lin C, He Y. Joint sentiment/topic model for sentiment analysis. 2009 Presented at: CIKM '09: Conference on Information and Knowledge Management; November 2-6, 2009; Hong Kong. [CrossRef]
- Lin C, He Y, Everson R, Ruger S. Weakly supervised joint sentiment-topic detection from text. IEEE Trans. Knowl. Data Eng 2012 Jun;24(6):1134-1145. [CrossRef]
- Jo Y, Oh AH. Aspect and sentiment unification model for online review analysis. 2011 Presented at: WSDM'11: Fourth ACM International Conference on Web Search and Data Mining; February 9-12, 2011; Hong Kong. [CrossRef]
- Mei Q, Ling X, Wondra M, Su H, Zhai CX. Topic sentiment mixture: modeling facets and opinions in weblogs. 2007 Presented at: Topic sentiment mixture: modeling facets and opinions in weblogs; May 8-12, 2007; Banff, Alberta, Canada p. A. [CrossRef]
- Dermouche M, Kouas L, Velcin J, Loudcher S. A joint model for topic-sentiment modeling from text. 2015 Presented at: SAC 2015: Symposium on Applied Computing; April 13-17, 2015; Salamanca, Spain. [CrossRef]
- Dermouche M, Velcin J, Khouas L, Loudcher S. A Joint Model for Topic-Sentiment Evolution over Time. 2014 Presented at: IEEE International Conference on Data Mining; December 14-17, 2014; Shenzhen, China. [CrossRef]
- Giménez M, Palanca J, Botti V. Semantic-based padding in convolutional neural networks for improving the performance in natural language processing. A case of study in sentiment analysis. Neurocomputing 2020 Feb;378:315-323. [CrossRef]
- Park H, Song M, Shin K. Deep learning models and datasets for aspect term sentiment classification: Implementing holistic recurrent attention on target-dependent memories. Knowledge-Based Systems 2020 Jan;187:104825. [CrossRef]
- Abd Elaziz M, Al-qaness MAA, Ewees AA, Dahou A. Recent Advances in NLP: The Case of Arabic Language. In: Kacprzyk J, editor. Studies in Computational Intelligence. Cham, Switzerland: Springer International Publishing; 2020.
- twintproject / twint. GitHub. 2021 Mar 02. URL: https://github.com/twintproject/twint [accessed 2022-05-15]
- Porter MF. An algorithm for suffix stripping. Program: electronic library & information systems 2006;40(3):211-218. [CrossRef]
- Natural Language Toolkit. URL: http://nltk.org/ [accessed 2022-05-15]
- De Queiroz G, Fay C, Hvitfeldt E, Keyes O, Misra K, Mastny T, et al. tidytext: Text Mining using 'dplyr', 'ggplot2', and Other Tidy Tools. CRAN.R. 2022 May 09. URL: https://CRAN.R-project.org/package=tidytext [accessed 2022-05-15]
- Pennington J, Socher R, Manning C. GloVe: Global Vectors for Word Representation. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) 2014:1532-1543. [CrossRef]
- Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proc. IEEE 1998;86(11):2278-2324. [CrossRef]
- Chang W, Cheng J, Allaire J, Xie Y, McPherson J, RStudio. shiny: Web Application Framework for R. CRAN.R. 2021 Oct 02. URL: https://CRAN.R-project.org/package=shiny [accessed 2022-05-15]
- Tweeter Word Cloud Generator. URL: https://adrien-boukobza.shinyapps.io/word_cloud_shiny/ [accessed 2022-05-15]
- Rinker T. textstem: Tools for Stemming and Lemmatizing Text. CRAN.R. 2018 Apr 09. URL: https://CRAN.R-project.org/package=textstem [accessed 2022-05-15]
- Castelvecchi D. Can we open the black box of AI? Nature. 2016 Oct 5. URL: https://www.nature.com/news/can-we-open-the-black-box-of-ai-1.20731 [accessed 2022-05-15]
- Lamy J, Sedki K, Tsopra R. Explainable decision support through the learning and visualization of preferences from a formal ontology of antibiotic treatments. J Biomed Inform 2020 Apr;104:103407 [FREE Full text] [CrossRef] [Medline]
- AL-Rubaiee H, Qiu R, Li D. The importance of neutral class in sentiment analysis of Arabic tweets. IJCSIT 2016 Apr 30;8(2):17-31. [CrossRef]
- Park SH, Han K. Methodologic guide for evaluating clinical performance and effect of artificial intelligence technology for medical diagnosis and prediction. Radiology 2018 Mar;286(3):800-809. [CrossRef] [Medline]
