1. Introduction
Among the 17 Sustainable Development Goals (SDGs) adopted by the United Nations in 2015, at least 8 are climate change related. The SDGs urge action by all countries in a global partnership to end poverty, improve health and education, and reduce inequality all while tackling climate change and working to preserve oceans and forests (United Nations 2022a.). Studying the perception of climate change in social media is a matter of interest. This paper aims to analyze the sentiments expressed in Twitter interactions in relation to climate change (CC). The study is performed considering the geographical and gender perspectives as well as different user typologies (individual users or companies). A total of 92 474 Twitter messages were utilized for the study.
Several methods have been proposed for the realization of topic modeling, some of them are latent semantic analysis (Foltz 1996), probabilistic latent semantic analysis (Hoffman 2013), latent Dirichlet allocation (LDA; Jelodar et al. 2019), and correlated (Blei and Lafferty 2007) and biterm (Yan et al. 2013) topic models. Other algorithms also exist that propose modifications based on the aforementioned methods (Ye et al. 2018; He et al. 2017).
There is various research on topics modeling, which have been utilized in a wide variety of fields including political science, software engineering, geography, and social sciences among others (Jelodar et al. 2019). Some of the most recently carried out examples are the following. In the political sciences field, Fang et al. (2012) propose an unsupervised topic model based on LDA for contrastive opinion modeling, which allows us to obtain opinions from multiple views, according to a specific topic and their difference on the topic with qualifying criteria. Grimmer (2010) study the expressed priorities of political figures as they are described in political texts. Topic modeling has also been successfully applied in software engineering. In this field, Bi et al. (2018) analyze those textual artifacts (e.g., requirements documents) that are generated during the typical software life cycle, in order to gain a deeper understanding of the software systems architecture. Soliman et al. (2016) examine, classify, and structure the exchanged messages on StackOverflow community into different contents in order to make technological decisions concerning software architecture. In the area of geographical research, Lozano et al. (2017) implement a distributed geoaware streaming latent Dirichlet allocation model for topics’ locations recognition in an unstructured text. Ye et al. (2021) describe a data-driven approach centered on location and toponym extracted from social media data to estimate the asymmetric connectivity between cities. In the social science field, Buhin et al. (2020) compare the topics that result from the application of LDA topic modeling to scientific papers with the existing categories of Web of Science Core Collection for the social sciences area. Roberts et al. (2016) describe a generalized linear model, which makes it possible to detect topics in documents.
Exchanged messages and users’ data that can be retrieved from Twitter are a very useful source for the execution of various types of analyses (Anber et al. 2016) such as ranking users (Nebot et al. 2018; De Silva and Riloff 2014), inspection of homophily and reciprocity characteristics (Singh et al. 2021), and examination of information diffusion (including event life cycle analysis, network-topology checking, retweetability) (Török and Kertész 2017; Iacopini et al. 2019; Kobayashi and Génois 2021).
Various mechanisms have been used to study the polarity of exchanged messages on social media focusing on certain topics. Certain analyses apply supervised machine learning techniques (Fernádez et al. 2014; Hamad et al. 2017; Deshwal and Sharma 2016; An et al. 2014). In particular, the latter performs polarity modeling using support vector machines (SVM) and naive Bayes methods on manually labeled tweets (prior classification as subjective or factual). Unsupervised machine learning approaches are also applied in Cigarrán et al. (2016), Pandarachalil et al. (2014), Suresh and Gladston (2016), and Veltri and Atanasova (2015). Suresh and Gladston (2016) use a fuzzy clustering model to perform sentiment analysis and compares the results with those obtained using k-means and expectation maximization algorithms. Veltri and Atanasova (2015) carries out automatic thematic and semantic network analysis as well as text classification according to psychological categories. Fernández-Gavilanes et al. (2015) propose an alternative procedure to capitalize the information obtained from a parsing analysis without requiring training. Additionally, some other studies are based on word lexicons whose polarity has been previously identified (Khoo and Johnkhan 2018; Ahmed et al. 2021).
Some pieces of the research perform sentiment analysis in order to compare opinions on a given topic in different countries. In Dahal et al. (2019), the authors use the LDA algorithm (Blei et al. 2003) to perform topic modeling as well as Valence Aware Dictionary and Sentiment Reasoner (VADER; Hutto 2022) algorithm for sentiment analysis. The study is done in several countries, with a special focus on the United States. There is also research that analyzes the evolution of sentiment on a particular topic over time (Cody et al. 2015).1
This paper aims to answer the following research questions, considering the interactions occurring on Twitter from February 2021 to May 2021:
-
For the corpus under analysis, taking new aspects in tweet processing into consideration,2 which of the two selected methods (LDA and biterm) provides better results?
-
Based on the above, with the method finally chosen and applying mechanisms to determine the location and gender of each tweet as well as user typology, what are the topics surrounding the interactions on climate change that occur on this social network?
-
In view of the above, what are the differences in the number and percentage of tweets in each topic when considered globally and taking into account the dimensions country, gender, and type of user (persons and organizations)?
-
Can the polarity detection mechanism be improved and applied to the Twitter social network? Are there differences in the polarity presented by each topic globally? Are there differences in the polarities per topic according to the dimensions of country, gender, and user type?
The objectives of this work are derived from the answers to the questions previously posed. Regarding the study of interactions on Twitter about climate change, the novelties achieved, based on the corpus used, are as follows:
-
Inclusion of hashtags and multiword expressions as n-grams in the corpus to be considered in the topic modeling. The incorporation of n-grams in order to improve the topic modeling performance has been used in other corpus, which do not refer to climate change.
-
Evaluation of tweet localization through examination of two fields in tweets (location and geolocation), in order to obtain the largest number of tweets with the home country identified.
-
Building of a mechanism that through Word2Vec, and k-means methods, is able to detect the polarity of each tweet. To adjust the hyperparameters of the Word2Vec model, which uses a neural network, a similarity and correlation check was carried out with the words included in WordSim353 (Agirre et al. 2009). To our knowledge this mechanism has not been used in the analysis of interactions on Twitter that relate to climate change.
-
The previous method is used as an additional mechanism to VADER and TextBlob lexicons for polarity detection. This makes it possible to obtain greater precision in the polarity determination (Bonta et al. 2019). Four metrics are used to evaluate the analogy between the three procedures: Jaccard and cosine similarity coefficients as well as Manhattan and Euclidean distances.
-
Considering the above, identification of new topics as well as characterization of the interactions globally and according to nations, gender, and user types (persons and organizations) is performed. Various statistical tests and similarity metrics have been used in the analysis. These methods have not been used, to our knowledge, in previous research concerning the examination of interactions occurring on Twitter related to climate change. We also think that the analysis considering the dimensions of gender and type of user has not been previously performed in the aforementioned area.
Climate change and its effects is one of the most relevant problems facing citizens today (European Commission 2021). To this extent, some studies based on surveys have been carried out in order to explore the gap between the types of behavioral and policy changes required to effectively mitigate the effects of climate change (Eichhorn et al. 2020). Social networks offer an excellent opportunity to gain a larger understanding about citizens’ opinions on certain current issues (Edwards and Santos 2014). The discovery and characterization of the existing perception of Twitter users regarding the effects of climate change, with a multidimensional vision, is of great interest. This analysis can help to adopt comprehensive regional and local strategies as well as to adequately orient international cooperation.
2. Materials and methods
a. Overview of used resources
1) T-Hoarder tool
For the extraction of information from Twitter the T-Hoarder tool (Congosto et al. 2017) was used. It is a framework that is capable of carrying out tweet crawling and data filtering as well as displaying summary information about Twitter activity on a given topic. The tool is implemented in UNIX as an operating system and uses Python as its programming language. T-Hoarder3 uses two application programming interfaces (APIs) to retrieve data from Twitter: the Rest API and the Streaming API. The Rest API allows us to carry out all type of queries on the data of Twitter in a synchronous form, with the limitation of searching for information from the previous week, not earlier in time. The Streaming API allows us to perform real-time data downloading.
