Twitter and the circular economy: examining the public discourse

4.1 Frequency and correlation analysis

The first step of investigation concerns the frequency of occurrence of various text elements, as recommended, e.g. by Kwartler (2017). The rationale is that text elements occurring most frequently represent the most relevant themes. In this section, we describe the tools we employ to: (1) perform a visual analysis of the occurrences of single words through word clouds; (2) compute the frequency of appearance of individual words as well as bigrams and trigrams; (3) analyse the co-occurrence of words in the text through association analysis.

The first tool we employ for that purpose is the word cloud. We employ word clouds here as a qualitative starting point for a deeper analysis, as suggested by Sinclair and Cardew-Hall (2008) and Burch et al. (2013), though they may evolve into a more detailed text analytics tool — see, e.g. the word cloud explorer proposed by Heimerl et al. (2014).

We then compute the frequency of appearance of individual words in the text, followed by the computation of frequency of bigrams and trigrams, i.e. sequences of two or three adjacent words. Adjacency is detected after removing stopwords. We perform a similar analysis for hashtags as well.

In addition to examining the close occurrence of different words, we also investigate the context where words appear by observing the co-occurrence of words in the same tweet, though not so close as in bigrams and trigrams. We carry out that task by computing the association score. Association is computed as the correlation of the word of interest with any other word in a Document-Term matrix (DTM). Scores range from 0 to 1. A score of 1 means that the two words (the word of interest and the other word appearing in the DTM) always appear together in documents, whilst a score approaching 0 means that those words seldom appear in the same document (see Section 3.2.2 of Kwartler (2017) for a more thorough definition with examples).

4.2 Topic modelling

Aside from the occurrence of single words or groups of two or three words, we are also interested in examining the topics that spur the interest of twitterers. We have already sketched a brief classification of topics in Section 5.1. In this section, we describe three algorithms we employ to identify topics automatically:

1. Latent Dirichlet Allocation (LDA);

2. BiTerm Topic Modelling (BTM);

3. Global Vectors for Word Representation (GloVe).

LDA is frequently used to capture the set of latent topics over documents within a corpus. LDA models the generation of documents according to the following process (Jónsson and Stolee, 2015; Silge and Robinson, 2017): (1) a mixture of k topics, θ, is sampled from a Dirichlet prior, which is parameterised by α; (2) a topic z_n is sampled from the multinomial distribution, p(θ; α), which models pzn=i∣θ; (3) given the topic z_n, a word, w_n, is then sampled via the multinomial distribution pw∣zn.

The R library provided by Grün and Hornik (2011), which implements the variational expectation-maximisation (VEM) algorithm for the latent Dirichlet allocation (LDA) model, was used to gather experimental data and compare it to other models.

However, LDA shows poor performance on shorter documents. One method to address this issue is aggregating documents into longer pseudo-documents. For Twitter data, this means combining tweets created by the same author (Jónsson and Stolee, 2015). Then a standard LDA model is trained on this modified data set. Yan et al. (2013) denoted this variation of the LDA model as LDA-U.

The second algorithm we employ is BTM, which was proposed by Yan et al. (2013).

Unlike conventional topic models, which learn topics based on document-level word co-occurrence patterns, where the word co-occurrence patterns become very sparse in each document, Yan et al. (2013) proposed a model which learns topics over short texts by directly modelling the generation of all the biterms in the whole corpus.

BTM directly models the word co-occurrence patterns based on biterms. We extract any two distinct words in a tweet as a biterm.

The biterms extracted from all the tweets in the dataset compose the training data of BTM. Exploiting the whole corpus allows us to tackle the sparsity problem in a single tweet. Hence, the whole corpus is a mixture of topics, where each biterm is drawn from a specific topic independently (Wijffels, 2020).

As discussed by Yan et al. (2013), BTM is different from conventional topic models because, e.g. LDA will suffer from the sparsity problem due to its excessive reliance on local observations for the inference of word topic assignment z, which in turn hurts the learning of topics ϕ.