- Kent EE, Prestin A, Gaysynsky A, Galica K, Rinker R, Graff K, et al. "Obesity is the New Major Cause of Cancer": connections between obesity and cancer on Facebook and Twitter. J Cancer Educ 2016 Sep 14;31(3):453-459. [CrossRef] [Medline]
- Giles EL, Adams JM. Capturing public opinion on public health topics: a comparison of experiences from a systematic review, focus group study, and analysis of online, user-generated content. Front Public Health 2015 Aug 24;3:200 [FREE Full text] [CrossRef] [Medline]
- Rowe G, Hawkes G, Houghton J. Initial UK public reaction to avian influenza: Analysis of opinions posted on the BBC website. Health, Risk & Society 2008 Aug;10(4):361-384. [CrossRef]
- Lipsitch M, Finelli L, Heffernan RT, Leung GM, Redd SC, 2009 H1n1 Surveillance Group. Improving the evidence base for decision making during a pandemic: the example of 2009 influenza A/H1N1. Biosecur Bioterror 2011 Jun;9(2):89-115 [FREE Full text] [CrossRef] [Medline]
- Barros JM, Duggan J, Rebholz-Schuhmann D. The application of internet-based sources for public health surveillance (infoveillance): systematic review. J Med Internet Res 2020 Mar 13;22(3):e13680 [FREE Full text] [CrossRef] [Medline]
- Chou WS, Hunt YM, Beckjord EB, Moser RP, Hesse BW. Social media use in the United States: implications for health communication. J Med Internet Res 2009 Nov 27;11(4):e48 [FREE Full text] [CrossRef] [Medline]
- Burton SH, Tanner KW, Giraud-Carrier CG, West JH, Barnes MD. "Right time, right place" health communication on Twitter: value and accuracy of location information. J Med Internet Res 2012 Nov 15;14(6):e156 [FREE Full text] [CrossRef] [Medline]
- LIWC. URL: https://www.liwc.app/ [accessed 2022-05-15]
- Dandala B, Joopudi V, Devarakonda M. Adverse drug events detection in clinical notes by jointly modeling entities and relations using neural networks. Drug Saf 2019 Jan 16;42(1):135-146. [CrossRef] [Medline]
- Vydiswaran VGV, Romero DM, Zhao X, Yu D, Gomez-Lopez I, Lu JX, et al. Uncovering the relationship between food-related discussion on Twitter and neighborhood characteristics. J Am Med Inform Assoc 2020 Feb 01;27(2):254-264 [FREE Full text] [CrossRef] [Medline]
- Liao Q, Yuan J, Dong M, Yang L, Fielding R, Lam WWT. Public engagement and government responsiveness in the communications about COVID-19 during the early epidemic stage in China: infodemiology study on social media data. J Med Internet Res 2020 May 26;22(5):e18796 [FREE Full text] [CrossRef] [Medline]
- Yeung D. Social media as a catalyst for policy action and social change for health and well-being: viewpoint. J Med Internet Res 2018 Mar 19;20(3):e94 [FREE Full text] [CrossRef] [Medline]
- McClellan C, Ali MM, Mutter R, Kroutil L, Landwehr J. Using social media to monitor mental health discussions - evidence from Twitter. J Am Med Inform Assoc 2017 May 01;24(3):496-502 [FREE Full text] [CrossRef] [Medline]
- Weissenbacher D, Sarker A, Klein A, O'Connor K, Magge A, Gonzalez-Hernandez G. Deep neural networks ensemble for detecting medication mentions in tweets. J Am Med Inform Assoc 2019 Dec 01;26(12):1618-1626 [FREE Full text] [CrossRef] [Medline]
|ASUM: Aspect and Sentiment Unification Model|
|BERT: Bidirectional Encoder Representation from Transformers|
|CNN: convolutional neural network|
|ELMo: Embeddings from Language Models|
|GloVe: Global Vector for word representation|
|JST: joint sentiment topic|
|LDA: latent Dirichlet allocation|
|LIWC: Linguistic Inquiry and Word Count|
|NLP: natural language processing|
|NLTK: Natural Language Toolkit|
|SVM: support vector machine|
|TSM: Topic-Sentiment Mixture|
|TTTS: Time-aware Topic Sentiment|
|WHO: World Health Organization|
Edited by C Lovis; submitted 16.10.21; peer-reviewed by J Chen, R Benson; comments to author 31.01.22; revised version received 14.02.22; accepted 21.04.22; published 25.05.22Copyright
©Adrien Boukobza, Anita Burgun, Bertrand Roudier, Rosy Tsopra. Originally published in JMIR Medical Informatics (https://medinform.jmir.org), 25.05.2022.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on https://medinform.jmir.org/, as well as this copyright and license information must be included.