With the aim of obtaining the keywords that make it possible to filter those tweets not related to climate change a small group of people was brought together for an interactive meeting in order to carry out this selection. If more than 15 keyword candidates were suggested, they were clustered using an affinity diagram (IEEE 2022). The keywords thus obtained were then filtered using the multiple voting system, linking together those that reinforce each other. Finally, a decision was made as to which keywords to use, and seven keywords were taken. These are climate change, global warming, greenhouse effect, climate emergency, climate crisis, climate disaster, and climate action with and without spacing (to include hashtags).
2) Software programs
Several programs in Python and R language were implemented, the functions they perform are summarized in the supplemental materials.
b. Overview of used methods
1) Download tweets
Twitter is a social site where people communicate through “tweets” or messages containing text, photos, GIFs, and/or videos up to 280 characters long. Besides posting, Twitter users can perform different interactions such as tweeting, replying, retweeting, and quoting.4 However, as this research aims to identify individuals’ polarity or sentiment toward climate change, it analyzes only the tweets corresponding to tweeting interaction. The replied, retweeted, and quoted interactions are not considered because they do not bring any new personal opinion. It should be noted that humans produce much more novel, responsive content, while bots generate much more retweets. The latter tend to tweet more uniform resource locators (URLs) and larger multimedia elements (Gilani et al. 2017; Himelein-Wachowiak et al. 2021). Given the above information, we do not perform any special treatment to identify bots, as we neglect all interactions. Additionally, only English-language tweets are considered. The number of tweets left for analysis are 92 474 posted by 56 866 unique users. To provide insights into the data under study, Fig. 1 depicts a histogram of unique users and their tweeting frequency. Additionally, Table 1 shows the number of tweets by type of interaction.
Occurrences of unique users that tweet with certain frequency. For visualization purposes, those users tweeting over 100 tweets and those tweeting 1 are removed. The unique number of users in these cases is 17 and 44 558, respectively.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Downloaded data in numbers. Number of tweets per type of interaction.
2) Preparation of tweets for analysis
A process of cleaning up tweets is carried out in which links, usernames, punctuation, and emojis are removed. To reduce the dimension of the vocabulary,5 the text is tokenized, stemmed, lower-cased, and only words with a length longer than two are kept for the analysis. The Natural Language Toolkit (NLTK) Python library (Bird 2021) was used to perform this task. Because of the integration of multiple features, especially unigrams and bigrams, there has been a proven increase in classification accuracy (Anjaria and Guddeti 2014; Alsaeedi and Khan 2019; Nokel and Loukachevitch 2016), n-grams are used to refine the vocabulary.6 The construction of n-grams words is done as follows:
-
The hashtags present in the extracted tweets are formatted to look like unigrams, bigrams, trigrams, and larger n-grams,7 then incorporated into the text of the tweet.
-
A detection of multiword expressions and bigrams from a stream of sentences is done. All words with a total collected count of less than 20 are ignored. The gensim phrase model (Bouma 2009; Mikolov et al. 2013b) is utilized to perform this identification task.8 The n-grams word combination a1…am is formatted to a1_…_am, analogously to hashtags, and replaced in the tweet text. A total of 12 397 n-grams were produced. Further information regarding hyper-parameter selection of the gensim phrase model can be found in the supplemental material.
3) Building of the topic model
In this task, two methods are used:
The LDA topic model, which is an unsupervised machine learning algorithm used to capture a set of latent topics over a collection of documents, utilizing a “bag of words” (BOW) approach.9 In this work the implementation of LDA used is Mallet 2.0.8 through the corresponding Python wrapper in gensim library (McCallum 2002).
A dictionary is produced from a tokenized word list obtained from the processed text corresponding to each tweet.10 Terms are filtered out so that each term does not appear fewer than 20 times and no more than 2.5% over the total corpus,11 in which each tweet is represented as a list of tuples (a BOW).12 To construct the model, an LdaMallet object is created from the Mallet library, providing it with the corpus, the dictionary, the number of topics, and the density of topics per document as well as the number of training iterations. A value equal to 0.1 and 1000 is taken for topic density and number of iterations, respectively.
The biterm topic model, which is a word co-occurrence based topic model that learns topics by modeling word–word co-occurrence patterns (e.g., biterms) in a corpus.13,14 In this work, the biterm library in Python is used.
The processed text15 of each tweet is vectorized in order to obtain the vocabulary and the biterms. A biterm topic model (BTM) object is created, providing it with the number of topics and vocabulary as input data. The model is trained from biterms, considering a number of iterations equal to 1000.
To select the appropriate number of topics, the coherence is evaluated per topic using the UMASS coherence score (Mimno et al. 2011), which estimates how often two words, w_i and w_j appear together in the corpus.16,17 The obtained results are shown in Table 2.
UMASS coherence score for LDA and biterm algorithms.
It can be noted that the LDA algorithm shows a higher coherence for the used corpus, similarly to Dahal et al. (2019). Based on these results, we use the LDA algorithm for the rest of the analyses performed in this research. It was decided that 10 topics would prove the best choice, being that this number has an average coherence close to the maximum and it is a value low enough to have representative topics that can be manually labeled with meaning.
4) Polarity analysis
To determine polarity three methods are applied: VADER, TextBlob (Loria 2020), and a noncommercial polarity approach based on word embeddings and k-means clustering.
VADER, which is an open source lexicon and rule-based sentiment analysis tool (Hutto and Gilbert 2014). This lexicon comprehends a list of features (words) that have been labeled according to their polarity. Taking an average of the word polarities included in the processed text,18 a polarity in the range [−1, 1] is assigned to each tweet.
TextBlob, which is a Python library makes it possible to carry out sentiment analysis. From the processed text obtained for each tweet,19 two different values are obtained: polarity and subjectivity. The former ranges between [−1, 1] and gives the actual sentiment, negative or positive. The latter ranges between [0, 1] and it is a measure of the amount of personal opinion or factual information present in the sentence. The higher the subjectivity score, the more opinionated the sentence is.
A noncommercial polarity algorithm, which uses Word2Vec as a technique to learn word embeddings (Mikolov et al. 2013a,b), and k-means clustering to identify sentiment in the tweets. Word2Vec is a model that embeds words in a lower-dimensional vector space utilizing a neural network. To achieve the above, Word2Vec uses a one-hot vector representation for each word in the corpus.20 This numerical word representation permits us to map out each word in V to a vector in an N dimensional space. It is expected that the embeddings for words with a similar context are closer to each other than to unrelated words in vector space. We use the continuous bag of words (CBOW) architecture provided by Word2Vec.
To determine the values for the hyperparameters indicated above, the WordSim353 dataset was used (WordSim353 2022; Finkelstein et al. 2002; Agirre et al. 2009), which consists of 353 pairs of words and their similarity scores rated by participants of a psycholinguistic experiment.22,23
Once the embeddings for the corpus had been determined, the k-means algorithm was used to establish the polarity of each tweet; given a set of data points xi, x2, xn, where xi = (xi1, xi2, …, xir) is a vector in Rr, in which r is the space dimension. This method creates a partition of the aforementioned data, putting it into k clusters, where k is a free input parameter. For the purpose of determining the polarity only two clusters, positive and negative, are required.24
Once the dataset is separated into two clusters, we examine which one is relatively positive, and which one is predominantly negative.25 For each cluster, the distances of the word vectors (di) to the centroid are estimated and the minimum value of all of them (dmin) is obtained. Each word is assigned an
The similarity between these three polarity algorithms is evaluated through the Jaccard and cosine similarity coefficients as well as the Manhattan and Euclidean distances.27
5) Determining the location of the author of each tweet
To extract the location from which each tweet was posted, the GeoPy, particularly Open Street Map (OSM) nominatim geolocation service was used. According to Kumar et al. (2014), approximately 1% of all tweets published on Twitter are geolocated. This represents a very small proportion and requires the use of other profile information to extract the location. We assess the tweeting location inspecting two fields (location and geolocation) to maximize retrieval. With this procedure, the location could be extracted from 47 471 tweets.
6) Determining the gender of the author of each tweet
To automatically extract the gender for each tweet, the online service “genderize.io” was used. It is a simple API that predicts the gender of a person given their name.28 By using the name category the gender could be extracted from 66 764 tweets out of a total of 92 474 tweets.