We refer the reader to the paper by Yan et al. (2013) for a detailed description of the algorithm to infer the BTM parameters.

In addition to grouping posts based on keywords' frequency and similarity, we employed a different approach to clustering those posts.

Semantic word embedding language modelling techniques represent each word within a corpus with a real-valued vector. These modelling techniques capture the semantic structure of the word vector space by examining not the scalar distance between word vectors but rather the context in which the words appear. Examples of such word embedding algorithms include word2vec (Mikolov et al., 2013a, b, c) and GloVe (Pennington et al., 2014).

As discussed by Jónsson and Stolee (2015), the semantic space created by such word embedding algorithms may be used as an alternative to traditional topic models. A possible approach is to use word2vec (and the continuous-bag-of-words or CBOW model) to construct a semantic vector space from a corpus and cluster words in the resulting vector space using a Gaussian mixture model (GMM) (Sridhar, 2015; Jónsson and Stolee, 2015). Jónsson and Stolee (2015) denoted this word2vec and GMMs by W2V-GMM. Using the same notation, we instead follow a GloVe-GMM approach [4].

Nguyen et al. (2015) shows that, on large datasets (i.e. above 20,000 documents or tweets, as in our case), GloVe produces higher scores than using word2vec for a small number of topics, for example, k ≤ 20, as in our case. Rather than just demonstrating which method applies better to the problem under consideration, we also aim to describe each topic. For this reason, we are not looking for many topics, which would make their interpretation too abstract or redundant (imagine having to label a hundred different topics).

The intuition behind GMM is that its components can be viewed as abstract topics within the word embedding's semantic vector space and each component z within the GMM has a certain probability of generating a word seen in the corpus (Jónsson and Stolee, 2015).

5.1 Frequency analysis

We start by examining the frequency of individual words (see Section 4.1). In Figure 1, we show the word cloud we have obtained for our dataset. The prevailing words would be circular and economy, confirming the correctness of the extraction process as expected. Since they do not have further discriminatory power, we have removed both terms and their combinations for the subsequent frequency analysis.

Starting from the frequency chart of Figure 2, we can try to get a sense of the most concerning issues. It is to be noted that we have employed lemmatisation, rather than stemming, to avoid considering words sharing the same semantics separately (Balakrishnan and Lloyd-Yemoh, 2014).

Terms such as reuse, waste, sustainability and recycling suggest that there is a need to reduce waste and create a CE that values biological and non-organic materials as much as possible. Today, future and climate could be a signal that there is a certain urgency to raise awareness of eco-sustainability in society, and to do this, the commitment of large industrial groups is necessary to respect the environment (business, industry). Note, linked to this theme, the term covid between the most frequent terms, but its presence could also be due to another theme (as we will confirm later through topics modelling): in fact, many see a rare opportunity beyond the pandemic, to build a resilient and low carbon economic recovery (Ellen MacArthur Foundation, 2020).

Know and webinar indicate that there is an opportunity to learn to better understand this environmental emergency in which we find ourselves and to play an active role, also thanks to the use of online resources such as webinars, which play an increasingly central function for learning and comparison in this singular historical period.

We have then three terms (resources, plastic and carbon) directly related to the materials that spur greatest concern for the environment.

Finally, we have two terms (global and world), which highlight the supernational characteristic of the theme.

After considering the occurrence of words, we can perform a frequency analysis of hashtags. In Figure 3 we report the most frequent hashtags. First, the hashtags #circulareconomy, #circular and #economy were removed, which, as one could assume, were the most frequent. After this step, the most popular hashtags become #sustainability, #earthovershootday and #covid19, in line with what was noted in the most frequent words, which also include webinar, plastic and recycling (also present amongst the hashtags in the barplot pictured in Figure 3). The most interesting elements (which did not emerge from previous studies) are the hashtags #covid19, #WCEFonline, #g20 and #eugreendeal. Those hashtags place the CE in a context of European and global dialogue, characterised by the health emergency we are facing. We also observe the hashtags #environment and #bitcoin, which relate the CE to the climate changes that today's society is causing and to the topic of cryptocurrencies, which represents an expanding sharing and CE.