Those tweets in which the gender of their author could not be identified were examined to detect whether they came from organizations, companies, or entities [see section 2b(7)].
7) NER
To establish the presence of an entity, organization, or company within the tweets with no gender assigned, the account names are compared to the Forbes top 2000 world largest companies in 2020 (Murphy et al. 2021). To do this, the Levenshtein distance between each list item and the “name” field in each tweet was calculated.29 This is followed by further analysis using the SpaCy Named Entity Recognition (NER) package, which makes it possible to provide the bio field of each tweet, to detect a company name in its text.30 A total of 773 companies that posted 1507 tweets were detected.
8) Analysis of the proportions per topic (both number and percentage of tweets)
To study whether the proportions related to number and percentage of tweets per topic are different between each pair of topics, the Fisher exact test is used (Sachs 1984).31 This test is utilized if several nominal variables exist, in order to investigate if the proportions of each are different. The concept of a contingency table is used.32
The hypotheses are as follows:
-
H0: “Relative proportions of variables are equal.”
-
Ha: “Relative proportion of variables are not analogous.”
The Fisher exact test is performed with a confidence interval α = 0.95. If p value ≤ 0.05, H0 is rejected and Ha is taken.
In this research, r = 2 and c = 11. Since it carries out a pairwise comparison between countries in relation to the number and percentage of the tweets in each topic (10). Only those countries with a tweeting frequency larger than 200 are assessed. These are sorted in descending order, and the top 20 are selected.
9) Study of gender dependence
It should be noted that some differences have been found between genders regarding opinions on certain topics (Arganini et al. 2012; Huddy et al. 2008). Addressing the gender dimension may be of interest in such a hot topic as climate change.
The influence of gender on variables such as: number and percentage of interactions, and polarity of the Twitter messages was studied. A statistical analysis was done applying nonparametric methods (Neuhauser 2017). As a hypothesis test, the Kruskal–Wallis (Verma and Abdel-Salam 2019) test was utilized, and the value of p value was estimated with a confidence interval of 0.95.33 The hypotheses are as follows:
-
Null hypothesis (H0): “The gender of the users does not impact on the variable.”
-
Alternative hypothesis (Ha): “The gender of the users impacts on the variable.”
If p value ≤ 0.05, H0 is rejected and Ha is taken; otherwise, H0 is accepted and Ha is rejected.
3. Results
a. General view
1) Hashtags and topic modeling
(i) Hashtags analysis
Hashtags, as already mentioned, usually represent information about the tweet in a condensed manner. These can be a single word or sets of them and thus they can be used as expert manually crafted n-grams to help in topic identification. Figure 2 shows a word cloud representation of the hashtags found in text. The size of the words is directly proportional to the frequency of appearance within the tweets, so a larger word represents a more used hashtag. This picture will be further explained in the discussion (section 4).
Word cloud image representation of hashtags present in text (formatted as n-grams).
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
(ii) Topic modeling
The most relevant words generated by each topic are retrieved and displayed in Table 3 along with their labels. Labels are manually assigned by inspection looking for the concept or relation between the most important words for each topic. As a form of visual validation, it can be observed that not only can the found topics relate to the top words, but also these could be matched to the most typed hashtags within tweets. Furthermore, to illustrate how the topics are distributed over the extracted data, the number of tweets per topic is displayed in Fig. 3. All of the above will be further described in the discussion (section 4).
Tweet distribution per generated topic with its assigned label.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
LDA 10 topic distribution with associated top words and labels.
2) Opinion mining
As already mentioned, the final polarity is derived from the three algorithms by means of a polling operation. With each algorithm carrying the same weight toward the decision, the most voted polarity is selected. Out of the 90 000 tweets, approximately 30% results in a perfect polling, that is, all algorithms agree. Language has many complications such as negations and ironies, and unsupervised sentiment analysis is a very uncertain and nontrivial task. Although the commercial models VADER and TextBlob have been widely tested in several datasets prior to their publication, their performance may vary when applied to different data. For this reason, a third polarity algorithm is developed so as to have a tiebreaker when the two algorithms are in disagreement.
To compare the performance of the three polarity algorithms, two distance based metrics and two similarity based metrics were calculated to each of the possible pairs. The results are displayed in Table 4. It can be observed that for both types of metrics, the VADER–TextBlob pair slightly outperforms the other pairs.
Measured similarity metrics between polarity algorithms.
Furthermore, the three polarity algorithms are also compared using a confusion matrix approach. The confusion matrix is used to evaluate the accuracy of a classification. It returns the count of true negatives, false negatives, true positives, and false positives. This can be observed in Fig. 4. Since the problem is unsupervised and there are not true labels, the true positives and true negatives are simply those in which two algorithms predict equally and the false positive or false negatives occur when the prediction differs. In all cases, the total number of coincidences is larger than the number of disagreement being the best pair the VADER–TextBlob with over 60 000 coincidences, followed by Word2Vec–VADER and ultimately Word2Vec–TextBlob. Unfortunately, it seems the individual algorithms differ in many cases, so to maximize the performance of the polarity classifier, the three algorithms are accounted for.
Confusion matrices between polarity algorithms.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Finally, positive/negative distribution across tweets according to combined polarity is displayed in Table 5. In general, the negative polarity is slightly predominant (54.8% of all tweets), that is, more people post negatively about climate change regardless of the topic, nationality, or gender.
Tweet distribution per polarity.
3) Genderize and GeoPy
Although the Genderize API counts with 6 084 389 validated names from 191 different countries, the task of gender identification on Twitter is extremely difficult because the username can be anything besides the name, such as an alias or a nickname. Nonetheless, as it can be observed in Table 6, out of the 92 474 tweets analyzed, the gender could be extracted for 67 422. From these, 51 783 tweets were from male (30 704 unique users) and 15 639 from female (11 239 unique users). Although the percentage over the gender-identified tweets is significantly higher for males (76.8%) suggesting that males are more prone to expressing their opinion on social media, this cannot be directly presumed because according to Tankovska (2021b), the number of male users on Twitter is as well far greater (63.7%) than that of females (36.3%). The implications this might have will be further addressed in the discussion (section 4).
Tweet distribution per gender.
Regarding geography, we provide insights into patterns connected to Twitter interactions mainly according to the location the users provide (since those who accept geolocation services are few). Bearing this in mind, out of the 92 474 tweets, nationality was successfully extracted from 47 471. For these subset of the tweets, the total number of unique users is found to be 29 550. Tweets across all continents were found among our data, particularly from 179 different countries worldwide (see Fig. 5). It can be observed that the most numerous tweeting countries are those whose official language is English, for example, the United States, the United Kingdom, and Australia. The tweets for which nationality could not be extracted were neglected for further nationality related analysis.34 Moreover, to provide more insights into the data under analysis, Table 7 provides the number of tweets and of unique users for the more relevant countries identified (top 20).
Frequency of tweets in each country. Approximate scale: 1:230,000,000.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Number of unique users per country.
b. Gender view
1) Number and percentage of tweets per topic
The gender dependence both in number and percentage of tweets in general and per topic was analyzed. Not all these variables present a p value ≥ 0.05 in the Shapiro–Wilks and Fligner–Killeen tests. Therefore, not all have homoscedasticity of variances nor are they normally distributed. In view of the above, the Kruskal–Wallis test is used as a hypothesis test. Both the total number of tweets per country and the number of tweets per country and topic have a p value of less than 0.05. However, the total percentage of tweets per country has a higher p value.35
2) Polarity of tweets per topic
At the level of the entire population both in number and percentage of total tweets and per topic, the influence of the gender on the polarity of the interactions is also checked. As in the analysis carried out in the previous section, this variable does not show a p value > 0.05 in the Shapiro–Wilks or Fligner–Killeen tests. Therefore, there is not homoscedasticity of its variance or normality in its distribution. The Kruskal–Wallis test showed a p value ≤ 0.05. Figures 6 and 7 show the number of tweets by topic and gender or polarity.