From the analysis of the most frequent terms, we can spot the presence of four macro-themes: materials (as represented by the words carbon, plastic, resources and waste); actions and practices (represented by the words recycling, reuse and sustainability); time and space (today, world, global, future, year); context (covid, climate, industry, energy and business). In Figure 4, we show the frequency of these macro-themes.

As can be seen, the distribution is not overly unbalanced. The biggest share is related to materials.

We can now examine which words are typically associated with each category's most representative words.

We start with materials, considering plastic. As shown in Figure 5, the words exhibiting the highest association with plastic are pollution, recycling, waste and packaging. The most striking concerns are the use of plastic in packaging and the growing amount of waste, which is difficult to recycle and keeps polluting the environment.

In this sense, the two tweets depicted in Figure 6, containing the words above, are very clear.

The timing issue is most represented by the word today. In Figure 7, we see that the most recurring words associated to today are consumed, Earth Overshoot Day, regenerate, year and earth. We find another time reference (year) and the Earth Overshoot Day [5], also known in the past as Ecological Debt Day (EDD), which indicates the day on which humanity consumes all the natural resources produced by the planet in the space of a year (Wackernagel and Pearce, 2018). In the days following this, we consume more resources than the planet produces, drawing on a reserve that sooner or later will run out. The earlier that date, the more we are running out of resources. In 2020, Earth Overshoot Day fell on August 22. It can be estimated that, by proceeding at this rate, around 2050, humanity will consume twice as many resources as produced by the Earth.

Regarding the macro-theme of actions and practices, we focus on the words recycling and sustainability.

The words exhibiting the most significant association with recycling are shown in Figure 8. In particular, the words potential and opportunity suggest how the recycling process of materials is still lagging behind its potential.

As for sustainability, as shown in Figure 9, the association with the words results, helped, customers and tons immediately stands out. Those words could indicate the effectiveness of raising the awareness of eco-sustainability on the side of customers and consumers to reduce tons of natural resources (see Figure 10).

Finally, we consider the context, where the most surprising appearance in the list of most frequent words is covid. We can spot the presence of that term in Figure 9. We have met many tweets concerning the relationship between sustainability and covid in our dataset. In most cases, they convey the opinion that the CE paradigm can help the economic recovery after the disaster caused by the pandemic. The tweet by World Economic Forum (WEF) depicted in Figure 11 is just an example of such tweets.

A different way of exploring the context in which words appear is to consider bigrams and trigrams, i.e. groups of two or three words that appear together. In analysing bigrams and trigrams, we confirm some aspects already identified in the associations between words. We report the Top 20 bigrams and trigrams in Figures 12 and 13, respectively.

The top bigrams refer to an August 2020 tweet from Schneider Electric, which has managed to save its customers 126,000 tons of natural resources in one year. Those bigrams highlight the focus on the results obtained by CE practices, namely the bigrams natural resources that have been saved, helped customers, results helped and sustainability results, as well as the trigrams resulting from their extension, i.e. customers tons natural, helped customer tons, results helped customers, sustainability results helped.

Another significant group of word combinations refers instead to the fossil fuel par excellence that is likely abated as a result of embracing a CE, i.e. carbon. We find the bigrams concept carbon, carbon management, approach carbon, carbon holistic; and again the trigrams resulting from their extension approach carbon management, carbon holistic approach, concept carbon holistic and holistic approach carbon.

We also find the single bigram holistic approach, which refers most often to a tweet from Daimler (a German manufacturer of cars and transport vehicles) about the holistic approach to designing a vehicle, as shown in Figure 14, where the focus is on the innovations brought about the design process by the CE approach.