Distribution of number of tweets by topic and gender.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Distribution of number of tweets per polarity.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Also, the gender dependence of polarity is examined per nation. According to Kruskal–Wallis test, positive and negative polarities in number of tweets per country present gender dependence (p value ≤ 0.05). This is in opposition to the percentage of tweets where only the positive polarity shows this interrelation.36 All of the above will be further explained in the discussion (section 4).
The number of tweets per polarity and topic, considering the total population is shown in Table 8. The polarity of Twitter messages per topic will be examined in detail in section 4 (discussion).
Number of tweets per polarity and topic.
c. Country/nationality view
For visualization purposes, only the countries with a statistically representative number of tweets are displayed and further analyzed in this section. The selected number of countries is 20 and the minimum number of tweets posted by each country should be larger than 200.
1) Number of tweets per topic
Considering the number of tweets per topic, the nationality dependence on the topics discussed was analyzed. The results can be observed in Fig. 8. Generally, topic 7 (CC activism) is the most discussed. Activism is a broad topic, which could include pollution reduction measures, recycling challenges, webinars, community projects or policy proposals, among others. Thus, it would be reasonable to assume that individuals post about this the most.
Distribution of number of tweets by country and topic.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
In the Fisher exact test a p value > 0.05 was obtained in the pairwise comparison between topics in relation to the number of tweets per country. Hence, the number of tweets per topic are dissimilar for all nationalities.
Figure 9 shows a heat map representing the Euclidean distance between countries, considering the number of tweets per topic (values have been rounded to nearest whole number).37
Euclidean distance between countries. The number of tweets per topic in each nation are considered for the calculation. 1: India, 2: United States, 3: China 4: Spain, 5: United Kingdom, 6: Australia, 7: Canada, 8: Germany, 9: Ireland, 10: Belgium, 11: France, 12: Kenya, 13: Netherlands, 14: South Africa, 15: Switzerland, 16: Nigeria, 17: New Zealand, 18: Pakistan, 19: Italy, 20: Sweden.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
2) Percentage of tweets per topic
Taking into consideration the percentage of tweets per topic, the nationality dependence on the topics discussed was also analyzed. The results are shown in Fig. 10.
Distribution of percentage of tweets by country and topic.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Regarding the pairwise comparison of topics through the Fisher exact test, the obtained p value is shown in Table 9.
Considering the percentage of tweets in each country, obtained p value in the pairwise comparison of topics through the Fisher exact test. In bold are those p values < 0.05.
It can be observed that the largest differences with the rest of the topics occur in topic T7 (CC activism), since a higher number of differences in the pairwise comparisons with the rest of topics with p value ≤ 0.05 exist. This topic is followed by topic T9 (biodiversity).
Figure 11 shows a heat map representing the Euclidean distance between countries, considering the percentage of tweets per topic (values have been rounded up to nearest whole number).38
Euclidean distance between countries. The percentage of tweets per topic in each nation are considered for the calculation. 1: India, 2: United States, 3: China 4: Spain, 5: United Kingdom, 6: Australia, 7: Canada, 8: Germany, 9: Ireland, 10: Belgium, 11: France, 12: Kenya, 13: Netherlands, 14: South Africa, 15: Switzerland, 16: Nigeria, 17: New Zealand, 18: Pakistan, 19: Italy, 20: Sweden.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
It should be noted that the highest differences (greatest distances) occur between country 6 (Australia), and countries 15 (Switzerland), 9 (Ireland), and 12 (Kenya). The countries with the highest differences to the rest are 6 (Australia) and 7 (Canada). The greatest analogies happen between 11 (France) and 20 (Sweden) as well as between 4 (Spain) and 20 (Sweden), 4 (Spain) and 11 (France), and 4 (Spain) and 13 (the Netherlands).
3) Number and percentage of tweets per topic and gender
The nationality dependence both in number and percentage of tweets per topic and gender were analyzed. The results can be observed in Figs. 12 and 13. When talking about absolute values, it can be noted that especially males in the United Kingdom and United States post the most tweets. This is not surprising since the statistics confirm that more users on Twitter are male (Tankovska 2021b) and more accounts belong to U.S. citizens (Tankovska 2021a). Exploring the percentages in more detail, one can draw more insightful conclusions. These will be further addressed in section 4 (discussion).
Distribution of number of tweets by country, topic, and gender.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Percentage of tweets by country, topic, and gender.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
4) Number of tweets per topic and polarity
The nationality dependence both in topic and polarity was analyzed. The results can be observed in Fig. 14. In general, regardless of the topic and nationality, the general trend is to post more negatively about climate change. Further discussion regarding this matter appears in section 4.
Distribution of number of tweets by country, topic, and polarity.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
d. Enterprises view
Around 25 000 tweets from the extracted data could not be assigned a gender, suggesting the presence of business accounts among other possibilities. These data represent a large percentage (30% of the analyzed tweets), thus further assessment must be done to identify the presence of business accounts. The percentage of tweets found to be from organizations, companies or entities is very small (2.18%) in comparison with the percentage of interactions made from individual users (97.82%). Since only those tweets containing a set of climate related keywords were retrieved during data crawling, it seems plausible that people might be more concerned about climate change and that individuals are more prone to giving their opinion on social networks than companies are.
Moreover, the presence of entities was also analyzed with respect to the tweeting country, the topic discussion, and the sentiment polarity (see Figs. 15–17). These will be further addressed in the discussion (section 4).
Distribution of percentage of tweets by topic, from entities and people.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Distribution of percentage of tweets from entities by country.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
Distribution of number of tweets by polarity, from entities and people.
Citation: Weather, Climate, and Society 14, 4; 10.1175/WCAS-D-21-0163.1
The number of tweets with positive and negative polarity from companies and individuals is depicted in Table 10. Moreover, some tweeting examples from entities per country, topic and polarity are presented in Table 11. All this will be further checked in section 4 (discussion).
Polarity of tweets per topic for tweets from companies and persons.
Tweeting examples from entities by country, topic, and polarity (removed URLs and emojis).
4. Discussion
a. General view
The very existence of life on our planet is conditioned by a climate that has changed within a relatively narrow range over hundreds of millions of years (Haywood et al. 2019). However, Earth’s climate is modifying as a result of human actions. Some evidence of these changes exists, such as warming of the atmosphere39 and oceans as well as an increase in sea levels, and a relevant sharp reduction in Arctic sea ice (United Nations 2022a.). Because of this, for some years now, there has been an international concern to achieve net zero emissions.
In 2015, the United Nations established the 2030 Agenda for Sustainable Development. It includes 17 goals that aim to eradicate the poverty, protect the planet, and improve the quality of life of all citizens (United Nations 2020). The Paris Agreement on climate change urged for keeping eventual warming “well below” 2°C (United Nations 2022b; National Academy of Sciences and Royal Society 2020). A strategic long-term vision was adopted by the European Commission in 2018. The objective is to achieve a climatic-neutral economy, which is boom-town, modern, and competitive by 2050 (European Commission 2018). Additionally, the European Parliament declared climate emergency in 2019 (European Parliament 2019). At the end of last year, the United States announced the Net Zero World initiative, a new cross-country partnership that aims to work to move faster the transition to net zero emissions, resilient and inclusive of alternative energy systems (EnergyGov 2021).40 India and China also made a commitment to reduce its emissions to net zero by 2070 (BBC 2021) and before 2060 (Xie 2021) in the United Nations Climate Change Conference (COP26).41
The hashtag analysis and topic modeling performed below reflects a public concern about the effects of climate change, in line with what has been explained previously concerning governments and international organizations.
Regarding hashtags analysis, from Fig. 2 one can observe different subtopics within the data and could broadly identify what these are. For example, it can be noted that words such as environment, Earth Day, energy, and COP26 are frequently mentioned, and the underlying topics such as renewable energies, sustainability, the environment, pollution, or the pandemic can be identified.
It should also be noted that the economic slowdown produced by the COVID-19 pandemic has had an impact on the environment. The amount of oil consumed in all transport sectors was drastically reduced (Rugani and Caro 2020). Industrial production and electricity demand also suffered a significant decrease, while there was an increase in the utilization of renewable energy sources in some countries (Halbrügge et al. 2021). As result, satellite images from the National Aeronautics and Space Administration (NASA) and the European Space Agency (ESA) showed significant improvements in air quality worldwide (Mulvaney et al. 2020).
With respect to topic modeling, in Fig. 3 it can be observed that the most discussed topic is CC activism with approximately 20 000 tweets, followed by biodiversity or CC evidence, each with around 13 000 tweets. In accordance with this, looking at the word cloud representation in Fig. 2 it can be seen that the hashtags related to these topics are the most present. For instance, words such as Earth Day, SDG, COP, environmental social and corporate governance (ESG), and more can be associated to activism, and words such as nature, environment, forest, and water, among others are very related to biodiversity. Certainly, the climate activist events are powerful drivers of attention compared to political events and temperature anomalies (Sisco et al. 2021).
A smaller number of tweets are found for the topics government action, policies and finances, and climate emergency. However, one of the most critical issues in combating the effects of climate change is funding because large-scale investments are needed to significantly reduce emissions, particularly in those sectors that are major producers of GHG. A report prepared by Climate Finance Advisors for the United Nations Environment Program Finance Initiative (UNEP FI) analyzed in detail these financial aspects in 2019 (Miller et al. 2019).
To achieve the mentioned international objectives, internationally coordinated actions need to be undertaken (IRENA 2019): governments must increase and reprioritize public spending on research and development, paying more attention to areas such as electrification, hydrogen, bioenergy and carbon capture, utilization, and storage, while encouraging projects and developing policies that facilitate private investment. They must also take into account the social and economic impacts of the energy transition process and encourage the participation of all citizens (IRENA 2019).
Concerning opinion, it was observed that generally what is talked about on Twitter regarding climate change tends to be more negative than positive. Some of the identified topics are, for example, sustainability or CC activism, which could be positive when discussed in isolation. However, there exist more negatively associated topics exist such as the pandemic, climate emergency, and CC evidence. Moreover, other topics such as biodiversity or net zero are more neutral by themselves, showing a similar level of probability to be spoken of either positively or negatively.
b. Gender view
The p value obtained in Kruskal–Wallis test is lower than 0.05 in all cases. Therefore, both the percentage and the number of tweets per country and topic depends on gender.
Considering the total population and all topics, the polarity regarding what it is talked about in relation to climate change is more negative (54.8%) than positive (45.2%). Also, at this general level, the tendency of both males and females is to post more negatively about climate change than positively: 54.90% of tweets from men are positive and 45.10% are negative. Regarding women, 54.45% and 45.55% are negative and positive, respectively. However, if polarity is studied over the entire population per topic, some differences can be observed. This is described below.
As can be observed in Table 8, the topics with a percentage of tweets larger than 10% are CC activism, CC evidence, biodiversity, sustainability, and CC awareness. Note the pandemic also presents a percentage equal to 10 for the female sex. The topics CC activism, CC awareness, biodiversity, sustainability, and the pandemic have a larger positive polarity for the male sex, whereas for the female sex these are policies and finances, biodiversity, government action, climate emergency, and the pandemic. In contrast, the topics CC evidence, biodiversity, CC activism, sustainability, and the pandemic have the largest negative polarity for the male sex. Policies and finances, climate emergency, CC activism, government action, and CC evidence show the highest negative polarity for the female sex. It can be noted that CC activism and biodiversity seem to be those that generate the most contentious debates among men. This happened for policies and finances and climate emergency) among women. For CC activism, negative opinions predominate among women.
Next, the gender dependence of polarity is examined per nation. As is shown in the supplemental material, positive and negative polarities in number of tweets per country present gender dependence (p value ≤ 0.05). This is in opposition to the percentage of tweets where only the positive polarity shows this interrelation. Regarding the positive polarity per country and topic, net zero and government action do not exhibit association with gender, in contrast with the rest of the topics (p value > 0.05). The negative polarity only presents a relation (p value ≤ 0.05) to gender in topics CC evidence, CC awareness, sustainability, and biodiversity.
c. Country/nationality view
Regarding the absolute values, that is, the number of tweets identified per country, the countries with the largest presence over the data analyzed were United States and United Kingdom by far with a total number of identified tweets between 11 000 and 13 000. Following this, Australia, Canada, and India can be identified, with around 3000 tweets. Bearing in mind that the study is performed in English, it seems understandable that the dominant countries are those in which English is the official language. Information on number of tweets per country can be seen in Fig. 5. Numerical information on frequency and percentage of tweets is also included in the supplemental materials.
When the topic distribution is analyzed in percentage, that is, number of tweets per topic over the total number of tweets from that country, several differences can be appreciated. For example, in general, the United States is the leading country in terms of tweet quantity. However, it is interesting that looking at CC activism (T7), which is the most discussed topic, only 20% of tweets posted in the United States are found to be related to this topic. On the contrary, Thailand, Switzerland, and Austria show almost twice the level of interest in it (around 40% of their total number of tweets).
As mentioned in section 3 (results), in the Fisher exact test a p value > 0.05 was obtained in the pairwise comparison between countries regarding their ratio of tweets per topic. It can be observed in Fig. 9, that the countries with the largest differences with the rest are the United States and the United Kingdom and, to a lesser extent, India, Australia, and Canada.
Regarding polarity, if all topics are considered globally, negative polarity predominates in all countries. This also occurs for the topics T0 (net zero), T5 (policies and finances), T6 (CC awareness), T7 (CC activism), and T9 (biodiversity). The topic T1 (pandemic) only has higher positive polarity in China and Italy. T2 (climate emergency) presents only larger positive polarity in Belgium and the United States. It is important to note that in 2021, the U.S. Congress introduced a regulation that would force the declaration of a climate emergency (U.S. Congress 2021) in the country. Also, the municipality of Koekelberg (Galindo 2019a) and the city of Brussels stated a climate emergency in 2019 (Galindo 2019b).42 The strong willingness in these countries to take action against the damage caused by climate change is aligned with the positive perception toward emergency actions detected on Twitter.
A few countries exhibit analogous positive and negative polarities. T3 (CC evidence) and T4 (government action) have a higher positive polarity in Switzerland. With respect to T8 (sustainability) only in Nigeria, Spain, and Sweden does a positive polarity predominate. Also, some countries show identical magnitudes in positive and negative polarities. It is relevant to note that, according to the tenth edition of the global energy trilemma index, published in 2020 (World Energy Council and Oliver Wyman 2020), which analyzed 128 countries’ energy systems, Sweden ranked second only to Switzerland in the trilemma 2020 score. This assessed energy security, environmental sustainability, and energy equity. Similar to what was happening in the specific area of sustainability. Spain ranked 15th, and Nigeria ranks very low in the overall index.
d. Enterprises view
From the Global Forbes 2000 largest companies worldwide, only 12 companies could be matched to our data. By inspection, it was observed that most of the 773 companies found among our data represent small businesses. At least 75 organizations are local news journals or magazines. Several statistics from Dudar (2012) point out the usage of Twitter by small businesses to boost their sales and revenues. It is worth noting that 70% of Small Businesses are on Twitter, and this might be because 67% of Twitter users are more likely to buy from the brands they follow.
Interestingly, countries such as the United States, the United Kingdom, and Australia, which have the largest tweeting presence, show a smaller percentage of tweets from business accounts than other countries (see Fig. 16). This suggests that the individuals from these countries are more concerned about climate change than companies are. Instead, countries such as Italy, Kenya, and Belgium whose tweeting frequency is smaller, have a larger percentage of tweets from businesses. In turn, this suggests that companies from these countries are more prone to discussing climate change.
The finding regarding Italy seems to be in line with Balluchi et al. (2020), which explains that Italian Legislative Decree 254/2016 established for the first time that “public interest entities” would be required to integrate statutory financial statements with disclosure of environmental, social, and governance strategies from the financial year 2017 onward. Balluchi et al. (2020) show that, although not mandatory, the number of corporate social responsibility (CSR) reports published grew steadily from 2007 until 2016. Besides, there existed a notable increase in the number of companies that, in 2017, generated a nonfinancial report after the introduction of the new regulation.