Finally, we have a group of word combinations that refer to an effective metric to gauge the deterioration of our environment, namely the Earth Overshoot Day that we have already met. Combinations related to this concept are the bigrams today earthovershootday, consumed far, earth regenerate, earthovershootday consumed, regenerate year and far earth, as well as the trigrams earth regenerate year, earthovershootday consumed far, today earthovershootday consumed, consumed far earth and far earth regenerate. Those words are present in a tweet about the Earth Overshoot Day posted by the President of the European Commission, Ursula von der Leyen, which is depicted in Figure 15.

Summing up, we find a pretty balanced mix of words concerning the desirable results of the CE and easily interpretable metrics to convey a dramatic picture of where the earth is going if we do not act.

Sustainability is the most frequent word, by far and large. Sustainability is the overall long-term goal, has become a buzzword (a search on Google returns over 13 billion documents) and was expected to rank at the top position in discussions on the CE.

However, what comes after reveals what twitterers are most concerned about. All the runner-ups in the list of most frequent words have a negative connotation. We observe the negatively-connotated habit of waste in the second position and two concerning chemicals, carbon and plastic, respectively, in the third and fifth position. Hence, it looks like discussants are primarily concerned with what should be avoided.

There are just two concerning means to achieve sustainability in the list of the twenty most frequent words, i.e. recycling and reuse. Though recycling ranks fourth, reuse lies at the bottom of the list. These can be considered positive statements, i.e. indicating how the CE can be accomplished and sustainability can be achieved. However, the list of words most frequently associated with recycling reveals that although potential and opportunity appear respectively as the first and third most frequent ones, the word critical is present in the second position. Hence, even the positive attitudes towards the CE are marred by concerns.

Some sense of time pressure further adds to the list of critical issues, emerging through the presence of the words today, future and year. Among the words most associated with today, we also see far, which may be seen as pushing the time horizon for the realisation of the CE.

People also appear concerned about the environmental consequences of poor conduct since the word climate is present, but it appears in a relatively lower-ranked position (13th).

Finally, we also observe a fair awareness of the economic side of the issue, since they mention business and industry more frequently (respectively in ninth and 12th position).

In comparison with the past literature, we must again notice that no attempts have been made so far to elicit people's opinions expressed on social media about what is meant by the CE. Liu et al. (2009) have performed a similar analysis through questionnaires, with a more limited number of participants. Our results are partially in line with theirs: the recycling aspect is prevalent in our analysis as in theirs, but the discussion on Twitter seems to ignore the concern for the day-by-day activities such as garbage sorting and collection, which was instead quite present in the answers obtained in Liu et al. (2009). Most efforts in the literature were instead aimed at extracting the response of participants to surveys about preset aspects of the CE. Let alone the papers by Liu and Bai (2014), Adams et al. (2017) and Xue et al. (2010), which were directed at professionals and public officers, our work confirms that the items concerning reusing, recycling and zero waste are prominent in the public discourse on Twitter, thus confirming the selection proposed by Lakatos et al. (2016) and Kirchherr et al. (2017). What is relatively new is a feeling of urgency, conveyed by the time-related terms, the attention devoted to specific chemicals (carbon and plastic) and the economic side of the matter.

5.2 Topic modelling

We now describe the results obtained with the three approaches we described in Section 4.

With LDA-U, we have experimented with different numbers of topics to find the right compromise between granularity of description, internal topic coherence and ease of topic labelling. We found the best compromise to be represented by eight topics after testing cases with up to 15 topics. In Figure 16, we show the terms most frequently associated with each topic.

Hereafter, we try to label each topic based on its associated terms. Though automatic methods have been proposed to label topics, as in Lau et al. (2011), we opt for a manual approach. The association between topics and labels is shown in Table 1.

Topic 1 is devoted to sustainable technologies, which are proposed to radically reduce the environmental burden without sacrificing societal and economic standards, as investigated, e.g. by George et al. (2020), Hoosain et al. (2020).