Belgium also has a strong commitment with the so-called sustainable development, both at the state and the federated dimensions. The concern for sustainability is even reflected in the Belgian constitution. Additionally, several plans on the implementation of the European 2030 Agenda have been built in order to establish a suitable support for sustainable development policies (Government of Belgium 2017). Muthuri and Gilbert (2011) describe that firm-related drivers such as public relations and performance as well as global institutional pressures justify the focal point and configuration of CSR in Kenya.
Furthermore, the distribution of tweets by topic from entities and people is as well analyzed (see Fig. 15). It can be observed that organizations post with more frequency about sustainability (T8), CC awareness (T6), and net zero (T0), whereas individuals post more about the pandemic (T1), net zero, and policies and finances (T5). It can also be observed that companies pay very little attention to topics such as climate emergency (T2) and CC evidence (T3), showing that organizations are more concerned about reducing pollution. As a result, they wish to communicate this in order to encourage action, rather than actually show evidence about it. This behavior may be motivated by CSR, which is commonly utilized by enterprises as a component in building up their media image (Andrejczuk 2010). Through CSR, companies can achieve a leading position in the sector (Księżak 2017). A positive company image is a key aspect in the achievement of success because it results in a high reputation among its customers. This increases the probability of both strengthening its position and expanding in the market (Andrejczuk 2010). Also, certain CSR practices, such as the minimization of waste, strongly impact operational efficiency and thus achieving profit maximization (Perry and Towers 2013).
In contrast, individuals on Twitter seem more prone to discussing climate change evidence so as to encourage or to persuade people of the imminent need for measures while also calling on governments and large companies to make policies and to allocate money toward this matter. Finally, also the distribution of tweets by polarity from entities and people is studied. As depicted in Fig. 6 and Table 10, the organizations tend to speak more positively about climate change than individuals. If topic and polarity were jointly analyzed, one could observe, as Fig. 15 shows, that sustainability and net zero are more likely to be expressed positively because they would most likely focus on emissions reduction goals. Companies could elect to send this message consequently gain popularity among individual users. Additionally, there has been a rise of companies with vegan offerings—for example, in cosmetics, food, or fashion products (Göbel et al. 2017; Manyukhina and Middlemiss 2017)—that can also be exploited positively by the producers. On the contrary, individual posts are more negative. Since the most discussed topic is the pandemic, it is understandable that the polarity appears more negative since death rates have been high and citizen restrictions have been tough and strict.
5. Conclusions and future research
The conclusions obtained from the characterization of the interactions per topic and polarity by gender, nationality and account typology are detailed below.
a. Topic modeling
For this corpus, having formatted both hashtags and multiword expressions into single words prior to topic detection, and in order to incorporate them into the text of each tweet, LDA yields better results than biterm, according to the UMASS coherence index. The 10 topics detected in the interactions are net zero, pandemic, climate emergency, CC evidence, government action, policies and finances, CC awareness, CC activism, sustainability, and biodiversity.
b. Interactions per topic
The most discussed topic is CC activism, possibly because it is a broad topic that could include others such as webinars, pollution reduction measures, recycling challenges, community projects, or policy proposals. Followed by biodiversity and CC evidence. The least discussed topics are government action and climate emergency. The previous results are similar between men and women, with both genders tweeting similarly regarding the topic.
For entities, the topic discussion concentrates more on sustainability, CC awareness, and net zero, probably due to the need for enterprises to keep a clean image (corporate social responsibility). It is also worth noting that companies pay less attention to topics such as climate emergency and CC evidence, showing that organizations appear more concerned about reducing pollution than they are about showing evidence.
In each country, the proportion of tweets per topic is different. The most discussed topic in terms of nationality varies from place to place, with the exemption of CC activism, which is predominant everywhere. For example, the United Kingdom appears to be more concerned about CC awareness than the United States is, whereas the United States posts more than the United Kingdom regarding biodiversity. India and the United States have the most relevant differences in relation to the proportion of tweets per topic. In terms of percentage of total tweets, the most discussed topic is T7 (CC activism) and the least discussed is T2 (climate emergency).
c. Polarity of the interactions
At the level of the population as a whole and of all interactions, negative polarity predominates. This is also the case when distinguishing by gender. However, considering the entire population and the topics separately, there are some differences in polarity according to gender. This is also the case if countries and topics are treated separately. T1 (pandemic) has only a higher positive polarity in China and Italy. For T2 (climate emergency) this is the same in Belgium and the United States. T3 (CC evidence) and T4 (government action) only exhibit a higher positive polarity in Switzerland. T8 (sustainability) only presents a larger positive polarity in Sweden, Spain, and Nigeria.
Regarding topic and polarity some imbalances can be observed. For example, topics such as CC evidence and biodiversity are spoken of more negatively (over two-thirds of them are negative), whereas CC activism is spoken of more positively.
It is worth mentioning that organizations post more positively regarding climate change than individuals do. This seems reasonable and is in accordance with the nature of the topics most discussed. For example, entities talk about sustainability, which is usually positive by itself.
Finally, it can be concluded that the computing applications developed here are a great help toward gaining a deeper understanding of the perceptions on climate change.
We also point out that all currently available tools and software packages, as well as the software programs developed and used in this research by the authors, can also be useful for sentiment analysis on Twitter regarding other issues of social interest.
d. Future research
The analysis of other social networks or online media can be incorporated into the research. With specific regard to interactions on Twitter: a subjectivity classifier and a misspelling corrector can be implemented in order to improve the polarity detection accuracy. Besides this, to further validate the coherence of the identified topics, two human evaluation tasks (word intrusion and topic intrusion; Chang et al. 2009) could be carried out. As a means to find out whether the perception of climate change depends on the language of the target audience, a similar study to this research but considering other languages can be done. Also, the perception of enterprises could correlate with their legal form, operational sector, and size. Moreover, a cross-perspective topic model approach similar to Fang et al. (2012) could be incorporated.
Additionally, a study of the relationships between users interacting can be performed, which would include the following:
-
A topological examination of betweenness centrality, degree, eigenvector centrality, and page rank (Newman 2003). In addition to checking both assortativity and community structure (Cherifi et al. 2019). The following community detection algorithms could be applied: Girvan–Newman procedure (Girvan and Newman 2002), fluid communities method (Parés et al. 2017), label propagation algorithm (Zhu and Ghahramani 2002), the clique percolation method (Kumpula et al. 2008), and the Kernighan–Lin algorithm (Kernighan and Lin 1970).
-
Development of algorithms for predicting hidden or new links in the interaction networks (Malhotra and Goyal 2020).
The number of tweets used for sentiment analysis greatly varies between types of research. Rasool et al. (2019) analyze the public perception of two relevant international clothing brands, Adidas and Nike. The authors retrieved 99 850 tweets on irregular dates in 2018 by utilizing the apparel brand’s name as a keyword. They used 17 006 tweets, which were the result of eliminating non-English tweets as well as the retweets. Hidroy et al. (2015) analyze the popularity/opinion/sentiment towards iPhone 6 product considering 940 tweets from users in New York (182), Los Angeles (89), Boston (103), Chicago (143), Dallas (156), San Francisco (138), and Philadelphia (129).
Such as incorporation of hashtags and multiword expression as n-grams to the corpus.
See supplemental materials.
A reply is when the user responds to another person’s tweet, a retweet is the way one forwards another user’s tweet to their followers, and a quote is a retweet posted with an additional comment.
List of unique words from all tweets.
An n-gram is a subsequence of n elements of a given sequence. For our purpose, n-grams will be constructed considering words as the smallest unit.
For example, the hashtag #HumanSolidarity would be formatted to Human_Solidarity.
We consider that a sequence of words a1…am must be accepted as a multiword expression, if it verifies
More details on LDA method can be found in the supplemental material.
According to section 2a.
With these settings the dictionary is reduced from 64 530 to 6741 words. From the dictionary and the processed text of each tweet, a corpus is generated.
Where the first element symbolizes the numerical identifier of the word and the second focuses on the times that word appears in the processed text corresponding to each tweet.
In contrast, LDA learns topics by modeling word–documents co-occurrences.
More details about this method can be found in the supplemental material.
According to section 2a.
It is defined as
For LDA and biterm algorithms, the calculation was done using the gensim and biterm Python libraries, respectively.
According to section 2a.
According to section 2a.
A one-hot vector symbolizes a word as an R|V|×1 vector with all 0s and one 1 at the index of that word in the vocabulary V.
More details about Word2Vec can be found in the supplemental material.
We utilize WordSim353 (2022; Agirre et al. 2009; Finkelstein et al. 2002), which contains a split of the test set into two subsets, one for evaluating similarity and the other for evaluating relatedness (Agirre et al. 2009). For those word pairs included in the repository and also in the text of tweets, embeddings were calculated. Jaccard and cosine similarity coefficients as well as Manhattan and Euclidean distances between these word pairs was also estimated.23 Finally, the Spearman correlation between all these pairs and the similarities indicated for them in the WordSim353 dataset was computed. A correlation higher than 0.9 was obtained considering the previously mentioned parameter values.
All definitions of these metrics are included in the supplemental material.
More details about k-means algorithm can be found in the supplemental material.
The “most_similar” method of the gensim library was utilized. For the first 100 words with similarity in the range [0.9–1] ordered from highest to lowest, we manually check which was the predominant sentiment.
The value obtained is closer to one the closer the word vectors are to the centroid. This quantity is closer to 0 the farther away they are from the centroid.
All definitions of these metrics are included in the supplemental material.
It returns a dictionary containing the name, gender, probability, and count. The gender response is either male, female, or none. The probability indicates the certainty of the assigned gender, and the count represents the examined data used to calculate the response.
If this distance is 0, the tweet is considered to belong to the company that is a specific list item. As a result, 12 companies are matched, Nestlé, Ferrovial, and Medtronic, among others.
To avoid misclassification of companies’ accounts and to ensure that personal accounts are neglected, some criteria must be met prior to applying NER. From the 25 052 tweets with no gender assigned, those with less than 1000 followers and with a ratio followers/following smaller than 5 are removed. The remaining tweets (3782) are further analyzed and if the tag “ORG” is present, the account is considered as a candidate. From the candidates, all accounts containing hashtags or emojis in their usernames are removed, this results in a total of 773 companies posting 1507 tweets. From the assessed, that is, almost 50%.
Sometimes, the Fisher exact test on an r × c table is named Fisher–Freeman–Halton test (Freeman and Halton 1951).
An example of a contingency table is shown in the supplemental material.
This test does not require normality or homoscedasticity of variances in the distribution. The normality and the homoscedasticity of variances in the distribution is checked by applying the Shapiro–Wilks and the Fligner–Killeen tests, respectively. A confidence interval with a value of 0.95 is considered.
Numerical data related to frequency and percentage of tweets per nation are included in the supplemental materials. Only those countries with a tweeting frequency larger than the mean tweeting frequency (265 tweets) were considered.
The detailed results are depicted in the supplemental material.
Detailed test results are shown in the supplemental material.
Similar results are obtained when the Manhattan distance is used, a graphical representation is included in the supplemental material.
Analogous results are gotten if the Manhattan distance is utilized, a picture is included in the supplemental material.
The United Nations weather agency reported that last year joined the list of the seven warmest years on record (United Nations 2022c).
This proposal is led by the U.S. Department of Energy (DOE) as part of the Build Back Better World plan (EnergyGov 2021).
In 2015, the top 10 GHG emitting countries generated >60% of total emissions: China (21.1%), the United States (14.1%), and India (5.2%) are the greatest contributors (Althor et al. 2016).
Similarly, the European Union also issued an emergency declaration that year (European Parliament 2019).
Acknowledgments.
This work was carried out as a result of the project “Hopper: Women, Society, Technology and Education” that was granted in the internal call for research projects in 2021 at the Universidad Francisco de Vitoria. The authors thank Mari Luz Congosto Martínez for her guidance on the use of the T-Hoarder tool. The authors also express gratitude to Javier Gómez Sánchez-Seco for his assistance in the world map construction.
Data availability statement.
No data repositories are used.
REFERENCES
Agirre, E., E. Alfonseca, K. Hall, J. Kravalova, M. Pasca, and A. Soroa, 2009: A study on similarity and relatedness using distributional and WordNet-based approaches. Proc. NAACL-HLT 2009: The 2009 Annual Conf. of the North American Chapter of the Association for Computational Linguistics, Boulder, CO, Association for Computational Linguistics, 19–27.
Ahmed, M. S., T. T. Aurpa, and M. M. Anwar, 2021: Detecting sentiment dynamics and clusters of Twitter users for trending topics in COVID-19 pandemic. PLOS ONE, 16, e0253300, https://doi.org/10.1371/journal.pone.0253300.
Alsaeedi, A., and M. Z. Khan, 2019: A study on sentiment analysis techniques of Twitter data. Int. J. Adv. Comput. Sci. Appl., 10, 361–374, https://doi.org/10.14569/IJACSA.2019.0100248.
Althor, G., J. Watson, and R. Fuller, 2016: Global mismatch between greenhouse gas emissions and the burden of climate change. Sci. Rep., 6, 20281, https://doi.org/10.1038/srep20281.
An, X., A. Ganguly, Y. Fang, S. Scyphers, A. Hunter, and J. Dy, 2014: Tracking climate change opinions from Twitter data. Workshop on Data Science for Social Good, New York City, NY, KDD, http://cobweb.cs.uga.edu/∼squinn/mmd_s15/papers/KDD_Twitter_ClimateChange.pdf.
Anber, H., S. Akram, and A. A. Abd El-Aziz, 2016: A literature review on Twitter data analysis. Int. J. Comput. Electr. Eng., 8, 241–249, https://doi.org/10.17706/IJCEE.2016.8.3.241-249.
Andrejczuk, M., 2010: Corporate social responsibility in shaping the media image of the company. J. Educ. Culture Soc., 1, 19–25, https://doi.org/10.15503/jecs20101.19.25.
Anjaria, M., and R. M. R. Guddeti, 2014: Influence factor based opinion mining of Twitter data using supervised learning. 2014 Sixth Int. Conf. on Communication Systems and Networks, Bangalore, India, IEEE, 1–8, https://doi.org/10.1109/COMSNETS.2014.6734907.
Arganini, C., A. Saba, R. Comitato, F. Virgili, and A. Turrini, 2012: Gender differences in food choice and dietary intake in modern western societies. Public Health: Social and Behavioral Health, J. Maddock, Ed., IntechOpen, https://doi.org/10.5772/37886.
Balluchi, F., K. Furlotti, and R. Torelli, 2020: Italy towards mandatory sustainability reporting voluntary corporate social responsibility disclosure of Italian companies and legislative decree 254/2016 statements: A quantitative analysis of the last 10 years. Sustainability and Law, V. Mauerhofer, D. Rupo, and L. Tarquinio, Eds., Springer, 257–274, https://doi.org/10.1007/978-3-030-42630-9_14.
BBC, 2021: COP26: India PM Narendra Modi pledges net zero by 2070. BBC News, 2 November, https://www.bbc.com/news/world-asia-india-59125143.
Bi, T., P. Liang, A. Tang, and C. Yang, 2018: A systematic mapping study on text analysis techniques in software architecture. J. Syst. Software, 144, 533–558, https://doi.org/10.1016/j.jss.2018.07.055.
Bird, S., 2021: The Natural Language Toolkit (NLTK) PyPI, accessed 18 May 2022, https://pypi.org/project/nltk/.
Blei, D. M., and J. D. Lafferty, 2007: A correlated topic model of science. Ann. Appl. Stat., 1, 17–35, https://doi.org/10.1214/07-AOAS114.
Blei, D. M., A. I. Ng, and M. I. Jordan, 2003: Latent dirichlet allocation. J. Mach. Learn. Res., 3, 993–1022.
Bonta, V., N. Kumaresh, and N. Janardhan, 2019: A comprehensive study on lexicon based approaches for sentiment analysis. Asian J. Comput. Sci. Technol., 8 (Suppl. 2), 1–6, https://doi.org/10.51983/ajcst-2019.8.S2.2037.
Bouma, G., 2009: Normalized (pointwise) mutual information in collocation extraction. Proc. Int. Conf. of the German Society for Computational Linguistics and Language Technology, Potsdam, Germany, German Society for Computational Linguistics and Language Technology, 43–53.