Topic 2 deals with the goal of zero waste, whereby all resources can be resumed fully back into the manufacturing system. This issue is particularly critical for plastic packaging, as highlighted, e.g. in Walker and Xanthos (2018), Meys et al. (2020), Leal Filho et al. (2019), Jang et al. (2020).

Topic 3 is related to the social diffusion of CE issues. We have labelled it as Online connection and learning to show how citizens managed to avoid complete isolation during the pandemic, experiencing new learning approaches (in this case, the best practices of the CE).

Topic 4 deals with Bitcoin for CE, introducing the cryptocurrency and the blockchain technology as a potential enabler for many circular economic principles (Kouhizadeh et al., 2019; Nandi et al., 2020, 2021).

Topic 5 is about generating enough food at a productivity level to maintain the human population, hence the label Food sustainability (Morawicki and González, 2018). A related issue is local sustainability, which considers how local government can promote sustainable development by building partnerships with different groups and organisations in the local community. A systematic review of the literature on this subject can be found in Echebarria et al. (2018).

Topic 6 deals with Energy resources in the light of the CE action plan, as dealt with, e.g. by Meng et al. (2018), Gielen et al. (2019), European Commission (2020), Maranesi and De Giovanni (2020), Angelis et al. (2018).

Topic 7 can be considered as dealing with business models and Sustainable technology. By business models, we mean the set of methods and strategies that a business or organisation uses to promote the new paradigm of CE. This aspect of CE is dealt with, e.g. by Boons and Lüdeke-Freund (2013).

Topic 8 deals with the CE Action plan. The CE Action Plan, described by European Commission (2020), provides a future-oriented agenda for achieving a cleaner and more competitive society in co-creation with economic actors, consumers, citizens and civil society organisations. Battery Recycling, concerning end-of-life lithium-ion batteries, in particular, is one of the actions that can be taken amongst the recycling activities that aim to encourage a shift to more circular economy approaches.

We have carried out the same task with BTM, finding again the best compromise being represented by choosing 8 topics, which are reported in Table 2. The differences with the LDA results are: (1) the presence of a time-related issue (Topic no. 5); (2) the removal of the tight relationship with food in sustainable consumption (Topic no. 6); (3) the combination of the zero-waste goal and the business model instrument to achieve it (Topic no. 1).

A visual rendering of the eight topics is shown in Figure 17.

Finally, we consider the application of Glove.

By applying the same choice of 8 topics, we found the selection shown in Table 3, where issues related to the society-at-large appear (Topics no. 1, 3 and 7).

The topics identified through GloVe-GMM appear to be much dirtier than those identified by the other two methodologies. In all the topics within the GloVe-GMM, we notice the presence of words that are inconsistent with each other. We have almost always been forced to search for individual words to deduce the correct topic. For example abml in Topic 1 stands for American Battery Metals, which recently has developed a full recycling process that can recover over 95% of each of the elemental metals that are integral to manufacturing new batteries. Canoo is an American manufacturer of electric vehicles (EV). The hashtag #circularevbatteries stands for circular EV) batteries whilst rcrjournal is the Twitter account of Resources, Conservation & Recycling Journal. This method seems to be more specific and identifies the main subjects of the topic. This is achieved despite a set of words that are less coherent with each other, which represents a limitation of the GloVe-GMM, at least for this case study, is an advantage.

The results obtained through topic modelling confirm the themes coming from the frequency analysis of single words. Along with the expected Sustainable tech, Zero waste, Energy resources and Food and sustainability, we see the economic and financial issue emerging with the topic Business models and Bitcoin.

5.3 Who's behind top influencers

As we said, we focussed on the messages posted on Twitter, which can be considered representative of the opinions of people active on the internet. Though companies and governments are also active on Twitter, individual accounts largely dominate, so the opinions expressed on Twitter largely represent individual users. Who are those opinionators? It is impossible to explain who the various users in our database are, but we can understand who the most popular users are.