Buhin, P. M., J. Dobša, and L. Kronegger, 2020: Topic modeling in social sciences: Case study of Web of Science. Proc. Central European Conf. on Information and Intelligent Systems, Varaždin, Croatia, University of Zagreb Faculty of Organization and Informatics, 211–218.
Chang, J., S. Gerrish, C. Wang, J. L. Boyd-Graber, and D. M. Blei, 2009: Reading tea leaves: How humans interpret topic models. Advances in Neural Information Processing Systems, Vol. 22, Y. Bengio et al., Eds., Curran, 288–296.
Cherifi, H., G. Palla, and B. K. Szymanski, 2019: On community structure in complex networks: Challenges and opportunities. Appl. Network Sci., 4, 117, https://doi.org/10.1007/s41109-019-0238-9.
Cigarrán, J. M., A. Castellanos, and A. Garcia-Serrano, 2016: A step forward for topic detection in Twitter: An FCA-based approach. Expert Syst. Appl., 57, 21–36, https://doi.org/10.1016/j.eswa.2016.03.011.
Cody, E., A. Reagan, L. Mitchell, P. Dodds, and C. Danforth, 2015: Climate change sentiment on Twitter: An unsolicited public opinion poll. PLOS ONE, 10, e0136092, https://doi.org/10.1371/journal.pone.0136092.
Congosto, M. L., P. Basanta-Val, and L. Sanchez-Fernandez, 2017: T-Hoarder: A framework to process Twitter data streams. J. Network Comput. Appl., 83, 28–39, https://doi.org/10.1016/j.jnca.2017.01.029.
Dahal, B., S. A. P. Kumar, and Z. Li, 2019: Topic modeling and sentiment analysis of global climate change tweets. Soc. Network Anal. Min., 9, 24, https://doi.org/10.1007/s13278-019-0568-8.
Deshwal, A., and S. Sharma, 2016: Twitter sentiment analysis using various classification algorithms. Fifth Int. Conf. on Reliability, Infocom Technologies and Optimization (Trends and Future Directions), Noida, India, IEEE, 251–257, https://doi.org/10.1109/ICRITO.2016.7784960.
De Silva, L., and E. Riloff, 2014: User type classification of tweets with implications for event recognition. Joint Workshop on Social Dynamics and Personal Attributes in Social Media, Baltimore, MD, Association for Computational Linguistics, 98–108, https://doi.org/10.3115/v1/W14-2714.
Dudar, B., 2012: Twitter for small business. Search Engine People, accessed 18 May 2022, https://www.searchenginepeople.com/blog/twitter-for-small-business-infographic.html.
Edwards, S., and D. Santos, Eds., 2014: Revolutionizing the Interaction between State and Citizens through Digital Communications. IGI Global, 330 pp., https://doi.org/10.4018/978-1-4666-6292-6.
Eichhorn, J., L. Molthof, and S. Nicke, 2020: From climate change awareness to climate crisis action public perceptions in Europe and the United States. Open Society European Policy Institute, 46 pp., https://www.opensocietyfoundations.org/uploads/d366d228-3b62-45d3-8894-6a2b5d693803/from-climate-change-awareness-to-climate-crisis-action-20201120.pdf.
EnergyGov, 2021: U.S. Launches Net-Zero World Initiative to accelerate global energy system decarbonization. Accessed 18 May 2022, https://www.energy.gov/articles/us-launches-net-zero-world-initiative-accelerate-global-energy-system-decarbonization.
European Commission, 2018: The Commission calls for a climate neutral Europe by 2050. Accessed 18 May 2022, https://ec.europa.eu/clima/news-your-voice/news/commission-calls-climate-neutral-europe-2050-2018-11-28_es.
European Commission, 2021: Citizen support for climate action. Accessed 18 May 2022, https://ec.europa.eu/clima/citizens/citizen-support-climate-action_en.
European Parliament, 2019: The European Parliament declares climate emergency. Accessed 18 May 2022, https://www.europarl.europa.eu/news/en/press-room/20191121IPR67110/the-european-parliament-declares-climate-emergency.
Fang, Y., L. Si, N. Somasundaram, and Z. Yu, 2012: Mining contrastive opinions on political texts using cross-perspective topic model. Proc. Fifth ACM Int. Conf. on Web Search and Data Mining, Seattle, WA, Association for Computing Machinery, 63–72, https://doi.org/10.1145/2124295.2124306.
Fernádez, J., J. Gómez, and P. Martínez-Barco, 2014: A supervised approach for sentiment analysis using skipgrams. Proc. Workshop on Natural Language Processing in the Fifth Information Systems Research Working Days (JISIC), Quito, Ecuador, Association for Computational Linguistics, 30–36, https://doi.org/10.3115/v1/W14-6904.
Fernández-Gavilanes, M., T. Álvarez-López, J. Juncal-Martínez, E. Costa-Montenegro, and F. J. González-Castaño, 2015: GTI: An unsupervised approach for sentiment analysis in Twitter. Proc. Ninth Int. Workshop on Semantic Evaluation, Denver, CO, Association for Computational Linguistics, 533–538, https://doi.org/10.18653/v1/S15-2089.
Finkelstein, L., E. Gabrilovich, Y. Matias, E. Rivlin, Z. Solan, G. Wolfman, and E. Ruppin, 2002: Placing search in context: The concept revisited. ACM Trans. Inf. Syst., 20, 116–131, https://doi.org/10.1145/503104.503110.
Foltz, P., 1996: Latent semantic analysis for text-based research. Behav. Res. Methods Instrum. Comput., 28, 197–202, https://doi.org/10.3758/BF03204765.
Freeman, G. H., and J. H. Halton, 1951: Note on exact treatment of contingency, goodness-of-fit and other problems of significance. Biometrika, 38, 141–149, https://doi.org/10.1093/biomet/38.1-2.141.
Galindo, G., 2019a: Koekelberg becomes first Belgian municipality to declare ‘climate emergency.’ Brussels Times, 15 May, https://www.brusselstimes.com/news/belgium-all-news/56711/koekelberg-becomes-first-belgian-municipality-to-declare-climate-emergency/.
Galindo, G., 2019b: Brussels declares state of climate emergency. Brussels Times, 24 September, https://www.brusselstimes.com/brussels/69504/brussels-declares-state-of-climate-emergency/.
Gilani, Z., R. Farahbakhsh, G. Tyson, L. Wang, and J. Crowcroft, 2017: An in-depth characterisation of Bots and Humans on Twitter. arXiv, 1704.01508, https://arxiv.org/abs/1704.01508.
Girvan, M., and M. E. J. Newman, 2002: Community structure in social and biological networks. Proc. Natl. Acad. Sci. USA, 99, 7821–7826, https://doi.org/10.1073/pnas.122653799.
Göbel, C., P. Teitscheid, S. Friedrich, N. Langen, M. Speck, T. Engelmann, and H. Rohn, 2017: Implementing sustainable business models in the hospitality sector with the help of a mission statement. J. Food Syst. Dyn., 8, 165–169, https://doi.org/10.18461/ijfsd.v8i2.827.
Government of Belgium, 2017: Pathways to sustainable development: First Belgian national voluntary review on the implementation of the 2030 agenda. United Nations High Level Political Forum, 93 pp., https://sustainabledevelopment.un.org/content/documents/15721Belgium_Rev.pdf.
Grimmer, J., 2010: A Bayesian hierarchical topic model for political texts: Measuring expressed agendas in senate press releases. Polit. Anal., 18, 1–35, https://doi.org/10.1093/pan/mpp034.
Halbrügge, S., P. Schott, M. Weibelzahl, H. Ulrich, G. Fridgen, and M. Schöpf, 2021: How did the German and other European electricity systems react to the COVID-19 pandemic? Appl. Energy, 285, 116370, https://doi.org/10.1016/j.apenergy.2020.116370.
Hamad, R., S. Alqahtani, and M. Torres, 2017: Emotion and polarity prediction from Twitter. Computing Conf., London, United Kingdom, IEEE, 297–306, https://doi.org/10.1109/SAI.2017.8252118.