A possible measure of a user's popularity on Twitter is retweeting. Retweeting is a major form of providing support for the opinion expressed through a tweet. Table 4 shows the Top 20 most retweeted users. But who are they? Are they activists? Or do they work for a company? Or are they scientists? Knowing who is behind their screen names may help us uncover their intentions in writing posts and understand people's preferences. Hereafter we disclose their roles in society. To accomplish this task, we checked the Twitter bio of each most retweeted user; the Twitter bio of user x can be defined as a summary of x's Twitter profile [6]. Furthermore, in the bio, there may be links to the official websites of the companies or institutions represented on Twitter by the account. By examining the Twitter bio, we can establish whether an account belongs to a company representative or a private person. In particular, we used the information contained in the bio to go back to personal or institutional websites, LinkedIn profiles and even the Wikipedia pages of each user. As we can see in the following list, influencers in the Top 20 have non-anonymous accounts and their information is easily retrieved on the web.

1. @SchneiderElec is the Twitter account of Schneider Electric, a French multinational company active in the field of digital automation and energy management.

2. @wef stands for the World Economic Forum, the international organisation for public-private cooperation.

3. @CCE_Guide stands for CCE Guide, a series of reports on carbon management written by leading international organisations and commissioned by the King Abdullah Petroleum Studies and Research Centre (KAPSARC).

4. @circulareconomy is the Twitter account of Ellen MacArthur Foundation, which is a UK-registered charity which promotes the CE.

5. @ErikSolheim is the Twitter account of Erik Solheim, a Norwegian diplomat and former politician who served in the Norwegian government as Minister of the Environment.

6. @vonderleyen is the Twitter account of Ursula von der Leyen, the current President of the European Commission.

7. @anandmahindra is the Twitter account of Anand Mahindra, an Indian billionaire businessman and the chairman of Mahindra Group, a Mumbai-based business conglomerate.

8. @Aramco stands for Saudi Aramco, a Saudi Arabian public petroleum and natural gas company based in Dhahran.

9. @EU_Commission is the Twitter account of the European Commission.

10. @PMOIndia stands for the Twitter account of the Office of the Prime Minister of India.

11. @iingwen is the Twitter account of Tsai Ing-wen; she is a Taiwanese politician and academic who is the current President of the Republic of China (Taiwan).

12. @CEStakeholderEU stands for European Circular Economy Stakeholder Platform, a joint initiative by the European Commission and the European Economic and Social Committee to disseminate the CE approach.

13. @EU_ENV stands for the official account of the European Commission Directorate-General for Environment (DG ENV).

14. @circleeconomy stands for Circle Economy, a social enterprise organised as a cooperative that seeks to accelerate the transition to circularity.

15. @antonioguterres is the Twitter account of António Guterres, the current Secretary-General of the United Nations.

16. @PACEcircular stands for Platform for Accelerating the Circular Economy, a global community created by WEF with the aim of accelerating the transition to a CE.

17. @JohnRHewson is the Twitter account of John Hewson, an Australian former politician who served as leader of the Liberal Party from 1990 to 1994.

18. @tetrapak stands for Tetra Pak, a Swedish-Swiss multinational food packaging and processing company.

19. @winston_graf seems to be no longer available.

20. @JavierBlas is the Twitter account of Javier Blas, a Bloomberg Opinion columnist covering energy and commodities.

We can classify these screen names according to their role, finding a large presence of politicians, leading international or governmental agencies, associations, organisations and companies. This means that these stakeholders are steadily present on Twitter, voicing their opinion and also they are able to capture the attention and the consensus of the people of Twitter. These results are in contrast with others found for public opinion formation on wind energy (Mastroeni et al., n.d.) and climate issues (Iacomini and Vellucci, 2021). For example, on the subject of wind energy, companies and institutional organisations seem to exert a much lower influence (Mastroeni et al., n.d.) whereas, on the subject of climate change, twitterers seem to be polarised around a small number of different opinions (Iacomini and Vellucci, 2021). Our analysis revealed instead a certain consensus amongst Twitter users on the need to adopt a circular approach.